Welcome to MMDetection’s documentation!¶
Prerequisites¶
In this section we demonstrate how to prepare an environment with PyTorch.
MMDetection works on Linux, Windows and macOS. It requires Python 3.7+, CUDA 9.2+ and PyTorch 1.5+.
Note
If you are experienced with PyTorch and have already installed it, just skip this part and jump to the next section. Otherwise, you can follow these steps for the preparation.
Step 0. Download and install Miniconda from the official website.
Step 1. Create a conda environment and activate it.
conda create --name openmmlab python=3.8 -y
conda activate openmmlab
Step 2. Install PyTorch following official instructions, e.g.
On GPU platforms:
conda install pytorch torchvision -c pytorch
On CPU platforms:
conda install pytorch torchvision cpuonly -c pytorch
Installation¶
We recommend that users follow our best practices to install MMDetection. However, the whole process is highly customizable. See Customize Installation section for more information.
Best Practices¶
Step 0. Install MMCV using MIM.
pip install -U openmim
mim install mmcv-full
Step 1. Install MMDetection.
Case a: If you develop and run mmdet directly, install it from source:
git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection
pip install -v -e .
# "-v" means verbose, or more output
# "-e" means installing a project in editable mode,
# thus any local modifications made to the code will take effect without reinstallation.
Case b: If you use mmdet as a dependency or third-party package, install it with pip:
pip install mmdet
Verify the installation¶
To verify whether MMDetection is installed correctly, we provide some sample codes to run an inference demo.
Step 1. We need to download config and checkpoint files.
mim download mmdet --config yolov3_mobilenetv2_320_300e_coco --dest .
The downloading will take several seconds or more, depending on your network environment. When it is done, you will find two files yolov3_mobilenetv2_320_300e_coco.py and yolov3_mobilenetv2_320_300e_coco_20210719_215349-d18dff72.pth in your current folder.
Step 2. Verify the inference demo.
Option (a). If you install mmdetection from source, just run the following command.
python demo/image_demo.py demo/demo.jpg yolov3_mobilenetv2_320_300e_coco.py yolov3_mobilenetv2_320_300e_coco_20210719_215349-d18dff72.pth --device cpu --out-file result.jpg
You will see a new image result.jpg on your current folder, where bounding boxes are plotted on cars, benches, etc.
Option (b). If you install mmdetection with pip, open you python interpreter and copy&paste the following codes.
from mmdet.apis import init_detector, inference_detector
config_file = 'yolov3_mobilenetv2_320_300e_coco.py'
checkpoint_file = 'yolov3_mobilenetv2_320_300e_coco_20210719_215349-d18dff72.pth'
model = init_detector(config_file, checkpoint_file, device='cpu') # or device='cuda:0'
inference_detector(model, 'demo/demo.jpg')
You will see a list of arrays printed, indicating the detected bounding boxes.
Customize Installation¶
CUDA versions¶
When installing PyTorch, you need to specify the version of CUDA. If you are not clear on which to choose, follow our recommendations:
For Ampere-based NVIDIA GPUs, such as GeForce 30 series and NVIDIA A100, CUDA 11 is a must.
For older NVIDIA GPUs, CUDA 11 is backward compatible, but CUDA 10.2 offers better compatibility and is more lightweight.
Please make sure the GPU driver satisfies the minimum version requirements. See this table for more information.
Note
Installing CUDA runtime libraries is enough if you follow our best practices, because no CUDA code will be compiled locally. However if you hope to compile MMCV from source or develop other CUDA operators, you need to install the complete CUDA toolkit from NVIDIA’s website, and its version should match the CUDA version of PyTorch. i.e., the specified version of cudatoolkit in conda install command.
Install MMCV without MIM¶
MMCV contains C++ and CUDA extensions, thus depending on PyTorch in a complex way. MIM solves such dependencies automatically and makes the installation easier. However, it is not a must.
To install MMCV with pip instead of MIM, please follow MMCV installation guides. This requires manually specifying a find-url based on PyTorch version and its CUDA version.
For example, the following command install mmcv-full built for PyTorch 1.10.x and CUDA 11.3.
pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.10/index.html
Install on CPU-only platforms¶
MMDetection can be built for CPU only environment. In CPU mode you can train (requires MMCV version >= 1.4.4), test or inference a model.
However some functionalities are gone in this mode:
Deformable Convolution
Modulated Deformable Convolution
ROI pooling
Deformable ROI pooling
CARAFE
SyncBatchNorm
CrissCrossAttention
MaskedConv2d
Temporal Interlace Shift
nms_cuda
sigmoid_focal_loss_cuda
bbox_overlaps
If you try to train/test/inference a model containing above ops, an error will be raised. The following table lists affected algorithms.
| Operator | Model |
|---|---|
| Deformable Convolution/Modulated Deformable Convolution | DCN、Guided Anchoring、RepPoints、CentripetalNet、VFNet、CascadeRPN、NAS-FCOS、DetectoRS |
| MaskedConv2d | Guided Anchoring |
| CARAFE | CARAFE |
| SyncBatchNorm | ResNeSt |
Install on Google Colab¶
Google Colab usually has PyTorch installed, thus we only need to install MMCV and MMDetection with the following commands.
Step 1. Install MMCV using MIM.
!pip3 install openmim
!mim install mmcv-full
Step 2. Install MMDetection from the source.
!git clone https://github.com/open-mmlab/mmdetection.git
%cd mmdetection
!pip install -e .
Step 3. Verification.
import mmdet
print(mmdet.__version__)
# Example output: 2.23.0
Note
Within Jupyter, the exclamation mark ! is used to call external executables and %cd is a magic command to change the current working directory of Python.
Using MMDetection with Docker¶
We provide a Dockerfile to build an image. Ensure that your docker version >=19.03.
# build an image with PyTorch 1.6, CUDA 10.1
# If you prefer other versions, just modified the Dockerfile
docker build -t mmdetection docker/
Run it with
docker run --gpus all --shm-size=8g -it -v {DATA_DIR}:/mmdetection/data mmdetection
Trouble shooting¶
If you have some issues during the installation, please first view the FAQ page. You may open an issue on GitHub if no solution is found.
Benchmark and Model Zoo¶
Mirror sites¶
We only use aliyun to maintain the model zoo since MMDetection V2.0. The model zoo of V1.x has been deprecated.
Common settings¶
All models were trained on
coco_2017_train, and tested on thecoco_2017_val.We use distributed training.
All pytorch-style pretrained backbones on ImageNet are from PyTorch model zoo, caffe-style pretrained backbones are converted from the newly released model from detectron2.
For fair comparison with other codebases, we report the GPU memory as the maximum value of
torch.cuda.max_memory_allocated()for all 8 GPUs. Note that this value is usually less than whatnvidia-smishows.We report the inference time as the total time of network forwarding and post-processing, excluding the data loading time. Results are obtained with the script benchmark.py which computes the average time on 2000 images.
ImageNet Pretrained Models¶
It is common to initialize from backbone models pre-trained on ImageNet classification task. All pre-trained model links can be found at open_mmlab. According to img_norm_cfg and source of weight, we can divide all the ImageNet pre-trained model weights into some cases:
TorchVision: Corresponding to torchvision weight, including ResNet50, ResNet101. The
img_norm_cfgisdict(mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True).Pycls: Corresponding to pycls weight, including RegNetX. The
img_norm_cfgisdict( mean=[103.530, 116.280, 123.675], std=[57.375, 57.12, 58.395], to_rgb=False).MSRA styles: Corresponding to MSRA weights, including ResNet50_Caffe and ResNet101_Caffe. The
img_norm_cfgisdict( mean=[103.530, 116.280, 123.675], std=[1.0, 1.0, 1.0], to_rgb=False).Caffe2 styles: Currently only contains ResNext101_32x8d. The
img_norm_cfgisdict(mean=[103.530, 116.280, 123.675], std=[57.375, 57.120, 58.395], to_rgb=False).Other styles: E.g SSD which corresponds to
img_norm_cfgisdict(mean=[123.675, 116.28, 103.53], std=[1, 1, 1], to_rgb=True)and YOLOv3 which corresponds toimg_norm_cfgisdict(mean=[0, 0, 0], std=[255., 255., 255.], to_rgb=True).
The detailed table of the commonly used backbone models in MMDetection is listed below :
| model | source | link | description |
|---|---|---|---|
| ResNet50 | TorchVision | torchvision's ResNet-50 | From torchvision's ResNet-50. |
| ResNet101 | TorchVision | torchvision's ResNet-101 | From torchvision's ResNet-101. |
| RegNetX | Pycls | RegNetX_3.2gf, RegNetX_800mf. etc. | From pycls. |
| ResNet50_Caffe | MSRA | MSRA's ResNet-50 | Converted copy of Detectron2's R-50.pkl model. The original weight comes from MSRA's original ResNet-50. |
| ResNet101_Caffe | MSRA | MSRA's ResNet-101 | Converted copy of Detectron2's R-101.pkl model. The original weight comes from MSRA's original ResNet-101. |
| ResNext101_32x8d | Caffe2 | Caffe2 ResNext101_32x8d | Converted copy of Detectron2's X-101-32x8d.pkl model. The ResNeXt-101-32x8d model trained with Caffe2 at FB. |
Baselines¶
Faster R-CNN¶
Please refer to Faster R-CNN for details.
Mask R-CNN¶
Please refer to Mask R-CNN for details.
Fast R-CNN (with pre-computed proposals)¶
Please refer to Fast R-CNN for details.
Cascade R-CNN and Cascade Mask R-CNN¶
Please refer to Cascade R-CNN for details.
Group Normalization (GN)¶
Please refer to Group Normalization for details.
Weight Standardization¶
Please refer to Weight Standardization for details.
Deformable Convolution v2¶
Please refer to Deformable Convolutional Networks for details.
Instaboost¶
Please refer to Instaboost for details.
Libra R-CNN¶
Please refer to Libra R-CNN for details.
Guided Anchoring¶
Please refer to Guided Anchoring for details.
FreeAnchor¶
Please refer to FreeAnchor for details.
Grid R-CNN (plus)¶
Please refer to Grid R-CNN for details.
Mask Scoring R-CNN¶
Please refer to Mask Scoring R-CNN for details.
Train from Scratch¶
Please refer to Rethinking ImageNet Pre-training for details.
Dynamic R-CNN¶
Please refer to Dynamic R-CNN for details.
Generalized Focal Loss¶
Please refer to Generalized Focal Loss for details.
CentripetalNet¶
Please refer to CentripetalNet for details.
Deformable DETR¶
Please refer to Deformable DETR for details.
AutoAssign¶
Please refer to AutoAssign for details.
Seesaw Loss¶
Please refer to Seesaw Loss for details.
PanopticFPN¶
Please refer to PanopticFPN for details.
MaskFormer¶
Please refer to MaskFormer for details.
Mask2Former¶
Please refer to Mask2Former for details.
Efficientnet¶
Please refer to Efficientnet for details.
Other datasets¶
We also benchmark some methods on PASCAL VOC, Cityscapes, OpenImages and WIDER FACE.
Pre-trained Models¶
We also train Faster R-CNN and Mask R-CNN using ResNet-50 and RegNetX-3.2G with multi-scale training and longer schedules. These models serve as strong pre-trained models for downstream tasks for convenience.
Speed benchmark¶
Training Speed benchmark¶
We provide analyze_logs.py to get average time of iteration in training. You can find examples in Log Analysis.
We compare the training speed of Mask R-CNN with some other popular frameworks (The data is copied from detectron2). For mmdetection, we benchmark with mask_rcnn_r50_caffe_fpn_poly_1x_coco_v1.py, which should have the same setting with mask_rcnn_R_50_FPN_noaug_1x.yaml of detectron2. We also provide the checkpoint and training log for reference. The throughput is computed as the average throughput in iterations 100-500 to skip GPU warmup time.
| Implementation | Throughput (img/s) |
|---|---|
| Detectron2 | 62 |
| MMDetection | 61 |
| maskrcnn-benchmark | 53 |
| tensorpack | 50 |
| simpledet | 39 |
| Detectron | 19 |
| matterport/Mask_RCNN | 14 |
Inference Speed Benchmark¶
We provide benchmark.py to benchmark the inference latency.
The script benchmarkes the model with 2000 images and calculates the average time ignoring first 5 times. You can change the output log interval (defaults: 50) by setting LOG-INTERVAL.
python tools/benchmark.py ${CONFIG} ${CHECKPOINT} [--log-interval $[LOG-INTERVAL]] [--fuse-conv-bn]
The latency of all models in our model zoo is benchmarked without setting fuse-conv-bn, you can get a lower latency by setting it.
Comparison with Detectron2¶
We compare mmdetection with Detectron2 in terms of speed and performance. We use the commit id 185c27e(30/4/2020) of detectron. For fair comparison, we install and run both frameworks on the same machine.
Hardware¶
8 NVIDIA Tesla V100 (32G) GPUs
Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
Software environment¶
Python 3.7
PyTorch 1.4
CUDA 10.1
CUDNN 7.6.03
NCCL 2.4.08
Performance¶
| Type | Lr schd | Detectron2 | mmdetection | Download |
|---|---|---|---|---|
| Faster R-CNN | 1x | 37.9 | 38.0 | model | log |
| Mask R-CNN | 1x | 38.6 & 35.2 | 38.8 & 35.4 | model | log |
| Retinanet | 1x | 36.5 | 37.0 | model | log |
Training Speed¶
The training speed is measure with s/iter. The lower, the better.
| Type | Detectron2 | mmdetection |
|---|---|---|
| Faster R-CNN | 0.210 | 0.216 |
| Mask R-CNN | 0.261 | 0.265 |
| Retinanet | 0.200 | 0.205 |
Inference Speed¶
The inference speed is measured with fps (img/s) on a single GPU, the higher, the better. To be consistent with Detectron2, we report the pure inference speed (without the time of data loading). For Mask R-CNN, we exclude the time of RLE encoding in post-processing. We also include the officially reported speed in the parentheses, which is slightly higher than the results tested on our server due to differences of hardwares.
| Type | Detectron2 | mmdetection |
|---|---|---|
| Faster R-CNN | 25.6 (26.3) | 22.2 |
| Mask R-CNN | 22.5 (23.3) | 19.6 |
| Retinanet | 17.8 (18.2) | 20.6 |
Training memory¶
| Type | Detectron2 | mmdetection |
|---|---|---|
| Faster R-CNN | 3.0 | 3.8 |
| Mask R-CNN | 3.4 | 3.9 |
| Retinanet | 3.9 | 3.4 |
1: Inference and train with existing models and standard datasets¶
MMDetection provides hundreds of existing and existing detection models in Model Zoo), and supports multiple standard datasets, including Pascal VOC, COCO, CityScapes, LVIS, etc. This note will show how to perform common tasks on these existing models and standard datasets, including:
Use existing models to inference on given images.
Test existing models on standard datasets.
Train predefined models on standard datasets.
Inference with existing models¶
By inference, we mean using trained models to detect objects on images. In MMDetection, a model is defined by a configuration file and existing model parameters are save in a checkpoint file.
To start with, we recommend Faster RCNN with this configuration file and this checkpoint file. It is recommended to download the checkpoint file to checkpoints directory.
High-level APIs for inference¶
MMDetection provide high-level Python APIs for inference on images. Here is an example of building the model and inference on given images or videos.
from mmdet.apis import init_detector, inference_detector
import mmcv
# Specify the path to model config and checkpoint file
config_file = 'configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py'
checkpoint_file = 'checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth'
# build the model from a config file and a checkpoint file
model = init_detector(config_file, checkpoint_file, device='cuda:0')
# test a single image and show the results
img = 'test.jpg' # or img = mmcv.imread(img), which will only load it once
result = inference_detector(model, img)
# visualize the results in a new window
model.show_result(img, result)
# or save the visualization results to image files
model.show_result(img, result, out_file='result.jpg')
# test a video and show the results
video = mmcv.VideoReader('video.mp4')
for frame in video:
result = inference_detector(model, frame)
model.show_result(frame, result, wait_time=1)
A notebook demo can be found in demo/inference_demo.ipynb.
Note: inference_detector only supports single-image inference for now.
Asynchronous interface - supported for Python 3.7+¶
For Python 3.7+, MMDetection also supports async interfaces. By utilizing CUDA streams, it allows not to block CPU on GPU bound inference code and enables better CPU/GPU utilization for single-threaded application. Inference can be done concurrently either between different input data samples or between different models of some inference pipeline.
See tests/async_benchmark.py to compare the speed of synchronous and asynchronous interfaces.
import asyncio
import torch
from mmdet.apis import init_detector, async_inference_detector
from mmdet.utils.contextmanagers import concurrent
async def main():
config_file = 'configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py'
checkpoint_file = 'checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth'
device = 'cuda:0'
model = init_detector(config_file, checkpoint=checkpoint_file, device=device)
# queue is used for concurrent inference of multiple images
streamqueue = asyncio.Queue()
# queue size defines concurrency level
streamqueue_size = 3
for _ in range(streamqueue_size):
streamqueue.put_nowait(torch.cuda.Stream(device=device))
# test a single image and show the results
img = 'test.jpg' # or img = mmcv.imread(img), which will only load it once
async with concurrent(streamqueue):
result = await async_inference_detector(model, img)
# visualize the results in a new window
model.show_result(img, result)
# or save the visualization results to image files
model.show_result(img, result, out_file='result.jpg')
asyncio.run(main())
Demos¶
We also provide three demo scripts, implemented with high-level APIs and supporting functionality codes. Source codes are available here.
Image demo¶
This script performs inference on a single image.
python demo/image_demo.py \
${IMAGE_FILE} \
${CONFIG_FILE} \
${CHECKPOINT_FILE} \
[--device ${GPU_ID}] \
[--score-thr ${SCORE_THR}]
Examples:
python demo/image_demo.py demo/demo.jpg \
configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \
checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth \
--device cpu
Webcam demo¶
This is a live demo from a webcam.
python demo/webcam_demo.py \
${CONFIG_FILE} \
${CHECKPOINT_FILE} \
[--device ${GPU_ID}] \
[--camera-id ${CAMERA-ID}] \
[--score-thr ${SCORE_THR}]
Examples:
python demo/webcam_demo.py \
configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \
checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth
Video demo¶
This script performs inference on a video.
python demo/video_demo.py \
${VIDEO_FILE} \
${CONFIG_FILE} \
${CHECKPOINT_FILE} \
[--device ${GPU_ID}] \
[--score-thr ${SCORE_THR}] \
[--out ${OUT_FILE}] \
[--show] \
[--wait-time ${WAIT_TIME}]
Examples:
python demo/video_demo.py demo/demo.mp4 \
configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \
checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth \
--out result.mp4
Video demo with GPU acceleration¶
This script performs inference on a video with GPU acceleration.
python demo/video_gpuaccel_demo.py \
${VIDEO_FILE} \
${CONFIG_FILE} \
${CHECKPOINT_FILE} \
[--device ${GPU_ID}] \
[--score-thr ${SCORE_THR}] \
[--nvdecode] \
[--out ${OUT_FILE}] \
[--show] \
[--wait-time ${WAIT_TIME}]
Examples:
python demo/video_gpuaccel_demo.py demo/demo.mp4 \
configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \
checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth \
--nvdecode --out result.mp4
Test existing models on standard datasets¶
To evaluate a model’s accuracy, one usually tests the model on some standard datasets. MMDetection supports multiple public datasets including COCO, Pascal VOC, CityScapes, and more. This section will show how to test existing models on supported datasets.
Prepare datasets¶
Public datasets like Pascal VOC or mirror and COCO are available from official websites or mirrors. Note: In the detection task, Pascal VOC 2012 is an extension of Pascal VOC 2007 without overlap, and we usually use them together.
It is recommended to download and extract the dataset somewhere outside the project directory and symlink the dataset root to $MMDETECTION/data as below.
If your folder structure is different, you may need to change the corresponding paths in config files.
We provide a script to download datasets such as COCO , you can run python tools/misc/download_dataset.py --dataset-name coco2017 to download COCO dataset.
For more usage please refer to dataset-download
mmdetection
├── mmdet
├── tools
├── configs
├── data
│ ├── coco
│ │ ├── annotations
│ │ ├── train2017
│ │ ├── val2017
│ │ ├── test2017
│ ├── cityscapes
│ │ ├── annotations
│ │ ├── leftImg8bit
│ │ │ ├── train
│ │ │ ├── val
│ │ ├── gtFine
│ │ │ ├── train
│ │ │ ├── val
│ ├── VOCdevkit
│ │ ├── VOC2007
│ │ ├── VOC2012
Some models require additional COCO-stuff datasets, such as HTC, DetectoRS and SCNet, you can download and unzip then move to the coco folder. The directory should be like this.
mmdetection
├── data
│ ├── coco
│ │ ├── annotations
│ │ ├── train2017
│ │ ├── val2017
│ │ ├── test2017
│ │ ├── stuffthingmaps
Panoptic segmentation models like PanopticFPN require additional COCO Panoptic datasets, you can download and unzip then move to the coco annotation folder. The directory should be like this.
mmdetection
├── data
│ ├── coco
│ │ ├── annotations
│ │ │ ├── panoptic_train2017.json
│ │ │ ├── panoptic_train2017
│ │ │ ├── panoptic_val2017.json
│ │ │ ├── panoptic_val2017
│ │ ├── train2017
│ │ ├── val2017
│ │ ├── test2017
The cityscapes annotations need to be converted into the coco format using tools/dataset_converters/cityscapes.py:
pip install cityscapesscripts
python tools/dataset_converters/cityscapes.py \
./data/cityscapes \
--nproc 8 \
--out-dir ./data/cityscapes/annotations
TODO: CHANGE TO THE NEW PATH
Test existing models¶
We provide testing scripts for evaluating an existing model on the whole dataset (COCO, PASCAL VOC, Cityscapes, etc.). The following testing environments are supported:
single GPU
CPU
single node multiple GPUs
multiple nodes
Choose the proper script to perform testing depending on the testing environment.
# single-gpu testing
python tools/test.py \
${CONFIG_FILE} \
${CHECKPOINT_FILE} \
[--out ${RESULT_FILE}] \
[--eval ${EVAL_METRICS}] \
[--show]
# CPU: disable GPUs and run single-gpu testing script
export CUDA_VISIBLE_DEVICES=-1
python tools/test.py \
${CONFIG_FILE} \
${CHECKPOINT_FILE} \
[--out ${RESULT_FILE}] \
[--eval ${EVAL_METRICS}] \
[--show]
# multi-gpu testing
bash tools/dist_test.sh \
${CONFIG_FILE} \
${CHECKPOINT_FILE} \
${GPU_NUM} \
[--out ${RESULT_FILE}] \
[--eval ${EVAL_METRICS}]
tools/dist_test.sh also supports multi-node testing, but relies on PyTorch’s launch utility.
Optional arguments:
RESULT_FILE: Filename of the output results in pickle format. If not specified, the results will not be saved to a file.EVAL_METRICS: Items to be evaluated on the results. Allowed values depend on the dataset, e.g.,proposal_fast,proposal,bbox,segmare available for COCO,mAP,recallfor PASCAL VOC. Cityscapes could be evaluated bycityscapesas well as all COCO metrics.--show: If specified, detection results will be plotted on the images and shown in a new window. It is only applicable to single GPU testing and used for debugging and visualization. Please make sure that GUI is available in your environment. Otherwise, you may encounter an error likecannot connect to X server.--show-dir: If specified, detection results will be plotted on the images and saved to the specified directory. It is only applicable to single GPU testing and used for debugging and visualization. You do NOT need a GUI available in your environment for using this option.--show-score-thr: If specified, detections with scores below this threshold will be removed.--cfg-options: if specified, the key-value pair optional cfg will be merged into config file--eval-options: if specified, the key-value pair optional eval cfg will be kwargs for dataset.evaluate() function, it’s only for evaluation
Examples¶
Assuming that you have already downloaded the checkpoints to the directory checkpoints/.
Test Faster R-CNN and visualize the results. Press any key for the next image. Config and checkpoint files are available here.
python tools/test.py \ configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \ checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth \ --show
Test Faster R-CNN and save the painted images for future visualization. Config and checkpoint files are available here.
python tools/test.py \ configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \ checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth \ --show-dir faster_rcnn_r50_fpn_1x_results
Test Faster R-CNN on PASCAL VOC (without saving the test results) and evaluate the mAP. Config and checkpoint files are available here.
python tools/test.py \ configs/pascal_voc/faster_rcnn_r50_fpn_1x_voc.py \ checkpoints/faster_rcnn_r50_fpn_1x_voc0712_20200624-c9895d40.pth \ --eval mAP
Test Mask R-CNN with 8 GPUs, and evaluate the bbox and mask AP. Config and checkpoint files are available here.
./tools/dist_test.sh \ configs/mask_rcnn_r50_fpn_1x_coco.py \ checkpoints/mask_rcnn_r50_fpn_1x_coco_20200205-d4b0c5d6.pth \ 8 \ --out results.pkl \ --eval bbox segm
Test Mask R-CNN with 8 GPUs, and evaluate the classwise bbox and mask AP. Config and checkpoint files are available here.
./tools/dist_test.sh \ configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.py \ checkpoints/mask_rcnn_r50_fpn_1x_coco_20200205-d4b0c5d6.pth \ 8 \ --out results.pkl \ --eval bbox segm \ --options "classwise=True"
Test Mask R-CNN on COCO test-dev with 8 GPUs, and generate JSON files for submitting to the official evaluation server. Config and checkpoint files are available here.
./tools/dist_test.sh \ configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.py \ checkpoints/mask_rcnn_r50_fpn_1x_coco_20200205-d4b0c5d6.pth \ 8 \ --format-only \ --options "jsonfile_prefix=./mask_rcnn_test-dev_results"
This command generates two JSON files
mask_rcnn_test-dev_results.bbox.jsonandmask_rcnn_test-dev_results.segm.json.Test Mask R-CNN on Cityscapes test with 8 GPUs, and generate txt and png files for submitting to the official evaluation server. Config and checkpoint files are available here.
./tools/dist_test.sh \ configs/cityscapes/mask_rcnn_r50_fpn_1x_cityscapes.py \ checkpoints/mask_rcnn_r50_fpn_1x_cityscapes_20200227-afe51d5a.pth \ 8 \ --format-only \ --options "txtfile_prefix=./mask_rcnn_cityscapes_test_results"
The generated png and txt would be under
./mask_rcnn_cityscapes_test_resultsdirectory.
Test without Ground Truth Annotations¶
MMDetection supports to test models without ground-truth annotations using CocoDataset. If your dataset format is not in COCO format, please convert them to COCO format. For example, if your dataset format is VOC, you can directly convert it to COCO format by the script in tools. If your dataset format is Cityscapes, you can directly convert it to COCO format by the script in tools. The rest of the formats can be converted using this script.
python tools/dataset_converters/images2coco.py \
${IMG_PATH} \
${CLASSES} \
${OUT} \
[--exclude-extensions]
arguments：
IMG_PATH: The root path of images.CLASSES: The text file with a list of categories.OUT: The output annotation json file name. The save dir is in the same directory asIMG_PATH.exclude-extensions: The suffix of images to be excluded, such as ‘png’ and ‘bmp’.
After the conversion is complete, you can use the following command to test
# single-gpu testing
python tools/test.py \
${CONFIG_FILE} \
${CHECKPOINT_FILE} \
--format-only \
--options ${JSONFILE_PREFIX} \
[--show]
# CPU: disable GPUs and run single-gpu testing script
export CUDA_VISIBLE_DEVICES=-1
python tools/test.py \
${CONFIG_FILE} \
${CHECKPOINT_FILE} \
[--out ${RESULT_FILE}] \
[--eval ${EVAL_METRICS}] \
[--show]
# multi-gpu testing
bash tools/dist_test.sh \
${CONFIG_FILE} \
${CHECKPOINT_FILE} \
${GPU_NUM} \
--format-only \
--options ${JSONFILE_PREFIX} \
[--show]
Assuming that the checkpoints in the model zoo have been downloaded to the directory checkpoints/, we can test Mask R-CNN on COCO test-dev with 8 GPUs, and generate JSON files using the following command.
./tools/dist_test.sh \
configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.py \
checkpoints/mask_rcnn_r50_fpn_1x_coco_20200205-d4b0c5d6.pth \
8 \
--format-only \
--options "jsonfile_prefix=./mask_rcnn_test-dev_results"
This command generates two JSON files mask_rcnn_test-dev_results.bbox.json and mask_rcnn_test-dev_results.segm.json.
Batch Inference¶
MMDetection supports inference with a single image or batched images in test mode. By default, we use single-image inference and you can use batch inference by modifying samples_per_gpu in the config of test data. You can do that either by modifying the config as below.
data = dict(train=dict(...), val=dict(...), test=dict(samples_per_gpu=2, ...))
Or you can set it through --cfg-options as --cfg-options data.test.samples_per_gpu=2
Deprecated ImageToTensor¶
In test mode, ImageToTensor pipeline is deprecated, it’s replaced by DefaultFormatBundle that recommended to manually replace it in the test data pipeline in your config file. examples:
# use ImageToTensor (deprecated)
pipelines = [
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(1333, 800),
flip=False,
transforms=[
dict(type='Resize', keep_ratio=True),
dict(type='RandomFlip'),
dict(type='Normalize', mean=[0, 0, 0], std=[1, 1, 1]),
dict(type='Pad', size_divisor=32),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect', keys=['img']),
])
]
# manually replace ImageToTensor to DefaultFormatBundle (recommended)
pipelines = [
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(1333, 800),
flip=False,
transforms=[
dict(type='Resize', keep_ratio=True),
dict(type='RandomFlip'),
dict(type='Normalize', mean=[0, 0, 0], std=[1, 1, 1]),
dict(type='Pad', size_divisor=32),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img']),
])
]
Train predefined models on standard datasets¶
MMDetection also provides out-of-the-box tools for training detection models. This section will show how to train predefined models (under configs) on standard datasets i.e. COCO.
Prepare datasets¶
Training requires preparing datasets too. See section Prepare datasets above for details.
Note:
Currently, the config files under configs/cityscapes use COCO pretrained weights to initialize.
You could download the existing models in advance if the network connection is unavailable or slow. Otherwise, it would cause errors at the beginning of training.
Learning rate automatically scale¶
Important: The default learning rate in config files is for 8 GPUs and 2 sample per gpu (batch size = 8 * 2 = 16). And it had been set to auto_scale_lr.base_batch_size in config/_base_/default_runtime.py. Learning rate will be automatically scaled base on this value when the batch size is 16. Meanwhile, in order not to affect other codebase which based on mmdet, the flag auto_scale_lr.enable is set to False by default.
If you want to enable this feature, you need to add argument --auto-scale-lr. And you need to check the config name which you want to use before you process the command, because the config name indicates the default batch size.
By default, it is 8 x 2 = 16 batch size, like faster_rcnn_r50_caffe_fpn_90k_coco.py or pisa_faster_rcnn_x101_32x4d_fpn_1x_coco.py. In other cases, you will see the config file name have _NxM_ in dictating, like cornernet_hourglass104_mstest_32x3_210e_coco.py which batch size is 32 x 3 = 96, or scnet_x101_64x4d_fpn_8x1_20e_coco.py which batch size is 8 x 1 = 8.
Please remember to check the bottom of the specific config file you want to use, it will have auto_scale_lr.base_batch_size if the batch size is not 16. If you can’t find those values, check the config file which in _base_=[xxx] and you will find it. Please do not modify its values if you want to automatically scale the LR.
Learning rate automatically scale basic usage is as follows.
python tools/train.py \
${CONFIG_FILE} \
--auto-scale-lr \
[optional arguments]
If you enabled this feature, the learning rate will be automatically scaled according to the number of GPUs of the machine and the batch size of training. See linear scaling rule for details. For example, If there are 4 GPUs and 2 pictures on each GPU, lr = 0.01, then if there are 16 GPUs and 4 pictures on each GPU, it will automatically scale to lr = 0.08.
If you don’t want to use it, you need to calculate the learning rate according to the linear scaling rule manually then change optimizer.lr in specific config file.
Training on a single GPU¶
We provide tools/train.py to launch training jobs on a single GPU.
The basic usage is as follows.
python tools/train.py \
${CONFIG_FILE} \
[optional arguments]
During training, log files and checkpoints will be saved to the working directory, which is specified by work_dir in the config file or via CLI argument --work-dir.
By default, the model is evaluated on the validation set every epoch, the evaluation interval can be specified in the config file as shown below.
# evaluate the model every 12 epoch.
evaluation = dict(interval=12)
This tool accepts several optional arguments, including:
--no-validate(not suggested): Disable evaluation during training.--work-dir ${WORK_DIR}: Override the working directory.--resume-from ${CHECKPOINT_FILE}: Resume from a previous checkpoint file.--options 'Key=value': Overrides other settings in the used config.
Note:
Difference between resume-from and load-from:
resume-from loads both the model weights and optimizer status, and the epoch is also inherited from the specified checkpoint. It is usually used for resuming the training process that is interrupted accidentally.
load-from only loads the model weights and the training epoch starts from 0. It is usually used for finetuning.
Training on CPU¶
The process of training on the CPU is consistent with single GPU training. We just need to disable GPUs before the training process.
export CUDA_VISIBLE_DEVICES=-1
And then run the script above.
Note:
We do not recommend users to use CPU for training because it is too slow. We support this feature to allow users to debug on machines without GPU for convenience.
Training on multiple GPUs¶
We provide tools/dist_train.sh to launch training on multiple GPUs.
The basic usage is as follows.
bash ./tools/dist_train.sh \
${CONFIG_FILE} \
${GPU_NUM} \
[optional arguments]
Optional arguments remain the same as stated above.
Launch multiple jobs simultaneously¶
If you would like to launch multiple jobs on a single machine, e.g., 2 jobs of 4-GPU training on a machine with 8 GPUs, you need to specify different ports (29500 by default) for each job to avoid communication conflict.
If you use dist_train.sh to launch training jobs, you can set the port in commands.
CUDA_VISIBLE_DEVICES=0,1,2,3 PORT=29500 ./tools/dist_train.sh ${CONFIG_FILE} 4
CUDA_VISIBLE_DEVICES=4,5,6,7 PORT=29501 ./tools/dist_train.sh ${CONFIG_FILE} 4
Train with multiple machines¶
If you launch with multiple machines simply connected with ethernet, you can simply run following commands:
On the first machine:
NNODES=2 NODE_RANK=0 PORT=$MASTER_PORT MASTER_ADDR=$MASTER_ADDR sh tools/dist_train.sh $CONFIG $GPUS
On the second machine:
NNODES=2 NODE_RANK=1 PORT=$MASTER_PORT MASTER_ADDR=$MASTER_ADDR sh tools/dist_train.sh $CONFIG $GPUS
Usually it is slow if you do not have high speed networking like InfiniBand.
Manage jobs with Slurm¶
Slurm is a good job scheduling system for computing clusters.
On a cluster managed by Slurm, you can use slurm_train.sh to spawn training jobs. It supports both single-node and multi-node training.
The basic usage is as follows.
[GPUS=${GPUS}] ./tools/slurm_train.sh ${PARTITION} ${JOB_NAME} ${CONFIG_FILE} ${WORK_DIR}
Below is an example of using 16 GPUs to train Mask R-CNN on a Slurm partition named dev, and set the work-dir to some shared file systems.
GPUS=16 ./tools/slurm_train.sh dev mask_r50_1x configs/mask_rcnn_r50_fpn_1x_coco.py /nfs/xxxx/mask_rcnn_r50_fpn_1x
You can check the source code to review full arguments and environment variables.
When using Slurm, the port option need to be set in one of the following ways:
Set the port through
--options. This is more recommended since it does not change the original configs.CUDA_VISIBLE_DEVICES=0,1,2,3 GPUS=4 ./tools/slurm_train.sh ${PARTITION} ${JOB_NAME} config1.py ${WORK_DIR} --options 'dist_params.port=29500' CUDA_VISIBLE_DEVICES=4,5,6,7 GPUS=4 ./tools/slurm_train.sh ${PARTITION} ${JOB_NAME} config2.py ${WORK_DIR} --options 'dist_params.port=29501'
Modify the config files to set different communication ports.
In
config1.py, setdist_params = dict(backend='nccl', port=29500)
In
config2.py, setdist_params = dict(backend='nccl', port=29501)
Then you can launch two jobs with
config1.pyandconfig2.py.CUDA_VISIBLE_DEVICES=0,1,2,3 GPUS=4 ./tools/slurm_train.sh ${PARTITION} ${JOB_NAME} config1.py ${WORK_DIR} CUDA_VISIBLE_DEVICES=4,5,6,7 GPUS=4 ./tools/slurm_train.sh ${PARTITION} ${JOB_NAME} config2.py ${WORK_DIR}
2: Train with customized datasets¶
In this note, you will know how to inference, test, and train predefined models with customized datasets. We use the balloon dataset as an example to describe the whole process.
The basic steps are as below:
Prepare the customized dataset
Prepare a config
Train, test, inference models on the customized dataset.
Prepare the customized dataset¶
There are three ways to support a new dataset in MMDetection:
reorganize the dataset into COCO format.
reorganize the dataset into a middle format.
implement a new dataset.
Usually we recommend to use the first two methods which are usually easier than the third.
In this note, we give an example for converting the data into COCO format.
Note: MMDetection only supports evaluating mask AP of dataset in COCO format for now. So for instance segmentation task users should convert the data into coco format.
COCO annotation format¶
The necessary keys of COCO format for instance segmentation is as below, for the complete details, please refer here.
{
"images": [image],
"annotations": [annotation],
"categories": [category]
}
image = {
"id": int,
"width": int,
"height": int,
"file_name": str,
}
annotation = {
"id": int,
"image_id": int,
"category_id": int,
"segmentation": RLE or [polygon],
"area": float,
"bbox": [x,y,width,height],
"iscrowd": 0 or 1,
}
categories = [{
"id": int,
"name": str,
"supercategory": str,
}]
Assume we use the balloon dataset. After downloading the data, we need to implement a function to convert the annotation format into the COCO format. Then we can use implemented COCODataset to load the data and perform training and evaluation.
If you take a look at the dataset, you will find the dataset format is as below:
{'base64_img_data': '',
'file_attributes': {},
'filename': '34020010494_e5cb88e1c4_k.jpg',
'fileref': '',
'regions': {'0': {'region_attributes': {},
'shape_attributes': {'all_points_x': [1020,
1000,
994,
1003,
1023,
1050,
1089,
1134,
1190,
1265,
1321,
1361,
1403,
1428,
1442,
1445,
1441,
1427,
1400,
1361,
1316,
1269,
1228,
1198,
1207,
1210,
1190,
1177,
1172,
1174,
1170,
1153,
1127,
1104,
1061,
1032,
1020],
'all_points_y': [963,
899,
841,
787,
738,
700,
663,
638,
621,
619,
643,
672,
720,
765,
800,
860,
896,
942,
990,
1035,
1079,
1112,
1129,
1134,
1144,
1153,
1166,
1166,
1150,
1136,
1129,
1122,
1112,
1084,
1037,
989,
963],
'name': 'polygon'}}},
'size': 1115004}
The annotation is a JSON file where each key indicates an image’s all annotations. The code to convert the balloon dataset into coco format is as below.
import os.path as osp
import mmcv
def convert_balloon_to_coco(ann_file, out_file, image_prefix):
data_infos = mmcv.load(ann_file)
annotations = []
images = []
obj_count = 0
for idx, v in enumerate(mmcv.track_iter_progress(data_infos.values())):
filename = v['filename']
img_path = osp.join(image_prefix, filename)
height, width = mmcv.imread(img_path).shape[:2]
images.append(dict(
id=idx,
file_name=filename,
height=height,
width=width))
bboxes = []
labels = []
masks = []
for _, obj in v['regions'].items():
assert not obj['region_attributes']
obj = obj['shape_attributes']
px = obj['all_points_x']
py = obj['all_points_y']
poly = [(x + 0.5, y + 0.5) for x, y in zip(px, py)]
poly = [p for x in poly for p in x]
x_min, y_min, x_max, y_max = (
min(px), min(py), max(px), max(py))
data_anno = dict(
image_id=idx,
id=obj_count,
category_id=0,
bbox=[x_min, y_min, x_max - x_min, y_max - y_min],
area=(x_max - x_min) * (y_max - y_min),
segmentation=[poly],
iscrowd=0)
annotations.append(data_anno)
obj_count += 1
coco_format_json = dict(
images=images,
annotations=annotations,
categories=[{'id':0, 'name': 'balloon'}])
mmcv.dump(coco_format_json, out_file)
Using the function above, users can successfully convert the annotation file into json format, then we can use CocoDataset to train and evaluate the model.
Prepare a config¶
The second step is to prepare a config thus the dataset could be successfully loaded. Assume that we want to use Mask R-CNN with FPN, the config to train the detector on balloon dataset is as below. Assume the config is under directory configs/balloon/ and named as mask_rcnn_r50_caffe_fpn_mstrain-poly_1x_balloon.py, the config is as below.
# The new config inherits a base config to highlight the necessary modification
_base_ = 'mask_rcnn/mask_rcnn_r50_caffe_fpn_mstrain-poly_1x_coco.py'
# We also need to change the num_classes in head to match the dataset's annotation
model = dict(
roi_head=dict(
bbox_head=dict(num_classes=1),
mask_head=dict(num_classes=1)))
# Modify dataset related settings
dataset_type = 'COCODataset'
classes = ('balloon',)
data = dict(
train=dict(
img_prefix='balloon/train/',
classes=classes,
ann_file='balloon/train/annotation_coco.json'),
val=dict(
img_prefix='balloon/val/',
classes=classes,
ann_file='balloon/val/annotation_coco.json'),
test=dict(
img_prefix='balloon/val/',
classes=classes,
ann_file='balloon/val/annotation_coco.json'))
# We can use the pre-trained Mask RCNN model to obtain higher performance
load_from = 'checkpoints/mask_rcnn_r50_caffe_fpn_mstrain-poly_3x_coco_bbox_mAP-0.408__segm_mAP-0.37_20200504_163245-42aa3d00.pth'
This checkpoint file can be downloaded here
3: Train with customized models and standard datasets¶
In this note, you will know how to train, test and inference your own customized models under standard datasets. We use the cityscapes dataset to train a customized Cascade Mask R-CNN R50 model as an example to demonstrate the whole process, which using AugFPN to replace the default FPN as neck, and add Rotate or Translate as training-time auto augmentation.
The basic steps are as below:
Prepare the standard dataset
Prepare your own customized model
Prepare a config
Train, test, and inference models on the standard dataset.
Prepare the standard dataset¶
In this note, as we use the standard cityscapes dataset as an example.
It is recommended to symlink the dataset root to $MMDETECTION/data.
If your folder structure is different, you may need to change the corresponding paths in config files.
mmdetection
├── mmdet
├── tools
├── configs
├── data
│ ├── coco
│ │ ├── annotations
│ │ ├── train2017
│ │ ├── val2017
│ │ ├── test2017
│ ├── cityscapes
│ │ ├── annotations
│ │ ├── leftImg8bit
│ │ │ ├── train
│ │ │ ├── val
│ │ ├── gtFine
│ │ │ ├── train
│ │ │ ├── val
│ ├── VOCdevkit
│ │ ├── VOC2007
│ │ ├── VOC2012
Or you can set your dataset root through
export MMDET_DATASETS=$data_root
We will replace dataset root with $MMDET_DATASETS, so you don’t have to modify the corresponding path in config files.
The cityscapes annotations have to be converted into the coco format using tools/dataset_converters/cityscapes.py:
pip install cityscapesscripts
python tools/dataset_converters/cityscapes.py ./data/cityscapes --nproc 8 --out-dir ./data/cityscapes/annotations
Currently the config files in cityscapes use COCO pre-trained weights to initialize.
You could download the pre-trained models in advance if network is unavailable or slow, otherwise it would cause errors at the beginning of training.
Prepare your own customized model¶
The second step is to use your own module or training setting. Assume that we want to implement a new neck called AugFPN to replace with the default FPN under the existing detector Cascade Mask R-CNN R50. The following implementsAugFPN under MMDetection.
1. Define a new neck (e.g. AugFPN)¶
Firstly create a new file mmdet/models/necks/augfpn.py.
from ..builder import NECKS
@NECKS.register_module()
class AugFPN(nn.Module):
def __init__(self,
in_channels,
out_channels,
num_outs,
start_level=0,
end_level=-1,
add_extra_convs=False):
pass
def forward(self, inputs):
# implementation is ignored
pass
2. Import the module¶
You can either add the following line to mmdet/models/necks/__init__.py,
from .augfpn import AugFPN
or alternatively add
custom_imports = dict(
imports=['mmdet.models.necks.augfpn.py'],
allow_failed_imports=False)
to the config file and avoid modifying the original code.
3. Modify the config file¶
neck=dict(
type='AugFPN',
in_channels=[256, 512, 1024, 2048],
out_channels=256,
num_outs=5)
For more detailed usages about customize your own models (e.g. implement a new backbone, head, loss, etc) and runtime training settings (e.g. define a new optimizer, use gradient clip, customize training schedules and hooks, etc), please refer to the guideline Customize Models and Customize Runtime Settings respectively.
Prepare a config¶
The third step is to prepare a config for your own training setting. Assume that we want to add AugFPN and Rotate or Translate augmentation to existing Cascade Mask R-CNN R50 to train the cityscapes dataset, and assume the config is under directory configs/cityscapes/ and named as cascade_mask_rcnn_r50_augfpn_autoaug_10e_cityscapes.py, the config is as below.
# The new config inherits the base configs to highlight the necessary modification
_base_ = [
'../_base_/models/cascade_mask_rcnn_r50_fpn.py',
'../_base_/datasets/cityscapes_instance.py', '../_base_/default_runtime.py'
]
model = dict(
# set None to avoid loading ImageNet pretrained backbone,
# instead here we set `load_from` to load from COCO pretrained detectors.
backbone=dict(init_cfg=None),
# replace neck from defaultly `FPN` to our new implemented module `AugFPN`
neck=dict(
type='AugFPN',
in_channels=[256, 512, 1024, 2048],
out_channels=256,
num_outs=5),
# We also need to change the num_classes in head from 80 to 8, to match the
# cityscapes dataset's annotation. This modification involves `bbox_head` and `mask_head`.
roi_head=dict(
bbox_head=[
dict(
type='Shared2FCBBoxHead',
in_channels=256,
fc_out_channels=1024,
roi_feat_size=7,
# change the number of classes from defaultly COCO to cityscapes
num_classes=8,
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[0., 0., 0., 0.],
target_stds=[0.1, 0.1, 0.2, 0.2]),
reg_class_agnostic=True,
loss_cls=dict(
type='CrossEntropyLoss',
use_sigmoid=False,
loss_weight=1.0),
loss_bbox=dict(type='SmoothL1Loss', beta=1.0,
loss_weight=1.0)),
dict(
type='Shared2FCBBoxHead',
in_channels=256,
fc_out_channels=1024,
roi_feat_size=7,
# change the number of classes from defaultly COCO to cityscapes
num_classes=8,
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[0., 0., 0., 0.],
target_stds=[0.05, 0.05, 0.1, 0.1]),
reg_class_agnostic=True,
loss_cls=dict(
type='CrossEntropyLoss',
use_sigmoid=False,
loss_weight=1.0),
loss_bbox=dict(type='SmoothL1Loss', beta=1.0,
loss_weight=1.0)),
dict(
type='Shared2FCBBoxHead',
in_channels=256,
fc_out_channels=1024,
roi_feat_size=7,
# change the number of classes from defaultly COCO to cityscapes
num_classes=8,
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[0., 0., 0., 0.],
target_stds=[0.033, 0.033, 0.067, 0.067]),
reg_class_agnostic=True,
loss_cls=dict(
type='CrossEntropyLoss',
use_sigmoid=False,
loss_weight=1.0),
loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0))
],
mask_head=dict(
type='FCNMaskHead',
num_convs=4,
in_channels=256,
conv_out_channels=256,
# change the number of classes from defaultly COCO to cityscapes
num_classes=8,
loss_mask=dict(
type='CrossEntropyLoss', use_mask=True, loss_weight=1.0))))
# over-write `train_pipeline` for new added `AutoAugment` training setting
img_norm_cfg = dict(
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
train_pipeline = [
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=True, with_mask=True),
dict(
type='AutoAugment',
policies=[
[dict(
type='Rotate',
level=5,
img_fill_val=(124, 116, 104),
prob=0.5,
scale=1)
],
[dict(type='Rotate', level=7, img_fill_val=(124, 116, 104)),
dict(
type='Translate',
level=5,
prob=0.5,
img_fill_val=(124, 116, 104))
],
]),
dict(
type='Resize', img_scale=[(2048, 800), (2048, 1024)], keep_ratio=True),
dict(type='RandomFlip', flip_ratio=0.5),
dict(type='Normalize', **img_norm_cfg),
dict(type='Pad', size_divisor=32),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks']),
]
# set batch_size per gpu, and set new training pipeline
data = dict(
samples_per_gpu=1,
workers_per_gpu=3,
# over-write `pipeline` with new training pipeline setting
train=dict(dataset=dict(pipeline=train_pipeline)))
# Set optimizer
optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)
optimizer_config = dict(grad_clip=None)
# Set customized learning policy
lr_config = dict(
policy='step',
warmup='linear',
warmup_iters=500,
warmup_ratio=0.001,
step=[8])
runner = dict(type='EpochBasedRunner', max_epochs=10)
# We can use the COCO pretrained Cascade Mask R-CNN R50 model for more stable performance initialization
load_from = 'https://download.openmmlab.com/mmdetection/v2.0/cascade_rcnn/cascade_mask_rcnn_r50_fpn_1x_coco/cascade_mask_rcnn_r50_fpn_1x_coco_20200203-9d4dcb24.pth'
Train a new model¶
To train a model with the new config, you can simply run
python tools/train.py configs/cityscapes/cascade_mask_rcnn_r50_augfpn_autoaug_10e_cityscapes.py
For more detailed usages, please refer to the Case 1.
Test and inference¶
To test the trained model, you can simply run
python tools/test.py configs/cityscapes/cascade_mask_rcnn_r50_augfpn_autoaug_10e_cityscapes.py work_dirs/cascade_mask_rcnn_r50_augfpn_autoaug_10e_cityscapes.py/latest.pth --eval bbox segm
For more detailed usages, please refer to the Case 1.
Tutorial 1: Learn about Configs¶
We incorporate modular and inheritance design into our config system, which is convenient to conduct various experiments.
If you wish to inspect the config file, you may run python tools/misc/print_config.py /PATH/TO/CONFIG to see the complete config.
Modify config through script arguments¶
When submitting jobs using “tools/train.py” or “tools/test.py”, you may specify --cfg-options to in-place modify the config.
Update config keys of dict chains.
The config options can be specified following the order of the dict keys in the original config. For example,
--cfg-options model.backbone.norm_eval=Falsechanges the all BN modules in model backbones totrainmode.Update keys inside a list of configs.
Some config dicts are composed as a list in your config. For example, the training pipeline
data.train.pipelineis normally a list e.g.[dict(type='LoadImageFromFile'), ...]. If you want to change'LoadImageFromFile'to'LoadImageFromWebcam'in the pipeline, you may specify--cfg-options data.train.pipeline.0.type=LoadImageFromWebcam.Update values of list/tuples.
If the value to be updated is a list or a tuple. For example, the config file normally sets
workflow=[('train', 1)]. If you want to change this key, you may specify--cfg-options workflow="[(train,1),(val,1)]". Note that the quotation mark ” is necessary to support list/tuple data types, and that NO white space is allowed inside the quotation marks in the specified value.
Config File Structure¶
There are 4 basic component types under config/_base_, dataset, model, schedule, default_runtime.
Many methods could be easily constructed with one of each like Faster R-CNN, Mask R-CNN, Cascade R-CNN, RPN, SSD.
The configs that are composed by components from _base_ are called primitive.
For all configs under the same folder, it is recommended to have only one primitive config. All other configs should inherit from the primitive config. In this way, the maximum of inheritance level is 3.
For easy understanding, we recommend contributors to inherit from existing methods.
For example, if some modification is made base on Faster R-CNN, user may first inherit the basic Faster R-CNN structure by specifying _base_ = ../faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py, then modify the necessary fields in the config files.
If you are building an entirely new method that does not share the structure with any of the existing methods, you may create a folder xxx_rcnn under configs,
Please refer to mmcv for detailed documentation.
Config Name Style¶
We follow the below style to name config files. Contributors are advised to follow the same style.
{model}_[model setting]_{backbone}_{neck}_[norm setting]_[misc]_[gpu x batch_per_gpu]_{schedule}_{dataset}
{xxx} is required field and [yyy] is optional.
{model}: model type likefaster_rcnn,mask_rcnn, etc.[model setting]: specific setting for some model, likewithout_semanticforhtc,momentforreppoints, etc.{backbone}: backbone type liker50(ResNet-50),x101(ResNeXt-101).{neck}: neck type likefpn,pafpn,nasfpn,c4.[norm_setting]:bn(Batch Normalization) is used unless specified, other norm layer type could begn(Group Normalization),syncbn(Synchronized Batch Normalization).gn-head/gn-neckindicates GN is applied in head/neck only, whilegn-allmeans GN is applied in the entire model, e.g. backbone, neck, head.[misc]: miscellaneous setting/plugins of model, e.g.dconv,gcb,attention,albu,mstrain.[gpu x batch_per_gpu]: GPUs and samples per GPU,8x2is used by default.{schedule}: training schedule, options are1x,2x,20e, etc.1xand2xmeans 12 epochs and 24 epochs respectively.20eis adopted in cascade models, which denotes 20 epochs. For1x/2x, initial learning rate decays by a factor of 10 at the 8/16th and 11/22th epochs. For20e, initial learning rate decays by a factor of 10 at the 16th and 19th epochs.{dataset}: dataset likecoco,cityscapes,voc_0712,wider_face.
Deprecated train_cfg/test_cfg¶
The train_cfg and test_cfg are deprecated in config file, please specify them in the model config. The original config structure is as below.
# deprecated
model = dict(
type=...,
...
)
train_cfg=dict(...)
test_cfg=dict(...)
The migration example is as below.
# recommended
model = dict(
type=...,
...
train_cfg=dict(...),
test_cfg=dict(...),
)
An Example of Mask R-CNN¶
To help the users have a basic idea of a complete config and the modules in a modern detection system, we make brief comments on the config of Mask R-CNN using ResNet50 and FPN as the following. For more detailed usage and the corresponding alternative for each modules, please refer to the API documentation.
model = dict(
type='MaskRCNN', # The name of detector
backbone=dict( # The config of backbone
type='ResNet', # The type of the backbone, refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/backbones/resnet.py#L308 for more details.
depth=50, # The depth of backbone, usually it is 50 or 101 for ResNet and ResNext backbones.
num_stages=4, # Number of stages of the backbone.
out_indices=(0, 1, 2, 3), # The index of output feature maps produced in each stages
frozen_stages=1, # The weights in the first 1 stage are frozen
norm_cfg=dict( # The config of normalization layers.
type='BN', # Type of norm layer, usually it is BN or GN
requires_grad=True), # Whether to train the gamma and beta in BN
norm_eval=True, # Whether to freeze the statistics in BN
style='pytorch', # The style of backbone, 'pytorch' means that stride 2 layers are in 3x3 conv, 'caffe' means stride 2 layers are in 1x1 convs.
init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet50')), # The ImageNet pretrained backbone to be loaded
neck=dict(
type='FPN', # The neck of detector is FPN. We also support 'NASFPN', 'PAFPN', etc. Refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/necks/fpn.py#L10 for more details.
in_channels=[256, 512, 1024, 2048], # The input channels, this is consistent with the output channels of backbone
out_channels=256, # The output channels of each level of the pyramid feature map
num_outs=5), # The number of output scales
rpn_head=dict(
type='RPNHead', # The type of RPN head is 'RPNHead', we also support 'GARPNHead', etc. Refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/dense_heads/rpn_head.py#L12 for more details.
in_channels=256, # The input channels of each input feature map, this is consistent with the output channels of neck
feat_channels=256, # Feature channels of convolutional layers in the head.
anchor_generator=dict( # The config of anchor generator
type='AnchorGenerator', # Most of methods use AnchorGenerator, SSD Detectors uses `SSDAnchorGenerator`. Refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/core/anchor/anchor_generator.py#L10 for more details
scales=[8], # Basic scale of the anchor, the area of the anchor in one position of a feature map will be scale * base_sizes
ratios=[0.5, 1.0, 2.0], # The ratio between height and width.
strides=[4, 8, 16, 32, 64]), # The strides of the anchor generator. This is consistent with the FPN feature strides. The strides will be taken as base_sizes if base_sizes is not set.
bbox_coder=dict( # Config of box coder to encode and decode the boxes during training and testing
type='DeltaXYWHBBoxCoder', # Type of box coder. 'DeltaXYWHBBoxCoder' is applied for most of methods. Refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/core/bbox/coder/delta_xywh_bbox_coder.py#L9 for more details.
target_means=[0.0, 0.0, 0.0, 0.0], # The target means used to encode and decode boxes
target_stds=[1.0, 1.0, 1.0, 1.0]), # The standard variance used to encode and decode boxes
loss_cls=dict( # Config of loss function for the classification branch
type='CrossEntropyLoss', # Type of loss for classification branch, we also support FocalLoss etc.
use_sigmoid=True, # RPN usually perform two-class classification, so it usually uses sigmoid function.
loss_weight=1.0), # Loss weight of the classification branch.
loss_bbox=dict( # Config of loss function for the regression branch.
type='L1Loss', # Type of loss, we also support many IoU Losses and smooth L1-loss, etc. Refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/losses/smooth_l1_loss.py#L56 for implementation.
loss_weight=1.0)), # Loss weight of the regression branch.
roi_head=dict( # RoIHead encapsulates the second stage of two-stage/cascade detectors.
type='StandardRoIHead', # Type of the RoI head. Refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/roi_heads/standard_roi_head.py#L10 for implementation.
bbox_roi_extractor=dict( # RoI feature extractor for bbox regression.
type='SingleRoIExtractor', # Type of the RoI feature extractor, most of methods uses SingleRoIExtractor. Refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/roi_heads/roi_extractors/single_level.py#L10 for details.
roi_layer=dict( # Config of RoI Layer
type='RoIAlign', # Type of RoI Layer, DeformRoIPoolingPack and ModulatedDeformRoIPoolingPack are also supported. Refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/ops/roi_align/roi_align.py#L79 for details.
output_size=7, # The output size of feature maps.
sampling_ratio=0), # Sampling ratio when extracting the RoI features. 0 means adaptive ratio.
out_channels=256, # output channels of the extracted feature.
featmap_strides=[4, 8, 16, 32]), # Strides of multi-scale feature maps. It should be consistent to the architecture of the backbone.
bbox_head=dict( # Config of box head in the RoIHead.
type='Shared2FCBBoxHead', # Type of the bbox head, Refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/roi_heads/bbox_heads/convfc_bbox_head.py#L177 for implementation details.
in_channels=256, # Input channels for bbox head. This is consistent with the out_channels in roi_extractor
fc_out_channels=1024, # Output feature channels of FC layers.
roi_feat_size=7, # Size of RoI features
num_classes=80, # Number of classes for classification
bbox_coder=dict( # Box coder used in the second stage.
type='DeltaXYWHBBoxCoder', # Type of box coder. 'DeltaXYWHBBoxCoder' is applied for most of methods.
target_means=[0.0, 0.0, 0.0, 0.0], # Means used to encode and decode box
target_stds=[0.1, 0.1, 0.2, 0.2]), # Standard variance for encoding and decoding. It is smaller since the boxes are more accurate. [0.1, 0.1, 0.2, 0.2] is a conventional setting.
reg_class_agnostic=False, # Whether the regression is class agnostic.
loss_cls=dict( # Config of loss function for the classification branch
type='CrossEntropyLoss', # Type of loss for classification branch, we also support FocalLoss etc.
use_sigmoid=False, # Whether to use sigmoid.
loss_weight=1.0), # Loss weight of the classification branch.
loss_bbox=dict( # Config of loss function for the regression branch.
type='L1Loss', # Type of loss, we also support many IoU Losses and smooth L1-loss, etc.
loss_weight=1.0)), # Loss weight of the regression branch.
mask_roi_extractor=dict( # RoI feature extractor for mask generation.
type='SingleRoIExtractor', # Type of the RoI feature extractor, most of methods uses SingleRoIExtractor.
roi_layer=dict( # Config of RoI Layer that extracts features for instance segmentation
type='RoIAlign', # Type of RoI Layer, DeformRoIPoolingPack and ModulatedDeformRoIPoolingPack are also supported
output_size=14, # The output size of feature maps.
sampling_ratio=0), # Sampling ratio when extracting the RoI features.
out_channels=256, # Output channels of the extracted feature.
featmap_strides=[4, 8, 16, 32]), # Strides of multi-scale feature maps.
mask_head=dict( # Mask prediction head
type='FCNMaskHead', # Type of mask head, refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/roi_heads/mask_heads/fcn_mask_head.py#L21 for implementation details.
num_convs=4, # Number of convolutional layers in mask head.
in_channels=256, # Input channels, should be consistent with the output channels of mask roi extractor.
conv_out_channels=256, # Output channels of the convolutional layer.
num_classes=80, # Number of class to be segmented.
loss_mask=dict( # Config of loss function for the mask branch.
type='CrossEntropyLoss', # Type of loss used for segmentation
use_mask=True, # Whether to only train the mask in the correct class.
loss_weight=1.0))), # Loss weight of mask branch.
train_cfg = dict( # Config of training hyperparameters for rpn and rcnn
rpn=dict( # Training config of rpn
assigner=dict( # Config of assigner
type='MaxIoUAssigner', # Type of assigner, MaxIoUAssigner is used for many common detectors. Refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/core/bbox/assigners/max_iou_assigner.py#L10 for more details.
pos_iou_thr=0.7, # IoU >= threshold 0.7 will be taken as positive samples
neg_iou_thr=0.3, # IoU < threshold 0.3 will be taken as negative samples
min_pos_iou=0.3, # The minimal IoU threshold to take boxes as positive samples
match_low_quality=True, # Whether to match the boxes under low quality (see API doc for more details).
ignore_iof_thr=-1), # IoF threshold for ignoring bboxes
sampler=dict( # Config of positive/negative sampler
type='RandomSampler', # Type of sampler, PseudoSampler and other samplers are also supported. Refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/core/bbox/samplers/random_sampler.py#L8 for implementation details.
num=256, # Number of samples
pos_fraction=0.5, # The ratio of positive samples in the total samples.
neg_pos_ub=-1, # The upper bound of negative samples based on the number of positive samples.
add_gt_as_proposals=False), # Whether add GT as proposals after sampling.
allowed_border=-1, # The border allowed after padding for valid anchors.
pos_weight=-1, # The weight of positive samples during training.
debug=False), # Whether to set the debug mode
rpn_proposal=dict( # The config to generate proposals during training
nms_across_levels=False, # Whether to do NMS for boxes across levels. Only work in `GARPNHead`, naive rpn does not support do nms cross levels.
nms_pre=2000, # The number of boxes before NMS
nms_post=1000, # The number of boxes to be kept by NMS, Only work in `GARPNHead`.
max_per_img=1000, # The number of boxes to be kept after NMS.
nms=dict( # Config of NMS
type='nms', # Type of NMS
iou_threshold=0.7 # NMS threshold
),
min_bbox_size=0), # The allowed minimal box size
rcnn=dict( # The config for the roi heads.
assigner=dict( # Config of assigner for second stage, this is different for that in rpn
type='MaxIoUAssigner', # Type of assigner, MaxIoUAssigner is used for all roi_heads for now. Refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/core/bbox/assigners/max_iou_assigner.py#L10 for more details.
pos_iou_thr=0.5, # IoU >= threshold 0.5 will be taken as positive samples
neg_iou_thr=0.5, # IoU < threshold 0.5 will be taken as negative samples
min_pos_iou=0.5, # The minimal IoU threshold to take boxes as positive samples
match_low_quality=False, # Whether to match the boxes under low quality (see API doc for more details).
ignore_iof_thr=-1), # IoF threshold for ignoring bboxes
sampler=dict(
type='RandomSampler', # Type of sampler, PseudoSampler and other samplers are also supported. Refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/core/bbox/samplers/random_sampler.py#L8 for implementation details.
num=512, # Number of samples
pos_fraction=0.25, # The ratio of positive samples in the total samples.
neg_pos_ub=-1, # The upper bound of negative samples based on the number of positive samples.
add_gt_as_proposals=True
), # Whether add GT as proposals after sampling.
mask_size=28, # Size of mask
pos_weight=-1, # The weight of positive samples during training.
debug=False)), # Whether to set the debug mode
test_cfg = dict( # Config for testing hyperparameters for rpn and rcnn
rpn=dict( # The config to generate proposals during testing
nms_across_levels=False, # Whether to do NMS for boxes across levels. Only work in `GARPNHead`, naive rpn does not support do nms cross levels.
nms_pre=1000, # The number of boxes before NMS
nms_post=1000, # The number of boxes to be kept by NMS, Only work in `GARPNHead`.
max_per_img=1000, # The number of boxes to be kept after NMS.
nms=dict( # Config of NMS
type='nms', #Type of NMS
iou_threshold=0.7 # NMS threshold
),
min_bbox_size=0), # The allowed minimal box size
rcnn=dict( # The config for the roi heads.
score_thr=0.05, # Threshold to filter out boxes
nms=dict( # Config of NMS in the second stage
type='nms', # Type of NMS
iou_thr=0.5), # NMS threshold
max_per_img=100, # Max number of detections of each image
mask_thr_binary=0.5))) # Threshold of mask prediction
dataset_type = 'CocoDataset' # Dataset type, this will be used to define the dataset
data_root = 'data/coco/' # Root path of data
img_norm_cfg = dict( # Image normalization config to normalize the input images
mean=[123.675, 116.28, 103.53], # Mean values used to pre-training the pre-trained backbone models
std=[58.395, 57.12, 57.375], # Standard variance used to pre-training the pre-trained backbone models
to_rgb=True
) # The channel orders of image used to pre-training the pre-trained backbone models
train_pipeline = [ # Training pipeline
dict(type='LoadImageFromFile'), # First pipeline to load images from file path
dict(
type='LoadAnnotations', # Second pipeline to load annotations for current image
with_bbox=True, # Whether to use bounding box, True for detection
with_mask=True, # Whether to use instance mask, True for instance segmentation
poly2mask=False), # Whether to convert the polygon mask to instance mask, set False for acceleration and to save memory
dict(
type='Resize', # Augmentation pipeline that resize the images and their annotations
img_scale=(1333, 800), # The largest scale of image
keep_ratio=True
), # whether to keep the ratio between height and width.
dict(
type='RandomFlip', # Augmentation pipeline that flip the images and their annotations
flip_ratio=0.5), # The ratio or probability to flip
dict(
type='Normalize', # Augmentation pipeline that normalize the input images
mean=[123.675, 116.28, 103.53], # These keys are the same of img_norm_cfg since the
std=[58.395, 57.12, 57.375], # keys of img_norm_cfg are used here as arguments
to_rgb=True),
dict(
type='Pad', # Padding config
size_divisor=32), # The number the padded images should be divisible
dict(type='DefaultFormatBundle'), # Default format bundle to gather data in the pipeline
dict(
type='Collect', # Pipeline that decides which keys in the data should be passed to the detector
keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks'])
]
test_pipeline = [
dict(type='LoadImageFromFile'), # First pipeline to load images from file path
dict(
type='MultiScaleFlipAug', # An encapsulation that encapsulates the testing augmentations
img_scale=(1333, 800), # Decides the largest scale for testing, used for the Resize pipeline
flip=False, # Whether to flip images during testing
transforms=[
dict(type='Resize', # Use resize augmentation
keep_ratio=True), # Whether to keep the ratio between height and width, the img_scale set here will be suppressed by the img_scale set above.
dict(type='RandomFlip'), # Thought RandomFlip is added in pipeline, it is not used because flip=False
dict(
type='Normalize', # Normalization config, the values are from img_norm_cfg
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True),
dict(
type='Pad', # Padding config to pad images divisible by 32.
size_divisor=32),
dict(
type='ImageToTensor', # convert image to tensor
keys=['img']),
dict(
type='Collect', # Collect pipeline that collect necessary keys for testing.
keys=['img'])
])
]
data = dict(
samples_per_gpu=2, # Batch size of a single GPU
workers_per_gpu=2, # Worker to pre-fetch data for each single GPU
train=dict( # Train dataset config
type='CocoDataset', # Type of dataset, refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/datasets/coco.py#L19 for details.
ann_file='data/coco/annotations/instances_train2017.json', # Path of annotation file
img_prefix='data/coco/train2017/', # Prefix of image path
pipeline=[ # pipeline, this is passed by the train_pipeline created before.
dict(type='LoadImageFromFile'),
dict(
type='LoadAnnotations',
with_bbox=True,
with_mask=True,
poly2mask=False),
dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
dict(type='RandomFlip', flip_ratio=0.5),
dict(
type='Normalize',
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True),
dict(type='Pad', size_divisor=32),
dict(type='DefaultFormatBundle'),
dict(
type='Collect',
keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks'])
]),
val=dict( # Validation dataset config
type='CocoDataset',
ann_file='data/coco/annotations/instances_val2017.json',
img_prefix='data/coco/val2017/',
pipeline=[ # Pipeline is passed by test_pipeline created before
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(1333, 800),
flip=False,
transforms=[
dict(type='Resize', keep_ratio=True),
dict(type='RandomFlip'),
dict(
type='Normalize',
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True),
dict(type='Pad', size_divisor=32),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect', keys=['img'])
])
]),
test=dict( # Test dataset config, modify the ann_file for test-dev/test submission
type='CocoDataset',
ann_file='data/coco/annotations/instances_val2017.json',
img_prefix='data/coco/val2017/',
pipeline=[ # Pipeline is passed by test_pipeline created before
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(1333, 800),
flip=False,
transforms=[
dict(type='Resize', keep_ratio=True),
dict(type='RandomFlip'),
dict(
type='Normalize',
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True),
dict(type='Pad', size_divisor=32),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect', keys=['img'])
])
],
samples_per_gpu=2 # Batch size of a single GPU used in testing
))
evaluation = dict( # The config to build the evaluation hook, refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/core/evaluation/eval_hooks.py#L7 for more details.
interval=1, # Evaluation interval
metric=['bbox', 'segm']) # Metrics used during evaluation
optimizer = dict( # Config used to build optimizer, support all the optimizers in PyTorch whose arguments are also the same as those in PyTorch
type='SGD', # Type of optimizers, refer to https://github.com/open-mmlab/mmdetection/blob/master/mmdet/core/optimizer/default_constructor.py#L13 for more details
lr=0.02, # Learning rate of optimizers, see detail usages of the parameters in the documentation of PyTorch
momentum=0.9, # Momentum
weight_decay=0.0001) # Weight decay of SGD
optimizer_config = dict( # Config used to build the optimizer hook, refer to https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/hooks/optimizer.py#L8 for implementation details.
grad_clip=None) # Most of the methods do not use gradient clip
lr_config = dict( # Learning rate scheduler config used to register LrUpdater hook
policy='step', # The policy of scheduler, also support CosineAnnealing, Cyclic, etc. Refer to details of supported LrUpdater from https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/hooks/lr_updater.py#L9.
warmup='linear', # The warmup policy, also support `exp` and `constant`.
warmup_iters=500, # The number of iterations for warmup
warmup_ratio=
0.001, # The ratio of the starting learning rate used for warmup
step=[8, 11]) # Steps to decay the learning rate
runner = dict(
type='EpochBasedRunner', # Type of runner to use (i.e. IterBasedRunner or EpochBasedRunner)
max_epochs=12) # Runner that runs the workflow in total max_epochs. For IterBasedRunner use `max_iters`
checkpoint_config = dict( # Config to set the checkpoint hook, Refer to https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/hooks/checkpoint.py for implementation.
interval=1) # The save interval is 1
log_config = dict( # config to register logger hook
interval=50, # Interval to print the log
hooks=[
dict(type='TextLoggerHook', by_epoch=False),
dict(type='TensorboardLoggerHook', by_epoch=False),
dict(type='MMDetWandbHook', by_epoch=False, # The Wandb logger is also supported, It requires `wandb` to be installed.
init_kwargs={'entity': "OpenMMLab", # The entity used to log on Wandb
'project': "MMDet", # Project name in WandB
'config': cfg_dict}), # Check https://docs.wandb.ai/ref/python/init for more init arguments.
# MMDetWandbHook is mmdet implementation of WandbLoggerHook. ClearMLLoggerHook, DvcliveLoggerHook, MlflowLoggerHook, NeptuneLoggerHook, PaviLoggerHook, SegmindLoggerHook are also supported based on MMCV implementation.
]) # The logger used to record the training process.
dist_params = dict(backend='nccl') # Parameters to setup distributed training, the port can also be set.
log_level = 'INFO' # The level of logging.
load_from = None # load models as a pre-trained model from a given path. This will not resume training.
resume_from = None # Resume checkpoints from a given path, the training will be resumed from the epoch when the checkpoint's is saved.
workflow = [('train', 1)] # Workflow for runner. [('train', 1)] means there is only one workflow and the workflow named 'train' is executed once. The workflow trains the model by 12 epochs according to the total_epochs.
work_dir = 'work_dir' # Directory to save the model checkpoints and logs for the current experiments.
FAQ¶
Ignore some fields in the base configs¶
Sometimes, you may set _delete_=True to ignore some of fields in base configs.
You may refer to mmcv for simple illustration.
In MMDetection, for example, to change the backbone of Mask R-CNN with the following config.
model = dict(
type='MaskRCNN',
pretrained='torchvision://resnet50',
backbone=dict(
type='ResNet',
depth=50,
num_stages=4,
out_indices=(0, 1, 2, 3),
frozen_stages=1,
norm_cfg=dict(type='BN', requires_grad=True),
norm_eval=True,
style='pytorch'),
neck=dict(...),
rpn_head=dict(...),
roi_head=dict(...))
ResNet and HRNet use different keywords to construct.
_base_ = '../mask_rcnn/mask_rcnn_r50_fpn_1x_coco.py'
model = dict(
pretrained='open-mmlab://msra/hrnetv2_w32',
backbone=dict(
_delete_=True,
type='HRNet',
extra=dict(
stage1=dict(
num_modules=1,
num_branches=1,
block='BOTTLENECK',
num_blocks=(4, ),
num_channels=(64, )),
stage2=dict(
num_modules=1,
num_branches=2,
block='BASIC',
num_blocks=(4, 4),
num_channels=(32, 64)),
stage3=dict(
num_modules=4,
num_branches=3,
block='BASIC',
num_blocks=(4, 4, 4),
num_channels=(32, 64, 128)),
stage4=dict(
num_modules=3,
num_branches=4,
block='BASIC',
num_blocks=(4, 4, 4, 4),
num_channels=(32, 64, 128, 256)))),
neck=dict(...))
The _delete_=True would replace all old keys in backbone field with new keys.
Use intermediate variables in configs¶
Some intermediate variables are used in the configs files, like train_pipeline/test_pipeline in datasets.
It’s worth noting that when modifying intermediate variables in the children configs, user need to pass the intermediate variables into corresponding fields again.
For example, we would like to use multi scale strategy to train a Mask R-CNN. train_pipeline/test_pipeline are intermediate variable we would like modify.
_base_ = './mask_rcnn_r50_fpn_1x_coco.py'
img_norm_cfg = dict(
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
train_pipeline = [
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=True, with_mask=True),
dict(
type='Resize',
img_scale=[(1333, 640), (1333, 672), (1333, 704), (1333, 736),
(1333, 768), (1333, 800)],
multiscale_mode="value",
keep_ratio=True),
dict(type='RandomFlip', flip_ratio=0.5),
dict(type='Normalize', **img_norm_cfg),
dict(type='Pad', size_divisor=32),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks']),
]
test_pipeline = [
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(1333, 800),
flip=False,
transforms=[
dict(type='Resize', keep_ratio=True),
dict(type='RandomFlip'),
dict(type='Normalize', **img_norm_cfg),
dict(type='Pad', size_divisor=32),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect', keys=['img']),
])
]
data = dict(
train=dict(pipeline=train_pipeline),
val=dict(pipeline=test_pipeline),
test=dict(pipeline=test_pipeline))
We first define the new train_pipeline/test_pipeline and pass them into data.
Similarly, if we would like to switch from SyncBN to BN or MMSyncBN, we need to substitute every norm_cfg in the config.
_base_ = './mask_rcnn_r50_fpn_1x_coco.py'
norm_cfg = dict(type='BN', requires_grad=True)
model = dict(
backbone=dict(norm_cfg=norm_cfg),
neck=dict(norm_cfg=norm_cfg),
...)
Tutorial 2: Customize Datasets¶
Support new data format¶
To support a new data format, you can either convert them to existing formats (COCO format or PASCAL format) or directly convert them to the middle format. You could also choose to convert them offline (before training by a script) or online (implement a new dataset and do the conversion at training). In MMDetection, we recommend to convert the data into COCO formats and do the conversion offline, thus you only need to modify the config’s data annotation paths and classes after the conversion of your data.
Reorganize new data formats to existing format¶
The simplest way is to convert your dataset to existing dataset formats (COCO or PASCAL VOC).
The annotation json files in COCO format has the following necessary keys:
'images': [
{
'file_name': 'COCO_val2014_000000001268.jpg',
'height': 427,
'width': 640,
'id': 1268
},
...
],
'annotations': [
{
'segmentation': [[192.81,
247.09,
...
219.03,
249.06]], # if you have mask labels
'area': 1035.749,
'iscrowd': 0,
'image_id': 1268,
'bbox': [192.81, 224.8, 74.73, 33.43],
'category_id': 16,
'id': 42986
},
...
],
'categories': [
{'id': 0, 'name': 'car'},
]
There are three necessary keys in the json file:
images: contains a list of images with their information likefile_name,height,width, andid.annotations: contains the list of instance annotations.categories: contains the list of categories names and their ID.
After the data pre-processing, there are two steps for users to train the customized new dataset with existing format (e.g. COCO format):
Modify the config file for using the customized dataset.
Check the annotations of the customized dataset.
Here we give an example to show the above two steps, which uses a customized dataset of 5 classes with COCO format to train an existing Cascade Mask R-CNN R50-FPN detector.
1. Modify the config file for using the customized dataset¶
There are two aspects involved in the modification of config file:
The
datafield. Specifically, you need to explicitly add theclassesfields indata.train,data.valanddata.test.The
num_classesfield in themodelpart. Explicitly over-write all thenum_classesfrom default value (e.g. 80 in COCO) to your classes number.
In configs/my_custom_config.py:
# the new config inherits the base configs to highlight the necessary modification
_base_ = './cascade_mask_rcnn_r50_fpn_1x_coco.py'
# 1. dataset settings
dataset_type = 'CocoDataset'
classes = ('a', 'b', 'c', 'd', 'e')
data = dict(
samples_per_gpu=2,
workers_per_gpu=2,
train=dict(
type=dataset_type,
# explicitly add your class names to the field `classes`
classes=classes,
ann_file='path/to/your/train/annotation_data',
img_prefix='path/to/your/train/image_data'),
val=dict(
type=dataset_type,
# explicitly add your class names to the field `classes`
classes=classes,
ann_file='path/to/your/val/annotation_data',
img_prefix='path/to/your/val/image_data'),
test=dict(
type=dataset_type,
# explicitly add your class names to the field `classes`
classes=classes,
ann_file='path/to/your/test/annotation_data',
img_prefix='path/to/your/test/image_data'))
# 2. model settings
# explicitly over-write all the `num_classes` field from default 80 to 5.
model = dict(
roi_head=dict(
bbox_head=[
dict(
type='Shared2FCBBoxHead',
# explicitly over-write all the `num_classes` field from default 80 to 5.
num_classes=5),
dict(
type='Shared2FCBBoxHead',
# explicitly over-write all the `num_classes` field from default 80 to 5.
num_classes=5),
dict(
type='Shared2FCBBoxHead',
# explicitly over-write all the `num_classes` field from default 80 to 5.
num_classes=5)],
# explicitly over-write all the `num_classes` field from default 80 to 5.
mask_head=dict(num_classes=5)))
2. Check the annotations of the customized dataset¶
Assuming your customized dataset is COCO format, make sure you have the correct annotations in the customized dataset:
The length for
categoriesfield in annotations should exactly equal the tuple length ofclassesfields in your config, meaning the number of classes (e.g. 5 in this example).The
classesfields in your config file should have exactly the same elements and the same order with thenameincategoriesof annotations. MMDetection automatically maps the uncontinuousidincategoriesto the continuous label indices, so the string order ofnameincategoriesfield affects the order of label indices. Meanwhile, the string order ofclassesin config affects the label text during visualization of predicted bounding boxes.The
category_idinannotationsfield should be valid, i.e., all values incategory_idshould belong toidincategories.
Here is a valid example of annotations:
'annotations': [
{
'segmentation': [[192.81,
247.09,
...
219.03,
249.06]], # if you have mask labels
'area': 1035.749,
'iscrowd': 0,
'image_id': 1268,
'bbox': [192.81, 224.8, 74.73, 33.43],
'category_id': 16,
'id': 42986
},
...
],
# MMDetection automatically maps the uncontinuous `id` to the continuous label indices.
'categories': [
{'id': 1, 'name': 'a'}, {'id': 3, 'name': 'b'}, {'id': 4, 'name': 'c'}, {'id': 16, 'name': 'd'}, {'id': 17, 'name': 'e'},
]
We use this way to support CityScapes dataset. The script is in cityscapes.py and we also provide the finetuning configs.
Note
For instance segmentation datasets, MMDetection only supports evaluating mask AP of dataset in COCO format for now.
It is recommended to convert the data offline before training, thus you can still use
CocoDatasetand only need to modify the path of annotations and the training classes.
Reorganize new data format to middle format¶
It is also fine if you do not want to convert the annotation format to COCO or PASCAL format. Actually, we define a simple annotation format and all existing datasets are processed to be compatible with it, either online or offline.
The annotation of a dataset is a list of dict, each dict corresponds to an image.
There are 3 field filename (relative path), width, height for testing,
and an additional field ann for training. ann is also a dict containing at least 2 fields:
bboxes and labels, both of which are numpy arrays. Some datasets may provide
annotations like crowd/difficult/ignored bboxes, we use bboxes_ignore and labels_ignore
to cover them.
Here is an example.
[
{
'filename': 'a.jpg',
'width': 1280,
'height': 720,
'ann': {
'bboxes': <np.ndarray, float32> (n, 4),
'labels': <np.ndarray, int64> (n, ),
'bboxes_ignore': <np.ndarray, float32> (k, 4),
'labels_ignore': <np.ndarray, int64> (k, ) (optional field)
}
},
...
]
There are two ways to work with custom datasets.
online conversion
You can write a new Dataset class inherited from
CustomDataset, and overwrite two methodsload_annotations(self, ann_file)andget_ann_info(self, idx), like CocoDataset and VOCDataset.offline conversion
You can convert the annotation format to the expected format above and save it to a pickle or json file, like pascal_voc.py. Then you can simply use
CustomDataset.
An example of customized dataset¶
Assume the annotation is in a new format in text files.
The bounding boxes annotations are stored in text file annotation.txt as the following
#
000001.jpg
1280 720
2
10 20 40 60 1
20 40 50 60 2
#
000002.jpg
1280 720
3
50 20 40 60 2
20 40 30 45 2
30 40 50 60 3
We can create a new dataset in mmdet/datasets/my_dataset.py to load the data.
import mmcv
import numpy as np
from .builder import DATASETS
from .custom import CustomDataset
@DATASETS.register_module()
class MyDataset(CustomDataset):
CLASSES = ('person', 'bicycle', 'car', 'motorcycle')
def load_annotations(self, ann_file):
ann_list = mmcv.list_from_file(ann_file)
data_infos = []
for i, ann_line in enumerate(ann_list):
if ann_line != '#':
continue
img_shape = ann_list[i + 2].split(' ')
width = int(img_shape[0])
height = int(img_shape[1])
bbox_number = int(ann_list[i + 3])
anns = ann_line.split(' ')
bboxes = []
labels = []
for anns in ann_list[i + 4:i + 4 + bbox_number]:
bboxes.append([float(ann) for ann in anns[:4]])
labels.append(int(anns[4]))
data_infos.append(
dict(
filename=ann_list[i + 1],
width=width,
height=height,
ann=dict(
bboxes=np.array(bboxes).astype(np.float32),
labels=np.array(labels).astype(np.int64))
))
return data_infos
def get_ann_info(self, idx):
return self.data_infos[idx]['ann']
Then in the config, to use MyDataset you can modify the config as the following
dataset_A_train = dict(
type='MyDataset',
ann_file = 'image_list.txt',
pipeline=train_pipeline
)
Customize datasets by dataset wrappers¶
MMDetection also supports many dataset wrappers to mix the dataset or modify the dataset distribution for training. Currently it supports to three dataset wrappers as below:
RepeatDataset: simply repeat the whole dataset.ClassBalancedDataset: repeat dataset in a class balanced manner.ConcatDataset: concat datasets.
Repeat dataset¶
We use RepeatDataset as wrapper to repeat the dataset. For example, suppose the original dataset is Dataset_A, to repeat it, the config looks like the following
dataset_A_train = dict(
type='RepeatDataset',
times=N,
dataset=dict( # This is the original config of Dataset_A
type='Dataset_A',
...
pipeline=train_pipeline
)
)
Class balanced dataset¶
We use ClassBalancedDataset as wrapper to repeat the dataset based on category
frequency. The dataset to repeat needs to instantiate function self.get_cat_ids(idx)
to support ClassBalancedDataset.
For example, to repeat Dataset_A with oversample_thr=1e-3, the config looks like the following
dataset_A_train = dict(
type='ClassBalancedDataset',
oversample_thr=1e-3,
dataset=dict( # This is the original config of Dataset_A
type='Dataset_A',
...
pipeline=train_pipeline
)
)
You may refer to source code for details.
Concatenate dataset¶
There are three ways to concatenate the dataset.
If the datasets you want to concatenate are in the same type with different annotation files, you can concatenate the dataset configs like the following.
dataset_A_train = dict( type='Dataset_A', ann_file = ['anno_file_1', 'anno_file_2'], pipeline=train_pipeline )
If the concatenated dataset is used for test or evaluation, this manner supports to evaluate each dataset separately. To test the concatenated datasets as a whole, you can set
separate_eval=Falseas below.dataset_A_train = dict( type='Dataset_A', ann_file = ['anno_file_1', 'anno_file_2'], separate_eval=False, pipeline=train_pipeline )
In case the dataset you want to concatenate is different, you can concatenate the dataset configs like the following.
dataset_A_train = dict() dataset_B_train = dict() data = dict( imgs_per_gpu=2, workers_per_gpu=2, train = [ dataset_A_train, dataset_B_train ], val = dataset_A_val, test = dataset_A_test )
If the concatenated dataset is used for test or evaluation, this manner also supports to evaluate each dataset separately.
We also support to define
ConcatDatasetexplicitly as the following.dataset_A_val = dict() dataset_B_val = dict() data = dict( imgs_per_gpu=2, workers_per_gpu=2, train=dataset_A_train, val=dict( type='ConcatDataset', datasets=[dataset_A_val, dataset_B_val], separate_eval=False))
This manner allows users to evaluate all the datasets as a single one by setting
separate_eval=False.
Note:
The option
separate_eval=Falseassumes the datasets useself.data_infosduring evaluation. Therefore, COCO datasets do not support this behavior since COCO datasets do not fully rely onself.data_infosfor evaluation. Combining different types of datasets and evaluating them as a whole is not tested thus is not suggested.Evaluating
ClassBalancedDatasetandRepeatDatasetis not supported thus evaluating concatenated datasets of these types is also not supported.
A more complex example that repeats Dataset_A and Dataset_B by N and M times, respectively, and then concatenates the repeated datasets is as the following.
dataset_A_train = dict(
type='RepeatDataset',
times=N,
dataset=dict(
type='Dataset_A',
...
pipeline=train_pipeline
)
)
dataset_A_val = dict(
...
pipeline=test_pipeline
)
dataset_A_test = dict(
...
pipeline=test_pipeline
)
dataset_B_train = dict(
type='RepeatDataset',
times=M,
dataset=dict(
type='Dataset_B',
...
pipeline=train_pipeline
)
)
data = dict(
imgs_per_gpu=2,
workers_per_gpu=2,
train = [
dataset_A_train,
dataset_B_train
],
val = dataset_A_val,
test = dataset_A_test
)
Modify Dataset Classes¶
With existing dataset types, we can modify the class names of them to train subset of the annotations. For example, if you want to train only three classes of the current dataset, you can modify the classes of dataset. The dataset will filter out the ground truth boxes of other classes automatically.
classes = ('person', 'bicycle', 'car')
data = dict(
train=dict(classes=classes),
val=dict(classes=classes),
test=dict(classes=classes))
MMDetection V2.0 also supports to read the classes from a file, which is common in real applications.
For example, assume the classes.txt contains the name of classes as the following.
person
bicycle
car
Users can set the classes as a file path, the dataset will load it and convert it to a list automatically.
classes = 'path/to/classes.txt'
data = dict(
train=dict(classes=classes),
val=dict(classes=classes),
test=dict(classes=classes))
Note:
Before MMDetection v2.5.0, the dataset will filter out the empty GT images automatically if the classes are set and there is no way to disable that through config. This is an undesirable behavior and introduces confusion because if the classes are not set, the dataset only filter the empty GT images when
filter_empty_gt=Trueandtest_mode=False. After MMDetection v2.5.0, we decouple the image filtering process and the classes modification, i.e., the dataset will only filter empty GT images whenfilter_empty_gt=Trueandtest_mode=False, no matter whether the classes are set. Thus, setting the classes only influences the annotations of classes used for training and users could decide whether to filter empty GT images by themselves.Since the middle format only has box labels and does not contain the class names, when using
CustomDataset, users cannot filter out the empty GT images through configs but only do this offline.Please remember to modify the
num_classesin the head when specifyingclassesin dataset. We implemented NumClassCheckHook to check whether the numbers are consistent since v2.9.0(after PR#4508).The features for setting dataset classes and dataset filtering will be refactored to be more user-friendly in the future (depends on the progress).
COCO Panoptic Dataset¶
Now we support COCO Panoptic Dataset, the format of panoptic annotations is different from COCO format. Both the foreground and the background will exist in the annotation file. The annotation json files in COCO Panoptic format has the following necessary keys:
'images': [
{
'file_name': '000000001268.jpg',
'height': 427,
'width': 640,
'id': 1268
},
...
]
'annotations': [
{
'filename': '000000001268.jpg',
'image_id': 1268,
'segments_info': [
{
'id':8345037, # One-to-one correspondence with the id in the annotation map.
'category_id': 51,
'iscrowd': 0,
'bbox': (x1, y1, w, h), # The bbox of the background is the outer rectangle of its mask.
'area': 24315
},
...
]
},
...
]
'categories': [ # including both foreground categories and background categories
{'id': 0, 'name': 'person'},
...
]
Moreover, the seg_prefix must be set to the path of the panoptic annotation images.
data = dict(
type='CocoPanopticDataset',
train=dict(
seg_prefix = 'path/to/your/train/panoptic/image_annotation_data'
),
val=dict(
seg_prefix = 'path/to/your/train/panoptic/image_annotation_data'
)
)
Tutorial 3: Customize Data Pipelines¶
Design of Data pipelines¶
Following typical conventions, we use Dataset and DataLoader for data loading
with multiple workers. Dataset returns a dict of data items corresponding
the arguments of models’ forward method.
Since the data in object detection may not be the same size (image size, gt bbox size, etc.),
we introduce a new DataContainer type in MMCV to help collect and distribute
data of different size.
See here for more details.
The data preparation pipeline and the dataset is decomposed. Usually a dataset defines how to process the annotations and a data pipeline defines all the steps to prepare a data dict. A pipeline consists of a sequence of operations. Each operation takes a dict as input and also output a dict for the next transform.
We present a classical pipeline in the following figure. The blue blocks are pipeline operations. With the pipeline going on, each operator can add new keys (marked as green) to the result dict or update the existing keys (marked as orange).
The operations are categorized into data loading, pre-processing, formatting and test-time augmentation.
Here is a pipeline example for Faster R-CNN.
img_norm_cfg = dict(
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
train_pipeline = [
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=True),
dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
dict(type='RandomFlip', flip_ratio=0.5),
dict(type='Normalize', **img_norm_cfg),
dict(type='Pad', size_divisor=32),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
]
test_pipeline = [
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(1333, 800),
flip=False,
transforms=[
dict(type='Resize', keep_ratio=True),
dict(type='RandomFlip'),
dict(type='Normalize', **img_norm_cfg),
dict(type='Pad', size_divisor=32),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect', keys=['img']),
])
]
For each operation, we list the related dict fields that are added/updated/removed.
Data loading¶
LoadImageFromFile
add: img, img_shape, ori_shape
LoadAnnotations
add: gt_bboxes, gt_bboxes_ignore, gt_labels, gt_masks, gt_semantic_seg, bbox_fields, mask_fields
LoadProposals
add: proposals
Pre-processing¶
Resize
add: scale, scale_idx, pad_shape, scale_factor, keep_ratio
update: img, img_shape, *bbox_fields, *mask_fields, *seg_fields
RandomFlip
add: flip
update: img, *bbox_fields, *mask_fields, *seg_fields
Pad
add: pad_fixed_size, pad_size_divisor
update: img, pad_shape, *mask_fields, *seg_fields
RandomCrop
update: img, pad_shape, gt_bboxes, gt_labels, gt_masks, *bbox_fields
Normalize
add: img_norm_cfg
update: img
SegRescale
update: gt_semantic_seg
PhotoMetricDistortion
update: img
Expand
update: img, gt_bboxes
MinIoURandomCrop
update: img, gt_bboxes, gt_labels
Corrupt
update: img
Formatting¶
ToTensor
update: specified by
keys.
ImageToTensor
update: specified by
keys.
Transpose
update: specified by
keys.
ToDataContainer
update: specified by
fields.
DefaultFormatBundle
update: img, proposals, gt_bboxes, gt_bboxes_ignore, gt_labels, gt_masks, gt_semantic_seg
Collect
add: img_meta (the keys of img_meta is specified by
meta_keys)remove: all other keys except for those specified by
keys
Test time augmentation¶
MultiScaleFlipAug
Extend and use custom pipelines¶
Write a new pipeline in a file, e.g., in
my_pipeline.py. It takes a dict as input and returns a dict.import random from mmdet.datasets import PIPELINES @PIPELINES.register_module() class MyTransform: """Add your transform Args: p (float): Probability of shifts. Default 0.5. """ def __init__(self, p=0.5): self.p = p def __call__(self, results): if random.random() > self.p: results['dummy'] = True return results
Import and use the pipeline in your config file. Make sure the import is relative to where your train script is located.
custom_imports = dict(imports=['path.to.my_pipeline'], allow_failed_imports=False) img_norm_cfg = dict( mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True) train_pipeline = [ dict(type='LoadImageFromFile'), dict(type='LoadAnnotations', with_bbox=True), dict(type='Resize', img_scale=(1333, 800), keep_ratio=True), dict(type='RandomFlip', flip_ratio=0.5), dict(type='Normalize', **img_norm_cfg), dict(type='Pad', size_divisor=32), dict(type='MyTransform', p=0.2), dict(type='DefaultFormatBundle'), dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']), ]
Visualize the output of your augmentation pipeline
To visualize the output of your augmentation pipeline,
tools/misc/browse_dataset.pycan help the user to browse a detection dataset (both images and bounding box annotations) visually, or save the image to a designated directory. More details can refer to useful_tools
Tutorial 4: Customize Models¶
We basically categorize model components into 5 types.
backbone: usually an FCN network to extract feature maps, e.g., ResNet, MobileNet.
neck: the component between backbones and heads, e.g., FPN, PAFPN.
head: the component for specific tasks, e.g., bbox prediction and mask prediction.
roi extractor: the part for extracting RoI features from feature maps, e.g., RoI Align.
loss: the component in head for calculating losses, e.g., FocalLoss, L1Loss, and GHMLoss.
Develop new components¶
Add a new backbone¶
Here we show how to develop new components with an example of MobileNet.
1. Define a new backbone (e.g. MobileNet)¶
Create a new file mmdet/models/backbones/mobilenet.py.
import torch.nn as nn
from ..builder import BACKBONES
@BACKBONES.register_module()
class MobileNet(nn.Module):
def __init__(self, arg1, arg2):
pass
def forward(self, x): # should return a tuple
pass
2. Import the module¶
You can either add the following line to mmdet/models/backbones/__init__.py
from .mobilenet import MobileNet
or alternatively add
custom_imports = dict(
imports=['mmdet.models.backbones.mobilenet'],
allow_failed_imports=False)
to the config file to avoid modifying the original code.
3. Use the backbone in your config file¶
model = dict(
...
backbone=dict(
type='MobileNet',
arg1=xxx,
arg2=xxx),
...
Add new necks¶
1. Define a neck (e.g. PAFPN)¶
Create a new file mmdet/models/necks/pafpn.py.
from ..builder import NECKS
@NECKS.register_module()
class PAFPN(nn.Module):
def __init__(self,
in_channels,
out_channels,
num_outs,
start_level=0,
end_level=-1,
add_extra_convs=False):
pass
def forward(self, inputs):
# implementation is ignored
pass
2. Import the module¶
You can either add the following line to mmdet/models/necks/__init__.py,
from .pafpn import PAFPN
or alternatively add
custom_imports = dict(
imports=['mmdet.models.necks.pafpn.py'],
allow_failed_imports=False)
to the config file and avoid modifying the original code.
3. Modify the config file¶
neck=dict(
type='PAFPN',
in_channels=[256, 512, 1024, 2048],
out_channels=256,
num_outs=5)
Add new heads¶
Here we show how to develop a new head with the example of Double Head R-CNN as the following.
First, add a new bbox head in mmdet/models/roi_heads/bbox_heads/double_bbox_head.py.
Double Head R-CNN implements a new bbox head for object detection.
To implement a bbox head, basically we need to implement three functions of the new module as the following.
from mmdet.models.builder import HEADS
from .bbox_head import BBoxHead
@HEADS.register_module()
class DoubleConvFCBBoxHead(BBoxHead):
r"""Bbox head used in Double-Head R-CNN
/-> cls
/-> shared convs ->
\-> reg
roi features
/-> cls
\-> shared fc ->
\-> reg
""" # noqa: W605
def __init__(self,
num_convs=0,
num_fcs=0,
conv_out_channels=1024,
fc_out_channels=1024,
conv_cfg=None,
norm_cfg=dict(type='BN'),
**kwargs):
kwargs.setdefault('with_avg_pool', True)
super(DoubleConvFCBBoxHead, self).__init__(**kwargs)
def forward(self, x_cls, x_reg):
Second, implement a new RoI Head if it is necessary. We plan to inherit the new DoubleHeadRoIHead from StandardRoIHead. We can find that a StandardRoIHead already implements the following functions.
import torch
from mmdet.core import bbox2result, bbox2roi, build_assigner, build_sampler
from ..builder import HEADS, build_head, build_roi_extractor
from .base_roi_head import BaseRoIHead
from .test_mixins import BBoxTestMixin, MaskTestMixin
@HEADS.register_module()
class StandardRoIHead(BaseRoIHead, BBoxTestMixin, MaskTestMixin):
"""Simplest base roi head including one bbox head and one mask head.
"""
def init_assigner_sampler(self):
def init_bbox_head(self, bbox_roi_extractor, bbox_head):
def init_mask_head(self, mask_roi_extractor, mask_head):
def forward_dummy(self, x, proposals):
def forward_train(self,
x,
img_metas,
proposal_list,
gt_bboxes,
gt_labels,
gt_bboxes_ignore=None,
gt_masks=None):
def _bbox_forward(self, x, rois):
def _bbox_forward_train(self, x, sampling_results, gt_bboxes, gt_labels,
img_metas):
def _mask_forward_train(self, x, sampling_results, bbox_feats, gt_masks,
img_metas):
def _mask_forward(self, x, rois=None, pos_inds=None, bbox_feats=None):
def simple_test(self,
x,
proposal_list,
img_metas,
proposals=None,
rescale=False):
"""Test without augmentation."""
Double Head’s modification is mainly in the bbox_forward logic, and it inherits other logics from the StandardRoIHead.
In the mmdet/models/roi_heads/double_roi_head.py, we implement the new RoI Head as the following:
from ..builder import HEADS
from .standard_roi_head import StandardRoIHead
@HEADS.register_module()
class DoubleHeadRoIHead(StandardRoIHead):
"""RoI head for Double Head RCNN
https://arxiv.org/abs/1904.06493
"""
def __init__(self, reg_roi_scale_factor, **kwargs):
super(DoubleHeadRoIHead, self).__init__(**kwargs)
self.reg_roi_scale_factor = reg_roi_scale_factor
def _bbox_forward(self, x, rois):
bbox_cls_feats = self.bbox_roi_extractor(
x[:self.bbox_roi_extractor.num_inputs], rois)
bbox_reg_feats = self.bbox_roi_extractor(
x[:self.bbox_roi_extractor.num_inputs],
rois,
roi_scale_factor=self.reg_roi_scale_factor)
if self.with_shared_head:
bbox_cls_feats = self.shared_head(bbox_cls_feats)
bbox_reg_feats = self.shared_head(bbox_reg_feats)
cls_score, bbox_pred = self.bbox_head(bbox_cls_feats, bbox_reg_feats)
bbox_results = dict(
cls_score=cls_score,
bbox_pred=bbox_pred,
bbox_feats=bbox_cls_feats)
return bbox_results
Last, the users need to add the module in
mmdet/models/bbox_heads/__init__.py and mmdet/models/roi_heads/__init__.py thus the corresponding registry could find and load them.
Alternatively, the users can add
custom_imports=dict(
imports=['mmdet.models.roi_heads.double_roi_head', 'mmdet.models.bbox_heads.double_bbox_head'])
to the config file and achieve the same goal.
The config file of Double Head R-CNN is as the following
_base_ = '../faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py'
model = dict(
roi_head=dict(
type='DoubleHeadRoIHead',
reg_roi_scale_factor=1.3,
bbox_head=dict(
_delete_=True,
type='DoubleConvFCBBoxHead',
num_convs=4,
num_fcs=2,
in_channels=256,
conv_out_channels=1024,
fc_out_channels=1024,
roi_feat_size=7,
num_classes=80,
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[0., 0., 0., 0.],
target_stds=[0.1, 0.1, 0.2, 0.2]),
reg_class_agnostic=False,
loss_cls=dict(
type='CrossEntropyLoss', use_sigmoid=False, loss_weight=2.0),
loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=2.0))))
Since MMDetection 2.0, the config system supports to inherit configs such that the users can focus on the modification.
The Double Head R-CNN mainly uses a new DoubleHeadRoIHead and a new
DoubleConvFCBBoxHead, the arguments are set according to the __init__ function of each module.
Add new loss¶
Assume you want to add a new loss as MyLoss, for bounding box regression.
To add a new loss function, the users need implement it in mmdet/models/losses/my_loss.py.
The decorator weighted_loss enable the loss to be weighted for each element.
import torch
import torch.nn as nn
from ..builder import LOSSES
from .utils import weighted_loss
@weighted_loss
def my_loss(pred, target):
assert pred.size() == target.size() and target.numel() > 0
loss = torch.abs(pred - target)
return loss
@LOSSES.register_module()
class MyLoss(nn.Module):
def __init__(self, reduction='mean', loss_weight=1.0):
super(MyLoss, self).__init__()
self.reduction = reduction
self.loss_weight = loss_weight
def forward(self,
pred,
target,
weight=None,
avg_factor=None,
reduction_override=None):
assert reduction_override in (None, 'none', 'mean', 'sum')
reduction = (
reduction_override if reduction_override else self.reduction)
loss_bbox = self.loss_weight * my_loss(
pred, target, weight, reduction=reduction, avg_factor=avg_factor)
return loss_bbox
Then the users need to add it in the mmdet/models/losses/__init__.py.
from .my_loss import MyLoss, my_loss
Alternatively, you can add
custom_imports=dict(
imports=['mmdet.models.losses.my_loss'])
to the config file and achieve the same goal.
To use it, modify the loss_xxx field.
Since MyLoss is for regression, you need to modify the loss_bbox field in the head.
loss_bbox=dict(type='MyLoss', loss_weight=1.0))
Tutorial 5: Customize Runtime Settings¶
Customize optimization settings¶
Customize optimizer supported by Pytorch¶
We already support to use all the optimizers implemented by PyTorch, and the only modification is to change the optimizer field of config files.
For example, if you want to use ADAM (note that the performance could drop a lot), the modification could be as the following.
optimizer = dict(type='Adam', lr=0.0003, weight_decay=0.0001)
To modify the learning rate of the model, the users only need to modify the lr in the config of optimizer. The users can directly set arguments following the API doc of PyTorch.
Customize self-implemented optimizer¶
1. Define a new optimizer¶
A customized optimizer could be defined as following.
Assume you want to add a optimizer named MyOptimizer, which has arguments a, b, and c.
You need to create a new directory named mmdet/core/optimizer.
And then implement the new optimizer in a file, e.g., in mmdet/core/optimizer/my_optimizer.py:
from .registry import OPTIMIZERS
from torch.optim import Optimizer
@OPTIMIZERS.register_module()
class MyOptimizer(Optimizer):
def __init__(self, a, b, c)
2. Add the optimizer to registry¶
To find the above module defined above, this module should be imported into the main namespace at first. There are two options to achieve it.
Modify
mmdet/core/optimizer/__init__.pyto import it.The newly defined module should be imported in
mmdet/core/optimizer/__init__.pyso that the registry will find the new module and add it:
from .my_optimizer import MyOptimizer
Use
custom_importsin the config to manually import it
custom_imports = dict(imports=['mmdet.core.optimizer.my_optimizer'], allow_failed_imports=False)
The module mmdet.core.optimizer.my_optimizer will be imported at the beginning of the program and the class MyOptimizer is then automatically registered.
Note that only the package containing the class MyOptimizer should be imported.
mmdet.core.optimizer.my_optimizer.MyOptimizer cannot be imported directly.
Actually users can use a totally different file directory structure using this importing method, as long as the module root can be located in PYTHONPATH.
3. Specify the optimizer in the config file¶
Then you can use MyOptimizer in optimizer field of config files.
In the configs, the optimizers are defined by the field optimizer like the following:
optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001)
To use your own optimizer, the field can be changed to
optimizer = dict(type='MyOptimizer', a=a_value, b=b_value, c=c_value)
Customize optimizer constructor¶
Some models may have some parameter-specific settings for optimization, e.g. weight decay for BatchNorm layers. The users can do those fine-grained parameter tuning through customizing optimizer constructor.
from mmcv.utils import build_from_cfg
from mmcv.runner.optimizer import OPTIMIZER_BUILDERS, OPTIMIZERS
from mmdet.utils import get_root_logger
from .my_optimizer import MyOptimizer
@OPTIMIZER_BUILDERS.register_module()
class MyOptimizerConstructor(object):
def __init__(self, optimizer_cfg, paramwise_cfg=None):
def __call__(self, model):
return my_optimizer
The default optimizer constructor is implemented here, which could also serve as a template for new optimizer constructor.
Additional settings¶
Tricks not implemented by the optimizer should be implemented through optimizer constructor (e.g., set parameter-wise learning rates) or hooks. We list some common settings that could stabilize the training or accelerate the training. Feel free to create PR, issue for more settings.
Use gradient clip to stabilize training: Some models need gradient clip to clip the gradients to stabilize the training process. An example is as below:
optimizer_config = dict( _delete_=True, grad_clip=dict(max_norm=35, norm_type=2))
If your config inherits the base config which already sets the
optimizer_config, you might need_delete_=Trueto override the unnecessary settings. See the config documentation for more details.Use momentum schedule to accelerate model convergence: We support momentum scheduler to modify model’s momentum according to learning rate, which could make the model converge in a faster way. Momentum scheduler is usually used with LR scheduler, for example, the following config is used in 3D detection to accelerate convergence. For more details, please refer to the implementation of CyclicLrUpdater and CyclicMomentumUpdater.
lr_config = dict( policy='cyclic', target_ratio=(10, 1e-4), cyclic_times=1, step_ratio_up=0.4, ) momentum_config = dict( policy='cyclic', target_ratio=(0.85 / 0.95, 1), cyclic_times=1, step_ratio_up=0.4, )
Customize training schedules¶
By default we use step learning rate with 1x schedule, this calls StepLRHook in MMCV.
We support many other learning rate schedule here, such as CosineAnnealing and Poly schedule. Here are some examples
Poly schedule:
lr_config = dict(policy='poly', power=0.9, min_lr=1e-4, by_epoch=False)
ConsineAnnealing schedule:
lr_config = dict( policy='CosineAnnealing', warmup='linear', warmup_iters=1000, warmup_ratio=1.0 / 10, min_lr_ratio=1e-5)
Customize workflow¶
Workflow is a list of (phase, epochs) to specify the running order and epochs. By default it is set to be
workflow = [('train', 1)]
which means running 1 epoch for training. Sometimes user may want to check some metrics (e.g. loss, accuracy) about the model on the validate set. In such case, we can set the workflow as
[('train', 1), ('val', 1)]
so that 1 epoch for training and 1 epoch for validation will be run iteratively.
Note:
The parameters of model will not be updated during val epoch.
Keyword
total_epochsin the config only controls the number of training epochs and will not affect the validation workflow.Workflows
[('train', 1), ('val', 1)]and[('train', 1)]will not change the behavior ofEvalHookbecauseEvalHookis called byafter_train_epochand validation workflow only affect hooks that are called throughafter_val_epoch. Therefore, the only difference between[('train', 1), ('val', 1)]and[('train', 1)]is that the runner will calculate losses on validation set after each training epoch.
Customize hooks¶
Customize self-implemented hooks¶
1. Implement a new hook¶
There are some occasions when the users might need to implement a new hook. MMDetection supports customized hooks in training (#3395) since v2.3.0. Thus the users could implement a hook directly in mmdet or their mmdet-based codebases and use the hook by only modifying the config in training. Before v2.3.0, the users need to modify the code to get the hook registered before training starts. Here we give an example of creating a new hook in mmdet and using it in training.
from mmcv.runner import HOOKS, Hook
@HOOKS.register_module()
class MyHook(Hook):
def __init__(self, a, b):
pass
def before_run(self, runner):
pass
def after_run(self, runner):
pass
def before_epoch(self, runner):
pass
def after_epoch(self, runner):
pass
def before_iter(self, runner):
pass
def after_iter(self, runner):
pass
Depending on the functionality of the hook, the users need to specify what the hook will do at each stage of the training in before_run, after_run, before_epoch, after_epoch, before_iter, and after_iter.
2. Register the new hook¶
Then we need to make MyHook imported. Assuming the file is in mmdet/core/utils/my_hook.py there are two ways to do that:
Modify
mmdet/core/utils/__init__.pyto import it.The newly defined module should be imported in
mmdet/core/utils/__init__.pyso that the registry will find the new module and add it:
from .my_hook import MyHook
Use
custom_importsin the config to manually import it
custom_imports = dict(imports=['mmdet.core.utils.my_hook'], allow_failed_imports=False)
3. Modify the config¶
custom_hooks = [
dict(type='MyHook', a=a_value, b=b_value)
]
You can also set the priority of the hook by adding key priority to 'NORMAL' or 'HIGHEST' as below
custom_hooks = [
dict(type='MyHook', a=a_value, b=b_value, priority='NORMAL')
]
By default the hook’s priority is set as NORMAL during registration.
Use hooks implemented in MMCV¶
If the hook is already implemented in MMCV, you can directly modify the config to use the hook as below
4. Example: NumClassCheckHook¶
We implement a customized hook named NumClassCheckHook to check whether the num_classes in head matches the length of CLASSES in dataset.
We set it in default_runtime.py.
custom_hooks = [dict(type='NumClassCheckHook')]
Modify default runtime hooks¶
There are some common hooks that are not registered through custom_hooks, they are
log_config
checkpoint_config
evaluation
lr_config
optimizer_config
momentum_config
In those hooks, only the logger hook has the VERY_LOW priority, others’ priority are NORMAL.
The above-mentioned tutorials already covers how to modify optimizer_config, momentum_config, and lr_config.
Here we reveals how what we can do with log_config, checkpoint_config, and evaluation.
Checkpoint config¶
The MMCV runner will use checkpoint_config to initialize CheckpointHook.
checkpoint_config = dict(interval=1)
The users could set max_keep_ckpts to only save only small number of checkpoints or decide whether to store state dict of optimizer by save_optimizer. More details of the arguments are here
Log config¶
The log_config wraps multiple logger hooks and enables to set intervals. Now MMCV supports WandbLoggerHook, MlflowLoggerHook, and TensorboardLoggerHook.
The detail usages can be found in the doc.
log_config = dict(
interval=50,
hooks=[
dict(type='TextLoggerHook'),
dict(type='TensorboardLoggerHook')
])
Tutorial 6: Customize Losses¶
MMDetection provides users with different loss functions. But the default configuration may be not applicable for different datasets or models, so users may want to modify a specific loss to adapt the new situation.
This tutorial first elaborate the computation pipeline of losses, then give some instructions about how to modify each step. The modification can be categorized as tweaking and weighting.
Computation pipeline of a loss¶
Given the input prediction and target, as well as the weights, a loss function maps the input tensor to the final loss scalar. The mapping can be divided into five steps:
Set the sampling method to sample positive and negative samples.
Get element-wise or sample-wise loss by the loss kernel function.
Weighting the loss with a weight tensor element-wisely.
Reduce the loss tensor to a scalar.
Weighting the loss with a scalar.
Set sampling method (step 1)¶
For some loss functions, sampling strategies are needed to avoid imbalance between positive and negative samples.
For example, when using CrossEntropyLoss in RPN head, we need to set RandomSampler in train_cfg
train_cfg=dict(
rpn=dict(
sampler=dict(
type='RandomSampler',
num=256,
pos_fraction=0.5,
neg_pos_ub=-1,
add_gt_as_proposals=False))
For some other losses which have positive and negative sample balance mechanism such as Focal Loss, GHMC, and QualityFocalLoss, the sampler is no more necessary.
Tweaking loss¶
Tweaking a loss is more related with step 2, 4, 5, and most modifications can be specified in the config. Here we take Focal Loss (FL) as an example. The following code sniper are the construction method and config of FL respectively, they are actually one to one correspondence.
@LOSSES.register_module()
class FocalLoss(nn.Module):
def __init__(self,
use_sigmoid=True,
gamma=2.0,
alpha=0.25,
reduction='mean',
loss_weight=1.0):
loss_cls=dict(
type='FocalLoss',
use_sigmoid=True,
gamma=2.0,
alpha=0.25,
loss_weight=1.0)
Tweaking hyper-parameters (step 2)¶
gamma and beta are two hyper-parameters in the Focal Loss. Say if we want to change the value of gamma to be 1.5 and alpha to be 0.5, then we can specify them in the config as follows:
loss_cls=dict(
type='FocalLoss',
use_sigmoid=True,
gamma=1.5,
alpha=0.5,
loss_weight=1.0)
Tweaking the way of reduction (step 3)¶
The default way of reduction is mean for FL. Say if we want to change the reduction from mean to sum, we can specify it in the config as follows:
loss_cls=dict(
type='FocalLoss',
use_sigmoid=True,
gamma=2.0,
alpha=0.25,
loss_weight=1.0,
reduction='sum')
Tweaking loss weight (step 5)¶
The loss weight here is a scalar which controls the weight of different losses in multi-task learning, e.g. classification loss and regression loss. Say if we want to change to loss weight of classification loss to be 0.5, we can specify it in the config as follows:
loss_cls=dict(
type='FocalLoss',
use_sigmoid=True,
gamma=2.0,
alpha=0.25,
loss_weight=0.5)
Weighting loss (step 3)¶
Weighting loss means we re-weight the loss element-wisely. To be more specific, we multiply the loss tensor with a weight tensor which has the same shape. As a result, different entries of the loss can be scaled differently, and so called element-wisely.
The loss weight varies across different models and highly context related, but overall there are two kinds of loss weights, label_weights for classification loss and bbox_weights for bbox regression loss. You can find them in the get_target method of the corresponding head. Here we take ATSSHead as an example, which inherit AnchorHead but overwrite its get_targets method which yields different label_weights and bbox_weights.
class ATSSHead(AnchorHead):
...
def get_targets(self,
anchor_list,
valid_flag_list,
gt_bboxes_list,
img_metas,
gt_bboxes_ignore_list=None,
gt_labels_list=None,
label_channels=1,
unmap_outputs=True):
Tutorial 7: Finetuning Models¶
Detectors pre-trained on the COCO dataset can serve as a good pre-trained model for other datasets, e.g., CityScapes and KITTI Dataset. This tutorial provides instruction for users to use the models provided in the Model Zoo for other datasets to obtain better performance.
There are two steps to finetune a model on a new dataset.
Add support for the new dataset following Tutorial 2: Customize Datasets.
Modify the configs as will be discussed in this tutorial.
Take the finetuning process on Cityscapes Dataset as an example, the users need to modify five parts in the config.
Inherit base configs¶
To release the burden and reduce bugs in writing the whole configs, MMDetection V2.0 support inheriting configs from multiple existing configs. To finetune a Mask RCNN model, the new config needs to inherit
_base_/models/mask_rcnn_r50_fpn.py to build the basic structure of the model. To use the Cityscapes Dataset, the new config can also simply inherit _base_/datasets/cityscapes_instance.py. For runtime settings such as training schedules, the new config needs to inherit _base_/default_runtime.py. This configs are in the configs directory and the users can also choose to write the whole contents rather than use inheritance.
_base_ = [
'../_base_/models/mask_rcnn_r50_fpn.py',
'../_base_/datasets/cityscapes_instance.py', '../_base_/default_runtime.py'
]
Modify head¶
Then the new config needs to modify the head according to the class numbers of the new datasets. By only changing num_classes in the roi_head, the weights of the pre-trained models are mostly reused except the final prediction head.
model = dict(
pretrained=None,
roi_head=dict(
bbox_head=dict(
type='Shared2FCBBoxHead',
in_channels=256,
fc_out_channels=1024,
roi_feat_size=7,
num_classes=8,
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[0., 0., 0., 0.],
target_stds=[0.1, 0.1, 0.2, 0.2]),
reg_class_agnostic=False,
loss_cls=dict(
type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0)),
mask_head=dict(
type='FCNMaskHead',
num_convs=4,
in_channels=256,
conv_out_channels=256,
num_classes=8,
loss_mask=dict(
type='CrossEntropyLoss', use_mask=True, loss_weight=1.0))))
Modify dataset¶
The users may also need to prepare the dataset and write the configs about dataset. MMDetection V2.0 already support VOC, WIDER FACE, COCO and Cityscapes Dataset.
Modify training schedule¶
The finetuning hyperparameters vary from the default schedule. It usually requires smaller learning rate and less training epochs
# optimizer
# lr is set for a batch size of 8
optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)
optimizer_config = dict(grad_clip=None)
# learning policy
lr_config = dict(
policy='step',
warmup='linear',
warmup_iters=500,
warmup_ratio=0.001,
step=[7])
# the max_epochs and step in lr_config need specifically tuned for the customized dataset
runner = dict(max_epochs=8)
log_config = dict(interval=100)
Use pre-trained model¶
To use the pre-trained model, the new config add the link of pre-trained models in the load_from. The users might need to download the model weights before training to avoid the download time during training.
load_from = 'https://download.openmmlab.com/mmdetection/v2.0/mask_rcnn/mask_rcnn_r50_caffe_fpn_mstrain-poly_3x_coco/mask_rcnn_r50_caffe_fpn_mstrain-poly_3x_coco_bbox_mAP-0.408__segm_mAP-0.37_20200504_163245-42aa3d00.pth' # noqa
Tutorial 8: Pytorch to ONNX (Experimental)¶
How to convert models from Pytorch to ONNX¶
Prerequisite¶
Install the prerequisites following get_started.md/Prepare environment.
Build custom operators for ONNX Runtime and install MMCV manually following How to build custom operators for ONNX Runtime
Install MMdetection manually following steps 2-3 in get_started.md/Install MMdetection.
Usage¶
python tools/deployment/pytorch2onnx.py \
${CONFIG_FILE} \
${CHECKPOINT_FILE} \
--output-file ${OUTPUT_FILE} \
--input-img ${INPUT_IMAGE_PATH} \
--shape ${IMAGE_SHAPE} \
--test-img ${TEST_IMAGE_PATH} \
--opset-version ${OPSET_VERSION} \
--cfg-options ${CFG_OPTIONS}
--dynamic-export \
--show \
--verify \
--simplify \
Description of all arguments¶
config: The path of a model config file.checkpoint: The path of a model checkpoint file.--output-file: The path of output ONNX model. If not specified, it will be set totmp.onnx.--input-img: The path of an input image for tracing and conversion. By default, it will be set totests/data/color.jpg.--shape: The height and width of input tensor to the model. If not specified, it will be set to800 1216.--test-img: The path of an image to verify the exported ONNX model. By default, it will be set toNone, meaning it will use--input-imgfor verification.--opset-version: The opset version of ONNX. If not specified, it will be set to11.--dynamic-export: Determines whether to export ONNX model with dynamic input and output shapes. If not specified, it will be set toFalse.--show: Determines whether to print the architecture of the exported model and whether to show detection outputs when--verifyis set toTrue. If not specified, it will be set toFalse.--verify: Determines whether to verify the correctness of an exported model. If not specified, it will be set toFalse.--simplify: Determines whether to simplify the exported ONNX model. If not specified, it will be set toFalse.--cfg-options: Override some settings in the used config file, the key-value pair inxxx=yyyformat will be merged into config file.--skip-postprocess: Determines whether export model without post process. If not specified, it will be set toFalse. Notice: This is an experimental option. Only work for some single stage models. Users need to implement the post-process by themselves. We do not guarantee the correctness of the exported model.
Example:
python tools/deployment/pytorch2onnx.py \
configs/yolo/yolov3_d53_mstrain-608_273e_coco.py \
checkpoints/yolo/yolov3_d53_mstrain-608_273e_coco.pth \
--output-file checkpoints/yolo/yolov3_d53_mstrain-608_273e_coco.onnx \
--input-img demo/demo.jpg \
--test-img tests/data/color.jpg \
--shape 608 608 \
--show \
--verify \
--dynamic-export \
--cfg-options \
model.test_cfg.deploy_nms_pre=-1 \
How to evaluate the exported models¶
We prepare a tool tools/deplopyment/test.py to evaluate ONNX models with ONNXRuntime and TensorRT.
Prerequisite¶
Install onnx and onnxruntime (CPU version)
pip install onnx onnxruntime==1.5.1
If you want to run the model on GPU, please remove the CPU version before using the GPU version.
pip uninstall onnxruntime pip install onnxruntime-gpu
Note: onnxruntime-gpu is version-dependent on CUDA and CUDNN, please ensure that your environment meets the requirements.
Build custom operators for ONNX Runtime following How to build custom operators for ONNX Runtime
Install TensorRT by referring to How to build TensorRT plugins in MMCV (optional)
Usage¶
python tools/deployment/test.py \
${CONFIG_FILE} \
${MODEL_FILE} \
--out ${OUTPUT_FILE} \
--backend ${BACKEND} \
--format-only ${FORMAT_ONLY} \
--eval ${EVALUATION_METRICS} \
--show-dir ${SHOW_DIRECTORY} \
----show-score-thr ${SHOW_SCORE_THRESHOLD} \
----cfg-options ${CFG_OPTIONS} \
----eval-options ${EVALUATION_OPTIONS} \
Description of all arguments¶
config: The path of a model config file.model: The path of an input model file.--out: The path of output result file in pickle format.--backend: Backend for input model to run and should beonnxruntimeortensorrt.--format-only: Format the output results without perform evaluation. It is useful when you want to format the result to a specific format and submit it to the test server. If not specified, it will be set toFalse.--eval: Evaluation metrics, which depends on the dataset, e.g., “bbox”, “segm”, “proposal” for COCO, and “mAP”, “recall” for PASCAL VOC.--show-dir: Directory where painted images will be saved--show-score-thr: Score threshold. Default is set to0.3.--cfg-options: Override some settings in the used config file, the key-value pair inxxx=yyyformat will be merged into config file.--eval-options: Custom options for evaluation, the key-value pair inxxx=yyyformat will be kwargs fordataset.evaluate()function
Notes:
If the deployed backend platform is TensorRT, please add environment variables before running the file:
export ONNX_BACKEND=MMCVTensorRT
If you want to use the
--dynamic-exportparameter in the TensorRT backend to export ONNX, please remove the--simplifyparameter, and vice versa.
Results and Models¶
| Model | Config | Metric | PyTorch | ONNX Runtime | TensorRT |
|---|---|---|---|---|---|
| FCOS | configs/fcos/fcos_r50_caffe_fpn_gn-head_4x4_1x_coco.py |
Box AP | 36.6 | 36.5 | 36.3 |
| FSAF | configs/fsaf/fsaf_r50_fpn_1x_coco.py |
Box AP | 36.0 | 36.0 | 35.9 |
| RetinaNet | configs/retinanet/retinanet_r50_fpn_1x_coco.py |
Box AP | 36.5 | 36.4 | 36.3 |
| SSD | configs/ssd/ssd300_coco.py |
Box AP | 25.6 | 25.6 | 25.6 |
| YOLOv3 | configs/yolo/yolov3_d53_mstrain-608_273e_coco.py |
Box AP | 33.5 | 33.5 | 33.5 |
| Faster R-CNN | configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py |
Box AP | 37.4 | 37.4 | 37.0 |
| Cascade R-CNN | configs/cascade_rcnn/cascade_rcnn_r50_fpn_1x_coco.py |
Box AP | 40.3 | 40.3 | 40.1 |
| Mask R-CNN | configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.py |
Box AP | 38.2 | 38.1 | 37.7 |
| Mask AP | 34.7 | 33.7 | 33.3 | ||
| Cascade Mask R-CNN | configs/cascade_rcnn/cascade_mask_rcnn_r50_fpn_1x_coco.py |
Box AP | 41.2 | 41.2 | 40.9 |
| Mask AP | 35.9 | 34.8 | 34.5 | ||
| CornerNet | configs/cornernet/cornernet_hourglass104_mstest_10x5_210e_coco.py |
Box AP | 40.6 | 40.4 | - |
| DETR | configs/detr/detr_r50_8x2_150e_coco.py |
Box AP | 40.1 | 40.1 | - |
| PointRend | configs/point_rend/point_rend_r50_caffe_fpn_mstrain_1x_coco.py |
Box AP | 38.4 | 38.4 | - |
| Mask AP | 36.3 | 35.2 | - |
Notes:
All ONNX models are evaluated with dynamic shape on coco dataset and images are preprocessed according to the original config file. Note that CornerNet is evaluated without test-time flip, since currently only single-scale evaluation is supported with ONNX Runtime.
Mask AP of Mask R-CNN drops by 1% for ONNXRuntime. The main reason is that the predicted masks are directly interpolated to original image in PyTorch, while they are at first interpolated to the preprocessed input image of the model and then to original image in other backend.
List of supported models exportable to ONNX¶
The table below lists the models that are guaranteed to be exportable to ONNX and runnable in ONNX Runtime.
| Model | Config | Dynamic Shape | Batch Inference | Note |
|---|---|---|---|---|
| FCOS | configs/fcos/fcos_r50_caffe_fpn_gn-head_4x4_1x_coco.py |
Y | Y | |
| FSAF | configs/fsaf/fsaf_r50_fpn_1x_coco.py |
Y | Y | |
| RetinaNet | configs/retinanet/retinanet_r50_fpn_1x_coco.py |
Y | Y | |
| SSD | configs/ssd/ssd300_coco.py |
Y | Y | |
| YOLOv3 | configs/yolo/yolov3_d53_mstrain-608_273e_coco.py |
Y | Y | |
| Faster R-CNN | configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py |
Y | Y | |
| Cascade R-CNN | configs/cascade_rcnn/cascade_rcnn_r50_fpn_1x_coco.py |
Y | Y | |
| Mask R-CNN | configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.py |
Y | Y | |
| Cascade Mask R-CNN | configs/cascade_rcnn/cascade_mask_rcnn_r50_fpn_1x_coco.py |
Y | Y | |
| CornerNet | configs/cornernet/cornernet_hourglass104_mstest_10x5_210e_coco.py |
Y | N | no flip, no batch inference, tested with torch==1.7.0 and onnxruntime==1.5.1. |
| DETR | configs/detr/detr_r50_8x2_150e_coco.py |
Y | Y | batch inference is not recommended |
| PointRend | configs/point_rend/point_rend_r50_caffe_fpn_mstrain_1x_coco.py |
Y | Y |
Notes:
Minimum required version of MMCV is
1.3.5All models above are tested with Pytorch==1.6.0 and onnxruntime==1.5.1, except for CornerNet. For more details about the torch version when exporting CornerNet to ONNX, which involves
mmcv::cummax, please refer to the Known Issues in mmcv.Though supported, it is not recommended to use batch inference in onnxruntime for
DETR, because there is huge performance gap between ONNX and torch model (e.g. 33.5 vs 39.9 mAP on COCO for onnxruntime and torch respectively, with a batch size 2). The main reason for the gap is that these is non-negligible effect on the predicted regressions during batch inference for ONNX, since the predicted coordinates is normalized byimg_shape(without padding) and should be converted to absolute format, butimg_shapeis not dynamically traceable thus the paddedimg_shape_for_onnxis used.Currently only single-scale evaluation is supported with ONNX Runtime, also
mmcv::SoftNonMaxSuppressionis only supported for single image by now.
The Parameters of Non-Maximum Suppression in ONNX Export¶
In the process of exporting the ONNX model, we set some parameters for the NMS op to control the number of output bounding boxes. The following will introduce the parameter setting of the NMS op in the supported models. You can set these parameters through --cfg-options.
nms_pre: The number of boxes before NMS. The default setting is1000.deploy_nms_pre: The number of boxes before NMS when exporting to ONNX model. The default setting is0.max_per_img: The number of boxes to be kept after NMS. The default setting is100.max_output_boxes_per_class: Maximum number of output boxes per class of NMS. The default setting is200.
Reminders¶
When the input model has custom op such as
RoIAlignand if you want to verify the exported ONNX model, you may have to buildmmcvwith ONNXRuntime from source.mmcv.onnx.simplifyfeature is based on onnx-simplifier. If you want to try it, please refer to onnx inmmcvand onnxruntime op inmmcvfor more information.If you meet any problem with the listed models above, please create an issue and it would be taken care of soon. For models not included in the list, please try to dig a little deeper and debug a little bit more and hopefully solve them by yourself.
Because this feature is experimental and may change fast, please always try with the latest
mmcvandmmdetecion.
FAQs¶
None
Tutorial 9: ONNX to TensorRT (Experimental)¶
How to convert models from ONNX to TensorRT¶
Prerequisite¶
Please refer to get_started.md for installation of MMCV and MMDetection from source.
Please refer to ONNXRuntime in mmcv and TensorRT plugin in mmcv to install
mmcv-fullwith ONNXRuntime custom ops and TensorRT plugins.Use our tool pytorch2onnx to convert the model from PyTorch to ONNX.
Usage¶
python tools/deployment/onnx2tensorrt.py \
${CONFIG} \
${MODEL} \
--trt-file ${TRT_FILE} \
--input-img ${INPUT_IMAGE_PATH} \
--shape ${INPUT_IMAGE_SHAPE} \
--min-shape ${MIN_IMAGE_SHAPE} \
--max-shape ${MAX_IMAGE_SHAPE} \
--workspace-size {WORKSPACE_SIZE} \
--show \
--verify \
Description of all arguments:
config: The path of a model config file.model: The path of an ONNX model file.--trt-file: The Path of output TensorRT engine file. If not specified, it will be set totmp.trt.--input-img: The path of an input image for tracing and conversion. By default, it will be set todemo/demo.jpg.--shape: The height and width of model input. If not specified, it will be set to400 600.--min-shape: The minimum height and width of model input. If not specified, it will be set to the same as--shape.--max-shape: The maximum height and width of model input. If not specified, it will be set to the same as--shape.--workspace-size: The required GPU workspace size in GiB to build TensorRT engine. If not specified, it will be set to1GiB.--show: Determines whether to show the outputs of the model. If not specified, it will be set toFalse.--verify: Determines whether to verify the correctness of models between ONNXRuntime and TensorRT. If not specified, it will be set toFalse.--verbose: Determines whether to print logging messages. It’s useful for debugging. If not specified, it will be set toFalse.
Example:
python tools/deployment/onnx2tensorrt.py \
configs/retinanet/retinanet_r50_fpn_1x_coco.py \
checkpoints/retinanet_r50_fpn_1x_coco.onnx \
--trt-file checkpoints/retinanet_r50_fpn_1x_coco.trt \
--input-img demo/demo.jpg \
--shape 400 600 \
--show \
--verify \
How to evaluate the exported models¶
We prepare a tool tools/deplopyment/test.py to evaluate TensorRT models.
Please refer to following links for more information.
List of supported models convertible to TensorRT¶
The table below lists the models that are guaranteed to be convertible to TensorRT.
| Model | Config | Dynamic Shape | Batch Inference | Note |
|---|---|---|---|---|
| SSD | configs/ssd/ssd300_coco.py |
Y | Y | |
| FSAF | configs/fsaf/fsaf_r50_fpn_1x_coco.py |
Y | Y | |
| FCOS | configs/fcos/fcos_r50_caffe_fpn_4x4_1x_coco.py |
Y | Y | |
| YOLOv3 | configs/yolo/yolov3_d53_mstrain-608_273e_coco.py |
Y | Y | |
| RetinaNet | configs/retinanet/retinanet_r50_fpn_1x_coco.py |
Y | Y | |
| Faster R-CNN | configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py |
Y | Y | |
| Cascade R-CNN | configs/cascade_rcnn/cascade_rcnn_r50_fpn_1x_coco.py |
Y | Y | |
| Mask R-CNN | configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.py |
Y | Y | |
| Cascade Mask R-CNN | configs/cascade_rcnn/cascade_mask_rcnn_r50_fpn_1x_coco.py |
Y | Y | |
| PointRend | configs/point_rend/point_rend_r50_caffe_fpn_mstrain_1x_coco.py |
Y | Y |
Notes:
All models above are tested with Pytorch==1.6.0, onnx==1.7.0 and TensorRT-7.2.1.6.Ubuntu-16.04.x86_64-gnu.cuda-10.2.cudnn8.0
Reminders¶
If you meet any problem with the listed models above, please create an issue and it would be taken care of soon. For models not included in the list, we may not provide much help here due to the limited resources. Please try to dig a little deeper and debug by yourself.
Because this feature is experimental and may change fast, please always try with the latest
mmcvandmmdetecion.
FAQs¶
None
Tutorial 10: Weight initialization¶
During training, a proper initialization strategy is beneficial to speeding up the training or obtaining a higher performance. MMCV provide some commonly used methods for initializing modules like nn.Conv2d. Model initialization in MMdetection mainly uses init_cfg. Users can initialize models with following two steps:
Define
init_cfgfor a model or its components inmodel_cfg, butinit_cfgof children components have higher priority and will overrideinit_cfgof parents modules.Build model as usual, but call
model.init_weights()method explicitly, and model parameters will be initialized as configuration.
The high-level workflow of initialization in MMdetection is :
model_cfg(init_cfg) -> build_from_cfg -> model -> init_weight() -> initialize(self, self.init_cfg) -> children’s init_weight()
Description¶
It is dict or list[dict], and contains the following keys and values:
type(str), containing the initializer name inINTIALIZERS, and followed by arguments of the initializer.layer(str or list[str]), containing the names of basiclayers in Pytorch or MMCV with learnable parameters that will be initialized, e.g.'Conv2d','DeformConv2d'.override(dict or list[dict]), containing the sub-modules that not inherit from BaseModule and whose initialization configuration is different from other layers’ which are in'layer'key. Initializer defined intypewill work for all layers defined inlayer, so if sub-modules are not derived Classes ofBaseModulebut can be initialized as same ways of layers inlayer, it does not need to useoverride.overridecontains:typefollowed by arguments of initializer;nameto indicate sub-module which will be initialized.
Initialize parameters¶
Inherit a new model from mmcv.runner.BaseModule or mmdet.models Here we show an example of FooModel.
import torch.nn as nn
from mmcv.runner import BaseModule
class FooModel(BaseModule)
def __init__(self,
arg1,
arg2,
init_cfg=None):
super(FooModel, self).__init__(init_cfg)
...
Initialize model by using
init_cfgdirectly in codeimport torch.nn as nn from mmcv.runner import BaseModule # or directly inherit mmdet models class FooModel(BaseModule) def __init__(self, arg1, arg2, init_cfg=XXX): super(FooModel, self).__init__(init_cfg) ...
Initialize model by using
init_cfgdirectly inmmcv.Sequentialormmcv.ModuleListcodefrom mmcv.runner import BaseModule, ModuleList class FooModel(BaseModule) def __init__(self, arg1, arg2, init_cfg=None): super(FooModel, self).__init__(init_cfg) ... self.conv1 = ModuleList(init_cfg=XXX)
Initialize model by using
init_cfgin config filemodel = dict( ... model = dict( type='FooModel', arg1=XXX, arg2=XXX, init_cfg=XXX), ...
Usage of init_cfg¶
Initialize model by
layerkeyIf we only define
layer, it just initialize the layer inlayerkey.NOTE: Value of
layerkey is the class name with attributes weights and bias of Pytorch, (so such asMultiheadAttention layeris not supported).
Define
layerkey for initializing module with same configuration.init_cfg = dict(type='Constant', layer=['Conv1d', 'Conv2d', 'Linear'], val=1) # initialize whole module with same configuration
Define
layerkey for initializing layer with different configurations.
init_cfg = [dict(type='Constant', layer='Conv1d', val=1),
dict(type='Constant', layer='Conv2d', val=2),
dict(type='Constant', layer='Linear', val=3)]
# nn.Conv1d will be initialized with dict(type='Constant', val=1)
# nn.Conv2d will be initialized with dict(type='Constant', val=2)
# nn.Linear will be initialized with dict(type='Constant', val=3)
Initialize model by
overridekey
When initializing some specific part with its attribute name, we can use
overridekey, and the value inoverridewill ignore the value in init_cfg.# layers： # self.feat = nn.Conv1d(3, 1, 3) # self.reg = nn.Conv2d(3, 3, 3) # self.cls = nn.Linear(1,2) init_cfg = dict(type='Constant', layer=['Conv1d','Conv2d'], val=1, bias=2, override=dict(type='Constant', name='reg', val=3, bias=4)) # self.feat and self.cls will be initialized with dict(type='Constant', val=1, bias=2) # The module called 'reg' will be initialized with dict(type='Constant', val=3, bias=4)
If
layeris None in init_cfg, only sub-module with the name in override will be initialized, and type and other args in override can be omitted.# layers： # self.feat = nn.Conv1d(3, 1, 3) # self.reg = nn.Conv2d(3, 3, 3) # self.cls = nn.Linear(1,2) init_cfg = dict(type='Constant', val=1, bias=2, override=dict(name='reg')) # self.feat and self.cls will be initialized by Pytorch # The module called 'reg' will be initialized with dict(type='Constant', val=1, bias=2)
If we don’t define
layerkey oroverridekey, it will not initialize anything.Invalid usage
# It is invalid that override don't have name key init_cfg = dict(type='Constant', layer=['Conv1d','Conv2d'], val=1, bias=2, override=dict(type='Constant', val=3, bias=4)) # It is also invalid that override has name and other args except type init_cfg = dict(type='Constant', layer=['Conv1d','Conv2d'], val=1, bias=2, override=dict(name='reg', val=3, bias=4))
Initialize model with the pretrained model
init_cfg = dict(type='Pretrained', checkpoint='torchvision://resnet50')
More details can refer to the documentation in MMCV and MMCV PR #780
Tutorial 11: How to xxx¶
This tutorial collects answers to any How to xxx with MMDetection. Feel free to update this doc if you meet new questions about How to and find the answers!
Use backbone network through MMClassification¶
The model registry in MMDet, MMCls, MMSeg all inherit from the root registry in MMCV. This allows these repositories to directly use the modules already implemented by each other. Therefore, users can use backbone networks from MMClassification in MMDetection without implementing a network that already exists in MMClassification.
Use backbone network implemented in MMClassification¶
Suppose you want to use MobileNetV3-small as the backbone network of RetinaNet, the example config is as the following.
_base_ = [
'../_base_/models/retinanet_r50_fpn.py',
'../_base_/datasets/coco_detection.py',
'../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py'
]
# please install mmcls>=0.20.0
# import mmcls.models to trigger register_module in mmcls
custom_imports = dict(imports=['mmcls.models'], allow_failed_imports=False)
pretrained = 'https://download.openmmlab.com/mmclassification/v0/mobilenet_v3/convert/mobilenet_v3_small-8427ecf0.pth'
model = dict(
backbone=dict(
_delete_=True, # Delete the backbone field in _base_
type='mmcls.MobileNetV3', # Using MobileNetV3 from mmcls
arch='small',
out_indices=(3, 8, 11), # Modify out_indices
init_cfg=dict(
type='Pretrained',
checkpoint=pretrained,
prefix='backbone.')), # The pre-trained weights of backbone network in MMCls have prefix='backbone.'. The prefix in the keys will be removed so that these weights can be normally loaded.
# Modify in_channels
neck=dict(in_channels=[24, 48, 96], start_level=0))
Use backbone network in TIMM through MMClassification¶
MMClassification also provides a wrapper for the PyTorch Image Models (timm) backbone network, users can directly use the backbone network in timm through MMClassification. Suppose you want to use EfficientNet-B1 as the backbone network of RetinaNet, the example config is as the following.
# https://github.com/open-mmlab/mmdetection/blob/master/configs/timm_example/retinanet_timm_efficientnet_b1_fpn_1x_coco.py
_base_ = [
'../_base_/models/retinanet_r50_fpn.py',
'../_base_/datasets/coco_detection.py',
'../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py'
]
# please install mmcls>=0.20.0
# import mmcls.models to trigger register_module in mmcls
custom_imports = dict(imports=['mmcls.models'], allow_failed_imports=False)
model = dict(
backbone=dict(
_delete_=True, # Delete the backbone field in _base_
type='mmcls.TIMMBackbone', # Using timm from mmcls
model_name='efficientnet_b1',
features_only=True,
pretrained=True,
out_indices=(1, 2, 3, 4)), # Modify out_indices
neck=dict(in_channels=[24, 40, 112, 320])) # Modify in_channels
optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)
type='mmcls.TIMMBackbone' means use the TIMMBackbone class from MMClassification in MMDetection, and the model used is EfficientNet-B1, where mmcls means the MMClassification repo and TIMMBackbone means the TIMMBackbone wrapper implemented in MMClassification.
For the principle of the Hierarchy Registry, please refer to the MMCV document. For how to use other backbones in MMClassification, you can refer to the MMClassification document.
Use Mosaic augmentation¶
If you want to use Mosaic in training, please make sure that you use MultiImageMixDataset at the same time. Taking the ‘Faster R-CNN’ algorithm as an example, you should modify the values of train_pipeline and train_dataset in the config as below:
# Open configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py directly and add the following fields
data_root = 'data/coco/'
dataset_type = 'CocoDataset'
img_scale=(1333, 800)
img_norm_cfg = dict(
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
train_pipeline = [
dict(type='Mosaic', img_scale=img_scale, pad_val=114.0),
dict(
type='RandomAffine',
scaling_ratio_range=(0.1, 2),
border=(-img_scale[0] // 2, -img_scale[1] // 2)), # The image will be enlarged by 4 times after Mosaic processing,so we use affine transformation to restore the image size.
dict(type='RandomFlip', flip_ratio=0.5),
dict(type='Normalize', **img_norm_cfg),
dict(type='Pad', size_divisor=32),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
]
train_dataset = dict(
_delete_ = True, # remove unnecessary Settings
type='MultiImageMixDataset',
dataset=dict(
type=dataset_type,
ann_file=data_root + 'annotations/instances_train2017.json',
img_prefix=data_root + 'train2017/',
pipeline=[
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=True)
],
filter_empty_gt=False,
),
pipeline=train_pipeline
)
data = dict(
train=train_dataset
)
Unfreeze backbone network after freezing the backbone in the config¶
If you have freezed the backbone network in the config and want to unfreeze it after some epoches, you can write a hook function to do it. Taking the Faster R-CNN with the resnet backbone as an example, you can freeze one stage of the backbone network and add a custom_hooks in the config as below:
_base_ = [
'../_base_/models/faster_rcnn_r50_fpn.py',
'../_base_/datasets/coco_detection.py',
'../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py'
]
model = dict(
# freeze one stage of the backbone network.
backbone=dict(frozen_stages=1),
)
custom_hooks = [dict(type="UnfreezeBackboneEpochBasedHook", unfreeze_epoch=1)]
Meanwhile write the hook class UnfreezeBackboneEpochBasedHook in mmdet/core/hook/unfreeze_backbone_epoch_based_hook.py
from mmcv.parallel import is_module_wrapper
from mmcv.runner.hooks import HOOKS, Hook
@HOOKS.register_module()
class UnfreezeBackboneEpochBasedHook(Hook):
"""Unfreeze backbone network Hook.
Args:
unfreeze_epoch (int): The epoch unfreezing the backbone network.
"""
def __init__(self, unfreeze_epoch=1):
self.unfreeze_epoch = unfreeze_epoch
def before_train_epoch(self, runner):
# Unfreeze the backbone network.
# Only valid for resnet.
if runner.epoch == self.unfreeze_epoch:
model = runner.model
if is_module_wrapper(model):
model = model.module
backbone = model.backbone
if backbone.frozen_stages >= 0:
if backbone.deep_stem:
backbone.stem.train()
for param in backbone.stem.parameters():
param.requires_grad = True
else:
backbone.norm1.train()
for m in [backbone.conv1, backbone.norm1]:
for param in m.parameters():
param.requires_grad = True
for i in range(1, backbone.frozen_stages + 1):
m = getattr(backbone, f'layer{i}')
m.train()
for param in m.parameters():
param.requires_grad = True
Get the channels of a new backbone¶
If you want to get the channels of a new backbone, you can build this backbone alone and input a pseudo image to get each stage output.
Take ResNet as an example:
from mmdet.models import ResNet
import torch
self = ResNet(depth=18)
self.eval()
inputs = torch.rand(1, 3, 32, 32)
level_outputs = self.forward(inputs)
for level_out in level_outputs:
print(tuple(level_out.shape))
Output of the above script is as below:
(1, 64, 8, 8)
(1, 128, 4, 4)
(1, 256, 2, 2)
(1, 512, 1, 1)
Users can get the channels of the new backbone by Replacing the ResNet(depth=18) in this script with their customized backbone.
Tutorial 12: Test Results Submission¶
Panoptic segmentation test results submission¶
The following sections introduce how to produce the prediction results of panoptic segmentation models on the COCO test-dev set and submit the predictions to COCO evaluation server.
Prerequisites¶
Download COCO test dataset images, testing image info, and panoptic train/val annotations, then unzip them, put ‘test2017’ to
data/coco/, put json files and annotation files todata/coco/annotations/.
# suppose data/coco/ does not exist
mkdir -pv data/coco/
# download test2017
wget -P data/coco/ http://images.cocodataset.org/zips/test2017.zip
wget -P data/coco/ http://images.cocodataset.org/annotations/image_info_test2017.zip
wget -P data/coco/ http://images.cocodataset.org/annotations/panoptic_annotations_trainval2017.zip
# unzip them
unzip data/coco/test2017.zip -d data/coco/
unzip data/coco/image_info_test2017.zip -d data/coco/
unzip data/coco/panoptic_annotations_trainval2017.zip -d data/coco/
# remove zip files (optional)
rm -rf data/coco/test2017.zip data/coco/image_info_test2017.zip data/coco/panoptic_annotations_trainval2017.zip
Run the following code to update category information in testing image info. Since the attribute
isthingis missing in category information of ‘image_info_test-dev2017.json’, we need to update it with the category information in ‘panoptic_val2017.json’.
python tools/misc/gen_coco_panoptic_test_info.py data/coco/annotations
After completing the above preparations, your directory structure of data should be like this:
data
`-- coco
|-- annotations
| |-- image_info_test-dev2017.json
| |-- image_info_test2017.json
| |-- panoptic_image_info_test-dev2017.json
| |-- panoptic_train2017.json
| |-- panoptic_train2017.zip
| |-- panoptic_val2017.json
| `-- panoptic_val2017.zip
`-- test2017
Inference on coco test-dev¶
The commands to perform inference on test2017 are as below:
# test with single gpu
CUDA_VISIBLE_DEVICES=0 python tools/test.py \
${CONFIG_FILE} \
${CHECKPOINT_FILE} \
--format-only \
--cfg-options data.test.ann_file=data/coco/annotations/panoptic_image_info_test-dev2017.json data.test.img_prefix=data/coco/test2017 \
--eval-options jsonfile_prefix=${WORK_DIR}/results
# test with four gpus
CUDA_VISIBLE_DEVICES=0,1,3,4 bash tools/dist_test.sh \
${CONFIG_FILE} \
${CHECKPOINT_FILE} \
4 \ # four gpus
--format-only \
--cfg-options data.test.ann_file=data/coco/annotations/panoptic_image_info_test-dev2017.json data.test.img_prefix=data/coco/test2017 \
--eval-options jsonfile_prefix=${WORK_DIR}/results
# test with slurm
GPUS=8 tools/slurm_test.sh \
${Partition} \
${JOB_NAME} \
${CONFIG_FILE} \
${CHECKPOINT_FILE} \
--format-only \
--cfg-options data.test.ann_file=data/coco/annotations/panoptic_image_info_test-dev2017.json data.test.img_prefix=data/coco/test2017 \
--eval-options jsonfile_prefix=${WORK_DIR}/results
Example
Suppose we perform inference on test2017 using pretrained MaskFormer with ResNet-50 backbone.
# test with single gpu
CUDA_VISIBLE_DEVICES=0 python tools/test.py \
configs/maskformer/maskformer_r50_mstrain_16x1_75e_coco.py \
checkpoints/maskformer_r50_mstrain_16x1_75e_coco_20220221_141956-bc2699cb.pth \
--format-only \
--cfg-options data.test.ann_file=data/coco/annotations/panoptic_image_info_test-dev2017.json data.test.img_prefix=data/coco/test2017 \
--eval-options jsonfile_prefix=work_dirs/maskformer/results
Rename files and zip results¶
After inference, the panoptic segmentation results (a json file and a directory where the masks are stored) will be in WORK_DIR. We should rename them according to the naming convention described on COCO’s Website. Finally, we need to compress the json and the directory where the masks are stored into a zip file, and rename the zip file according to the naming convention. Note that the zip file should directly contains the above two files.
The commands to rename files and zip results:
# In WORK_DIR, we have panoptic segmentation results: 'panoptic' and 'results.panoptic.json'.
cd ${WORK_DIR}
# replace '[algorithm_name]' with the name of algorithm you used.
mv ./panoptic ./panoptic_test-dev2017_[algorithm_name]_results
mv ./results.panoptic.json ./panoptic_test-dev2017_[algorithm_name]_results.json
zip panoptic_test-dev2017_[algorithm_name]_results.zip -ur panoptic_test-dev2017_[algorithm_name]_results panoptic_test-dev2017_[algorithm_name]_results.json
Tutorial 13: Useful Hooks¶
MMDetection and MMCV provide users with various useful hooks including log hooks, evaluation hooks, NumClassCheckHook, etc. This tutorial introduces the functionalities and usages of hooks implemented in MMDetection. For using hooks in MMCV, please read the API documentation in MMCV.
CheckInvalidLossHook¶
EvalHook and DistEvalHook¶
ExpMomentumEMAHook and LinearMomentumEMAHook¶
NumClassCheckHook¶
MemoryProfilerHook¶
Memory profiler hook records memory information including virtual memory, swap memory, and the memory of the current process. This hook helps grasp the memory usage of the system and discover potential memory leak bugs. To use this hook, users should install memory_profiler and psutil by pip install memory_profiler psutil first.
Usage¶
To use this hook, users should add the following code to the config file.
custom_hooks = [
dict(type='MemoryProfilerHook', interval=50)
]
Result¶
During training, you can see the messages in the log recorded by MemoryProfilerHook as below. The system has 250 GB (246360 MB + 9407 MB) of memory and 8 GB (5740 MB + 2452 MB) of swap memory in total. Currently 9407 MB (4.4%) of memory and 5740 MB (29.9%) of swap memory were consumed. And the current training process consumed 5434 MB of memory.
2022-04-21 08:49:56,881 - mmdet - INFO - Memory information available_memory: 246360 MB, used_memory: 9407 MB, memory_utilization: 4.4 %, available_swap_memory: 5740 MB, used_swap_memory: 2452 MB, swap_memory_utilization: 29.9 %, current_process_memory: 5434 MB
SetEpochInfoHook¶
SyncNormHook¶
SyncRandomSizeHook¶
YOLOXLrUpdaterHook¶
YOLOXModeSwitchHook¶
How to implement a custom hook¶
In general, there are 10 points where hooks can be inserted from the beginning to the end of model training. The users can implement custom hooks and insert them at different points in the process of training to do what they want.
global points:
before_run,after_runpoints in training:
before_train_epoch,before_train_iter,after_train_iter,after_train_epochpoints in validation:
before_val_epoch,before_val_iter,after_val_iter,after_val_epoch
For example, users can implement a hook to check loss and terminate training when loss goes NaN. To achieve that, there are three steps to go:
Implement a new hook that inherits the
Hookclass in MMCV, and implementafter_train_itermethod which checks whether loss goes NaN after everyntraining iterations.The implemented hook should be registered in
HOOKSby@HOOKS.register_module()as shown in the code below.Add
custom_hooks = [dict(type='CheckInvalidLossHook', interval=50)]in the config file.
import torch
from mmcv.runner.hooks import HOOKS, Hook
@HOOKS.register_module()
class CheckInvalidLossHook(Hook):
"""Check invalid loss hook.
This hook will regularly check whether the loss is valid
during training.
Args:
interval (int): Checking interval (every k iterations).
Default: 50.
"""
def __init__(self, interval=50):
self.interval = interval
def after_train_iter(self, runner):
if self.every_n_iters(runner, self.interval):
assert torch.isfinite(runner.outputs['loss']), \
runner.logger.info('loss become infinite or NaN!')
Please read customize_runtime for more about implementing a custom hook.
Apart from training/testing scripts, We provide lots of useful tools under the
tools/ directory.
Log Analysis¶
tools/analysis_tools/analyze_logs.py plots loss/mAP curves given a training
log file. Run pip install seaborn first to install the dependency.
python tools/analysis_tools/analyze_logs.py plot_curve [--keys ${KEYS}] [--eval-interval ${EVALUATION_INTERVAL}] [--title ${TITLE}] [--legend ${LEGEND}] [--backend ${BACKEND}] [--style ${STYLE}] [--out ${OUT_FILE}]
Examples:
Plot the classification loss of some run.
python tools/analysis_tools/analyze_logs.py plot_curve log.json --keys loss_cls --legend loss_cls
Plot the classification and regression loss of some run, and save the figure to a pdf.
python tools/analysis_tools/analyze_logs.py plot_curve log.json --keys loss_cls loss_bbox --out losses.pdf
Compare the bbox mAP of two runs in the same figure.
python tools/analysis_tools/analyze_logs.py plot_curve log1.json log2.json --keys bbox_mAP --legend run1 run2
Compute the average training speed.
python tools/analysis_tools/analyze_logs.py cal_train_time log.json [--include-outliers]
The output is expected to be like the following.
-----Analyze train time of work_dirs/some_exp/20190611_192040.log.json----- slowest epoch 11, average time is 1.2024 fastest epoch 1, average time is 1.1909 time std over epochs is 0.0028 average iter time: 1.1959 s/iter
Result Analysis¶
tools/analysis_tools/analyze_results.py calculates single image mAP and saves or shows the topk images with the highest and lowest scores based on prediction results.
Usage
python tools/analysis_tools/analyze_results.py \
${CONFIG} \
${PREDICTION_PATH} \
${SHOW_DIR} \
[--show] \
[--wait-time ${WAIT_TIME}] \
[--topk ${TOPK}] \
[--show-score-thr ${SHOW_SCORE_THR}] \
[--cfg-options ${CFG_OPTIONS}]
Description of all arguments:
config: The path of a model config file.prediction_path: Output result file in pickle format fromtools/test.pyshow_dir: Directory where painted GT and detection images will be saved--show：Determines whether to show painted images, If not specified, it will be set toFalse--wait-time: The interval of show (s), 0 is block--topk: The number of saved images that have the highest and lowesttopkscores after sorting. If not specified, it will be set to20.--show-score-thr: Show score threshold. If not specified, it will be set to0.--cfg-options: If specified, the key-value pair optional cfg will be merged into config file
Examples:
Assume that you have got result file in pickle format from tools/test.py in the path ‘./result.pkl’.
Test Faster R-CNN and visualize the results, save images to the directory
results/
python tools/analysis_tools/analyze_results.py \
configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \
result.pkl \
results \
--show
Test Faster R-CNN and specified topk to 50, save images to the directory
results/
python tools/analysis_tools/analyze_results.py \
configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \
result.pkl \
results \
--topk 50
If you want to filter the low score prediction results, you can specify the
show-score-thrparameter
python tools/analysis_tools/analyze_results.py \
configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \
result.pkl \
results \
--show-score-thr 0.3
Visualization¶
Visualize Datasets¶
tools/misc/browse_dataset.py helps the user to browse a detection dataset (both
images and bounding box annotations) visually, or save the image to a
designated directory.
python tools/misc/browse_dataset.py ${CONFIG} [-h] [--skip-type ${SKIP_TYPE[SKIP_TYPE...]}] [--output-dir ${OUTPUT_DIR}] [--not-show] [--show-interval ${SHOW_INTERVAL}]
Visualize Models¶
First, convert the model to ONNX as described here. Note that currently only RetinaNet is supported, support for other models will be coming in later versions. The converted model could be visualized by tools like Netron.
Visualize Predictions¶
If you need a lightweight GUI for visualizing the detection results, you can refer DetVisGUI project.
Error Analysis¶
tools/analysis_tools/coco_error_analysis.py analyzes COCO results per category and by
different criterion. It can also make a plot to provide useful information.
python tools/analysis_tools/coco_error_analysis.py ${RESULT} ${OUT_DIR} [-h] [--ann ${ANN}] [--types ${TYPES[TYPES...]}]
Example:
Assume that you have got Mask R-CNN checkpoint file in the path ‘checkpoint’. For other checkpoints, please refer to our model zoo. You can use the following command to get the results bbox and segmentation json file.
# out: results.bbox.json and results.segm.json
python tools/test.py \
configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.py \
checkpoint/mask_rcnn_r50_fpn_1x_coco_20200205-d4b0c5d6.pth \
--format-only \
--options "jsonfile_prefix=./results"
Get COCO bbox error results per category , save analyze result images to the directory
results/
python tools/analysis_tools/coco_error_analysis.py \
results.bbox.json \
results \
--ann=data/coco/annotations/instances_val2017.json \
Get COCO segmentation error results per category , save analyze result images to the directory
results/
python tools/analysis_tools/coco_error_analysis.py \
results.segm.json \
results \
--ann=data/coco/annotations/instances_val2017.json \
--types='segm'
Model Serving¶
In order to serve an MMDetection model with TorchServe, you can follow the steps:
1. Convert model from MMDetection to TorchServe¶
python tools/deployment/mmdet2torchserve.py ${CONFIG_FILE} ${CHECKPOINT_FILE} \
--output-folder ${MODEL_STORE} \
--model-name ${MODEL_NAME}
Note: ${MODEL_STORE} needs to be an absolute path to a folder.
2. Build mmdet-serve docker image¶
docker build -t mmdet-serve:latest docker/serve/
3. Run mmdet-serve¶
Check the official docs for running TorchServe with docker.
In order to run in GPU, you need to install nvidia-docker. You can omit the --gpus argument in order to run in CPU.
Example:
docker run --rm \
--cpus 8 \
--gpus device=0 \
-p8080:8080 -p8081:8081 -p8082:8082 \
--mount type=bind,source=$MODEL_STORE,target=/home/model-server/model-store \
mmdet-serve:latest
Read the docs about the Inference (8080), Management (8081) and Metrics (8082) APis
4. Test deployment¶
curl -O curl -O https://raw.githubusercontent.com/pytorch/serve/master/docs/images/3dogs.jpg
curl http://127.0.0.1:8080/predictions/${MODEL_NAME} -T 3dogs.jpg
You should obtain a response similar to:
[
{
"class_name": "dog",
"bbox": [
294.63409423828125,
203.99111938476562,
417.048583984375,
281.62744140625
],
"score": 0.9987992644309998
},
{
"class_name": "dog",
"bbox": [
404.26019287109375,
126.0080795288086,
574.5091552734375,
293.6662292480469
],
"score": 0.9979367256164551
},
{
"class_name": "dog",
"bbox": [
197.2144775390625,
93.3067855834961,
307.8505554199219,
276.7560119628906
],
"score": 0.993338406085968
}
]
And you can use test_torchserver.py to compare result of torchserver and pytorch, and visualize them.
python tools/deployment/test_torchserver.py ${IMAGE_FILE} ${CONFIG_FILE} ${CHECKPOINT_FILE} ${MODEL_NAME}
[--inference-addr ${INFERENCE_ADDR}] [--device ${DEVICE}] [--score-thr ${SCORE_THR}]
Example:
python tools/deployment/test_torchserver.py \
demo/demo.jpg \
configs/yolo/yolov3_d53_320_273e_coco.py \
checkpoint/yolov3_d53_320_273e_coco-421362b6.pth \
yolov3
Model Complexity¶
tools/analysis_tools/get_flops.py is a script adapted from flops-counter.pytorch to compute the FLOPs and params of a given model.
python tools/analysis_tools/get_flops.py ${CONFIG_FILE} [--shape ${INPUT_SHAPE}]
You will get the results like this.
==============================
Input shape: (3, 1280, 800)
Flops: 239.32 GFLOPs
Params: 37.74 M
==============================
Note: This tool is still experimental and we do not guarantee that the number is absolutely correct. You may well use the result for simple comparisons, but double check it before you adopt it in technical reports or papers.
FLOPs are related to the input shape while parameters are not. The default input shape is (1, 3, 1280, 800).
Some operators are not counted into FLOPs like GN and custom operators. Refer to
mmcv.cnn.get_model_complexity_info()for details.The FLOPs of two-stage detectors is dependent on the number of proposals.
Model conversion¶
MMDetection model to ONNX (experimental)¶
We provide a script to convert model to ONNX format. We also support comparing the output results between Pytorch and ONNX model for verification.
python tools/deployment/pytorch2onnx.py ${CONFIG_FILE} ${CHECKPOINT_FILE} --output-file ${ONNX_FILE} [--shape ${INPUT_SHAPE} --verify]
Note: This tool is still experimental. Some customized operators are not supported for now. For a detailed description of the usage and the list of supported models, please refer to pytorch2onnx.
MMDetection 1.x model to MMDetection 2.x¶
tools/model_converters/upgrade_model_version.py upgrades a previous MMDetection checkpoint
to the new version. Note that this script is not guaranteed to work as some
breaking changes are introduced in the new version. It is recommended to
directly use the new checkpoints.
python tools/model_converters/upgrade_model_version.py ${IN_FILE} ${OUT_FILE} [-h] [--num-classes NUM_CLASSES]
RegNet model to MMDetection¶
tools/model_converters/regnet2mmdet.py convert keys in pycls pretrained RegNet models to
MMDetection style.
python tools/model_converters/regnet2mmdet.py ${SRC} ${DST} [-h]
Detectron ResNet to Pytorch¶
tools/model_converters/detectron2pytorch.py converts keys in the original detectron pretrained
ResNet models to PyTorch style.
python tools/model_converters/detectron2pytorch.py ${SRC} ${DST} ${DEPTH} [-h]
Prepare a model for publishing¶
tools/model_converters/publish_model.py helps users to prepare their model for publishing.
Before you upload a model to AWS, you may want to
convert model weights to CPU tensors
delete the optimizer states and
compute the hash of the checkpoint file and append the hash id to the filename.
python tools/model_converters/publish_model.py ${INPUT_FILENAME} ${OUTPUT_FILENAME}
E.g.,
python tools/model_converters/publish_model.py work_dirs/faster_rcnn/latest.pth faster_rcnn_r50_fpn_1x_20190801.pth
The final output filename will be faster_rcnn_r50_fpn_1x_20190801-{hash id}.pth.
Dataset Conversion¶
tools/data_converters/ contains tools to convert the Cityscapes dataset
and Pascal VOC dataset to the COCO format.
python tools/dataset_converters/cityscapes.py ${CITYSCAPES_PATH} [-h] [--img-dir ${IMG_DIR}] [--gt-dir ${GT_DIR}] [-o ${OUT_DIR}] [--nproc ${NPROC}]
python tools/dataset_converters/pascal_voc.py ${DEVKIT_PATH} [-h] [-o ${OUT_DIR}]
Dataset Download¶
tools/misc/download_dataset.py supports downloading datasets such as COCO, VOC, and LVIS.
python tools/misc/download_dataset.py --dataset-name coco2017
python tools/misc/download_dataset.py --dataset-name voc2007
python tools/misc/download_dataset.py --dataset-name lvis
Benchmark¶
Robust Detection Benchmark¶
tools/analysis_tools/test_robustness.py andtools/analysis_tools/robustness_eval.py helps users to evaluate model robustness. The core idea comes from Benchmarking Robustness in Object Detection: Autonomous Driving when Winter is Coming. For more information how to evaluate models on corrupted images and results for a set of standard models please refer to robustness_benchmarking.md.
FPS Benchmark¶
tools/analysis_tools/benchmark.py helps users to calculate FPS. The FPS value includes model forward and post-processing. In order to get a more accurate value, currently only supports single GPU distributed startup mode.
python -m torch.distributed.launch --nproc_per_node=1 --master_port=${PORT} tools/analysis_tools/benchmark.py \
${CONFIG} \
${CHECKPOINT} \
[--repeat-num ${REPEAT_NUM}] \
[--max-iter ${MAX_ITER}] \
[--log-interval ${LOG_INTERVAL}] \
--launcher pytorch
Examples: Assuming that you have already downloaded the Faster R-CNN model checkpoint to the directory checkpoints/.
python -m torch.distributed.launch --nproc_per_node=1 --master_port=29500 tools/analysis_tools/benchmark.py \
configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \
checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth \
--launcher pytorch
Miscellaneous¶
Evaluating a metric¶
tools/analysis_tools/eval_metric.py evaluates certain metrics of a pkl result file
according to a config file.
python tools/analysis_tools/eval_metric.py ${CONFIG} ${PKL_RESULTS} [-h] [--format-only] [--eval ${EVAL[EVAL ...]}]
[--cfg-options ${CFG_OPTIONS [CFG_OPTIONS ...]}]
[--eval-options ${EVAL_OPTIONS [EVAL_OPTIONS ...]}]
Print the entire config¶
tools/misc/print_config.py prints the whole config verbatim, expanding all its
imports.
python tools/misc/print_config.py ${CONFIG} [-h] [--options ${OPTIONS [OPTIONS...]}]
Hyper-parameter Optimization¶
YOLO Anchor Optimization¶
tools/analysis_tools/optimize_anchors.py provides two method to optimize YOLO anchors.
One is k-means anchor cluster which refers from darknet.
python tools/analysis_tools/optimize_anchors.py ${CONFIG} --algorithm k-means --input-shape ${INPUT_SHAPE [WIDTH HEIGHT]} --output-dir ${OUTPUT_DIR}
Another is using differential evolution to optimize anchors.
python tools/analysis_tools/optimize_anchors.py ${CONFIG} --algorithm differential_evolution --input-shape ${INPUT_SHAPE [WIDTH HEIGHT]} --output-dir ${OUTPUT_DIR}
E.g.,
python tools/analysis_tools/optimize_anchors.py configs/yolo/yolov3_d53_320_273e_coco.py --algorithm differential_evolution --input-shape 608 608 --device cuda --output-dir work_dirs
You will get:
loading annotations into memory...
Done (t=9.70s)
creating index...
index created!
2021-07-19 19:37:20,951 - mmdet - INFO - Collecting bboxes from annotation...
[>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 117266/117266, 15874.5 task/s, elapsed: 7s, ETA: 0s
2021-07-19 19:37:28,753 - mmdet - INFO - Collected 849902 bboxes.
differential_evolution step 1: f(x)= 0.506055
differential_evolution step 2: f(x)= 0.506055
......
differential_evolution step 489: f(x)= 0.386625
2021-07-19 19:46:40,775 - mmdet - INFO Anchor evolution finish. Average IOU: 0.6133754253387451
2021-07-19 19:46:40,776 - mmdet - INFO Anchor differential evolution result:[[10, 12], [15, 30], [32, 22], [29, 59], [61, 46], [57, 116], [112, 89], [154, 198], [349, 336]]
2021-07-19 19:46:40,798 - mmdet - INFO Result saved in work_dirs/anchor_optimize_result.json
Confusion Matrix¶
A confusion matrix is a summary of prediction results.
tools/analysis_tools/confusion_matrix.py can analyze the prediction results and plot a confusion matrix table.
First, run tools/test.py to save the .pkl detection results.
Then, run
python tools/analysis_tools/confusion_matrix.py ${CONFIG} ${DETECTION_RESULTS} ${SAVE_DIR} --show
And you will get a confusion matrix like this:
COCO Separated & Occluded Mask Metric¶
Detecting occluded objects still remains a challenge for state-of-the-art object detectors. We implemented the metric presented in paper A Tri-Layer Plugin to Improve Occluded Detection to calculate the recall of separated and occluded masks.
There are two ways to use this metric:
Offline evaluation¶
We provide a script to calculate the metric with a dumped prediction file.
First, use the tools/test.py script to dump the detection results:
python tools/test.py ${CONFIG} ${MODEL_PATH} --out results.pkl
Then, run the tools/analysis_tools/coco_occluded_separated_recall.py script to get the recall of separated and occluded masks:
python tools/analysis_tools/coco_occluded_separated_recall.py results.pkl --out occluded_separated_recall.json
The output should be like this:
loading annotations into memory...
Done (t=0.51s)
creating index...
index created!
processing detection results...
[>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 5000/5000, 109.3 task/s, elapsed: 46s, ETA: 0s
computing occluded mask recall...
[>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 5550/5550, 780.5 task/s, elapsed: 7s, ETA: 0s
COCO occluded mask recall: 58.79%
COCO occluded mask success num: 3263
computing separated mask recall...
[>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 3522/3522, 778.3 task/s, elapsed: 5s, ETA: 0s
COCO separated mask recall: 31.94%
COCO separated mask success num: 1125
+-----------+--------+-------------+
| mask type | recall | num correct |
+-----------+--------+-------------+
| occluded | 58.79% | 3263 |
| separated | 31.94% | 1125 |
+-----------+--------+-------------+
Evaluation results have been saved to occluded_separated_recall.json.
Online evaluation¶
We implement OccludedSeparatedCocoDataset which inherited from the CocoDataset.
To evaluate the recall of separated and occluded masks during training, just replace the validation dataset type with 'OccludedSeparatedCocoDataset' in your config:
data = dict(
samples_per_gpu=2,
workers_per_gpu=2,
train=dict(
type=dataset_type,
ann_file=data_root + 'annotations/instances_train2017.json',
img_prefix=data_root + 'train2017/',
pipeline=train_pipeline),
val=dict(
type='OccludedSeparatedCocoDataset', # modify this
ann_file=data_root + 'annotations/instances_val2017.json',
img_prefix=data_root + 'val2017/',
pipeline=test_pipeline),
test=dict(
type='OccludedSeparatedCocoDataset', # modify this
ann_file=data_root + 'annotations/instances_val2017.json',
img_prefix=data_root + 'val2017/',
pipeline=test_pipeline))
Please cite the paper if you use this metric:
@article{zhan2022triocc,
title={A Tri-Layer Plugin to Improve Occluded Detection},
author={Zhan, Guanqi and Xie, Weidi and Zisserman, Andrew},
journal={British Machine Vision Conference},
year={2022}
}
Conventions¶
Please check the following conventions if you would like to modify MMDetection as your own project.
Loss¶
In MMDetection, a dict containing losses and metrics will be returned by model(**data).
For example, in bbox head,
class BBoxHead(nn.Module):
...
def loss(self, ...):
losses = dict()
# classification loss
losses['loss_cls'] = self.loss_cls(...)
# classification accuracy
losses['acc'] = accuracy(...)
# bbox regression loss
losses['loss_bbox'] = self.loss_bbox(...)
return losses
bbox_head.loss() will be called during model forward.
The returned dict contains 'loss_bbox', 'loss_cls', 'acc' .
Only 'loss_bbox', 'loss_cls' will be used during back propagation,
'acc' will only be used as a metric to monitor training process.
By default, only values whose keys contain 'loss' will be back propagated.
This behavior could be changed by modifying BaseDetector.train_step().
Empty Proposals¶
In MMDetection, We have added special handling and unit test for empty proposals of two-stage. We need to deal with the empty proposals of the entire batch and single image at the same time. For example, in CascadeRoIHead,
# simple_test method
...
# There is no proposal in the whole batch
if rois.shape[0] == 0:
bbox_results = [[
np.zeros((0, 5), dtype=np.float32)
for _ in range(self.bbox_head[-1].num_classes)
]] * num_imgs
if self.with_mask:
mask_classes = self.mask_head[-1].num_classes
segm_results = [[[] for _ in range(mask_classes)]
for _ in range(num_imgs)]
results = list(zip(bbox_results, segm_results))
else:
results = bbox_results
return results
...
# There is no proposal in the single image
for i in range(self.num_stages):
...
if i < self.num_stages - 1:
for j in range(num_imgs):
# Handle empty proposal
if rois[j].shape[0] > 0:
bbox_label = cls_score[j][:, :-1].argmax(dim=1)
refine_roi = self.bbox_head[i].regress_by_class(
rois[j], bbox_label, bbox_pred[j], img_metas[j])
refine_roi_list.append(refine_roi)
If you have customized RoIHead, you can refer to the above method to deal with empty proposals.
Coco Panoptic Dataset¶
In MMDetection, we have supported COCO Panoptic dataset. We clarify a few conventions about the implementation of CocoPanopticDataset here.
For mmdet<=2.16.0, the range of foreground and background labels in semantic segmentation are different from the default setting of MMDetection. The label
0stands forVOIDlabel and the category labels start from1. Since mmdet=2.17.0, the category labels of semantic segmentation start from0and label255stands forVOIDfor consistency with labels of bounding boxes. To achieve that, thePadpipeline supports setting the padding value forseg.In the evaluation, the panoptic result is a map with the same shape as the original image. Each value in the result map has the format of
instance_id * INSTANCE_OFFSET + category_id.
Compatibility of MMDetection 2.x¶
MMDetection 2.25.0¶
In order to support Mask2Former for instance segmentation, the original config files of Mask2Former for panpotic segmentation need to be renamed PR #7571.
| before v2.25.0 | after v2.25.0 |
'mask2former_xxx_coco.py' represents config files for **panoptic segmentation**.
|
'mask2former_xxx_coco.py' represents config files for **instance segmentation**.
'mask2former_xxx_coco-panoptic.py' represents config files for **panoptic segmentation**.
|
|---|
MMDetection 2.21.0¶
In order to support CPU training, the logic of scatter in batch collating has been changed. We recommend to use MMCV v1.4.4 or higher. For more details, please refer to MMCV PR #1621.
MMDetection 2.18.1¶
MMCV compatibility¶
In order to fix the wrong weight reference bug in BaseTransformerLayer, the logic in batch first mode of MultiheadAttention has been changed. We recommend to use MMCV v1.3.17 or higher. For more details, please refer to MMCV PR #1418.
MMDetection 2.18.0¶
DIIHead compatibility¶
In order to support QueryInst, attn_feats is added into the returned tuple of DIIHead.
MMDetection 2.14.0¶
MMCV Version¶
In order to fix the problem that the priority of EvalHook is too low, all hook priorities have been re-adjusted in 1.3.8, so MMDetection 2.14.0 needs to rely on the latest MMCV 1.3.8 version. For related information, please refer to #1120, for related issues, please refer to #5343.
SSD compatibility¶
In v2.14.0, to make SSD more flexible to use, PR5291 refactored its backbone, neck and head. The users can use the script tools/model_converters/upgrade_ssd_version.py to convert their models.
python tools/model_converters/upgrade_ssd_version.py ${OLD_MODEL_PATH} ${NEW_MODEL_PATH}
OLD_MODEL_PATH: the path to load the old version SSD model.
NEW_MODEL_PATH: the path to save the converted model weights.
MMDetection 2.12.0¶
MMDetection is going through big refactoring for more general and convenient usages during the releases from v2.12.0 to v2.18.0 (maybe longer). In v2.12.0 MMDetection inevitably brings some BC-breakings, including the MMCV dependency, model initialization, model registry, and mask AP evaluation.
MMCV Version¶
MMDetection v2.12.0 relies on the newest features in MMCV 1.3.3, including BaseModule for unified parameter initialization, model registry, and the CUDA operator MultiScaleDeformableAttn for Deformable DETR. Note that MMCV 1.3.2 already contains all the features used by MMDet but has known issues. Therefore, we recommend users to skip MMCV v1.3.2 and use v1.3.2, though v1.3.2 might work for most of the cases.
Unified model initialization¶
To unify the parameter initialization in OpenMMLab projects, MMCV supports BaseModule that accepts init_cfg to allow the modules’ parameters initialized in a flexible and unified manner. Now the users need to explicitly call model.init_weights() in the training script to initialize the model (as in here, previously this was handled by the detector. The downstream projects must update their model initialization accordingly to use MMDetection v2.12.0. Please refer to PR #4750 for details.
Unified model registry¶
To easily use backbones implemented in other OpenMMLab projects, MMDetection v2.12.0 inherits the model registry created in MMCV (#760). In this way, as long as the backbone is supported in an OpenMMLab project and that project also uses the registry in MMCV, users can use that backbone in MMDetection by simply modifying the config without copying the code of that backbone into MMDetection. Please refer to PR #5059 for more details.
Mask AP evaluation¶
Before PR 4898 and V2.12.0, the mask AP of small, medium, and large instances is calculated based on the bounding box area rather than the real mask area. This leads to higher APs and APm but lower APl but will not affect the overall mask AP. PR 4898 change it to use mask areas by deleting bbox in mask AP calculation.
The new calculation does not affect the overall mask AP evaluation and is consistent with Detectron2.
Compatibility with MMDetection 1.x¶
MMDetection 2.0 goes through a big refactoring and addresses many legacy issues. It is not compatible with the 1.x version, i.e., running inference with the same model weights in these two versions will produce different results. Thus, MMDetection 2.0 re-benchmarks all the models and provides their links and logs in the model zoo.
The major differences are in four folds: coordinate system, codebase conventions, training hyperparameters, and modular design.
Coordinate System¶
The new coordinate system is consistent with Detectron2 and treats the center of the most left-top pixel as (0, 0) rather than the left-top corner of that pixel.
Accordingly, the system interprets the coordinates in COCO bounding box and segmentation annotations as coordinates in range [0, width] or [0, height].
This modification affects all the computation related to the bbox and pixel selection,
which is more natural and accurate.
The height and width of a box with corners (x1, y1) and (x2, y2) in the new coordinate system is computed as
width = x2 - x1andheight = y2 - y1. In MMDetection 1.x and previous version, a “+ 1” was added both height and width. This modification are in three folds:Box transformation and encoding/decoding in regression.
IoU calculation. This affects the matching process between ground truth and bounding box and the NMS process. The effect to compatibility is very negligible, though.
The corners of bounding box is in float type and no longer quantized. This should provide more accurate bounding box results. This also makes the bounding box and RoIs not required to have minimum size of 1, whose effect is small, though.
The anchors are center-aligned to feature grid points and in float type. In MMDetection 1.x and previous version, the anchors are in
inttype and not center-aligned. This affects the anchor generation in RPN and all the anchor-based methods.ROIAlign is better aligned with the image coordinate system. The new implementation is adopted from Detectron2. The RoIs are shifted by half a pixel by default when they are used to cropping RoI features, compared to MMDetection 1.x. The old behavior is still available by setting
aligned=Falseinstead ofaligned=True.Mask cropping and pasting are more accurate.
We use the new RoIAlign to crop mask targets. In MMDetection 1.x, the bounding box is quantized before it is used to crop mask target, and the crop process is implemented by numpy. In new implementation, the bounding box for crop is not quantized and sent to RoIAlign. This implementation accelerates the training speed by a large margin (~0.1s per iter, ~2 hour when training Mask R50 for 1x schedule) and should be more accurate.
In MMDetection 2.0, the “
paste_mask()” function is different and should be more accurate than those in previous versions. This change follows the modification in Detectron2 and can improve mask AP on COCO by ~0.5% absolute.
Codebase Conventions¶
MMDetection 2.0 changes the order of class labels to reduce unused parameters in regression and mask branch more naturally (without +1 and -1). This effect all the classification layers of the model to have a different ordering of class labels. The final layers of regression branch and mask head no longer keep K+1 channels for K categories, and their class orders are consistent with the classification branch.
In MMDetection 2.0, label “K” means background, and labels [0, K-1] correspond to the K = num_categories object categories.
In MMDetection 1.x and previous version, label “0” means background, and labels [1, K] correspond to the K categories.
Note: The class order of softmax RPN is still the same as that in 1.x in versions<=2.4.0 while sigmoid RPN is not affected. The class orders in all heads are unified since MMDetection v2.5.0.
Low quality matching in R-CNN is not used. In MMDetection 1.x and previous versions, the
max_iou_assignerwill match low quality boxes for each ground truth box in both RPN and R-CNN training. We observe this sometimes does not assign the most perfect GT box to some bounding boxes, thus MMDetection 2.0 do not allow low quality matching by default in R-CNN training in the new system. This sometimes may slightly improve the box AP (~0.1% absolute).Separate scale factors for width and height. In MMDetection 1.x and previous versions, the scale factor is a single float in mode
keep_ratio=True. This is slightly inaccurate because the scale factors for width and height have slight difference. MMDetection 2.0 adopts separate scale factors for width and height, the improvement on AP ~0.1% absolute.Configs name conventions are changed. MMDetection V2.0 adopts the new name convention to maintain the gradually growing model zoo as the following:
[model]_(model setting)_[backbone]_[neck]_(norm setting)_(misc)_(gpu x batch)_[schedule]_[dataset].py,
where the (
misc) includes DCN and GCBlock, etc. More details are illustrated in the documentation for configMMDetection V2.0 uses new ResNet Caffe backbones to reduce warnings when loading pre-trained models. Most of the new backbones’ weights are the same as the former ones but do not have
conv.bias, except that they use a differentimg_norm_cfg. Thus, the new backbone will not cause warning of unexpected keys.
Training Hyperparameters¶
The change in training hyperparameters does not affect model-level compatibility but slightly improves the performance. The major ones are:
The number of proposals after nms is changed from 2000 to 1000 by setting
nms_post=1000andmax_num=1000. This slightly improves both mask AP and bbox AP by ~0.2% absolute.The default box regression losses for Mask R-CNN, Faster R-CNN and RetinaNet are changed from smooth L1 Loss to L1 loss. This leads to an overall improvement in box AP (~0.6% absolute). However, using L1-loss for other methods such as Cascade R-CNN and HTC does not improve the performance, so we keep the original settings for these methods.
The sample num of RoIAlign layer is set to be 0 for simplicity. This leads to slightly improvement on mask AP (~0.2% absolute).
The default setting does not use gradient clipping anymore during training for faster training speed. This does not degrade performance of the most of models. For some models such as RepPoints we keep using gradient clipping to stabilize the training process and to obtain better performance.
The default warmup ratio is changed from 1/3 to 0.001 for a more smooth warming up process since the gradient clipping is usually not used. The effect is found negligible during our re-benchmarking, though.
Upgrade Models from 1.x to 2.0¶
To convert the models trained by MMDetection V1.x to MMDetection V2.0, the users can use the script tools/model_converters/upgrade_model_version.py to convert
their models. The converted models can be run in MMDetection V2.0 with slightly dropped performance (less than 1% AP absolute).
Details can be found in configs/legacy.
pycocotools compatibility¶
mmpycocotools is the OpenMMlab’s fork of official pycocotools, which works for both MMDetection and Detectron2.
Before PR 4939, since pycocotools and mmpycocotool have the same package name, if users already installed pycocotools (installed Detectron2 first under the same environment), then the setup of MMDetection will skip installing mmpycocotool. Thus MMDetection fails due to the missing mmpycocotools.
If MMDetection is installed before Detectron2, they could work under the same environment.
PR 4939 deprecates mmpycocotools in favor of official pycocotools.
Users may install MMDetection and Detectron2 under the same environment after PR 4939, no matter what the installation order is.
Projects based on MMDetection¶
There are many projects built upon MMDetection. We list some of them as examples of how to extend MMDetection for your own projects. As the page might not be completed, please feel free to create a PR to update this page.
Projects as an extension¶
Some projects extend the boundary of MMDetection for deployment or other research fields. They reveal the potential of what MMDetection can do. We list several of them as below.
OTEDetection: OpenVINO training extensions for object detection.
MMDetection3d: OpenMMLab’s next-generation platform for general 3D object detection.
Projects of papers¶
There are also projects released with papers. Some of the papers are published in top-tier conferences (CVPR, ICCV, and ECCV), the others are also highly influential. To make this list also a reference for the community to develop and compare new object detection algorithms, we list them following the time order of top-tier conferences. Methods already supported and maintained by MMDetection are not listed.
Anchor Pruning for Object Detection, CVIU 2022. [paper][github]
Involution: Inverting the Inherence of Convolution for Visual Recognition, CVPR21. [paper][github]
Multiple Instance Active Learning for Object Detection, CVPR 2021. [paper][github]
Adaptive Class Suppression Loss for Long-Tail Object Detection, CVPR 2021. [paper][github]
Generalizable Pedestrian Detection: The Elephant In The Room, CVPR2021. [paper][github]
Group Fisher Pruning for Practical Network Compression, ICML2021. [paper][github]
Overcoming Classifier Imbalance for Long-tail Object Detection with Balanced Group Softmax, CVPR2020. [paper][github]
Coherent Reconstruction of Multiple Humans from a Single Image, CVPR2020. [paper][github]
Look-into-Object: Self-supervised Structure Modeling for Object Recognition, CVPR 2020. [paper][github]
D2Det: Towards High Quality Object Detection and Instance Segmentation, CVPR2020. [paper][github]
CentripetalNet: Pursuing High-quality Keypoint Pairs for Object Detection, CVPR2020. [paper][github]
Learning a Unified Sample Weighting Network for Object Detection, CVPR 2020. [paper][github]
Scale-equalizing Pyramid Convolution for Object Detection, CVPR2020. [paper] [github]
Revisiting the Sibling Head in Object Detector, CVPR2020. [paper][github]
PolarMask: Single Shot Instance Segmentation with Polar Representation, CVPR2020. [paper][github]
Hit-Detector: Hierarchical Trinity Architecture Search for Object Detection, CVPR2020. [paper][github]
ZeroQ: A Novel Zero Shot Quantization Framework, CVPR2020. [paper][github]
CBNet: A Novel Composite Backbone Network Architecture for Object Detection, AAAI2020. [paper][github]
RDSNet: A New Deep Architecture for Reciprocal Object Detection and Instance Segmentation, AAAI2020. [paper][github]
Training-Time-Friendly Network for Real-Time Object Detection, AAAI2020. [paper][github]
Cascade RPN: Delving into High-Quality Region Proposal Network with Adaptive Convolution, NeurIPS 2019. [paper][github]
Reasoning R-CNN: Unifying Adaptive Global Reasoning into Large-scale Object Detection, CVPR2019. [paper][github]
Learning RoI Transformer for Oriented Object Detection in Aerial Images, CVPR2019. [paper][github]
Dense Peppoints: Representing Visual Objects with Dense Point Sets. [paper][github]
IterDet: Iterative Scheme for Object Detection in Crowded Environments. [paper][github]
A Ranking-based, Balanced Loss Function Unifying Classification and Localisation in Object Detection, NeurIPS2020 [paper][github]
RelationNet++: Bridging Visual Representations for Object Detection via Transformer Decoder, NeurIPS2020 [paper][github]
Generalized Focal Loss V2: Learning Reliable Localization Quality Estimation for Dense Object Detection, CVPR2021[paper][github]
Swin Transformer: Hierarchical Vision Transformer using Shifted Windows, ICCV2021[paper][github]
Focal Transformer: Focal Self-attention for Local-Global Interactions in Vision Transformers, NeurIPS2021[paper][github]
End-to-End Semi-Supervised Object Detection with Soft Teacher, ICCV2021[paper][github]
CBNetV2: A Novel Composite Backbone Network Architecture for Object Detection [paper][github]
Changelog¶
v2.28.0 (28/1/2023)¶
Highlights¶
Support Objects365 Dataset and Separated and Occluded COCO metric
Support acceleration of RetinaNet and SSD on Ascend
Deprecate the support of Python 3.6
New Features and Improvements¶
Support Objects365 Dataset (#7525)
Support Separated and Occluded COCO metric (#9574)
Support acceleration of RetinaNet and SSD on Ascend with documentation (#9648, #9614)
Added missing
-to--format-onlyin documentation.
Deprecations¶
Upgrade the minimum Python version to 3.7, the support of Python 3.6 is no longer guaranteed (#9604)
Bug Fixes¶
Fix validation loss logging by (#9663)
Fix inconsistent float precision between mmdet and mmcv (#9570)
Fix argument name for fp32 in
DeformableDETRHead(#9607)Fix typo of all config file path in Metafile.yml (#9627)
Contributors¶
A total of 11 developers contributed to this release. Thanks @eantono, @akstt, @@lpizzinidev, @RangiLyu, @kbumsik, @tianleiSHI, @nijkah, @BIGWangYuDong, @wangjiangben-hw, @@jamiechoi1995, @ZwwWayne
New Contributors¶
@kbumsik made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9627
@akstt made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9614
@lpizzinidev made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9649
@eantono made their first contribution in https://github.com/open-mmlab/mmdetection/pull/9663
v2.27.0 (5/1/2023)¶
Highlights¶
Support receptive field search of CNN models(TPAMI 2022: RF-Next) (#8191)
Bug Fixes¶
Fix deadlock issue related with MMDetWandbHook (#9476)
Improvements¶
Add minimum GitHub token permissions for workflows (#8928)
Delete compatible code for parrots in roi extractor (#9503)
Deprecate np.bool Type Alias (#9498)
Replace numpy transpose with torch permute to speed-up data pre-processing (#9533)
Documents¶
Fix typo in docs/zh_cn/tutorials/config.md (#9416)
Fix Faster RCNN FP16 config link in README (#9366)
Contributors¶
A total of 12 developers contributed to this release. Thanks @Min-Sheng, @gasvn, @lzyhha, @jbwang1997, @zachcoleman, @chenyuwang814, @MilkClouds, @Fizzez, @boahc077, @apatsekin, @zytx121, @DonggeunYu
v2.26.0 (23/11/2022)¶
Highlights¶
Support training on NPU (#9267)
Bug Fixes¶
Fix RPN visualization (#9151)
Fix readthedocs by freezing the dependency versions (#9154)
Fix device argument error in MMDet_Tutorial.ipynb (#9112)
Fix solov2 cannot dealing with empty gt image (#9185)
Fix random flipping ratio comparison of mixup image (#9336)
Improvements¶
Complement necessary argument of seg_suffix of cityscapes (#9330)
Support copy paste based on bbox when there is no gt mask (#8905)
Make scipy as a default dependency in runtime (#9186)
Documents¶
Delete redundant Chinese characters in docs (#9175)
Add MMEval in README (#9217)
Contributors¶
A total of 11 developers contributed to this release. Thanks @wangjiangben-hw, @motokimura, @AdorableJiang, @BainOuO, @JarvisKevin, @wanghonglie, @zytx121, @BIGWangYuDong, @hhaAndroid, @RangiLyu, @ZwwWayne
v2.25.3 (25/10/2022)¶
Bug Fixes¶
Skip remote sync when wandb is offline (#8755)
Fix jpg to png bug when using seg maps (#9078)
Improvements¶
Fix typo in warning (#8844)
Fix CI for timm, pycocotools, onnx (#9034)
Upgrade pre-commit hooks (#8964)
Documents¶
Update BoundedIoULoss config in readme (#8808)
Fix Faster R-CNN Readme (#8803)
Update location of test_cfg and train_cfg (#8792)
Fix issue template (#8966)
Update random sampler docstring (#9033)
Fix wrong image link (#9054)
Fix FPG readme (#9041)
Contributors¶
A total of 13 developers contributed to this release. Thanks @Zheng-LinXiao, @i-aki-y, @fbagci, @sudoAimer, @Czm369, @DrRyanHuang, @RangiLyu, @wanghonglie, @shinya7y, @Ryoo72, @akshaygulabrao, @gy-7, @Neesky
v2.25.2 (15/9/2022)¶
Bug Fixes¶
Fix DyDCNv2 RuntimeError (#8485)
Fix repeated import of CascadeRPNHead (#8578)
Fix absolute positional embedding of swin backbone (#8127)
Fix get train_pipeline method of val workflow (#8575)
Improvements¶
Upgrade onnxsim to at least 0.4.0 (#8383)
Support tuple format in analyze_results script (#8549)
Fix floordiv warning (#8648)
Documents¶
Fix typo in HTC link (#8487)
Fix docstring of
BboxOverlaps2D(#8512)Added missed Chinese tutorial link (#8564)
Fix mistakes in gaussian radius formula (#8607)
Update config documentation about how to Add WandB Hook (#8663)
Add mmengine link in readme (#8799)
Update issue template (#8802)
Contributors¶
A total of 16 developers contributed to this release. Thanks @daquexian, @lyq10085, @ZwwWayne, @fbagci, @BubblyYi, @fathomson, @ShunchiZhang, @ceasona, @Happylkx, @normster, @chhluo, @Lehsuby, @JiayuXu0, @Nourollah, @hewanru-bit, @RangiLyu
v2.25.1 (29/7/2022)¶
Bug Fixes¶
Fix single GPU distributed training of cuda device specifying (#8176)
Fix PolygonMask bug in FilterAnnotations (#8136)
Fix mdformat version to support python3.6 (#8195)
Fix GPG key error in Dockerfile (#8215)
Fix
WandbLoggerHookerror (#8273)Fix Pytorch 1.10 incompatibility issues (#8439)
Improvements¶
Add
mimtoextras_requirein setup.py (#8194)Support get image shape on macOS (#8434)
Add test commands of
mimin CI (#8230 & #8240)Update
maskformerto be compatible when cfg is a dictionary (#8263)Clean
Pillowversion check in CI (#8229)
Documents¶
Change example hook name in tutorials (#8118)
Update projects (#8120)
Update metafile and release new models (#8294)
Add download link in tutorials (#8391)
Contributors¶
A total of 15 developers contributed to this release. Thanks @ZwwWayne, @ayulockin, @Mxbonn, @p-mishra1, @Youth-Got, @MiXaiLL76, @chhluo, @jbwang1997, @atinfinity, @shinya7y, @duanzhihua, @STLAND-admin, @BIGWangYuDong, @grimoire, @xiaoyuan0203
v2.25.0 (31/5/2022)¶
Highlights¶
Support dedicated
WandbLoggerhookSupport ConvNeXt, DDOD, SOLOv2
Support Mask2Former for instance segmentation
Rename config files of Mask2Former
Backwards incompatible changes¶
Rename config files of Mask2Former (#7571)
before v2.25.0 after v2.25.0 mask2former_xxx_coco.pyrepresents config files for panoptic segmentation.
mask2former_xxx_coco.pyrepresents config files for instance segmentation.mask2former_xxx_coco-panoptic.pyrepresents config files for panoptic segmentation.
New Features¶
Support ConvNeXt (#7281)
Support DDOD (#7279)
Support SOLOv2 (#7441)
Support Mask2Former for instance segmentation (#7571, #8032)
Bug Fixes¶
Enable YOLOX training on different devices (#7912)
Fix the log plot error when evaluation with
interval != 1(#7784)Fix RuntimeError of HTC (#8083)
Improvements¶
Support dedicated
WandbLoggerhook (#7459)Users can set
cfg.log_config.hooks = [ dict(type='MMDetWandbHook', init_kwargs={'project': 'MMDetection-tutorial'}, interval=10, log_checkpoint=True, log_checkpoint_metadata=True, num_eval_images=10)]
in the config to use
MMDetWandbHook. Example can be found in this colab tutorialAdd
AvoidOOMto avoid OOM (#7434, #8091)Try to use
AvoidCUDAOOMto avoid GPU out of memory. It will first retry after callingtorch.cuda.empty_cache(). If it still fails, it will then retry by converting the type of inputs to FP16 format. If it still fails, it will try to copy inputs from GPUs to CPUs to continue computing. Try AvoidOOM in code to make the code continue to run when GPU memory runs out:from mmdet.utils import AvoidCUDAOOM output = AvoidCUDAOOM.retry_if_cuda_oom(some_function)(input1, input2)
Users can also try
AvoidCUDAOOMas a decorator to make the code continue to run when GPU memory runs out:from mmdet.utils import AvoidCUDAOOM @AvoidCUDAOOM.retry_if_cuda_oom def function(*args, **kwargs): ... return xxx
Support reading
gpu_collectfromcfg.evaluation.gpu_collect(#7672)Speedup the Video Inference by Accelerating data-loading Stage (#7832)
Support replacing the
${key}with the value ofcfg.key(#7492)Accelerate result analysis in
analyze_result.py. The evaluation time is speedup by 10 ~ 15 times and only tasks 10 ~ 15 minutes now. (#7891)Support to set
block_dilationsinDilatedEncoder(#7812)Support panoptic segmentation result analysis (#7922)
Release DyHead with Swin-Large backbone (#7733)
Documentations updating and adding
Fix wrong default type of
act_cfginSwinTransformer(#7794)Fix text errors in the tutorials (#7959)
Rewrite the installation guide (#7897)
Useful hooks (#7810)
Fix heading anchor in documentation (#8006)
Replace
markdownlintwithmdformatfor avoiding installing ruby (#8009)
Contributors¶
A total of 20 developers contributed to this release.
Thanks @ZwwWayne, @DarthThomas, @solyaH, @LutingWang, @chenxinfeng4, @Czm369, @Chenastron, @chhluo, @austinmw, @Shanyaliux @hellock, @Y-M-Y, @jbwang1997, @hhaAndroid, @Irvingao, @zhanggefan, @BIGWangYuDong, @Keiku, @PeterVennerstrom, @ayulockin
v2.24.0 (26/4/2022)¶
Highlights¶
Support Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation
Support automatically scaling LR according to GPU number and samples per GPU
Support Class Aware Sampler that improves performance on OpenImages Dataset
New Features¶
Support Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation, see example configs (#7501)
Support Class Aware Sampler, users can set
data=dict(train_dataloader=dict(class_aware_sampler=dict(num_sample_class=1))))
in the config to use
ClassAwareSampler. Examples can be found in the configs of OpenImages Dataset. (#7436)Support automatically scaling LR according to GPU number and samples per GPU. (#7482) In each config, there is a corresponding config of auto-scaling LR as below,
auto_scale_lr = dict(enable=True, base_batch_size=N)
where
Nis the batch size used for the current learning rate in the config (also equals tosamples_per_gpu* gpu number to train this config). By default, we setenable=Falseso that the original usages will not be affected. Users can setenable=Truein each config or add--auto-scale-lrafter the command line to enable this feature and should check the correctness ofbase_batch_sizein customized configs.Support setting dataloader arguments in config and add functions to handle config compatibility. (#7668) The comparison between the old and new usages is as below.
v2.23.0 v2.24.0 data = dict( samples_per_gpu=64, workers_per_gpu=4, train=dict(type='xxx', ...), val=dict(type='xxx', samples_per_gpu=4, ...), test=dict(type='xxx', ...), )
# A recommended config that is clear data = dict( train=dict(type='xxx', ...), val=dict(type='xxx', ...), test=dict(type='xxx', ...), # Use different batch size during inference. train_dataloader=dict(samples_per_gpu=64, workers_per_gpu=4), val_dataloader=dict(samples_per_gpu=8, workers_per_gpu=2), test_dataloader=dict(samples_per_gpu=8, workers_per_gpu=2), ) # Old style still works but allows to set more arguments about data loaders data = dict( samples_per_gpu=64, # only works for train_dataloader workers_per_gpu=4, # only works for train_dataloader train=dict(type='xxx', ...), val=dict(type='xxx', ...), test=dict(type='xxx', ...), # Use different batch size during inference. val_dataloader=dict(samples_per_gpu=8, workers_per_gpu=2), test_dataloader=dict(samples_per_gpu=8, workers_per_gpu=2), )
Support memory profile hook. Users can use it to monitor the memory usages during training as below (#7560)
custom_hooks = [ dict(type='MemoryProfilerHook', interval=50) ]
Support to run on PyTorch with MLU chip (#7578)
Support re-spliting data batch with tag (#7641)
Support the
DiceCostused by K-Net inMaskHungarianAssigner(#7716)Support splitting COCO data for Semi-supervised object detection (#7431)
Support Pathlib for Config.fromfile (#7685)
Support to use file client in OpenImages dataset (#7433)
Add a probability parameter to Mosaic transformation (#7371)
Support specifying interpolation mode in
Resizepipeline (#7585)
Bug Fixes¶
Avoid invalid bbox after deform_sampling (#7567)
Fix the issue that argument color_theme does not take effect when exporting confusion matrix (#7701)
Fix the
end_levelin Necks, which should be the index of the end input backbone level (#7502)Fix the bug that
mix_resultsmay be None inMultiImageMixDataset(#7530)Fix the bug in ResNet plugin when two plugins are used (#7797)
Improvements¶
Enhance
load_json_logsof analyze_logs.py for resumed training logs (#7732)Add argument
out_filein image_demo.py (#7676)Allow mixed precision training with
SimOTAAssigner(#7516)Updated INF to 100000.0 to be the same as that in the official YOLOX (#7778)
Add documentations of:
how to get channels of a new backbone (#7642)
how to unfreeze the backbone network (#7570)
how to train fast_rcnn model (#7549)
proposals in Deformable DETR (#7690)
from-scratch install script in get_started.md (#7575)
Release pre-trained models of
Mask2Former (#7595, #7709)
RetinaNet with ResNet-18 and release models (#7387)
RetinaNet with EfficientNet backbone (#7646)
Contributors¶
A total of 27 developers contributed to this release. Thanks @jovialio, @zhangsanfeng2022, @HarryZJ, @jamiechoi1995, @nestiank, @PeterH0323, @RangeKing, @Y-M-Y, @mattcasey02, @weiji14, @Yulv-git, @xiefeifeihu, @FANG-MING, @meng976537406, @nijkah, @sudz123, @CCODING04, @SheffieldCao, @Czm369, @BIGWangYuDong, @zytx121, @jbwang1997, @chhluo, @jshilong, @RangiLyu, @hhaAndroid, @ZwwWayne
v2.23.0 (28/3/2022)¶
Highlights¶
Support Mask2Former: Masked-attention Mask Transformer for Universal Image Segmentation
Support EfficientNet: EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
Support setting data root through environment variable
MMDET_DATASETS, users don’t have to modify the corresponding path in config files anymore.Find a good recipe for fine-tuning high precision ResNet backbone pre-trained by Torchvision.
New Features¶
Support Mask2Former(#6938)(#7466)(#7471)
Support EfficientNet (#7514)
Support setting data root through environment variable
MMDET_DATASETS, users don’t have to modify the corresponding path in config files anymore. (#7386)Support setting different seeds to different ranks (#7432)
Update the
dist_train.shso that the script can be used to support launching multi-node training on machines without slurm (#7415)Find a good recipe for fine-tuning high precision ResNet backbone pre-trained by Torchvision (#7489)
Bug Fixes¶
Fix bug in VOC unit test which removes the data directory (#7270)
Adjust the order of
get_classesandFileClient(#7276)Force the inputs of
get_bboxesin yolox_head to float32 (#7324)Fix misplaced arguments in LoadPanopticAnnotations (#7388)
Fix reduction=mean in CELoss. (#7449)
Update unit test of CrossEntropyCost (#7537)
Fix memory leaking in panpotic segmentation evaluation (#7538)
Fix the bug of shape broadcast in YOLOv3 (#7551)
Improvements¶
Add Chinese version of onnx2tensorrt.md (#7219)
Update colab tutorials (#7310)
Update information about Localization Distillation (#7350)
Add Chinese version of
finetune.md(#7178)Update YOLOX log for non square input (#7235)
Add
nprocincoco_panoptic.pyfor panoptic quality computing (#7315)Allow to set channel_order in LoadImageFromFile (#7258)
Take point sample related functions out of mask_point_head (#7353)
Add instance evaluation for coco_panoptic (#7313)
Enhance the robustness of analyze_logs.py (#7407)
Supplementary notes of sync_random_seed (#7440)
Update docstring of cross entropy loss (#7472)
Update pascal voc result (#7503)
We create How-to documentation to record any questions about How to xxx. In this version, we added
How to use Mosaic augmentation (#7507)
How to use backbone in mmcls (#7438)
How to produce and submit the prediction results of panoptic segmentation models on COCO test-dev set (#7430))
Contributors¶
A total of 27 developers contributed to this release. Thanks @ZwwWayne, @haofanwang, @shinya7y, @chhluo, @yangrisheng, @triple-Mu, @jbwang1997, @HikariTJU, @imflash217, @274869388, @zytx121, @matrixgame2018, @jamiechoi1995, @BIGWangYuDong, @JingweiZhang12, @Xiangxu-0103, @hhaAndroid, @jshilong, @osbm, @ceroytres, @bunge-bedstraw-herb, @Youth-Got, @daavoo, @jiangyitong, @RangiLyu, @CCODING04, @yarkable
v2.22.0 (24/2/2022)¶
Highlights¶
Support MaskFormer: Per-Pixel Classification is Not All You Need for Semantic Segmentation (#7212)
Support DyHead: Dynamic Head: Unifying Object Detection Heads with Attentions (#6823)
Release a good recipe of using ResNet in object detectors pre-trained by ResNet Strikes Back, which consistently brings about 3~4 mAP improvements over RetinaNet, Faster/Mask/Cascade Mask R-CNN (#7001)
Support Open Images Dataset (#6331)
Support TIMM backbone: PyTorch Image Models (#7020)
New Features¶
Support MaskFormer (#7212)
Support DyHead (#6823)
Support ResNet Strikes Back (#7001)
Support OpenImages Dataset (#6331)
Support TIMM backbone (#7020)
Support visualization for Panoptic Segmentation (#7041)
Breaking Changes¶
In order to support the visualization for Panoptic Segmentation, the num_classes can not be None when using the get_palette function to determine whether to use the panoptic palette.
Bug Fixes¶
Fix bug for the best checkpoints can not be saved when the
key_scoreis None (#7101)Fix MixUp transform filter boxes failing case (#7080)
Add missing properties in SABLHead (#7091)
Fix bug when NaNs exist in confusion matrix (#7147)
Fix PALETTE AttributeError in downstream task (#7230)
Improvements¶
Speed up SimOTA matching (#7098)
Add Chinese translation of
docs_zh-CN/tutorials/init_cfg.md(#7188)
Contributors¶
A total of 20 developers contributed to this release. Thanks @ZwwWayne, @hhaAndroid, @RangiLyu, @AronLin, @BIGWangYuDong, @jbwang1997, @zytx121, @chhluo, @shinya7y, @LuooChen, @dvansa, @siatwangmin, @del-zhenwu, @vikashranjan26, @haofanwang, @jamiechoi1995, @HJoonKwon, @yarkable, @zhijian-liu, @RangeKing
v2.21.0 (8/2/2022)¶
Breaking Changes¶
To standardize the contents in config READMEs and meta files of OpenMMLab projects, the READMEs and meta files in each config directory have been significantly changed. The template will be released in the future, for now, you can refer to the examples of README for algorithm, dataset and backbone. To align with the standard, the configs in dcn are put into to two directories named dcn and dcnv2.
New Features¶
Allow to customize colors of different classes during visualization (#6716)
Support CPU training (#7016)
Add download script of COCO, LVIS, and VOC dataset (#7015)
Bug Fixes¶
Fix weight conversion issue of RetinaNet with Swin-S (#6973)
Update
__repr__ofCompose(#6951)Fix BadZipFile Error when build docker (#6966)
Fix bug in non-distributed multi-gpu training/testing (#7019)
Fix bbox clamp in PyTorch 1.10 (#7074)
Relax the requirement of PALETTE in dataset wrappers (#7085)
Keep the same weights before reassign in the PAA head (#7032)
Update code demo in doc (#7092)
Improvements¶
Speed-up training by allow to set variables of multi-processing (#6974, #7036)
Add links of Chinese tutorials in readme (#6897)
Disable cv2 multiprocessing by default for acceleration (#6867)
Deprecate the support for “python setup.py test” (#6998)
Re-organize metafiles and config readmes (#7051)
Fix None grad problem during training TOOD by adding
SigmoidGeometricMean(#7090)
Contributors¶
A total of 26 developers contributed to this release. Thanks @del-zhenwu, @zimoqingfeng, @srishilesh, @imyhxy, @jenhaoyang, @jliu-ac, @kimnamu, @ShengliLiu, @garvan2021, @ciusji, @DIYer22, @kimnamu, @q3394101, @zhouzaida, @gaotongxiao, @topsy404, @AntoAndGar, @jbwang1997, @nijkah, @ZwwWayne, @Czm369, @jshilong, @RangiLyu, @BIGWangYuDong, @hhaAndroid, @AronLin
v2.20.0 (27/12/2021)¶
New Features¶
Support TOOD: Task-aligned One-stage Object Detection (ICCV 2021 Oral) (#6746)
Support resuming from the latest checkpoint automatically (#6727)
Bug Fixes¶
Fix wrong bbox
loss_weightof the PAA head (#6744)Fix the padding value of
gt_semantic_segin batch collating (#6837)Fix test error of lvis when using
classwise(#6845)Avoid BC-breaking of
get_local_path(#6719)Fix bug in
sync_norm_hookwhen the BN layer does not exist (#6852)Use pycocotools directly no matter what platform it is (#6838)
Improvements¶
Add unit test for SimOTA with no valid bbox (#6770)
Use precommit to check readme (#6802)
Support selecting GPU-ids in non-distributed testing time (#6781)
Contributors¶
A total of 16 developers contributed to this release. Thanks @ZwwWayne, @Czm369, @jshilong, @RangiLyu, @BIGWangYuDong, @hhaAndroid, @jamiechoi1995, @AronLin, @Keiku, @gkagkos, @fcakyon, @www516717402, @vansin, @zactodd, @kimnamu, @jenhaoyang
v2.19.1 (14/12/2021)¶
New Features¶
Release YOLOX COCO pretrained models (#6698)
Bug Fixes¶
Fix DCN initialization in DenseHead (#6625)
Fix initialization of ConvFCHead (#6624)
Fix PseudoSampler in RCNN (#6622)
Fix weight initialization in Swin and PVT (#6663)
Fix dtype bug in BaseDenseHead (#6767)
Fix SimOTA with no valid bbox (#6733)
Improvements¶
Add an example of combining swin and one-stage models (#6621)
Add
get_ann_infoto dataset_wrappers (#6526)Support keeping image ratio in the multi-scale training of YOLOX (#6732)
Support
bbox_clip_borderfor the augmentations of YOLOX (#6730)
Documents¶
Update metafile (#6717)
Add mmhuman3d in readme (#6699)
Update FAQ docs (#6587)
Add doc for
detect_anomalous_params(#6697)
Contributors¶
A total of 11 developers contributed to this release. Thanks @ZwwWayne, @LJoson, @Czm369, @jshilong, @ZCMax, @RangiLyu, @BIGWangYuDong, @hhaAndroid, @zhaoxin111, @GT9505, @shinya7y
v2.19.0 (29/11/2021)¶
Highlights¶
Support Label Assignment Distillation
Support
persistent_workersfor Pytorch >= 1.7Align accuracy to the updated official YOLOX
New Features¶
Support Label Assignment Distillation (#6342)
Support
persistent_workersfor Pytorch >= 1.7 (#6435)
Bug Fixes¶
Fix repeatedly output warning message (#6584)
Avoid infinite GPU waiting in dist training (#6501)
Fix SSD512 config error (#6574)
Fix MMDetection model to ONNX command (#6558)
Improvements¶
Refactor configs of FP16 models (#6592)
Align accuracy to the updated official YOLOX (#6443)
Speed up training and reduce memory cost when using PhotoMetricDistortion. (#6442)
Make OHEM work with seesaw loss (#6514)
Documents¶
Update README.md (#6567)
Contributors¶
A total of 11 developers contributed to this release. Thanks @FloydHsiu, @RangiLyu, @ZwwWayne, @AndreaPi, @st9007a, @hachreak, @BIGWangYuDong, @hhaAndroid, @AronLin, @chhluo, @vealocia, @HarborYuan, @st9007a, @jshilong
v2.18.1 (15/11/2021)¶
Bug Fixes¶
Fix aug test error when the number of prediction bboxes is 0 (#6398)
Fix SpatialReductionAttention in PVT (#6488)
Fix wrong use of
trunc_normal_initin PVT and Swin-Transformer (#6432)
Improvements¶
Save the printed AP information of COCO API to logger (#6505)
Always map location to cpu when load checkpoint (#6405)
Set a random seed when the user does not set a seed (#6457)
Documents¶
Chinese version of Corruption Benchmarking (#6375)
Fix config path in docs (#6396)
Update GRoIE readme (#6401)
Contributors¶
A total of 11 developers contributed to this release. Thanks @st9007a, @hachreak, @HarborYuan, @vealocia, @chhluo, @AndreaPi, @AronLin, @BIGWangYuDong, @hhaAndroid, @RangiLyu, @ZwwWayne
v2.18.0 (27/10/2021)¶
Highlights¶
Support QueryInst (#6050)
Refactor dense heads to decouple onnx export logics from
get_bboxesand speed up inference (#5317, #6003, #6369, #6268, #6315)
Bug Fixes¶
Fix init_weight in fcn_mask_head (#6378)
Fix type error in imshow_bboxes of RPN (#6386)
Fix broken colab link in MMDetection Tutorial (#6382)
Make sure the device and dtype of scale_factor are the same as bboxes (#6374)
Remove sampling hardcode (#6317)
Fix RandomAffine bbox coordinate recorrection (#6293)
Fix init bug of final cls/reg layer in convfc head (#6279)
Fix img_shape broken in auto_augment (#6259)
Fix kwargs parameter missing error in two_stage (#6256)
Improvements¶
Unify the interface of stuff head and panoptic head (#6308)
Polish readme (#6243)
Add code-spell pre-commit hook and fix a typo (#6306)
Fix typo (#6245, #6190)
Fix sampler unit test (#6284)
Fix
forward_dummyof YOLACT to enableget_flops(#6079)Fix link error in the config documentation (#6252)
Adjust the order to beautify the document (#6195)
Refactors¶
Refactor one-stage get_bboxes logic (#5317)
Refactor ONNX export of One-Stage models (#6003, #6369)
Refactor dense_head and speedup (#6268)
Migrate to use prior_generator in training of dense heads (#6315)
Contributors¶
A total of 18 developers contributed to this release. Thanks @Boyden, @onnkeat, @st9007a, @vealocia, @yhcao6, @DapangpangX, @yellowdolphin, @cclauss, @kennymckormick, @pingguokiller, @collinzrj, @AndreaPi, @AronLin, @BIGWangYuDong, @hhaAndroid, @jshilong, @RangiLyu, @ZwwWayne
v2.17.0 (28/9/2021)¶
Highlights¶
New Features¶
Support SOLO (#5832)
Support large scale jittering and New Mask R-CNN baselines (#6132)
Add a general data structure for the results of models (#5508)
Added a base class for one-stage instance segmentation (#5904)
Speed up
YOLOv3inference (#5991)Release Swin Transformer pre-trained models (#6100)
Support mixed precision training in
YOLOX(#5983)Support
valworkflow inYOLACT(#5986)Add script to test
torchserve(#5936)Support
onnxsimwith dynamic input shape (#6117)
Bug Fixes¶
Fix the function naming errors in
model_wrappers(#5975)Fix regression loss bug when the input is an empty tensor (#5976)
Fix scores not contiguous error in
centernet_head(#6016)Fix missing parameters bug in
imshow_bboxes(#6034)Fix bug in
aug_testofHTCwhen the length ofdet_bboxesis 0 (#6088)Fix empty proposal errors in the training of some two-stage models (#5941)
Fix
dynamic_axesparameter error inONNXdynamic shape export (#6104)Fix
dynamic_shapebug ofSyncRandomSizeHook(#6144)Fix the Swin Transformer config link error in the configuration (#6172)
Improvements¶
Add filter rules in
Mosaictransform (#5897)Add size divisor in get flops to avoid some potential bugs (#6076)
Add Chinese translation of
docs_zh-CN/tutorials/customize_dataset.md(#5915)Add Chinese translation of
conventions.md(#5825)Add description of the output of data pipeline (#5886)
Add dataset information in the README file for
PanopticFPN(#5996)Add
extra_reprforDropBlocklayer to get details in the model printing (#6140)Fix CI out of memory and add PyTorch1.9 Python3.9 unit tests (#5862)
Fix download links error of some model (#6069)
Improve the generalization of XML dataset (#5943)
Polish assertion error messages (#6017)
Remove
opencv-python-headlessdependency byalbumentations(#5868)Check dtype in transform unit tests (#5969)
Replace the default theme of documentation with PyTorch Sphinx Theme (#6146)
Update the paper and code fields in the metafile (#6043)
Support to customize padding value of segmentation map (#6152)
Support to resize multiple segmentation maps (#5747)
Contributors¶
A total of 24 developers contributed to this release. Thanks @morkovka1337, @HarborYuan, @guillaumefrd, @guigarfr, @www516717402, @gaotongxiao, @ypwhs, @MartaYang, @shinya7y, @justiceeem, @zhaojinjian0000, @VVsssssk, @aravind-anantha, @wangbo-zhao, @czczup, @whai362, @czczup, @marijnl, @AronLin, @BIGWangYuDong, @hhaAndroid, @jshilong, @RangiLyu, @ZwwWayne
v2.16.0 (30/8/2021)¶
Highlights¶
Support Panoptic FPN and Swin Transformer
New Features¶
Support Panoptic FPN and release models (#5577, #5902)
Support Swin Transformer backbone (#5748)
Release RetinaNet models pre-trained with multi-scale 3x schedule (#5636)
Add script to convert unlabeled image list to coco format (#5643)
Add hook to check whether the loss value is valid (#5674)
Add YOLO anchor optimizing tool (#5644)
Support export onnx models without post process. (#5851)
Support classwise evaluation in CocoPanopticDataset (#5896)
Adapt browse_dataset for concatenated datasets. (#5935)
Add
PatchEmbedandPatchMergingwithAdaptivePadding(#5952)
Bug Fixes¶
Fix unit tests of YOLOX (#5859)
Fix lose randomness in
imshow_det_bboxes(#5845)Make output result of
ImageToTensorcontiguous (#5756)Fix inference bug when calling
regress_by_classin RoIHead in some cases (#5884)Fix bug in CIoU loss where alpha should not have gradient. (#5835)
Fix the bug that
multiscale_outputis defined but not used in HRNet (#5887)Set the priority of EvalHook to LOW. (#5882)
Fix a YOLOX bug when applying bbox rescaling in test mode (#5899)
Fix mosaic coordinate error (#5947)
Fix dtype of bbox in RandomAffine. (#5930)
Improvements¶
Add Chinese version of
data_pipelineand (#5662)Support to remove state dicts of EMA when publishing models. (#5858)
Refactor the loss function in HTC and SCNet (#5881)
Use warnings instead of logger.warning (#5540)
Use legacy coordinate in metric of VOC (#5627)
Add Chinese version of customize_losses (#5826)
Add Chinese version of model_zoo (#5827)
Contributors¶
A total of 19 developers contributed to this release. Thanks @ypwhs, @zywvvd, @collinzrj, @OceanPang, @ddonatien, @@haotian-liu, @viibridges, @Muyun99, @guigarfr, @zhaojinjian0000, @jbwang1997,@wangbo-zhao, @xvjiarui, @RangiLyu, @jshilong, @AronLin, @BIGWangYuDong, @hhaAndroid, @ZwwWayne
v2.15.1 (11/8/2021)¶
Bug Fixes¶
Update correct SSD models. (#5789)
Fix casting error in mask structure (#5820)
Fix MMCV deployment documentation links. (#5790)
Improvements¶
Use dynamic MMCV download link in TorchServe dockerfile (#5779)
Rename the function
upsample_liketointerpolate_asfor more general usage (#5788)
Contributors¶
A total of 14 developers contributed to this release. Thanks @HAOCHENYE, @xiaohu2015, @HsLOL, @zhiqwang, @Adamdad, @shinya7y, @Johnson-Wang, @RangiLyu, @jshilong, @mmeendez8, @AronLin, @BIGWangYuDong, @hhaAndroid, @ZwwWayne
v2.15.0 (02/8/2021)¶
Highlights¶
Support adding MIM dependencies during pip installation
Support MobileNetV2 for SSD-Lite and YOLOv3
Support Chinese Documentation
New Features¶
Add function
upsample_like(#5732)Support to output pdf and epub format documentation (#5738)
Support and release Cascade Mask R-CNN 3x pre-trained models (#5645)
Add
ignore_indexto CrossEntropyLoss (#5646)Support adding MIM dependencies during pip installation (#5676)
Add MobileNetV2 config and models for YOLOv3 (#5510)
Support COCO Panoptic Dataset (#5231)
Support ONNX export of cascade models (#5486)
Support DropBlock with RetinaNet (#5544)
Support MobileNetV2 SSD-Lite (#5526)
Bug Fixes¶
Fix the device of label in multiclass_nms (#5673)
Fix error of backbone initialization from pre-trained checkpoint in config file (#5603, #5550)
Fix download links of RegNet pretrained weights (#5655)
Fix two-stage runtime error given empty proposal (#5559)
Fix flops count error in DETR (#5654)
Fix unittest for
NumClassCheckHookwhen it is not used. (#5626)Fix description bug of using custom dataset (#5546)
Fix bug of
multiclass_nmsthat returns the global indices (#5592)Fix
valid_masklogic error in RPNHead (#5562)Fix unit test error of pretrained configs (#5561)
Fix typo error in anchor_head.py (#5555)
Fix bug when using dataset wrappers (#5552)
Fix a typo error in demo/MMDet_Tutorial.ipynb (#5511)
Fixing crash in
get_root_loggerwhencfg.log_levelis not None (#5521)Fix docker version (#5502)
Fix optimizer parameter error when using
IterBasedRunner(#5490)
Improvements¶
Add unit tests for MMTracking (#5620)
Add Chinese translation of documentation (#5718, #5618, #5558, #5423, #5593, #5421, #5408. #5369, #5419, #5530, #5531)
Update resource limit (#5697)
Update docstring for InstaBoost (#5640)
Support key
reduction_overridein all loss functions (#5515)Use repeatdataset to accelerate CenterNet training (#5509)
Remove unnecessary code in autoassign (#5519)
Add documentation about
init_cfg(#5273)
Contributors¶
A total of 18 developers contributed to this release. Thanks @OceanPang, @AronLin, @hellock, @Outsider565, @RangiLyu, @ElectronicElephant, @likyoo, @BIGWangYuDong, @hhaAndroid, @noobying, @yyz561, @likyoo, @zeakey, @ZwwWayne, @ChenyangLiu, @johnson-magic, @qingswu, @BuxianChen
v2.14.0 (29/6/2021)¶
Highlights¶
Add
simple_testto dense heads to improve the consistency of single-stage and two-stage detectorsRevert the
test_mixinsto single image test to improve efficiency and readabilityAdd Faster R-CNN and Mask R-CNN config using multi-scale training with 3x schedule
New Features¶
Support pretrained models from MoCo v2 and SwAV (#5286)
Add Faster R-CNN and Mask R-CNN config using multi-scale training with 3x schedule (#5179, #5233)
Add
reduction_overridein MSELoss (#5437)Stable support of exporting DETR to ONNX with dynamic shapes and batch inference (#5168)
Stable support of exporting PointRend to ONNX with dynamic shapes and batch inference (#5440)
Bug Fixes¶
Fix size mismatch bug in
multiclass_nms(#4980)Fix the import path of
MultiScaleDeformableAttention(#5338)Fix errors in config of GCNet ResNext101 models (#5360)
Fix Grid-RCNN error when there is no bbox result (#5357)
Fix errors in
onnx_exportof bbox_head when setting reg_class_agnostic (#5468)Fix type error of AutoAssign in the document (#5478)
Fix web links ending with
.md(#5315)
Improvements¶
Add
simple_testto dense heads to improve the consistency of single-stage and two-stage detectors (#5264)Add support for mask diagonal flip in TTA (#5403)
Revert the
test_mixinsto single image test to improve efficiency and readability (#5249)Make YOLOv3 Neck more flexible (#5218)
Refactor SSD to make it more general (#5291)
Refactor
anchor_generatorandpoint_generator(#5349)Allow to configure out the
mask_headof the HTC algorithm (#5389)Delete deprecated warning in FPN (#5311)
Move
model.pretrainedtomodel.backbone.init_cfg(#5370)Make deployment tools more friendly to use (#5280)
Clarify installation documentation (#5316)
Add ImageNet Pretrained Models docs (#5268)
Add FAQ about training loss=nan solution and COCO AP or AR =-1 (# 5312, #5313)
Change all weight links of http to https (#5328)
v2.13.0 (01/6/2021)¶
Highlights¶
Support new methods: CenterNet, Seesaw Loss, MobileNetV2
New Features¶
Support paper Objects as Points (#4602)
Support paper Seesaw Loss for Long-Tailed Instance Segmentation (CVPR 2021) (#5128)
Support MobileNetV2 backbone and inverted residual block (#5122)
Support MIM (#5143)
ONNX exportation with dynamic shapes of CornerNet (#5136)
Add
mask_softconfig option to allow non-binary masks (#4615)Add PWC metafile (#5135)
Bug Fixes¶
Fix YOLOv3 FP16 training error (#5172)
Fix Cacscade R-CNN TTA test error when
det_bboxeslength is 0 (#5221)Fix
iou_thrvariable naming errors in VOC recall calculation function (#5195)Fix Faster R-CNN performance dropped in ONNX Runtime (#5197)
Fix DETR dict changed error when using python 3.8 during iteration (#5226)
Improvements¶
Refactor ONNX export of two stage detector (#5205)
Replace MMDetection’s EvalHook with MMCV’s EvalHook for consistency (#4806)
Update RoI extractor for ONNX (#5194)
Use better parameter initialization in YOLOv3 head for higher performance (#5181)
Release new DCN models of Mask R-CNN by mixed-precision training (#5201)
Update YOLOv3 model weights (#5229)
Add DetectoRS ResNet-101 model weights (#4960)
Discard bboxes with sizes equals to
min_bbox_size(#5011)Remove duplicated code in DETR head (#5129)
Remove unnecessary object in class definition (#5180)
Fix doc link (#5192)
v2.12.0 (01/5/2021)¶
Highlights¶
Support new methods: AutoAssign, YOLOF, and Deformable DETR
Stable support of exporting models to ONNX with batched images and dynamic shape (#5039)
Backwards Incompatible Changes¶
MMDetection is going through big refactoring for more general and convenient usages during the releases from v2.12.0 to v2.15.0 (maybe longer). In v2.12.0 MMDetection inevitably brings some BC-breakings, including the MMCV dependency, model initialization, model registry, and mask AP evaluation.
MMCV version. MMDetection v2.12.0 relies on the newest features in MMCV 1.3.3, including
BaseModulefor unified parameter initialization, model registry, and the CUDA operatorMultiScaleDeformableAttnfor Deformable DETR. Note that MMCV 1.3.2 already contains all the features used by MMDet but has known issues. Therefore, we recommend users skip MMCV v1.3.2 and use v1.3.3, though v1.3.2 might work for most cases.Unified model initialization (#4750). To unify the parameter initialization in OpenMMLab projects, MMCV supports
BaseModulethat acceptsinit_cfgto allow the modules’ parameters initialized in a flexible and unified manner. Now the users need to explicitly callmodel.init_weights()in the training script to initialize the model (as in here, previously this was handled by the detector. The models in MMDetection have been re-benchmarked to ensure accuracy based on PR #4750. The downstream projects should update their code accordingly to use MMDetection v2.12.0.Unified model registry (#5059). To easily use backbones implemented in other OpenMMLab projects, MMDetection migrates to inherit the model registry created in MMCV (#760). In this way, as long as the backbone is supported in an OpenMMLab project and that project also uses the registry in MMCV, users can use that backbone in MMDetection by simply modifying the config without copying the code of that backbone into MMDetection.
Mask AP evaluation (#4898). Previous versions calculate the areas of masks through the bounding boxes when calculating the mask AP of small, medium, and large instances. To indeed use the areas of masks, we pop the key
bboxduring mask AP calculation. This change does not affect the overall mask AP evaluation and aligns the mask AP of similar models in other projects like Detectron2.
New Features¶
Support paper AutoAssign: Differentiable Label Assignment for Dense Object Detection (#4295)
Support paper You Only Look One-level Feature (#4295)
Support paper Deformable DETR: Deformable Transformers for End-to-End Object Detection (#4778)
Support calculating IoU with FP16 tensor in
bbox_overlapsto save memory and keep speed (#4889)Add
__repr__in custom dataset to count the number of instances (#4756)Add windows support by updating requirements.txt (#5052)
Stable support of exporting models to ONNX with batched images and dynamic shape, including SSD, FSAF,FCOS, YOLOv3, RetinaNet, Faster R-CNN, and Mask R-CNN (#5039)
Improvements¶
Use MMCV
MODEL_REGISTRY(#5059)Unified parameter initialization for more flexible usage (#4750)
Rename variable names and fix docstring in anchor head (#4883)
Support training with empty GT in Cascade RPN (#4928)
Add more details of usage of
test_robustnessin documentation (#4917)Changing to use
pycocotoolsinstead ofmmpycocotoolsto fully support Detectron2 and MMDetection in one environment (#4939)Update torch serve dockerfile to support dockers of more versions (#4954)
Add check for training with single class dataset (#4973)
Refactor transformer and DETR Head (#4763)
Update FPG model zoo (#5079)
More accurate mask AP of small/medium/large instances (#4898)
Bug Fixes¶
Fix bug in mean_ap.py when calculating mAP by 11 points (#4875)
Fix error when key
metais not in old checkpoints (#4936)Fix hanging bug when training with empty GT in VFNet, GFL, and FCOS by changing the place of
reduce_mean(#4923, #4978, #5058)Fix asyncronized inference error and provide related demo (#4941)
Fix IoU losses dimensionality unmatch error (#4982)
Fix torch.randperm whtn using PyTorch 1.8 (#5014)
Fix empty bbox error in
mask_headwhen using CARAFE (#5062)Fix
supplement_maskbug when there are zero-size RoIs (#5065)Fix testing with empty rois in RoI Heads (#5081)
v2.11.0 (01/4/2021)¶
Highlights
Support new method: Localization Distillation for Object Detection
Support Pytorch2ONNX with batch inference and dynamic shape
New Features
Support Localization Distillation for Object Detection (#4758)
Support Pytorch2ONNX with batch inference and dynamic shape for Faster-RCNN and mainstream one-stage detectors (#4796)
Improvements
Support batch inference in head of RetinaNet (#4699)
Add batch dimension in second stage of Faster-RCNN (#4785)
Support batch inference in bbox coder (#4721)
Add check for
ann_idsinCOCODatasetto ensure it is unique (#4789)support for showing the FPN results (#4716)
support dynamic shape for grid_anchor (#4684)
Move pycocotools version check to when it is used (#4880)
Bug Fixes
Fix a bug of TridentNet when doing the batch inference (#4717)
Fix a bug of Pytorch2ONNX in FASF (#4735)
Fix a bug when show the image with float type (#4732)
v2.10.0 (01/03/2021)¶
Highlights¶
Support new methods: FPG
Support ONNX2TensorRT for SSD, FSAF, FCOS, YOLOv3, and Faster R-CNN.
New Features¶
Support ONNX2TensorRT for SSD, FSAF, FCOS, YOLOv3, and Faster R-CNN (#4569)
Support Feature Pyramid Grids (FPG) (#4645)
Support video demo (#4420)
Add seed option for sampler (#4665)
Support to customize type of runner (#4570, #4669)
Support synchronizing BN buffer in
EvalHook(#4582)Add script for GIF demo (#4573)
Bug Fixes¶
Fix ConfigDict AttributeError and add Colab link (#4643)
Avoid crash in empty gt training of GFL head (#4631)
Fix
iou_thrsbug in RPN evaluation (#4581)Fix syntax error of config when upgrading model version (#4584)
Improvements¶
Refactor unit test file structures (#4600)
Refactor nms config (#4636)
Get loading pipeline by checking the class directly rather than through config strings (#4619)
Add doctests for mask target generation and mask structures (#4614)
Use deep copy when copying pipeline arguments (#4621)
Update documentations (#4642, #4650, #4620, #4630)
Remove redundant code calling
import_modules_from_strings(#4601)Clean deprecated FP16 API (#4571)
Check whether
CLASSESis correctly initialized in the initialization ofXMLDataset(#4555)Support batch inference in the inference API (#4462, #4526)
Clean deprecated warning and fix ‘meta’ error (#4695)
v2.9.0 (01/02/2021)¶
Highlights¶
Support new methods: SCNet, Sparse R-CNN
Move
train_cfgandtest_cfginto model in configsSupport to visualize results based on prediction quality
New Features¶
Support SCNet (#4356)
Support Sparse R-CNN (#4219)
Support evaluate mAP by multiple IoUs (#4398)
Support concatenate dataset for testing (#4452)
Support to visualize results based on prediction quality (#4441)
Add ONNX simplify option to Pytorch2ONNX script (#4468)
Add hook for checking compatibility of class numbers in heads and datasets (#4508)
Bug Fixes¶
Fix CPU inference bug of Cascade RPN (#4410)
Fix NMS error of CornerNet when there is no prediction box (#4409)
Fix TypeError in CornerNet inference (#4411)
Fix bug of PAA when training with background images (#4391)
Fix the error that the window data is not destroyed when
out_file is not Noneandshow==False(#4442)Fix order of NMS
score_factorthat will decrease the performance of YOLOv3 (#4473)Fix bug in HTC TTA when the number of detection boxes is 0 (#4516)
Fix resize error in mask data structures (#4520)
Improvements¶
Allow to customize classes in LVIS dataset (#4382)
Add tutorials for building new models with existing datasets (#4396)
Add CPU compatibility information in documentation (#4405)
Add documentation of deprecated
ImageToTensorfor batch inference (#4408)Add more details in documentation for customizing dataset (#4430)
Switch
imshow_det_bboxesvisualization backend from OpenCV to Matplotlib (#4389)Deprecate
ImageToTensorinimage_demo.py(#4400)Move train_cfg/test_cfg into model (#4347, #4489)
Update docstring for
reg_decoded_bboxoption in bbox heads (#4467)Update dataset information in documentation (#4525)
Release pre-trained R50 and R101 PAA detectors with multi-scale 3x training schedules (#4495)
Add guidance for speed benchmark (#4537)
v2.8.0 (04/01/2021)¶
Highlights¶
Support new methods: Cascade RPN, TridentNet
New Features¶
Support Cascade RPN (#1900)
Support TridentNet (#3313)
Bug Fixes¶
Fix bug of show result in async_benchmark (#4367)
Fix scale factor in MaskTestMixin (#4366)
Fix but when returning indices in
multiclass_nms(#4362)Fix bug of empirical attention in resnext backbone error (#4300)
Fix bug of
img_norm_cfgin FCOS-HRNet models with updated performance and models (#4250)Fix invalid checkpoint and log in Mask R-CNN models on Cityscapes dataset (#4287)
Fix bug in distributed sampler when dataset is too small (#4257)
Fix bug of ‘PAFPN has no attribute extra_convs_on_inputs’ (#4235)
Improvements¶
Update model url from aws to aliyun (#4349)
Update ATSS for PyTorch 1.6+ (#4359)
Update script to install ruby in pre-commit installation (#4360)
Delete deprecated
mmdet.ops(#4325)Refactor hungarian assigner for more general usage in Sparse R-CNN (#4259)
Handle scipy import in DETR to reduce package dependencies (#4339)
Update documentation of usages for config options after MMCV (1.2.3) supports overriding list in config (#4326)
Update pre-train models of faster rcnn trained on COCO subsets (#4307)
Avoid zero or too small value for beta in Dynamic R-CNN (#4303)
Add doccumentation for Pytorch2ONNX (#4271)
Add deprecated warning FPN arguments (#4264)
Support returning indices of kept bboxes when using nms (#4251)
Update type and device requirements when creating tensors
GFLHead(#4210)Update device requirements when creating tensors in
CrossEntropyLoss(#4224)
v2.7.0 (30/11/2020)¶
Support YOLO, Mask R-CNN, and Cascade R-CNN models exportable to ONNX.
New Features¶
Support DETR (#4201, #4206)
Support to link the best checkpoint in training (#3773)
Support to override config through options in inference.py (#4175)
Support YOLO, Mask R-CNN, and Cascade R-CNN models exportable to ONNX (#4087, #4083)
Support ResNeSt backbone (#2959)
Support unclip border bbox regression (#4076)
Add tpfp func in evaluating AP (#4069)
Support mixed precision training of SSD detector with other backbones (#4081)
Add Faster R-CNN DC5 models (#4043)
Bug Fixes¶
Fix bug of
gpu_idin distributed training mode (#4163)Support Albumentations with version higher than 0.5 (#4032)
Fix num_classes bug in faster rcnn config (#4088)
Update code in docs/2_new_data_model.md (#4041)
Improvements¶
Ensure DCN offset to have similar type as features in VFNet (#4198)
Add config links in README files of models (#4190)
Add tutorials for loss conventions (#3818)
Add solution to installation issues in 30-series GPUs (#4176)
Update docker version in get_started.md (#4145)
Add model statistics and polish some titles in configs README (#4140)
Clamp neg probability in FreeAnchor (#4082)
Speed up expanding large images (#4089)
Fix Pytorch 1.7 incompatibility issues (#4103)
Update trouble shooting page to resolve segmentation fault (#4055)
Update aLRP-Loss in project page (#4078)
Clean duplicated
reduce_meanfunction (#4056)Refactor Q&A (#4045)
v2.6.0 (1/11/2020)¶
Support new method: VarifocalNet.
Refactored documentation with more tutorials.
New Features¶
Support GIoU calculation in
BboxOverlaps2D, and re-implementgiou_lossusingbbox_overlaps(#3936)Support random sampling in CPU mode (#3948)
Support VarifocalNet (#3666, #4024)
Bug Fixes¶
Fix SABL validating bug in Cascade R-CNN (#3913)
Avoid division by zero in PAA head when num_pos=0 (#3938)
Fix temporary directory bug of multi-node testing error (#4034, #4017)
Fix
--show-diroption in test script (#4025)Fix GA-RetinaNet r50 model url (#3983)
Update code in docs and fix broken urls (#3947)
Improvements¶
Refactor pytorch2onnx API into
mmdet.core.exportand usegenerate_inputs_and_wrap_modelfor pytorch2onnx (#3857, #3912)Update RPN upgrade scripts for v2.5.0 compatibility (#3986)
Use mmcv
tensor2imgs(#4010)Update test robustness (#4000)
Update trouble shooting page (#3994)
Accelerate PAA training speed (#3985)
Support batch_size > 1 in validation (#3966)
Use RoIAlign implemented in MMCV for inference in CPU mode (#3930)
Documentation refactoring (#4031)
v2.5.0 (5/10/2020)¶
Highlights¶
Support new methods: YOLACT, CentripetalNet.
Add more documentations for easier and more clear usage.
Backwards Incompatible Changes¶
FP16 related methods are imported from mmcv instead of mmdet. (#3766, #3822)
Mixed precision training utils in mmdet.core.fp16 are moved to mmcv.runner, including force_fp32, auto_fp16, wrap_fp16_model, and Fp16OptimizerHook. A deprecation warning will be raised if users attempt to import those methods from mmdet.core.fp16, and will be finally removed in V2.10.0.
[0, N-1] represents foreground classes and N indicates background classes for all models. (#3221)
Before v2.5.0, the background label for RPN is 0, and N for other heads. Now the behavior is consistent for all models. Thus self.background_labels in dense_heads is removed and all heads use self.num_classes to indicate the class index of background labels.
This change has no effect on the pre-trained models in the v2.x model zoo, but will affect the training of all models with RPN heads. Two-stage detectors whose RPN head uses softmax will be affected because the order of categories is changed.
Only call get_subset_by_classes when test_mode=True and self.filter_empty_gt=True (#3695)
Function get_subset_by_classes in dataset is refactored and only filters out images when test_mode=True and self.filter_empty_gt=True.
In the original implementation, get_subset_by_classes is not related to the flag self.filter_empty_gt and will only be called when the classes is set during initialization no matter test_mode is True or False. This brings ambiguous behavior and potential bugs in many cases. After v2.5.0, if filter_empty_gt=False, no matter whether the classes are specified in a dataset, the dataset will use all the images in the annotations. If filter_empty_gt=True and test_mode=True, no matter whether the classes are specified, the dataset will call ``get_subset_by_classes` to check the images and filter out images containing no GT boxes. Therefore, the users should be responsible for the data filtering/cleaning process for the test dataset.
New Features¶
Test time augmentation for single stage detectors (#3844, #3638)
Support to show the name of experiments during training (#3764)
Add
Shear,Rotate,TranslateAugmentation (#3656, #3619, #3687)Add image-only transformations including
Constrast,Equalize,Color, andBrightness. (#3643)Support YOLACT (#3456)
Support CentripetalNet (#3390)
Support PyTorch 1.6 in docker (#3905)
Bug Fixes¶
Fix the bug of training ATSS when there is no ground truth boxes (#3702)
Fix the bug of using Focal Loss when there is
num_posis 0 (#3702)Fix the label index mapping in dataset browser (#3708)
Fix Mask R-CNN training stuck problem when their is no positive rois (#3713)
Fix the bug of
self.rpn_head.test_cfginRPNTestMixinby usingself.rpn_headin rpn head (#3808)Fix deprecated
Conv2dfrom mmcv.ops (#3791)Fix device bug in RepPoints (#3836)
Fix SABL validating bug (#3849)
Use
https://download.openmmlab.com/mmcv/dist/index.htmlfor installing MMCV (#3840)Fix nonzero in NMS for PyTorch 1.6.0 (#3867)
Fix the API change bug of PAA (#3883)
Fix typo in bbox_flip (#3886)
Fix cv2 import error of ligGL.so.1 in Dockerfile (#3891)
Improvements¶
Change to use
mmcv.utils.collect_envfor collecting environment information to avoid duplicate codes (#3779)Update checkpoint file names to v2.0 models in documentation (#3795)
Update tutorials for changing runtime settings (#3778), modifying loss (#3777)
Improve the function of
simple_test_bboxesin SABL (#3853)Convert mask to bool before using it as img’s index for robustness and speedup (#3870)
Improve documentation of modules and dataset customization (#3821)
v2.4.0 (5/9/2020)¶
Highlights
Fix lots of issues/bugs and reorganize the trouble shooting page
Support new methods SABL, YOLOv3, and PAA Assign
Support Batch Inference
Start to publish
mmdetpackage to PyPI since v2.3.0Switch model zoo to download.openmmlab.com
Backwards Incompatible Changes
Support Batch Inference (#3564, #3686, #3705): Since v2.4.0, MMDetection could inference model with multiple images in a single GPU. This change influences all the test APIs in MMDetection and downstream codebases. To help the users migrate their code, we use
replace_ImageToTensor(#3686) to convert legacy test data pipelines during dataset initialization.Support RandomFlip with horizontal/vertical/diagonal direction (#3608): Since v2.4.0, MMDetection supports horizontal/vertical/diagonal flip in the data augmentation. This influences bounding box, mask, and image transformations in data augmentation process and the process that will map those data back to the original format.
Migrate to use
mmlvisandmmpycocotoolsfor COCO and LVIS dataset (#3727). The APIs are fully compatible with the originallvisandpycocotools. Users need to uninstall the existing pycocotools and lvis packages in their environment first and installmmlvis&mmpycocotools.
Bug Fixes
Fix default mean/std for onnx (#3491)
Fix coco evaluation and add metric items (#3497)
Fix typo for install.md (#3516)
Fix atss when sampler per gpu is 1 (#3528)
Fix import of fuse_conv_bn (#3529)
Fix bug of gaussian_target, update unittest of heatmap (#3543)
Fixed VOC2012 evaluate (#3553)
Fix scale factor bug of rescale (#3566)
Fix with_xxx_attributes in base detector (#3567)
Fix boxes scaling when number is 0 (#3575)
Fix rfp check when neck config is a list (#3591)
Fix import of fuse conv bn in benchmark.py (#3606)
Fix webcam demo (#3634)
Fix typo and itemize issues in tutorial (#3658)
Fix error in distributed training when some levels of FPN are not assigned with bounding boxes (#3670)
Fix the width and height orders of stride in valid flag generation (#3685)
Fix weight initialization bug in Res2Net DCN (#3714)
Fix bug in OHEMSampler (#3677)
New Features
Support Cutout augmentation (#3521)
Support evaluation on multiple datasets through ConcatDataset (#3522)
Support PAA assign #(3547)
Support eval metric with pickle results (#3607)
Support YOLOv3 (#3083)
Support SABL (#3603)
Support to publish to Pypi in github-action (#3510)
Support custom imports (#3641)
Improvements
Refactor common issues in documentation (#3530)
Add pytorch 1.6 to CI config (#3532)
Add config to runner meta (#3534)
Add eval-option flag for testing (#3537)
Add init_eval to evaluation hook (#3550)
Add include_bkg in ClassBalancedDataset (#3577)
Using config’s loading in inference_detector (#3611)
Add ATSS ResNet-101 models in model zoo (#3639)
Update urls to download.openmmlab.com (#3665)
Support non-mask training for CocoDataset (#3711)
v2.3.0 (5/8/2020)¶
Highlights
The CUDA/C++ operators have been moved to
mmcv.ops. For backward compatibilitymmdet.opsis kept as warppers ofmmcv.ops.Support new methods CornerNet, DIOU/CIOU loss, and new dataset: LVIS V1
Provide more detailed colab training tutorials and more complete documentation.
Support to convert RetinaNet from Pytorch to ONNX.
Bug Fixes
Fix the model initialization bug of DetectoRS (#3187)
Fix the bug of module names in NASFCOSHead (#3205)
Fix the filename bug in publish_model.py (#3237)
Fix the dimensionality bug when
inside_flags.any()isFalsein dense heads (#3242)Fix the bug of forgetting to pass flip directions in
MultiScaleFlipAug(#3262)Fixed the bug caused by default value of
stem_channels(#3333)Fix the bug of model checkpoint loading for CPU inference (#3318, #3316)
Fix topk bug when box number is smaller than the expected topk number in ATSSAssigner (#3361)
Fix the gt priority bug in center_region_assigner.py (#3208)
Fix NaN issue of iou calculation in iou_loss.py (#3394)
Fix the bug that
iou_thrsis not actually used during evaluation in coco.py (#3407)Fix test-time augmentation of RepPoints (#3435)
Fix runtimeError caused by incontiguous tensor in Res2Net+DCN (#3412)
New Features
Support CornerNet (#3036)
Support LVIS V1 dataset (#)
Support customized hooks in training (#3395)
Support fp16 training of generalized focal loss (#3410)
Support to convert RetinaNet from Pytorch to ONNX (#3075)
Improvements
Support to process ignore boxes in ATSS assigner (#3082)
Allow to crop images without ground truth in
RandomCrop(#3153)Enable the the
Accuracymodule to set threshold (#3155)Refactoring unit tests (#3206)
Unify the training settings of
to_float32andnorm_cfgin RegNets configs (#3210)Add colab training tutorials for beginners (#3213, #3273)
Move CUDA/C++ operators into
mmcv.opsand keepmmdet.opsas warppers for backward compatibility (#3232)(#3457)Update installation scripts in documentation (#3290) and dockerfile (#3320)
Support to set image resize backend (#3392)
Remove git hash in version file (#3466)
Check mmcv version to force version compatibility (#3460)
v2.2.0 (1/7/2020)¶
Highlights
Support new methods: DetectoRS, PointRend, Generalized Focal Loss, Dynamic R-CNN
Bug Fixes
Fix FreeAnchor when no gt in image (#3176)
Clean up deprecated usage of
register_module()(#3092, #3161)Fix pretrain bug in NAS FCOS (#3145)
Fix
num_classesin SSD (#3142)Fix FCOS warmup (#3119)
Fix
rstripintools/publish_model.pyFix
flip_ratiodefault value in RandomFLip pipeline (#3106)Fix cityscapes eval with ms_rcnn (#3112)
Fix RPN softmax (#3056)
Fix filename of LVIS@v0.5 (#2998)
Fix nan loss by filtering out-of-frame gt_bboxes in COCO (#2999)
Fix bug in FSAF (#3018)
Add FocalLoss
num_classescheck (#2964)Fix PISA Loss when there are no gts (#2992)
Avoid nan in
iou_calculator(#2975)Prevent possible bugs in loading and transforms caused by shallow copy (#2967)
New Features
Add DetectoRS (#3064)
Support Generalize Focal Loss (#3097)
Support PointRend (#2752)
Support Dynamic R-CNN (#3040)
Add DeepFashion dataset (#2968)
Implement FCOS training tricks (#2935)
Use BaseDenseHead as base class for anchor-base heads (#2963)
Add
with_cpfor BasicBlock (#2891)Add
stem_channelsargument for ResNet (#2954)
Improvements
Add anchor free base head (#2867)
Migrate to github action (#3137)
Add docstring for datasets, pipelines, core modules and methods (#3130, #3125, #3120)
Add VOC benchmark (#3060)
Add
concatmode in GRoI (#3098)Remove cmd arg
autorescale-lr(#3080)Use
len(data['img_metas'])to indicatenum_samples(#3073, #3053)Switch to EpochBasedRunner (#2976)
v2.1.0 (8/6/2020)¶
Highlights
Support new dataset: LVIS
Bug Fixes
Change the CLI argument
--validateto--no-validateto enable validation after training epochs by default. (#2651)Add missing cython to docker file (#2713)
Fix bug in nms cpu implementation (#2754)
Fix bug when showing mask results (#2763)
Fix gcc requirement (#2806)
Fix bug in async test (#2820)
Fix mask encoding-decoding bugs in test API (#2824)
Fix bug in test time augmentation (#2858, #2921, #2944)
Fix a typo in comment of apis/train (#2877)
Fix the bug of returning None when no gt bboxes are in the original image in
RandomCrop. Fix the bug that misses to handlegt_bboxes_ignore,gt_label_ignore, andgt_masks_ignoreinRandomCrop,MinIoURandomCropandExpandmodules. (#2810)Fix bug of
base_channelsof regnet (#2917)Fix the bug of logger when loading pre-trained weights in base detector (#2936)
New Features
Add IoU models (#2666)
Add colab demo for inference
Support class agnostic nms (#2553)
Add benchmark gathering scripts for development only (#2676)
Add mmdet-based project links (#2736, #2767, #2895)
Add config dump in training (#2779)
Add ClassBalancedDataset (#2721)
Add res2net backbone (#2237)
Support RegNetX models (#2710)
Use
mmcv.FileClientto support different storage backends (#2712)Add ClassBalancedDataset (#2721)
Code Release: Prime Sample Attention in Object Detection (CVPR 2020) (#2626)
Implement NASFCOS (#2682)
Add class weight in CrossEntropyLoss (#2797)
Support LVIS dataset (#2088)
Support GRoIE (#2584)
Improvements
Allow different x and y strides in anchor heads. (#2629)
Make FSAF loss more robust to no gt (#2680)
Compute pure inference time instead (#2657) and update inference speed (#2730)
Avoided the possibility that a patch with 0 area is cropped. (#2704)
Add warnings when deprecated
imgs_per_gpuis used. (#2700)Add a mask rcnn example for config (#2645)
Update model zoo (#2762, #2866, #2876, #2879, #2831)
Add
ori_filenameto img_metas and use it in test show-dir (#2612)Use
img_fieldsto handle multiple images during image transform (#2800)Add upsample_cfg support in FPN (#2787)
Add
['img']as defaultimg_fieldsfor back compatibility (#2809)Rename the pretrained model from
open-mmlab://resnet50_caffeandopen-mmlab://resnet50_caffe_bgrtoopen-mmlab://detectron/resnet50_caffeandopen-mmlab://detectron2/resnet50_caffe. (#2832)Added sleep(2) in test.py to reduce hanging problem (#2847)
Support
c10::halfin CARAFE (#2890)Improve documentations (#2918, #2714)
Use optimizer constructor in mmcv and clean the original implementation in
mmdet.core.optimizer(#2947)
v2.0.0 (6/5/2020)¶
In this release, we made lots of major refactoring and modifications.
Faster speed. We optimize the training and inference speed for common models, achieving up to 30% speedup for training and 25% for inference. Please refer to model zoo for details.
Higher performance. We change some default hyperparameters with no additional cost, which leads to a gain of performance for most models. Please refer to compatibility for details.
More documentation and tutorials. We add a bunch of documentation and tutorials to help users get started more smoothly. Read it here.
Support PyTorch 1.5. The support for 1.1 and 1.2 is dropped, and we switch to some new APIs.
Better configuration system. Inheritance is supported to reduce the redundancy of configs.
Better modular design. Towards the goal of simplicity and flexibility, we simplify some encapsulation while add more other configurable modules like BBoxCoder, IoUCalculator, OptimizerConstructor, RoIHead. Target computation is also included in heads and the call hierarchy is simpler.
Breaking Changes Models training with MMDetection 1.x are not fully compatible with 2.0, please refer to the compatibility doc for the details and how to migrate to the new version.
Improvements
Unify cuda and cpp API for custom ops. (#2277)
New config files with inheritance. (#2216)
Encapsulate the second stage into RoI heads. (#1999)
Refactor GCNet/EmpericalAttention into plugins. (#2345)
Set low quality match as an option in IoU-based bbox assigners. (#2375)
Change the codebase’s coordinate system. (#2380)
Refactor the category order in heads. 0 means the first positive class instead of background now. (#2374)
Add bbox sampler and assigner registry. (#2419)
Speed up the inference of RPN. (#2420)
Add
train_cfgandtest_cfgas class members in all anchor heads. (#2422)Merge target computation methods into heads. (#2429)
Add bbox coder to support different bbox encoding and losses. (#2480)
Unify the API for regression loss. (#2156)
Refactor Anchor Generator. (#2474)
Make
lran optional argument for optimizers. (#2509)Migrate to modules and methods in MMCV. (#2502, #2511, #2569, #2572)
Support PyTorch 1.5. (#2524)
Drop the support for Python 3.5 and use F-string in the codebase. (#2531)
Bug Fixes
Fix the scale factors for resized images without keep the aspect ratio. (#2039)
Check if max_num > 0 before slicing in NMS. (#2486)
Fix Deformable RoIPool when there is no instance. (#2490)
Fix the default value of assigned labels. (#2536)
Fix the evaluation of Cityscapes. (#2578)
New Features
Add deep_stem and avg_down option to ResNet, i.e., support ResNetV1d. (#2252)
Add L1 loss. (#2376)
Support both polygon and bitmap for instance masks. (#2353, #2540)
Support CPU mode for inference. (#2385)
Add optimizer constructor for complicated configuration of optimizers. (#2397, #2488)
Implement PAFPN. (#2392)
Support empty tensor input for some modules. (#2280)
Support for custom dataset classes without overriding it. (#2408, #2443)
Support to train subsets of coco dataset. (#2340)
Add iou_calculator to potentially support more IoU calculation methods. (2405)
Support class wise mean AP (was removed in the last version). (#2459)
Add option to save the testing result images. (#2414)
Support MomentumUpdaterHook. (#2571)
Add a demo to inference a single image. (#2605)
v1.1.0 (24/2/2020)¶
Highlights
Dataset evaluation is rewritten with a unified api, which is used by both evaluation hooks and test scripts.
Support new methods: CARAFE.
Breaking Changes
The new MMDDP inherits from the official DDP, thus the
__init__api is changed to be the same as official DDP.The
mask_headfield in HTC config files is modified.The evaluation and testing script is updated.
In all transforms, instance masks are stored as a numpy array shaped (n, h, w) instead of a list of (h, w) arrays, where n is the number of instances.
Bug Fixes
Fix IOU assigners when ignore_iof_thr > 0 and there is no pred boxes. (#2135)
Fix mAP evaluation when there are no ignored boxes. (#2116)
Fix the empty RoI input for Deformable RoI Pooling. (#2099)
Fix the dataset settings for multiple workflows. (#2103)
Fix the warning related to
torch.uint8in PyTorch 1.4. (#2105)Fix the inference demo on devices other than gpu:0. (#2098)
Fix Dockerfile. (#2097)
Fix the bug that
pad_valis unused in Pad transform. (#2093)Fix the albumentation transform when there is no ground truth bbox. (#2032)
Improvements
Use torch instead of numpy for random sampling. (#2094)
Migrate to the new MMDDP implementation in MMCV v0.3. (#2090)
Add meta information in logs. (#2086)
Rewrite Soft NMS with pytorch extension and remove cython as a dependency. (#2056)
Rewrite dataset evaluation. (#2042, #2087, #2114, #2128)
Use numpy array for masks in transforms. (#2030)
New Features
Implement “CARAFE: Content-Aware ReAssembly of FEatures”. (#1583)
Add
worker_init_fn()in data_loader when seed is set. (#2066, #2111)Add logging utils. (#2035)
v1.0.0 (30/1/2020)¶
This release mainly improves the code quality and add more docstrings.
Highlights
Documentation is online now: https://mmdetection.readthedocs.io.
Support new models: ATSS.
DCN is now available with the api
build_conv_layerandConvModulelike the normal conv layer.A tool to collect environment information is available for trouble shooting.
Bug Fixes
Fix the incompatibility of the latest numpy and pycocotools. (#2024)
Fix the case when distributed package is unavailable, e.g., on Windows. (#1985)
Fix the dimension issue for
refine_bboxes(). (#1962)Fix the typo when
seg_prefixis a list. (#1906)Add segmentation map cropping to RandomCrop. (#1880)
Fix the return value of
ga_shape_target_single(). (#1853)Fix the loaded shape of empty proposals. (#1819)
Fix the mask data type when using albumentation. (#1818)
Improvements
Enhance AssignResult and SamplingResult. (#1995)
Add ability to overwrite existing module in Registry. (#1982)
Reorganize requirements and make albumentations and imagecorruptions optional. (#1969)
Check NaN in
SSDHead. (#1935)Encapsulate the DCN in ResNe(X)t into a ConvModule & Conv_layers. (#1894)
Refactoring for mAP evaluation and support multiprocessing and logging. (#1889)
Init the root logger before constructing Runner to log more information. (#1865)
Split
SegResizeFlipPadRescaleinto different existing transforms. (#1852)Move
init_dist()to MMCV. (#1851)Documentation and docstring improvements. (#1971, #1938, #1869, #1838)
Fix the color of the same class for mask visualization. (#1834)
Remove the option
keep_all_stagesin HTC and Cascade R-CNN. (#1806)
New Features
Add two test-time options
crop_maskandrle_mask_encodefor mask heads. (#2013)Support loading grayscale images as single channel. (#1975)
Implement “Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection”. (#1872)
Add sphinx generated docs. (#1859, #1864)
Add GN support for flops computation. (#1850)
Collect env info for trouble shooting. (#1812)
v1.0rc1 (13/12/2019)¶
The RC1 release mainly focuses on improving the user experience, and fixing bugs.
Highlights
Support new models: FoveaBox, RepPoints and FreeAnchor.
Add a Dockerfile.
Add a jupyter notebook demo and a webcam demo.
Setup the code style and CI.
Add lots of docstrings and unit tests.
Fix lots of bugs.
Breaking Changes
There was a bug for computing COCO-style mAP w.r.t different scales (AP_s, AP_m, AP_l), introduced by #621. (#1679)
Bug Fixes
Fix a sampling interval bug in Libra R-CNN. (#1800)
Fix the learning rate in SSD300 WIDER FACE. (#1781)
Fix the scaling issue when
keep_ratio=False. (#1730)Fix typos. (#1721, #1492, #1242, #1108, #1107)
Fix the shuffle argument in
build_dataloader. (#1693)Clip the proposal when computing mask targets. (#1688)
Fix the “index out of range” bug for samplers in some corner cases. (#1610, #1404)
Fix the NMS issue on devices other than GPU:0. (#1603)
Fix SSD Head and GHM Loss on CPU. (#1578)
Fix the OOM error when there are too many gt bboxes. (#1575)
Fix the wrong keyword argument
nms_cfgin HTC. (#1573)Process masks and semantic segmentation in Expand and MinIoUCrop transforms. (#1550, #1361)
Fix a scale bug in the Non Local op. (#1528)
Fix a bug in transforms when
gt_bboxes_ignoreis None. (#1498)Fix a bug when
img_prefixis None. (#1497)Pass the device argument to
grid_anchorsandvalid_flags. (#1478)Fix the data pipeline for test_robustness. (#1476)
Fix the argument type of deformable pooling. (#1390)
Fix the coco_eval when there are only two classes. (#1376)
Fix a bug in Modulated DeformableConv when deformable_group>1. (#1359)
Fix the mask cropping in RandomCrop. (#1333)
Fix zero outputs in DeformConv when not running on cuda:0. (#1326)
Fix the type issue in Expand. (#1288)
Fix the inference API. (#1255)
Fix the inplace operation in Expand. (#1249)
Fix the from-scratch training config. (#1196)
Fix inplace add in RoIExtractor which cause an error in PyTorch 1.2. (#1160)
Fix FCOS when input images has no positive sample. (#1136)
Fix recursive imports. (#1099)
Improvements
Print the config file and mmdet version in the log. (#1721)
Lint the code before compiling in travis CI. (#1715)
Add a probability argument for the
Expandtransform. (#1651)Update the PyTorch and CUDA version in the docker file. (#1615)
Raise a warning when specifying
--validatein non-distributed training. (#1624, #1651)Beautify the mAP printing. (#1614)
Add pre-commit hook. (#1536)
Add the argument
in_channelsto backbones. (#1475)Add lots of docstrings and unit tests, thanks to @Erotemic. (#1603, #1517, #1506, #1505, #1491, #1479, #1477, #1475, #1474)
Add support for multi-node distributed test when there is no shared storage. (#1399)
Optimize Dockerfile to reduce the image size. (#1306)
Update new results of HRNet. (#1284, #1182)
Add an argument
no_norm_on_lateralin FPN. (#1240)Test the compiling in CI. (#1235)
Move docs to a separate folder. (#1233)
Add a jupyter notebook demo. (#1158)
Support different type of dataset for training. (#1133)
Use int64_t instead of long in cuda kernels. (#1131)
Support unsquare RoIs for bbox and mask heads. (#1128)
Manually add type promotion to make compatible to PyTorch 1.2. (#1114)
Allowing validation dataset for computing validation loss. (#1093)
Use
.scalar_type()instead of.type()to suppress some warnings. (#1070)
New Features
Add an option
--with_apto compute the AP for each class. (#1549)Implement “FreeAnchor: Learning to Match Anchors for Visual Object Detection”. (#1391)
Support Albumentations for augmentations in the data pipeline. (#1354)
Implement “FoveaBox: Beyond Anchor-based Object Detector”. (#1339)
Support horizontal and vertical flipping. (#1273, #1115)
Implement “RepPoints: Point Set Representation for Object Detection”. (#1265)
Add test-time augmentation to HTC and Cascade R-CNN. (#1251)
Add a COCO result analysis tool. (#1228)
Add Dockerfile. (#1168)
Add a webcam demo. (#1155, #1150)
Add FLOPs counter. (#1127)
Allow arbitrary layer order for ConvModule. (#1078)
v1.0rc0 (27/07/2019)¶
Implement lots of new methods and components (Mixed Precision Training, HTC, Libra R-CNN, Guided Anchoring, Empirical Attention, Mask Scoring R-CNN, Grid R-CNN (Plus), GHM, GCNet, FCOS, HRNet, Weight Standardization, etc.). Thank all collaborators!
Support two additional datasets: WIDER FACE and Cityscapes.
Refactoring for loss APIs and make it more flexible to adopt different losses and related hyper-parameters.
Speed up multi-gpu testing.
Integrate all compiling and installing in a single script.
v0.6.0 (14/04/2019)¶
Up to 30% speedup compared to the model zoo.
Support both PyTorch stable and nightly version.
Replace NMS and SigmoidFocalLoss with Pytorch CUDA extensions.
v0.6rc0(06/02/2019)¶
Migrate to PyTorch 1.0.
v0.5.7 (06/02/2019)¶
Add support for Deformable ConvNet v2. (Many thanks to the authors and @chengdazhi)
This is the last release based on PyTorch 0.4.1.
v0.5.6 (17/01/2019)¶
Add support for Group Normalization.
Unify RPNHead and single stage heads (RetinaHead, SSDHead) with AnchorHead.
v0.5.5 (22/12/2018)¶
Add SSD for COCO and PASCAL VOC.
Add ResNeXt backbones and detection models.
Refactoring for Samplers/Assigners and add OHEM.
Add VOC dataset and evaluation scripts.
v0.5.4 (27/11/2018)¶
Add SingleStageDetector and RetinaNet.
v0.5.3 (26/11/2018)¶
Add Cascade R-CNN and Cascade Mask R-CNN.
Add support for Soft-NMS in config files.
v0.5.2 (21/10/2018)¶
Add support for custom datasets.
Add a script to convert PASCAL VOC annotations to the expected format.
v0.5.1 (20/10/2018)¶
Add BBoxAssigner and BBoxSampler, the
train_cfgfield in config files are restructured.ConvFCRoIHead/SharedFCRoIHeadare renamed toConvFCBBoxHead/SharedFCBBoxHeadfor consistency.
Frequently Asked Questions¶
We list some common troubles faced by many users and their corresponding solutions here. Feel free to enrich the list if you find any frequent issues and have ways to help others to solve them. If the contents here do not cover your issue, please create an issue using the provided templates and make sure you fill in all required information in the template.
Installation¶
Compatibility issue between MMCV and MMDetection; “ConvWS is already registered in conv layer”; “AssertionError: MMCV==xxx is used but incompatible. Please install mmcv>=xxx, <=xxx.”
Compatible MMDetection and MMCV versions are shown as below. Please choose the correct version of MMCV to avoid installation issues.
| MMDetection version | MMCV version |
|---|---|
| master | mmcv-full>=1.3.17, \<1.8.0 |
| 2.28.0 | mmcv-full>=1.3.17, \<1.8.0 |
| 2.27.0 | mmcv-full>=1.3.17, \<1.8.0 |
| 2.26.0 | mmcv-full>=1.3.17, \<1.8.0 |
| 2.25.3 | mmcv-full>=1.3.17, \<1.7.0 |
| 2.25.2 | mmcv-full>=1.3.17, \<1.7.0 |
| 2.25.1 | mmcv-full>=1.3.17, \<1.6.0 |
| 2.25.0 | mmcv-full>=1.3.17, \<1.6.0 |
| 2.24.1 | mmcv-full>=1.3.17, \<1.6.0 |
| 2.24.0 | mmcv-full>=1.3.17, \<1.6.0 |
| 2.23.0 | mmcv-full>=1.3.17, \<1.5.0 |
| 2.22.0 | mmcv-full>=1.3.17, \<1.5.0 |
| 2.21.0 | mmcv-full>=1.3.17, \<1.5.0 |
| 2.20.0 | mmcv-full>=1.3.17, \<1.5.0 |
| 2.19.1 | mmcv-full>=1.3.17, \<1.5.0 |
| 2.19.0 | mmcv-full>=1.3.17, \<1.5.0 |
| 2.18.0 | mmcv-full>=1.3.17, \<1.4.0 |
| 2.17.0 | mmcv-full>=1.3.14, \<1.4.0 |
| 2.16.0 | mmcv-full>=1.3.8, \<1.4.0 |
| 2.15.1 | mmcv-full>=1.3.8, \<1.4.0 |
| 2.15.0 | mmcv-full>=1.3.8, \<1.4.0 |
| 2.14.0 | mmcv-full>=1.3.8, \<1.4.0 |
| 2.13.0 | mmcv-full>=1.3.3, \<1.4.0 |
| 2.12.0 | mmcv-full>=1.3.3, \<1.4.0 |
| 2.11.0 | mmcv-full>=1.2.4, \<1.4.0 |
| 2.10.0 | mmcv-full>=1.2.4, \<1.4.0 |
| 2.9.0 | mmcv-full>=1.2.4, \<1.4.0 |
| 2.8.0 | mmcv-full>=1.2.4, \<1.4.0 |
| 2.7.0 | mmcv-full>=1.1.5, \<1.4.0 |
| 2.6.0 | mmcv-full>=1.1.5, \<1.4.0 |
| 2.5.0 | mmcv-full>=1.1.5, \<1.4.0 |
| 2.4.0 | mmcv-full>=1.1.1, \<1.4.0 |
| 2.3.0 | mmcv-full==1.0.5 |
| 2.3.0rc0 | mmcv-full>=1.0.2 |
| 2.2.1 | mmcv==0.6.2 |
| 2.2.0 | mmcv==0.6.2 |
| 2.1.0 | mmcv>=0.5.9, \<=0.6.1 |
| 2.0.0 | mmcv>=0.5.1, \<=0.5.8 |
“No module named ‘mmcv.ops’”; “No module named ‘mmcv._ext’”.
Uninstall existing mmcv in the environment using
pip uninstall mmcv.Install mmcv-full following the installation instruction.
Using albumentations
If you would like to use
albumentations, we suggest usingpip install -r requirements/albu.txtorpip install -U albumentations --no-binary qudida,albumentations. If you simply usepip install albumentations>=0.3.2, it will installopencv-python-headlesssimultaneously (even though you have already installedopencv-python). Please refer to the official documentation for details.ModuleNotFoundError is raised when using some algorithms
Some extra dependencies are required for Instaboost, Panoptic Segmentation, LVIS dataset, etc. Please note the error message and install corresponding packages, e.g.,
# for instaboost pip install instaboostfast # for panoptic segmentation pip install git+https://github.com/cocodataset/panopticapi.git # for LVIS dataset pip install git+https://github.com/lvis-dataset/lvis-api.git
Coding¶
Do I need to reinstall mmdet after some code modifications
If you follow the best practice and install mmdet with
pip install -e ., any local modifications made to the code will take effect without reinstallation.How to develop with multiple MMDetection versions
You can have multiple folders like mmdet-2.21, mmdet-2.22. When you run the train or test script, it will adopt the mmdet package in the current folder.
To use the default MMDetection installed in the environment rather than the one you are working with, you can remove the following line in those scripts:
PYTHONPATH="$(dirname $0)/..":$PYTHONPATH
PyTorch/CUDA Environment¶
“RTX 30 series card fails when building MMCV or MMDet”
Temporary work-around: do
MMCV_WITH_OPS=1 MMCV_CUDA_ARGS='-gencode=arch=compute_80,code=sm_80' pip install -e .. The common issue isnvcc fatal : Unsupported gpu architecture 'compute_86'. This means that the compiler should optimize for sm_86, i.e., nvidia 30 series card, but such optimizations have not been supported by CUDA toolkit 11.0. This work-around modifies the compile flag by addingMMCV_CUDA_ARGS='-gencode=arch=compute_80,code=sm_80', which tellsnvccto optimize for sm_80, i.e., Nvidia A100. Although A100 is different from the 30 series card, they use similar ampere architecture. This may hurt the performance but it works.PyTorch developers have updated that the default compiler flags should be fixed by pytorch/pytorch#47585. So using PyTorch-nightly may also be able to solve the problem, though we have not tested it yet.
“invalid device function” or “no kernel image is available for execution”.
Check if your cuda runtime version (under
/usr/local/),nvcc --versionandconda list cudatoolkitversion match.Run
python mmdet/utils/collect_env.pyto check whether PyTorch, torchvision, and MMCV are built for the correct GPU architecture. You may need to setTORCH_CUDA_ARCH_LISTto reinstall MMCV. The GPU arch table could be found here, i.e. runTORCH_CUDA_ARCH_LIST=7.0 pip install mmcv-fullto build MMCV for Volta GPUs. The compatibility issue could happen when using old GPUS, e.g., Tesla K80 (3.7) on colab.Check whether the running environment is the same as that when mmcv/mmdet has compiled. For example, you may compile mmcv using CUDA 10.0 but run it on CUDA 9.0 environments.
“undefined symbol” or “cannot open xxx.so”.
If those symbols are CUDA/C++ symbols (e.g., libcudart.so or GLIBCXX), check whether the CUDA/GCC runtimes are the same as those used for compiling mmcv, i.e. run
python mmdet/utils/collect_env.pyto see if"MMCV Compiler"/"MMCV CUDA Compiler"is the same as"GCC"/"CUDA_HOME".If those symbols are PyTorch symbols (e.g., symbols containing caffe, aten, and TH), check whether the PyTorch version is the same as that used for compiling mmcv.
Run
python mmdet/utils/collect_env.pyto check whether PyTorch, torchvision, and MMCV are built by and running on the same environment.
setuptools.sandbox.UnpickleableException: DistutilsSetupError(“each element of ‘ext_modules’ option must be an Extension instance or 2-tuple”)
If you are using miniconda rather than anaconda, check whether Cython is installed as indicated in #3379. You need to manually install Cython first and then run command
pip install -r requirements.txt.You may also need to check the compatibility between the
setuptools,Cython, andPyTorchin your environment.
“Segmentation fault”.
Check you GCC version and use GCC 5.4. This usually caused by the incompatibility between PyTorch and the environment (e.g., GCC < 4.9 for PyTorch). We also recommend the users to avoid using GCC 5.5 because many feedbacks report that GCC 5.5 will cause “segmentation fault” and simply changing it to GCC 5.4 could solve the problem.
Check whether PyTorch is correctly installed and could use CUDA op, e.g. type the following command in your terminal.
python -c 'import torch; print(torch.cuda.is_available())'
And see whether they could correctly output results.
If Pytorch is correctly installed, check whether MMCV is correctly installed.
python -c 'import mmcv; import mmcv.ops'
If MMCV is correctly installed, then there will be no issue of the above two commands.
If MMCV and Pytorch is correctly installed, you man use
ipdb,pdbto set breakpoints or directly add ‘print’ in mmdetection code and see which part leads the segmentation fault.
Training¶
“Loss goes Nan”
Check if the dataset annotations are valid: zero-size bounding boxes will cause the regression loss to be Nan due to the commonly used transformation for box regression. Some small size (width or height are smaller than 1) boxes will also cause this problem after data augmentation (e.g., instaboost). So check the data and try to filter out those zero-size boxes and skip some risky augmentations on the small-size boxes when you face the problem.
Reduce the learning rate: the learning rate might be too large due to some reasons, e.g., change of batch size. You can rescale them to the value that could stably train the model.
Extend the warmup iterations: some models are sensitive to the learning rate at the start of the training. You can extend the warmup iterations, e.g., change the
warmup_itersfrom 500 to 1000 or 2000.Add gradient clipping: some models requires gradient clipping to stabilize the training process. The default of
grad_clipisNone, you can add gradient clippint to avoid gradients that are too large, i.e., setoptimizer_config=dict(_delete_=True, grad_clip=dict(max_norm=35, norm_type=2))in your config file. If your config does not inherits from any basic config that containsoptimizer_config=dict(grad_clip=None), you can simply addoptimizer_config=dict(grad_clip=dict(max_norm=35, norm_type=2)).
“GPU out of memory”
There are some scenarios when there are large amount of ground truth boxes, which may cause OOM during target assignment. You can set
gpu_assign_thr=Nin the config of assigner thus the assigner will calculate box overlaps through CPU when there are more than N GT boxes.Set
with_cp=Truein the backbone. This uses the sublinear strategy in PyTorch to reduce GPU memory cost in the backbone.Try mixed precision training using following the examples in
config/fp16. Theloss_scalemight need further tuning for different models.Try to use
AvoidCUDAOOMto avoid GPU out of memory. It will first retry after callingtorch.cuda.empty_cache(). If it still fails, it will then retry by converting the type of inputs to FP16 format. If it still fails, it will try to copy inputs from GPUs to CPUs to continue computing. Try AvoidOOM in you code to make the code continue to run when GPU memory runs out:from mmdet.utils import AvoidCUDAOOM output = AvoidCUDAOOM.retry_if_cuda_oom(some_function)(input1, input2)
You can also try
AvoidCUDAOOMas a decorator to make the code continue to run when GPU memory runs out:from mmdet.utils import AvoidCUDAOOM @AvoidCUDAOOM.retry_if_cuda_oom def function(*args, **kwargs): ... return xxx
“RuntimeError: Expected to have finished reduction in the prior iteration before starting a new one”
This error indicates that your module has parameters that were not used in producing loss. This phenomenon may be caused by running different branches in your code in DDP mode.
You can set
find_unused_parameters = Truein the config to solve the above problems(but this will slow down the training speed.If the version of your MMCV >= 1.4.1, you can get the name of those unused parameters with
detect_anomalous_params=Trueinoptimizer_configof config.
Save the best model
It can be turned on by configuring
evaluation = dict(save_best=‘auto’). In the case of theautoparameter, the first key in the returned evaluation result will be used as the basis for selecting the best model. You can also directly set the key in the evaluation result to manually set it, for example,evaluation = dict(save_best='mAP' ).Resume training with
ExpMomentumEMAHookIf you use
ExpMomentumEMAHookin training, you can’t just use command line parameters--resume-fromnor--cfg-options resume_fromto restore model parameters during resume, i.e., the commandpython tools/train.py configs/yolox/yolox_s_8x8_300e_coco.py --resume-from ./work_dir/yolox_s_8x8_300e_coco/epoch_x.pthwill not work. SinceExpMomentumEMAHookneeds to reload the weights, taking theyolox_salgorithm as an example, you should modify the values ofresume_fromin two places of the config as below:# Open configs/yolox/yolox_s_8x8_300e_coco.py directly and modify all resume_from fields resume_from=./work_dir/yolox_s_8x8_300e_coco/epoch_x.pth custom_hooks=[... dict( type='ExpMomentumEMAHook', resume_from=./work_dir/yolox_s_8x8_300e_coco/epoch_x.pth, momentum=0.0001, priority=49) ]
Evaluation¶
COCO Dataset, AP or AR = -1
According to the definition of COCO dataset, the small and medium areas in an image are less than 1024 (32*32), 9216 (96*96), respectively.
If the corresponding area has no object, the result of AP and AR will set to -1.
Model¶
stylein ResNetThe
styleparameter in ResNet allows eitherpytorchorcaffestyle. It indicates the difference in the Bottleneck module. Bottleneck is a stacking structure of1x1-3x3-1x1convolutional layers. In the case ofcaffemode, the convolution layer withstride=2is the first1x1convolution, while inpyorchmode, it is the second3x3convolution hasstride=2. A sample code is as below:if self.style == 'pytorch': self.conv1_stride = 1 self.conv2_stride = stride else: self.conv1_stride = stride self.conv2_stride = 1
ResNeXt parameter description
ResNeXt comes from the paper
Aggregated Residual Transformations for Deep Neural Networks. It introduces group and uses “cardinality” to control the number of groups to achieve a balance between accuracy and complexity. It controls the basic width and grouping parameters of the internal Bottleneck module through two hyperparametersbaseWidthandcardinality. An example configuration name in MMDetection ismask_rcnn_x101_64x4d_fpn_mstrain-poly_3x_coco.py, wheremask_rcnnrepresents the algorithm using Mask R-CNN,x101represents the backbone network using ResNeXt-101, and64x4drepresents that the bottleneck block has 64 group and each group has basic width of 4.norm_evalin backboneSince the detection model is usually large and the input image resolution is high, this will result in a small batch of the detection model, which will make the variance of the statistics calculated by BatchNorm during the training process very large and not as stable as the statistics obtained during the pre-training of the backbone network . Therefore, the
norm_eval=Truemode is generally used in training, and the BatchNorm statistics in the pre-trained backbone network are directly used. The few algorithms that use large batches are thenorm_eval=Falsemode, such as NASFPN. For the backbone network without ImageNet pre-training and the batch is relatively small, you can consider usingSyncBN.
mmdet.apis¶
mmdet.core¶
anchor¶
- class mmdet.core.anchor.AnchorGenerator(strides, ratios, scales=None, base_sizes=None, scale_major=True, octave_base_scale=None, scales_per_octave=None, centers=None, center_offset=0.0)[source]¶
Standard anchor generator for 2D anchor-based detectors.
- Parameters
strides (list[int] | list[tuple[int, int]]) – Strides of anchors in multiple feature levels in order (w, h).
ratios (list[float]) – The list of ratios between the height and width of anchors in a single level.
scales (list[int] | None) – Anchor scales for anchors in a single level. It cannot be set at the same time if octave_base_scale and scales_per_octave are set.
base_sizes (list[int] | None) – The basic sizes of anchors in multiple levels. If None is given, strides will be used as base_sizes. (If strides are non square, the shortest stride is taken.)
scale_major (bool) – Whether to multiply scales first when generating base anchors. If true, the anchors in the same row will have the same scales. By default it is True in V2.0
octave_base_scale (int) – The base scale of octave.
scales_per_octave (int) – Number of scales for each octave. octave_base_scale and scales_per_octave are usually used in retinanet and the scales should be None when they are set.
centers (list[tuple[float, float]] | None) – The centers of the anchor relative to the feature grid center in multiple feature levels. By default it is set to be None and not used. If a list of tuple of float is given, they will be used to shift the centers of anchors.
center_offset (float) – The offset of center in proportion to anchors’ width and height. By default it is 0 in V2.0.
Examples
>>> from mmdet.core import AnchorGenerator >>> self = AnchorGenerator([16], [1.], [1.], [9]) >>> all_anchors = self.grid_priors([(2, 2)], device='cpu') >>> print(all_anchors) [tensor([[-4.5000, -4.5000, 4.5000, 4.5000], [11.5000, -4.5000, 20.5000, 4.5000], [-4.5000, 11.5000, 4.5000, 20.5000], [11.5000, 11.5000, 20.5000, 20.5000]])] >>> self = AnchorGenerator([16, 32], [1.], [1.], [9, 18]) >>> all_anchors = self.grid_priors([(2, 2), (1, 1)], device='cpu') >>> print(all_anchors) [tensor([[-4.5000, -4.5000, 4.5000, 4.5000], [11.5000, -4.5000, 20.5000, 4.5000], [-4.5000, 11.5000, 4.5000, 20.5000], [11.5000, 11.5000, 20.5000, 20.5000]]), tensor([[-9., -9., 9., 9.]])]
- gen_base_anchors()[source]¶
Generate base anchors.
- Returns
Base anchors of a feature grid in multiple feature levels.
- Return type
list(torch.Tensor)
- gen_single_level_base_anchors(base_size, scales, ratios, center=None)[source]¶
Generate base anchors of a single level.
- Parameters
base_size (int | float) – Basic size of an anchor.
scales (torch.Tensor) – Scales of the anchor.
ratios (torch.Tensor) – The ratio between between the height and width of anchors in a single level.
center (tuple[float], optional) – The center of the base anchor related to a single feature grid. Defaults to None.
- Returns
Anchors in a single-level feature maps.
- Return type
torch.Tensor
- grid_anchors(featmap_sizes, device='cuda')[source]¶
Generate grid anchors in multiple feature levels.
- Parameters
featmap_sizes (list[tuple]) – List of feature map sizes in multiple feature levels.
device (str) – Device where the anchors will be put on.
- Returns
Anchors in multiple feature levels. The sizes of each tensor should be [N, 4], where N = width * height * num_base_anchors, width and height are the sizes of the corresponding feature level, num_base_anchors is the number of anchors for that level.
- Return type
list[torch.Tensor]
- grid_priors(featmap_sizes, dtype=torch.float32, device='cuda')[source]¶
Generate grid anchors in multiple feature levels.
- Parameters
featmap_sizes (list[tuple]) – List of feature map sizes in multiple feature levels.
dtype (
torch.dtype) – Dtype of priors. Default: torch.float32.device (str) – The device where the anchors will be put on.
- Returns
Anchors in multiple feature levels. The sizes of each tensor should be [N, 4], where N = width * height * num_base_anchors, width and height are the sizes of the corresponding feature level, num_base_anchors is the number of anchors for that level.
- Return type
list[torch.Tensor]
- property num_base_anchors¶
total number of base anchors in a feature grid
- Type
list[int]
- property num_base_priors¶
The number of priors (anchors) at a point on the feature grid
- Type
list[int]
- property num_levels¶
number of feature levels that the generator will be applied
- Type
int
- single_level_grid_anchors(base_anchors, featmap_size, stride=(16, 16), device='cuda')[source]¶
Generate grid anchors of a single level.
Note
This function is usually called by method
self.grid_anchors.- Parameters
base_anchors (torch.Tensor) – The base anchors of a feature grid.
featmap_size (tuple[int]) – Size of the feature maps.
stride (tuple[int], optional) – Stride of the feature map in order (w, h). Defaults to (16, 16).
device (str, optional) – Device the tensor will be put on. Defaults to ‘cuda’.
- Returns
Anchors in the overall feature maps.
- Return type
torch.Tensor
- single_level_grid_priors(featmap_size, level_idx, dtype=torch.float32, device='cuda')[source]¶
Generate grid anchors of a single level.
Note
This function is usually called by method
self.grid_priors.- Parameters
featmap_size (tuple[int]) – Size of the feature maps.
level_idx (int) – The index of corresponding feature map level.
(obj (dtype) – torch.dtype): Date type of points.Defaults to
torch.float32.device (str, optional) – The device the tensor will be put on. Defaults to ‘cuda’.
- Returns
Anchors in the overall feature maps.
- Return type
torch.Tensor
- single_level_valid_flags(featmap_size, valid_size, num_base_anchors, device='cuda')[source]¶
Generate the valid flags of anchor in a single feature map.
- Parameters
featmap_size (tuple[int]) – The size of feature maps, arrange as (h, w).
valid_size (tuple[int]) – The valid size of the feature maps.
num_base_anchors (int) – The number of base anchors.
device (str, optional) – Device where the flags will be put on. Defaults to ‘cuda’.
- Returns
The valid flags of each anchor in a single level feature map.
- Return type
torch.Tensor
- sparse_priors(prior_idxs, featmap_size, level_idx, dtype=torch.float32, device='cuda')[source]¶
Generate sparse anchors according to the
prior_idxs.- Parameters
prior_idxs (Tensor) – The index of corresponding anchors in the feature map.
featmap_size (tuple[int]) – feature map size arrange as (h, w).
level_idx (int) – The level index of corresponding feature map.
(obj (device) – torch.dtype): Date type of points.Defaults to
torch.float32.(obj – torch.device): The device where the points is located.
- Returns
- Anchor with shape (N, 4), N should be equal to
the length of
prior_idxs.
- Return type
Tensor
- valid_flags(featmap_sizes, pad_shape, device='cuda')[source]¶
Generate valid flags of anchors in multiple feature levels.
- Parameters
featmap_sizes (list(tuple)) – List of feature map sizes in multiple feature levels.
pad_shape (tuple) – The padded shape of the image.
device (str) – Device where the anchors will be put on.
- Returns
Valid flags of anchors in multiple levels.
- Return type
list(torch.Tensor)
- class mmdet.core.anchor.LegacyAnchorGenerator(strides, ratios, scales=None, base_sizes=None, scale_major=True, octave_base_scale=None, scales_per_octave=None, centers=None, center_offset=0.0)[source]¶
Legacy anchor generator used in MMDetection V1.x.
Note
Difference to the V2.0 anchor generator:
The center offset of V1.x anchors are set to be 0.5 rather than 0.
The width/height are minused by 1 when calculating the anchors’ centers and corners to meet the V1.x coordinate system.
The anchors’ corners are quantized.
- Parameters
strides (list[int] | list[tuple[int]]) – Strides of anchors in multiple feature levels.
ratios (list[float]) – The list of ratios between the height and width of anchors in a single level.
scales (list[int] | None) – Anchor scales for anchors in a single level. It cannot be set at the same time if octave_base_scale and scales_per_octave are set.
base_sizes (list[int]) – The basic sizes of anchors in multiple levels. If None is given, strides will be used to generate base_sizes.
scale_major (bool) – Whether to multiply scales first when generating base anchors. If true, the anchors in the same row will have the same scales. By default it is True in V2.0
octave_base_scale (int) – The base scale of octave.
scales_per_octave (int) – Number of scales for each octave. octave_base_scale and scales_per_octave are usually used in retinanet and the scales should be None when they are set.
centers (list[tuple[float, float]] | None) – The centers of the anchor relative to the feature grid center in multiple feature levels. By default it is set to be None and not used. It a list of float is given, this list will be used to shift the centers of anchors.
center_offset (float) – The offset of center in proportion to anchors’ width and height. By default it is 0.5 in V2.0 but it should be 0.5 in v1.x models.
Examples
>>> from mmdet.core import LegacyAnchorGenerator >>> self = LegacyAnchorGenerator( >>> [16], [1.], [1.], [9], center_offset=0.5) >>> all_anchors = self.grid_anchors(((2, 2),), device='cpu') >>> print(all_anchors) [tensor([[ 0., 0., 8., 8.], [16., 0., 24., 8.], [ 0., 16., 8., 24.], [16., 16., 24., 24.]])]
- gen_single_level_base_anchors(base_size, scales, ratios, center=None)[source]¶
Generate base anchors of a single level.
Note
The width/height of anchors are minused by 1 when calculating the centers and corners to meet the V1.x coordinate system.
- Parameters
base_size (int | float) – Basic size of an anchor.
scales (torch.Tensor) – Scales of the anchor.
ratios (torch.Tensor) – The ratio between between the height. and width of anchors in a single level.
center (tuple[float], optional) – The center of the base anchor related to a single feature grid. Defaults to None.
- Returns
Anchors in a single-level feature map.
- Return type
torch.Tensor
- class mmdet.core.anchor.MlvlPointGenerator(strides, offset=0.5)[source]¶
Standard points generator for multi-level (Mlvl) feature maps in 2D points-based detectors.
- Parameters
strides (list[int] | list[tuple[int, int]]) – Strides of anchors in multiple feature levels in order (w, h).
offset (float) – The offset of points, the value is normalized with corresponding stride. Defaults to 0.5.
- grid_priors(featmap_sizes, dtype=torch.float32, device='cuda', with_stride=False)[source]¶
Generate grid points of multiple feature levels.
- Parameters
featmap_sizes (list[tuple]) – List of feature map sizes in multiple feature levels, each size arrange as as (h, w).
dtype (
dtype) – Dtype of priors. Default: torch.float32.device (str) – The device where the anchors will be put on.
with_stride (bool) – Whether to concatenate the stride to the last dimension of points.
- Returns
Points of multiple feature levels. The sizes of each tensor should be (N, 2) when with stride is
False, where N = width * height, width and height are the sizes of the corresponding feature level, and the last dimension 2 represent (coord_x, coord_y), otherwise the shape should be (N, 4), and the last dimension 4 represent (coord_x, coord_y, stride_w, stride_h).- Return type
list[torch.Tensor]
- property num_base_priors¶
The number of priors (points) at a point on the feature grid
- Type
list[int]
- property num_levels¶
number of feature levels that the generator will be applied
- Type
int
- single_level_grid_priors(featmap_size, level_idx, dtype=torch.float32, device='cuda', with_stride=False)[source]¶
Generate grid Points of a single level.
Note
This function is usually called by method
self.grid_priors.- Parameters
featmap_size (tuple[int]) – Size of the feature maps, arrange as (h, w).
level_idx (int) – The index of corresponding feature map level.
dtype (
dtype) – Dtype of priors. Default: torch.float32.device (str, optional) – The device the tensor will be put on. Defaults to ‘cuda’.
with_stride (bool) – Concatenate the stride to the last dimension of points.
- Returns
Points of single feature levels. The shape of tensor should be (N, 2) when with stride is
False, where N = width * height, width and height are the sizes of the corresponding feature level, and the last dimension 2 represent (coord_x, coord_y), otherwise the shape should be (N, 4), and the last dimension 4 represent (coord_x, coord_y, stride_w, stride_h).- Return type
Tensor
- single_level_valid_flags(featmap_size, valid_size, device='cuda')[source]¶
Generate the valid flags of points of a single feature map.
- Parameters
featmap_size (tuple[int]) – The size of feature maps, arrange as as (h, w).
valid_size (tuple[int]) – The valid size of the feature maps. The size arrange as as (h, w).
device (str, optional) – The device where the flags will be put on. Defaults to ‘cuda’.
- Returns
The valid flags of each points in a single level feature map.
- Return type
torch.Tensor
- sparse_priors(prior_idxs, featmap_size, level_idx, dtype=torch.float32, device='cuda')[source]¶
Generate sparse points according to the
prior_idxs.- Parameters
prior_idxs (Tensor) – The index of corresponding anchors in the feature map.
featmap_size (tuple[int]) – feature map size arrange as (w, h).
level_idx (int) – The level index of corresponding feature map.
(obj (device) – torch.dtype): Date type of points. Defaults to
torch.float32.(obj – torch.device): The device where the points is located.
- Returns
Anchor with shape (N, 2), N should be equal to the length of
prior_idxs. And last dimension 2 represent (coord_x, coord_y).- Return type
Tensor
- valid_flags(featmap_sizes, pad_shape, device='cuda')[source]¶
Generate valid flags of points of multiple feature levels.
- Parameters
featmap_sizes (list(tuple)) – List of feature map sizes in multiple feature levels, each size arrange as as (h, w).
pad_shape (tuple(int)) – The padded shape of the image, arrange as (h, w).
device (str) – The device where the anchors will be put on.
- Returns
Valid flags of points of multiple levels.
- Return type
list(torch.Tensor)
- class mmdet.core.anchor.YOLOAnchorGenerator(strides, base_sizes)[source]¶
Anchor generator for YOLO.
- Parameters
strides (list[int] | list[tuple[int, int]]) – Strides of anchors in multiple feature levels.
base_sizes (list[list[tuple[int, int]]]) – The basic sizes of anchors in multiple levels.
- gen_base_anchors()[source]¶
Generate base anchors.
- Returns
Base anchors of a feature grid in multiple feature levels.
- Return type
list(torch.Tensor)
- gen_single_level_base_anchors(base_sizes_per_level, center=None)[source]¶
Generate base anchors of a single level.
- Parameters
base_sizes_per_level (list[tuple[int, int]]) – Basic sizes of anchors.
center (tuple[float], optional) – The center of the base anchor related to a single feature grid. Defaults to None.
- Returns
Anchors in a single-level feature maps.
- Return type
torch.Tensor
- property num_levels¶
number of feature levels that the generator will be applied
- Type
int
- responsible_flags(featmap_sizes, gt_bboxes, device='cuda')[source]¶
Generate responsible anchor flags of grid cells in multiple scales.
- Parameters
featmap_sizes (list(tuple)) – List of feature map sizes in multiple feature levels.
gt_bboxes (Tensor) – Ground truth boxes, shape (n, 4).
device (str) – Device where the anchors will be put on.
- Returns
responsible flags of anchors in multiple level
- Return type
list(torch.Tensor)
- single_level_responsible_flags(featmap_size, gt_bboxes, stride, num_base_anchors, device='cuda')[source]¶
Generate the responsible flags of anchor in a single feature map.
- Parameters
featmap_size (tuple[int]) – The size of feature maps.
gt_bboxes (Tensor) – Ground truth boxes, shape (n, 4).
stride (tuple(int)) – stride of current level
num_base_anchors (int) – The number of base anchors.
device (str, optional) – Device where the flags will be put on. Defaults to ‘cuda’.
- Returns
The valid flags of each anchor in a single level feature map.
- Return type
torch.Tensor
- mmdet.core.anchor.anchor_inside_flags(flat_anchors, valid_flags, img_shape, allowed_border=0)[source]¶
Check whether the anchors are inside the border.
- Parameters
flat_anchors (torch.Tensor) – Flatten anchors, shape (n, 4).
valid_flags (torch.Tensor) – An existing valid flags of anchors.
img_shape (tuple(int)) – Shape of current image.
allowed_border (int, optional) – The border to allow the valid anchor. Defaults to 0.
- Returns
Flags indicating whether the anchors are inside a valid range.
- Return type
torch.Tensor
- mmdet.core.anchor.calc_region(bbox, ratio, featmap_size=None)[source]¶
Calculate a proportional bbox region.
The bbox center are fixed and the new h’ and w’ is h * ratio and w * ratio.
- Parameters
bbox (Tensor) – Bboxes to calculate regions, shape (n, 4).
ratio (float) – Ratio of the output region.
featmap_size (tuple) – Feature map size used for clipping the boundary.
- Returns
x1, y1, x2, y2
- Return type
tuple
bbox¶
export¶
mask¶
evaluation¶
post_processing¶
utils¶
mmdet.models¶
detectors¶
backbones¶
- class mmdet.models.backbones.CSPDarknet(arch='P5', deepen_factor=1.0, widen_factor=1.0, out_indices=(2, 3, 4), frozen_stages=-1, use_depthwise=False, arch_ovewrite=None, spp_kernal_sizes=(5, 9, 13), conv_cfg=None, norm_cfg={'eps': 0.001, 'momentum': 0.03, 'type': 'BN'}, act_cfg={'type': 'Swish'}, norm_eval=False, init_cfg={'a': 2.23606797749979, 'distribution': 'uniform', 'layer': 'Conv2d', 'mode': 'fan_in', 'nonlinearity': 'leaky_relu', 'type': 'Kaiming'})[source]¶
CSP-Darknet backbone used in YOLOv5 and YOLOX.
- Parameters
arch (str) – Architecture of CSP-Darknet, from {P5, P6}. Default: P5.
deepen_factor (float) – Depth multiplier, multiply number of blocks in CSP layer by this amount. Default: 1.0.
widen_factor (float) – Width multiplier, multiply number of channels in each layer by this amount. Default: 1.0.
out_indices (Sequence[int]) – Output from which stages. Default: (2, 3, 4).
frozen_stages (int) – Stages to be frozen (stop grad and set eval mode). -1 means not freezing any parameters. Default: -1.
use_depthwise (bool) – Whether to use depthwise separable convolution. Default: False.
arch_ovewrite (list) – Overwrite default arch settings. Default: None.
spp_kernal_sizes – (tuple[int]): Sequential of kernel sizes of SPP layers. Default: (5, 9, 13).
conv_cfg (dict) – Config dict for convolution layer. Default: None.
norm_cfg (dict) – Dictionary to construct and config norm layer. Default: dict(type=’BN’, requires_grad=True).
act_cfg (dict) – Config dict for activation layer. Default: dict(type=’LeakyReLU’, negative_slope=0.1).
norm_eval (bool) – Whether to set norm layers to eval mode, namely, freeze running stats (mean and var). Note: Effect on Batch Norm and its variants only.
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None.
Example
>>> from mmdet.models import CSPDarknet >>> import torch >>> self = CSPDarknet(depth=53) >>> self.eval() >>> inputs = torch.rand(1, 3, 416, 416) >>> level_outputs = self.forward(inputs) >>> for level_out in level_outputs: ... print(tuple(level_out.shape)) ... (1, 256, 52, 52) (1, 512, 26, 26) (1, 1024, 13, 13)
- forward(x)[source]¶
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- train(mode=True)[source]¶
Sets the module in training mode.
This has any effect only on certain modules. See documentations of particular modules for details of their behaviors in training/evaluation mode, if they are affected, e.g.
Dropout,BatchNorm, etc.- Parameters
mode (bool) – whether to set training mode (
True) or evaluation mode (False). Default:True.- Returns
self
- Return type
Module
- class mmdet.models.backbones.Darknet(depth=53, out_indices=(3, 4, 5), frozen_stages=-1, conv_cfg=None, norm_cfg={'requires_grad': True, 'type': 'BN'}, act_cfg={'negative_slope': 0.1, 'type': 'LeakyReLU'}, norm_eval=True, pretrained=None, init_cfg=None)[source]¶
Darknet backbone.
- Parameters
depth (int) – Depth of Darknet. Currently only support 53.
out_indices (Sequence[int]) – Output from which stages.
frozen_stages (int) – Stages to be frozen (stop grad and set eval mode). -1 means not freezing any parameters. Default: -1.
conv_cfg (dict) – Config dict for convolution layer. Default: None.
norm_cfg (dict) – Dictionary to construct and config norm layer. Default: dict(type=’BN’, requires_grad=True)
act_cfg (dict) – Config dict for activation layer. Default: dict(type=’LeakyReLU’, negative_slope=0.1).
norm_eval (bool) – Whether to set norm layers to eval mode, namely, freeze running stats (mean and var). Note: Effect on Batch Norm and its variants only.
pretrained (str, optional) – model pretrained path. Default: None
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None
Example
>>> from mmdet.models import Darknet >>> import torch >>> self = Darknet(depth=53) >>> self.eval() >>> inputs = torch.rand(1, 3, 416, 416) >>> level_outputs = self.forward(inputs) >>> for level_out in level_outputs: ... print(tuple(level_out.shape)) ... (1, 256, 52, 52) (1, 512, 26, 26) (1, 1024, 13, 13)
- forward(x)[source]¶
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- static make_conv_res_block(in_channels, out_channels, res_repeat, conv_cfg=None, norm_cfg={'requires_grad': True, 'type': 'BN'}, act_cfg={'negative_slope': 0.1, 'type': 'LeakyReLU'})[source]¶
In Darknet backbone, ConvLayer is usually followed by ResBlock. This function will make that. The Conv layers always have 3x3 filters with stride=2. The number of the filters in Conv layer is the same as the out channels of the ResBlock.
- Parameters
in_channels (int) – The number of input channels.
out_channels (int) – The number of output channels.
res_repeat (int) – The number of ResBlocks.
conv_cfg (dict) – Config dict for convolution layer. Default: None.
norm_cfg (dict) – Dictionary to construct and config norm layer. Default: dict(type=’BN’, requires_grad=True)
act_cfg (dict) – Config dict for activation layer. Default: dict(type=’LeakyReLU’, negative_slope=0.1).
- train(mode=True)[source]¶
Sets the module in training mode.
This has any effect only on certain modules. See documentations of particular modules for details of their behaviors in training/evaluation mode, if they are affected, e.g.
Dropout,BatchNorm, etc.- Parameters
mode (bool) – whether to set training mode (
True) or evaluation mode (False). Default:True.- Returns
self
- Return type
Module
- class mmdet.models.backbones.DetectoRS_ResNeXt(groups=1, base_width=4, **kwargs)[source]¶
ResNeXt backbone for DetectoRS.
- Parameters
groups (int) – The number of groups in ResNeXt.
base_width (int) – The base width of ResNeXt.
- class mmdet.models.backbones.DetectoRS_ResNet(sac=None, stage_with_sac=(False, False, False, False), rfp_inplanes=None, output_img=False, pretrained=None, init_cfg=None, **kwargs)[source]¶
ResNet backbone for DetectoRS.
- Parameters
sac (dict, optional) – Dictionary to construct SAC (Switchable Atrous Convolution). Default: None.
stage_with_sac (list) – Which stage to use sac. Default: (False, False, False, False).
rfp_inplanes (int, optional) – The number of channels from RFP. Default: None. If specified, an additional conv layer will be added for
rfp_feat. Otherwise, the structure is the same as base class.output_img (bool) – If
True, the input image will be inserted into the starting position of output. Default: False.
- class mmdet.models.backbones.EfficientNet(arch='b0', drop_path_rate=0.0, out_indices=(6,), frozen_stages=0, conv_cfg={'type': 'Conv2dAdaptivePadding'}, norm_cfg={'eps': 0.001, 'type': 'BN'}, act_cfg={'type': 'Swish'}, norm_eval=False, with_cp=False, init_cfg=[{'type': 'Kaiming', 'layer': 'Conv2d'}, {'type': 'Constant', 'layer': ['_BatchNorm', 'GroupNorm'], 'val': 1}])[source]¶
EfficientNet backbone.
- Parameters
arch (str) – Architecture of efficientnet. Defaults to b0.
out_indices (Sequence[int]) – Output from which stages. Defaults to (6, ).
frozen_stages (int) – Stages to be frozen (all param fixed). Defaults to 0, which means not freezing any parameters.
conv_cfg (dict) – Config dict for convolution layer. Defaults to None, which means using conv2d.
norm_cfg (dict) – Config dict for normalization layer. Defaults to dict(type=’BN’).
act_cfg (dict) – Config dict for activation layer. Defaults to dict(type=’Swish’).
norm_eval (bool) – Whether to set norm layers to eval mode, namely, freeze running stats (mean and var). Note: Effect on Batch Norm and its variants only. Defaults to False.
with_cp (bool) – Use checkpoint or not. Using checkpoint will save some memory while slowing down the training speed. Defaults to False.
- forward(x)[source]¶
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- train(mode=True)[source]¶
Sets the module in training mode.
This has any effect only on certain modules. See documentations of particular modules for details of their behaviors in training/evaluation mode, if they are affected, e.g.
Dropout,BatchNorm, etc.- Parameters
mode (bool) – whether to set training mode (
True) or evaluation mode (False). Default:True.- Returns
self
- Return type
Module
- class mmdet.models.backbones.HRNet(extra, in_channels=3, conv_cfg=None, norm_cfg={'type': 'BN'}, norm_eval=True, with_cp=False, zero_init_residual=False, multiscale_output=True, pretrained=None, init_cfg=None)[source]¶
HRNet backbone.
High-Resolution Representations for Labeling Pixels and Regions arXiv:.
- Parameters
extra (dict) –
Detailed configuration for each stage of HRNet. There must be 4 stages, the configuration for each stage must have 5 keys:
num_modules(int): The number of HRModule in this stage.
num_branches(int): The number of branches in the HRModule.
block(str): The type of convolution block.
- num_blocks(tuple): The number of blocks in each branch.
The length must be equal to num_branches.
- num_channels(tuple): The number of channels in each branch.
The length must be equal to num_branches.
in_channels (int) – Number of input image channels. Default: 3.
conv_cfg (dict) – Dictionary to construct and config conv layer.
norm_cfg (dict) – Dictionary to construct and config norm layer.
norm_eval (bool) – Whether to set norm layers to eval mode, namely, freeze running stats (mean and var). Note: Effect on Batch Norm and its variants only. Default: True.
with_cp (bool) – Use checkpoint or not. Using checkpoint will save some memory while slowing down the training speed. Default: False.
zero_init_residual (bool) – Whether to use zero init for last norm layer in resblocks to let them behave as identity. Default: False.
multiscale_output (bool) – Whether to output multi-level features produced by multiple branches. If False, only the first level feature will be output. Default: True.
pretrained (str, optional) – Model pretrained path. Default: None.
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None.
Example
>>> from mmdet.models import HRNet >>> import torch >>> extra = dict( >>> stage1=dict( >>> num_modules=1, >>> num_branches=1, >>> block='BOTTLENECK', >>> num_blocks=(4, ), >>> num_channels=(64, )), >>> stage2=dict( >>> num_modules=1, >>> num_branches=2, >>> block='BASIC', >>> num_blocks=(4, 4), >>> num_channels=(32, 64)), >>> stage3=dict( >>> num_modules=4, >>> num_branches=3, >>> block='BASIC', >>> num_blocks=(4, 4, 4), >>> num_channels=(32, 64, 128)), >>> stage4=dict( >>> num_modules=3, >>> num_branches=4, >>> block='BASIC', >>> num_blocks=(4, 4, 4, 4), >>> num_channels=(32, 64, 128, 256))) >>> self = HRNet(extra, in_channels=1) >>> self.eval() >>> inputs = torch.rand(1, 1, 32, 32) >>> level_outputs = self.forward(inputs) >>> for level_out in level_outputs: ... print(tuple(level_out.shape)) (1, 32, 8, 8) (1, 64, 4, 4) (1, 128, 2, 2) (1, 256, 1, 1)
- property norm1¶
the normalization layer named “norm1”
- Type
nn.Module
- property norm2¶
the normalization layer named “norm2”
- Type
nn.Module
- class mmdet.models.backbones.HourglassNet(downsample_times=5, num_stacks=2, stage_channels=(256, 256, 384, 384, 384, 512), stage_blocks=(2, 2, 2, 2, 2, 4), feat_channel=256, norm_cfg={'requires_grad': True, 'type': 'BN'}, pretrained=None, init_cfg=None)[source]¶
HourglassNet backbone.
Stacked Hourglass Networks for Human Pose Estimation. More details can be found in the paper .
- Parameters
downsample_times (int) – Downsample times in a HourglassModule.
num_stacks (int) – Number of HourglassModule modules stacked, 1 for Hourglass-52, 2 for Hourglass-104.
stage_channels (list[int]) – Feature channel of each sub-module in a HourglassModule.
stage_blocks (list[int]) – Number of sub-modules stacked in a HourglassModule.
feat_channel (int) – Feature channel of conv after a HourglassModule.
norm_cfg (dict) – Dictionary to construct and config norm layer.
pretrained (str, optional) – model pretrained path. Default: None
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None
Example
>>> from mmdet.models import HourglassNet >>> import torch >>> self = HourglassNet() >>> self.eval() >>> inputs = torch.rand(1, 3, 511, 511) >>> level_outputs = self.forward(inputs) >>> for level_output in level_outputs: ... print(tuple(level_output.shape)) (1, 256, 128, 128) (1, 256, 128, 128)
- class mmdet.models.backbones.MobileNetV2(widen_factor=1.0, out_indices=(1, 2, 4, 7), frozen_stages=-1, conv_cfg=None, norm_cfg={'type': 'BN'}, act_cfg={'type': 'ReLU6'}, norm_eval=False, with_cp=False, pretrained=None, init_cfg=None)[source]¶
MobileNetV2 backbone.
- Parameters
widen_factor (float) – Width multiplier, multiply number of channels in each layer by this amount. Default: 1.0.
out_indices (Sequence[int], optional) – Output from which stages. Default: (1, 2, 4, 7).
frozen_stages (int) – Stages to be frozen (all param fixed). Default: -1, which means not freezing any parameters.
conv_cfg (dict, optional) – Config dict for convolution layer. Default: None, which means using conv2d.
norm_cfg (dict) – Config dict for normalization layer. Default: dict(type=’BN’).
act_cfg (dict) – Config dict for activation layer. Default: dict(type=’ReLU6’).
norm_eval (bool) – Whether to set norm layers to eval mode, namely, freeze running stats (mean and var). Note: Effect on Batch Norm and its variants only. Default: False.
with_cp (bool) – Use checkpoint or not. Using checkpoint will save some memory while slowing down the training speed. Default: False.
pretrained (str, optional) – model pretrained path. Default: None
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None
- make_layer(out_channels, num_blocks, stride, expand_ratio)[source]¶
Stack InvertedResidual blocks to build a layer for MobileNetV2.
- Parameters
out_channels (int) – out_channels of block.
num_blocks (int) – number of blocks.
stride (int) – stride of the first block. Default: 1
expand_ratio (int) – Expand the number of channels of the hidden layer in InvertedResidual by this ratio. Default: 6.
- class mmdet.models.backbones.PyramidVisionTransformer(pretrain_img_size=224, in_channels=3, embed_dims=64, num_stages=4, num_layers=[3, 4, 6, 3], num_heads=[1, 2, 5, 8], patch_sizes=[4, 2, 2, 2], strides=[4, 2, 2, 2], paddings=[0, 0, 0, 0], sr_ratios=[8, 4, 2, 1], out_indices=(0, 1, 2, 3), mlp_ratios=[8, 8, 4, 4], qkv_bias=True, drop_rate=0.0, attn_drop_rate=0.0, drop_path_rate=0.1, use_abs_pos_embed=True, norm_after_stage=False, use_conv_ffn=False, act_cfg={'type': 'GELU'}, norm_cfg={'eps': 1e-06, 'type': 'LN'}, pretrained=None, convert_weights=True, init_cfg=None)[source]¶
Pyramid Vision Transformer (PVT)
Implementation of Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions.
- Parameters
pretrain_img_size (int | tuple[int]) – The size of input image when pretrain. Defaults: 224.
in_channels (int) – Number of input channels. Default: 3.
embed_dims (int) – Embedding dimension. Default: 64.
num_stags (int) – The num of stages. Default: 4.
num_layers (Sequence[int]) – The layer number of each transformer encode layer. Default: [3, 4, 6, 3].
num_heads (Sequence[int]) – The attention heads of each transformer encode layer. Default: [1, 2, 5, 8].
patch_sizes (Sequence[int]) – The patch_size of each patch embedding. Default: [4, 2, 2, 2].
strides (Sequence[int]) – The stride of each patch embedding. Default: [4, 2, 2, 2].
paddings (Sequence[int]) – The padding of each patch embedding. Default: [0, 0, 0, 0].
sr_ratios (Sequence[int]) – The spatial reduction rate of each transformer encode layer. Default: [8, 4, 2, 1].
out_indices (Sequence[int] | int) – Output from which stages. Default: (0, 1, 2, 3).
mlp_ratios (Sequence[int]) – The ratio of the mlp hidden dim to the embedding dim of each transformer encode layer. Default: [8, 8, 4, 4].
qkv_bias (bool) – Enable bias for qkv if True. Default: True.
drop_rate (float) – Probability of an element to be zeroed. Default 0.0.
attn_drop_rate (float) – The drop out rate for attention layer. Default 0.0.
drop_path_rate (float) – stochastic depth rate. Default 0.1.
use_abs_pos_embed (bool) – If True, add absolute position embedding to the patch embedding. Defaults: True.
use_conv_ffn (bool) – If True, use Convolutional FFN to replace FFN. Default: False.
act_cfg (dict) – The activation config for FFNs. Default: dict(type=’GELU’).
norm_cfg (dict) – Config dict for normalization layer. Default: dict(type=’LN’).
pretrained (str, optional) – model pretrained path. Default: None.
convert_weights (bool) – The flag indicates whether the pre-trained model is from the original repo. We may need to convert some keys to make it compatible. Default: True.
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None.
- forward(x)[source]¶
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- class mmdet.models.backbones.PyramidVisionTransformerV2(**kwargs)[source]¶
Implementation of PVTv2: Improved Baselines with Pyramid Vision Transformer.
- class mmdet.models.backbones.RegNet(arch, in_channels=3, stem_channels=32, base_channels=32, strides=(2, 2, 2, 2), dilations=(1, 1, 1, 1), out_indices=(0, 1, 2, 3), style='pytorch', deep_stem=False, avg_down=False, frozen_stages=-1, conv_cfg=None, norm_cfg={'requires_grad': True, 'type': 'BN'}, norm_eval=True, dcn=None, stage_with_dcn=(False, False, False, False), plugins=None, with_cp=False, zero_init_residual=True, pretrained=None, init_cfg=None)[source]¶
RegNet backbone.
More details can be found in paper .
- Parameters
arch (dict) –
The parameter of RegNets.
w0 (int): initial width
wa (float): slope of width
wm (float): quantization parameter to quantize the width
depth (int): depth of the backbone
group_w (int): width of group
bot_mul (float): bottleneck ratio, i.e. expansion of bottleneck.
strides (Sequence[int]) – Strides of the first block of each stage.
base_channels (int) – Base channels after stem layer.
in_channels (int) – Number of input image channels. Default: 3.
dilations (Sequence[int]) – Dilation of each stage.
out_indices (Sequence[int]) – Output from which stages.
style (str) – pytorch or caffe. If set to “pytorch”, the stride-two layer is the 3x3 conv layer, otherwise the stride-two layer is the first 1x1 conv layer.
frozen_stages (int) – Stages to be frozen (all param fixed). -1 means not freezing any parameters.
norm_cfg (dict) – dictionary to construct and config norm layer.
norm_eval (bool) – Whether to set norm layers to eval mode, namely, freeze running stats (mean and var). Note: Effect on Batch Norm and its variants only.
with_cp (bool) – Use checkpoint or not. Using checkpoint will save some memory while slowing down the training speed.
zero_init_residual (bool) – whether to use zero init for last norm layer in resblocks to let them behave as identity.
pretrained (str, optional) – model pretrained path. Default: None
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None
Example
>>> from mmdet.models import RegNet >>> import torch >>> self = RegNet( arch=dict( w0=88, wa=26.31, wm=2.25, group_w=48, depth=25, bot_mul=1.0)) >>> self.eval() >>> inputs = torch.rand(1, 3, 32, 32) >>> level_outputs = self.forward(inputs) >>> for level_out in level_outputs: ... print(tuple(level_out.shape)) (1, 96, 8, 8) (1, 192, 4, 4) (1, 432, 2, 2) (1, 1008, 1, 1)
- adjust_width_group(widths, bottleneck_ratio, groups)[source]¶
Adjusts the compatibility of widths and groups.
- Parameters
widths (list[int]) – Width of each stage.
bottleneck_ratio (float) – Bottleneck ratio.
groups (int) – number of groups in each stage
- Returns
The adjusted widths and groups of each stage.
- Return type
tuple(list)
- generate_regnet(initial_width, width_slope, width_parameter, depth, divisor=8)[source]¶
Generates per block width from RegNet parameters.
- Parameters
initial_width ([int]) – Initial width of the backbone
width_slope ([float]) – Slope of the quantized linear function
width_parameter ([int]) – Parameter used to quantize the width.
depth ([int]) – Depth of the backbone.
divisor (int, optional) – The divisor of channels. Defaults to 8.
- Returns
return a list of widths of each stage and the number of stages
- Return type
list, int
- class mmdet.models.backbones.Res2Net(scales=4, base_width=26, style='pytorch', deep_stem=True, avg_down=True, pretrained=None, init_cfg=None, **kwargs)[source]¶
Res2Net backbone.
- Parameters
scales (int) – Scales used in Res2Net. Default: 4
base_width (int) – Basic width of each scale. Default: 26
depth (int) – Depth of res2net, from {50, 101, 152}.
in_channels (int) – Number of input image channels. Default: 3.
num_stages (int) – Res2net stages. Default: 4.
strides (Sequence[int]) – Strides of the first block of each stage.
dilations (Sequence[int]) – Dilation of each stage.
out_indices (Sequence[int]) – Output from which stages.
style (str) – pytorch or caffe. If set to “pytorch”, the stride-two layer is the 3x3 conv layer, otherwise the stride-two layer is the first 1x1 conv layer.
deep_stem (bool) – Replace 7x7 conv in input stem with 3 3x3 conv
avg_down (bool) – Use AvgPool instead of stride conv when downsampling in the bottle2neck.
frozen_stages (int) – Stages to be frozen (stop grad and set eval mode). -1 means not freezing any parameters.
norm_cfg (dict) – Dictionary to construct and config norm layer.
norm_eval (bool) – Whether to set norm layers to eval mode, namely, freeze running stats (mean and var). Note: Effect on Batch Norm and its variants only.
plugins (list[dict]) –
List of plugins for stages, each dict contains:
cfg (dict, required): Cfg dict to build plugin.
position (str, required): Position inside block to insert plugin, options are ‘after_conv1’, ‘after_conv2’, ‘after_conv3’.
stages (tuple[bool], optional): Stages to apply plugin, length should be same as ‘num_stages’.
with_cp (bool) – Use checkpoint or not. Using checkpoint will save some memory while slowing down the training speed.
zero_init_residual (bool) – Whether to use zero init for last norm layer in resblocks to let them behave as identity.
pretrained (str, optional) – model pretrained path. Default: None
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None
Example
>>> from mmdet.models import Res2Net >>> import torch >>> self = Res2Net(depth=50, scales=4, base_width=26) >>> self.eval() >>> inputs = torch.rand(1, 3, 32, 32) >>> level_outputs = self.forward(inputs) >>> for level_out in level_outputs: ... print(tuple(level_out.shape)) (1, 256, 8, 8) (1, 512, 4, 4) (1, 1024, 2, 2) (1, 2048, 1, 1)
- class mmdet.models.backbones.ResNeSt(groups=1, base_width=4, radix=2, reduction_factor=4, avg_down_stride=True, **kwargs)[source]¶
ResNeSt backbone.
- Parameters
groups (int) – Number of groups of Bottleneck. Default: 1
base_width (int) – Base width of Bottleneck. Default: 4
radix (int) – Radix of SplitAttentionConv2d. Default: 2
reduction_factor (int) – Reduction factor of inter_channels in SplitAttentionConv2d. Default: 4.
avg_down_stride (bool) – Whether to use average pool for stride in Bottleneck. Default: True.
kwargs (dict) – Keyword arguments for ResNet.
- class mmdet.models.backbones.ResNeXt(groups=1, base_width=4, **kwargs)[source]¶
ResNeXt backbone.
- Parameters
depth (int) – Depth of resnet, from {18, 34, 50, 101, 152}.
in_channels (int) – Number of input image channels. Default: 3.
num_stages (int) – Resnet stages. Default: 4.
groups (int) – Group of resnext.
base_width (int) – Base width of resnext.
strides (Sequence[int]) – Strides of the first block of each stage.
dilations (Sequence[int]) – Dilation of each stage.
out_indices (Sequence[int]) – Output from which stages.
style (str) – pytorch or caffe. If set to “pytorch”, the stride-two layer is the 3x3 conv layer, otherwise the stride-two layer is the first 1x1 conv layer.
frozen_stages (int) – Stages to be frozen (all param fixed). -1 means not freezing any parameters.
norm_cfg (dict) – dictionary to construct and config norm layer.
norm_eval (bool) – Whether to set norm layers to eval mode, namely, freeze running stats (mean and var). Note: Effect on Batch Norm and its variants only.
with_cp (bool) – Use checkpoint or not. Using checkpoint will save some memory while slowing down the training speed.
zero_init_residual (bool) – whether to use zero init for last norm layer in resblocks to let them behave as identity.
- class mmdet.models.backbones.ResNet(depth, in_channels=3, stem_channels=None, base_channels=64, num_stages=4, strides=(1, 2, 2, 2), dilations=(1, 1, 1, 1), out_indices=(0, 1, 2, 3), style='pytorch', deep_stem=False, avg_down=False, frozen_stages=-1, conv_cfg=None, norm_cfg={'requires_grad': True, 'type': 'BN'}, norm_eval=True, dcn=None, stage_with_dcn=(False, False, False, False), plugins=None, with_cp=False, zero_init_residual=True, pretrained=None, init_cfg=None)[source]¶
ResNet backbone.
- Parameters
depth (int) – Depth of resnet, from {18, 34, 50, 101, 152}.
stem_channels (int | None) – Number of stem channels. If not specified, it will be the same as base_channels. Default: None.
base_channels (int) – Number of base channels of res layer. Default: 64.
in_channels (int) – Number of input image channels. Default: 3.
num_stages (int) – Resnet stages. Default: 4.
strides (Sequence[int]) – Strides of the first block of each stage.
dilations (Sequence[int]) – Dilation of each stage.
out_indices (Sequence[int]) – Output from which stages.
style (str) – pytorch or caffe. If set to “pytorch”, the stride-two layer is the 3x3 conv layer, otherwise the stride-two layer is the first 1x1 conv layer.
deep_stem (bool) – Replace 7x7 conv in input stem with 3 3x3 conv
avg_down (bool) – Use AvgPool instead of stride conv when downsampling in the bottleneck.
frozen_stages (int) – Stages to be frozen (stop grad and set eval mode). -1 means not freezing any parameters.
norm_cfg (dict) – Dictionary to construct and config norm layer.
norm_eval (bool) – Whether to set norm layers to eval mode, namely, freeze running stats (mean and var). Note: Effect on Batch Norm and its variants only.
plugins (list[dict]) –
List of plugins for stages, each dict contains:
cfg (dict, required): Cfg dict to build plugin.
position (str, required): Position inside block to insert plugin, options are ‘after_conv1’, ‘after_conv2’, ‘after_conv3’.
stages (tuple[bool], optional): Stages to apply plugin, length should be same as ‘num_stages’.
with_cp (bool) – Use checkpoint or not. Using checkpoint will save some memory while slowing down the training speed.
zero_init_residual (bool) – Whether to use zero init for last norm layer in resblocks to let them behave as identity.
pretrained (str, optional) – model pretrained path. Default: None
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None
Example
>>> from mmdet.models import ResNet >>> import torch >>> self = ResNet(depth=18) >>> self.eval() >>> inputs = torch.rand(1, 3, 32, 32) >>> level_outputs = self.forward(inputs) >>> for level_out in level_outputs: ... print(tuple(level_out.shape)) (1, 64, 8, 8) (1, 128, 4, 4) (1, 256, 2, 2) (1, 512, 1, 1)
- make_stage_plugins(plugins, stage_idx)[source]¶
Make plugins for ResNet
stage_idxth stage.Currently we support to insert
context_block,empirical_attention_block,nonlocal_blockinto the backbone like ResNet/ResNeXt. They could be inserted after conv1/conv2/conv3 of Bottleneck.An example of plugins format could be:
Examples
>>> plugins=[ ... dict(cfg=dict(type='xxx', arg1='xxx'), ... stages=(False, True, True, True), ... position='after_conv2'), ... dict(cfg=dict(type='yyy'), ... stages=(True, True, True, True), ... position='after_conv3'), ... dict(cfg=dict(type='zzz', postfix='1'), ... stages=(True, True, True, True), ... position='after_conv3'), ... dict(cfg=dict(type='zzz', postfix='2'), ... stages=(True, True, True, True), ... position='after_conv3') ... ] >>> self = ResNet(depth=18) >>> stage_plugins = self.make_stage_plugins(plugins, 0) >>> assert len(stage_plugins) == 3
Suppose
stage_idx=0, the structure of blocks in the stage would be:conv1-> conv2->conv3->yyy->zzz1->zzz2
Suppose ‘stage_idx=1’, the structure of blocks in the stage would be:
conv1-> conv2->xxx->conv3->yyy->zzz1->zzz2
If stages is missing, the plugin would be applied to all stages.
- Parameters
plugins (list[dict]) – List of plugins cfg to build. The postfix is required if multiple same type plugins are inserted.
stage_idx (int) – Index of stage to build
- Returns
Plugins for current stage
- Return type
list[dict]
- property norm1¶
the normalization layer named “norm1”
- Type
nn.Module
- class mmdet.models.backbones.ResNetV1d(**kwargs)[source]¶
ResNetV1d variant described in Bag of Tricks.
Compared with default ResNet(ResNetV1b), ResNetV1d replaces the 7x7 conv in the input stem with three 3x3 convs. And in the downsampling block, a 2x2 avg_pool with stride 2 is added before conv, whose stride is changed to 1.
- class mmdet.models.backbones.SSDVGG(depth, with_last_pool=False, ceil_mode=True, out_indices=(3, 4), out_feature_indices=(22, 34), pretrained=None, init_cfg=None, input_size=None, l2_norm_scale=None)[source]¶
VGG Backbone network for single-shot-detection.
- Parameters
depth (int) – Depth of vgg, from {11, 13, 16, 19}.
with_last_pool (bool) – Whether to add a pooling layer at the last of the model
ceil_mode (bool) – When True, will use ceil instead of floor to compute the output shape.
out_indices (Sequence[int]) – Output from which stages.
out_feature_indices (Sequence[int]) – Output from which feature map.
pretrained (str, optional) – model pretrained path. Default: None
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None
input_size (int, optional) – Deprecated argumment. Width and height of input, from {300, 512}.
l2_norm_scale (float, optional) – Deprecated argumment. L2 normalization layer init scale.
Example
>>> self = SSDVGG(input_size=300, depth=11) >>> self.eval() >>> inputs = torch.rand(1, 3, 300, 300) >>> level_outputs = self.forward(inputs) >>> for level_out in level_outputs: ... print(tuple(level_out.shape)) (1, 1024, 19, 19) (1, 512, 10, 10) (1, 256, 5, 5) (1, 256, 3, 3) (1, 256, 1, 1)
- class mmdet.models.backbones.SwinTransformer(pretrain_img_size=224, in_channels=3, embed_dims=96, patch_size=4, window_size=7, mlp_ratio=4, depths=(2, 2, 6, 2), num_heads=(3, 6, 12, 24), strides=(4, 2, 2, 2), out_indices=(0, 1, 2, 3), qkv_bias=True, qk_scale=None, patch_norm=True, drop_rate=0.0, attn_drop_rate=0.0, drop_path_rate=0.1, use_abs_pos_embed=False, act_cfg={'type': 'GELU'}, norm_cfg={'type': 'LN'}, with_cp=False, pretrained=None, convert_weights=False, frozen_stages=-1, init_cfg=None)[source]¶
Swin Transformer A PyTorch implement of : Swin Transformer: Hierarchical Vision Transformer using Shifted Windows -
Inspiration from https://github.com/microsoft/Swin-Transformer
- Parameters
pretrain_img_size (int | tuple[int]) – The size of input image when pretrain. Defaults: 224.
in_channels (int) – The num of input channels. Defaults: 3.
embed_dims (int) – The feature dimension. Default: 96.
patch_size (int | tuple[int]) – Patch size. Default: 4.
window_size (int) – Window size. Default: 7.
mlp_ratio (int) – Ratio of mlp hidden dim to embedding dim. Default: 4.
depths (tuple[int]) – Depths of each Swin Transformer stage. Default: (2, 2, 6, 2).
num_heads (tuple[int]) – Parallel attention heads of each Swin Transformer stage. Default: (3, 6, 12, 24).
strides (tuple[int]) – The patch merging or patch embedding stride of each Swin Transformer stage. (In swin, we set kernel size equal to stride.) Default: (4, 2, 2, 2).
out_indices (tuple[int]) – Output from which stages. Default: (0, 1, 2, 3).
qkv_bias (bool, optional) – If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None, optional) – Override default qk scale of head_dim ** -0.5 if set. Default: None.
patch_norm (bool) – If add a norm layer for patch embed and patch merging. Default: True.
drop_rate (float) – Dropout rate. Defaults: 0.
attn_drop_rate (float) – Attention dropout rate. Default: 0.
drop_path_rate (float) – Stochastic depth rate. Defaults: 0.1.
use_abs_pos_embed (bool) – If True, add absolute position embedding to the patch embedding. Defaults: False.
act_cfg (dict) – Config dict for activation layer. Default: dict(type=’GELU’).
norm_cfg (dict) – Config dict for normalization layer at output of backone. Defaults: dict(type=’LN’).
with_cp (bool, optional) – Use checkpoint or not. Using checkpoint will save some memory while slowing down the training speed. Default: False.
pretrained (str, optional) – model pretrained path. Default: None.
convert_weights (bool) – The flag indicates whether the pre-trained model is from the original repo. We may need to convert some keys to make it compatible. Default: False.
frozen_stages (int) – Stages to be frozen (stop grad and set eval mode). Default: -1 (-1 means not freezing any parameters).
init_cfg (dict, optional) – The Config for initialization. Defaults to None.
- forward(x)[source]¶
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- class mmdet.models.backbones.TridentResNet(depth, num_branch, test_branch_idx, trident_dilations, **kwargs)[source]¶
The stem layer, stage 1 and stage 2 in Trident ResNet are identical to ResNet, while in stage 3, Trident BottleBlock is utilized to replace the normal BottleBlock to yield trident output. Different branch shares the convolution weight but uses different dilations to achieve multi-scale output.
/ stage3(b0) x - stem - stage1 - stage2 - stage3(b1) - output stage3(b2) /
- Parameters
depth (int) – Depth of resnet, from {50, 101, 152}.
num_branch (int) – Number of branches in TridentNet.
test_branch_idx (int) – In inference, all 3 branches will be used if test_branch_idx==-1, otherwise only branch with index test_branch_idx will be used.
trident_dilations (tuple[int]) – Dilations of different trident branch. len(trident_dilations) should be equal to num_branch.
necks¶
- class mmdet.models.necks.BFP(Balanced Feature Pyramids)[source]¶
BFP takes multi-level features as inputs and gather them into a single one, then refine the gathered feature and scatter the refined results to multi-level features. This module is used in Libra R-CNN (CVPR 2019), see the paper Libra R-CNN: Towards Balanced Learning for Object Detection for details.
- Parameters
in_channels (int) – Number of input channels (feature maps of all levels should have the same channels).
num_levels (int) – Number of input feature levels.
conv_cfg (dict) – The config dict for convolution layers.
norm_cfg (dict) – The config dict for normalization layers.
refine_level (int) – Index of integration and refine level of BSF in multi-level features from bottom to top.
refine_type (str) – Type of the refine op, currently support [None, ‘conv’, ‘non_local’].
init_cfg (dict or list[dict], optional) – Initialization config dict.
- class mmdet.models.necks.CTResNetNeck(in_channel, num_deconv_filters, num_deconv_kernels, use_dcn=True, init_cfg=None)[source]¶
The neck used in CenterNet for object classification and box regression.
- Parameters
in_channel (int) – Number of input channels.
num_deconv_filters (tuple[int]) – Number of filters per stage.
num_deconv_kernels (tuple[int]) – Number of kernels per stage.
use_dcn (bool) – If True, use DCNv2. Default: True.
init_cfg (dict or list[dict], optional) – Initialization config dict.
- forward(inputs)[source]¶
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- class mmdet.models.necks.ChannelMapper(in_channels, out_channels, kernel_size=3, conv_cfg=None, norm_cfg=None, act_cfg={'type': 'ReLU'}, num_outs=None, init_cfg={'distribution': 'uniform', 'layer': 'Conv2d', 'type': 'Xavier'})[source]¶
Channel Mapper to reduce/increase channels of backbone features.
This is used to reduce/increase channels of backbone features.
- Parameters
in_channels (List[int]) – Number of input channels per scale.
out_channels (int) – Number of output channels (used at each scale).
kernel_size (int, optional) – kernel_size for reducing channels (used at each scale). Default: 3.
conv_cfg (dict, optional) – Config dict for convolution layer. Default: None.
norm_cfg (dict, optional) – Config dict for normalization layer. Default: None.
act_cfg (dict, optional) – Config dict for activation layer in ConvModule. Default: dict(type=’ReLU’).
num_outs (int, optional) – Number of output feature maps. There would be extra_convs when num_outs larger than the length of in_channels.
init_cfg (dict or list[dict], optional) – Initialization config dict.
Example
>>> import torch >>> in_channels = [2, 3, 5, 7] >>> scales = [340, 170, 84, 43] >>> inputs = [torch.rand(1, c, s, s) ... for c, s in zip(in_channels, scales)] >>> self = ChannelMapper(in_channels, 11, 3).eval() >>> outputs = self.forward(inputs) >>> for i in range(len(outputs)): ... print(f'outputs[{i}].shape = {outputs[i].shape}') outputs[0].shape = torch.Size([1, 11, 340, 340]) outputs[1].shape = torch.Size([1, 11, 170, 170]) outputs[2].shape = torch.Size([1, 11, 84, 84]) outputs[3].shape = torch.Size([1, 11, 43, 43])
- class mmdet.models.necks.DilatedEncoder(in_channels, out_channels, block_mid_channels, num_residual_blocks, block_dilations)[source]¶
Dilated Encoder for YOLOF <https://arxiv.org/abs/2103.09460>`.
- This module contains two types of components:
- the original FPN lateral convolution layer and fpn convolution layer,
which are 1x1 conv + 3x3 conv
the dilated residual block
- Parameters
in_channels (int) – The number of input channels.
out_channels (int) – The number of output channels.
block_mid_channels (int) – The number of middle block output channels
num_residual_blocks (int) – The number of residual blocks.
block_dilations (list) – The list of residual blocks dilation.
- forward(feature)[source]¶
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- class mmdet.models.necks.DyHead(in_channels, out_channels, num_blocks=6, zero_init_offset=True, init_cfg=None)[source]¶
DyHead neck consisting of multiple DyHead Blocks.
See Dynamic Head: Unifying Object Detection Heads with Attentions for details.
- Parameters
in_channels (int) – Number of input channels.
out_channels (int) – Number of output channels.
num_blocks (int, optional) – Number of DyHead Blocks. Default: 6.
zero_init_offset (bool, optional) – Whether to use zero init for spatial_conv_offset. Default: True.
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None.
- class mmdet.models.necks.FPG(in_channels, out_channels, num_outs, stack_times, paths, inter_channels=None, same_down_trans=None, same_up_trans={'kernel_size': 3, 'padding': 1, 'stride': 2, 'type': 'conv'}, across_lateral_trans={'kernel_size': 1, 'type': 'conv'}, across_down_trans={'kernel_size': 3, 'type': 'conv'}, across_up_trans=None, across_skip_trans={'type': 'identity'}, output_trans={'kernel_size': 3, 'type': 'last_conv'}, start_level=0, end_level=-1, add_extra_convs=False, norm_cfg=None, skip_inds=None, init_cfg=[{'type': 'Caffe2Xavier', 'layer': 'Conv2d'}, {'type': 'Constant', 'layer': ['_BatchNorm', '_InstanceNorm', 'GroupNorm', 'LayerNorm'], 'val': 1.0}])[source]¶
FPG.
Implementation of Feature Pyramid Grids (FPG). This implementation only gives the basic structure stated in the paper. But users can implement different type of transitions to fully explore the the potential power of the structure of FPG.
- Parameters
in_channels (int) – Number of input channels (feature maps of all levels should have the same channels).
out_channels (int) – Number of output channels (used at each scale)
num_outs (int) – Number of output scales.
stack_times (int) – The number of times the pyramid architecture will be stacked.
paths (list[str]) – Specify the path order of each stack level. Each element in the list should be either ‘bu’ (bottom-up) or ‘td’ (top-down).
inter_channels (int) – Number of inter channels.
same_up_trans (dict) – Transition that goes down at the same stage.
same_down_trans (dict) – Transition that goes up at the same stage.
across_lateral_trans (dict) – Across-pathway same-stage
across_down_trans (dict) – Across-pathway bottom-up connection.
across_up_trans (dict) – Across-pathway top-down connection.
across_skip_trans (dict) – Across-pathway skip connection.
output_trans (dict) – Transition that trans the output of the last stage.
start_level (int) – Index of the start input backbone level used to build the feature pyramid. Default: 0.
end_level (int) – Index of the end input backbone level (exclusive) to build the feature pyramid. Default: -1, which means the last level.
add_extra_convs (bool) – It decides whether to add conv layers on top of the original feature maps. Default to False. If True, its actual mode is specified by extra_convs_on_inputs.
norm_cfg (dict) – Config dict for normalization layer. Default: None.
init_cfg (dict or list[dict], optional) – Initialization config dict.
- forward(inputs)[source]¶
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- class mmdet.models.necks.FPN(in_channels, out_channels, num_outs, start_level=0, end_level=-1, add_extra_convs=False, relu_before_extra_convs=False, no_norm_on_lateral=False, conv_cfg=None, norm_cfg=None, act_cfg=None, upsample_cfg={'mode': 'nearest'}, init_cfg={'distribution': 'uniform', 'layer': 'Conv2d', 'type': 'Xavier'})[source]¶
Feature Pyramid Network.
This is an implementation of paper Feature Pyramid Networks for Object Detection.
- Parameters
in_channels (list[int]) – Number of input channels per scale.
out_channels (int) – Number of output channels (used at each scale).
num_outs (int) – Number of output scales.
start_level (int) – Index of the start input backbone level used to build the feature pyramid. Default: 0.
end_level (int) – Index of the end input backbone level (exclusive) to build the feature pyramid. Default: -1, which means the last level.
add_extra_convs (bool | str) –
If bool, it decides whether to add conv layers on top of the original feature maps. Default to False. If True, it is equivalent to add_extra_convs=’on_input’. If str, it specifies the source feature map of the extra convs. Only the following options are allowed
’on_input’: Last feat map of neck inputs (i.e. backbone feature).
’on_lateral’: Last feature map after lateral convs.
’on_output’: The last output feature map after fpn convs.
relu_before_extra_convs (bool) – Whether to apply relu before the extra conv. Default: False.
no_norm_on_lateral (bool) – Whether to apply norm on lateral. Default: False.
conv_cfg (dict) – Config dict for convolution layer. Default: None.
norm_cfg (dict) – Config dict for normalization layer. Default: None.
act_cfg (dict) – Config dict for activation layer in ConvModule. Default: None.
upsample_cfg (dict) – Config dict for interpolate layer. Default: dict(mode=’nearest’).
init_cfg (dict or list[dict], optional) – Initialization config dict.
Example
>>> import torch >>> in_channels = [2, 3, 5, 7] >>> scales = [340, 170, 84, 43] >>> inputs = [torch.rand(1, c, s, s) ... for c, s in zip(in_channels, scales)] >>> self = FPN(in_channels, 11, len(in_channels)).eval() >>> outputs = self.forward(inputs) >>> for i in range(len(outputs)): ... print(f'outputs[{i}].shape = {outputs[i].shape}') outputs[0].shape = torch.Size([1, 11, 340, 340]) outputs[1].shape = torch.Size([1, 11, 170, 170]) outputs[2].shape = torch.Size([1, 11, 84, 84]) outputs[3].shape = torch.Size([1, 11, 43, 43])
- class mmdet.models.necks.FPN_CARAFE(in_channels, out_channels, num_outs, start_level=0, end_level=-1, norm_cfg=None, act_cfg=None, order=('conv', 'norm', 'act'), upsample_cfg={'encoder_dilation': 1, 'encoder_kernel': 3, 'type': 'carafe', 'up_group': 1, 'up_kernel': 5}, init_cfg=None)[source]¶
FPN_CARAFE is a more flexible implementation of FPN. It allows more choice for upsample methods during the top-down pathway.
It can reproduce the performance of ICCV 2019 paper CARAFE: Content-Aware ReAssembly of FEatures Please refer to https://arxiv.org/abs/1905.02188 for more details.
- Parameters
in_channels (list[int]) – Number of channels for each input feature map.
out_channels (int) – Output channels of feature pyramids.
num_outs (int) – Number of output stages.
start_level (int) – Start level of feature pyramids. (Default: 0)
end_level (int) – End level of feature pyramids. (Default: -1 indicates the last level).
norm_cfg (dict) – Dictionary to construct and config norm layer.
activate (str) – Type of activation function in ConvModule (Default: None indicates w/o activation).
order (dict) – Order of components in ConvModule.
upsample (str) – Type of upsample layer.
upsample_cfg (dict) – Dictionary to construct and config upsample layer.
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None
- class mmdet.models.necks.HRFPN(High Resolution Feature Pyramids)[source]¶
paper: High-Resolution Representations for Labeling Pixels and Regions.
- Parameters
in_channels (list) – number of channels for each branch.
out_channels (int) – output channels of feature pyramids.
num_outs (int) – number of output stages.
pooling_type (str) – pooling for generating feature pyramids from {MAX, AVG}.
conv_cfg (dict) – dictionary to construct and config conv layer.
norm_cfg (dict) – dictionary to construct and config norm layer.
with_cp (bool) – Use checkpoint or not. Using checkpoint will save some memory while slowing down the training speed.
stride (int) – stride of 3x3 convolutional layers
init_cfg (dict or list[dict], optional) – Initialization config dict.
- class mmdet.models.necks.NASFCOS_FPN(in_channels, out_channels, num_outs, start_level=1, end_level=-1, add_extra_convs=False, conv_cfg=None, norm_cfg=None, init_cfg=None)[source]¶
FPN structure in NASFPN.
Implementation of paper NAS-FCOS: Fast Neural Architecture Search for Object Detection
- Parameters
in_channels (List[int]) – Number of input channels per scale.
out_channels (int) – Number of output channels (used at each scale)
num_outs (int) – Number of output scales.
start_level (int) – Index of the start input backbone level used to build the feature pyramid. Default: 0.
end_level (int) – Index of the end input backbone level (exclusive) to build the feature pyramid. Default: -1, which means the last level.
add_extra_convs (bool) – It decides whether to add conv layers on top of the original feature maps. Default to False. If True, its actual mode is specified by extra_convs_on_inputs.
conv_cfg (dict) – dictionary to construct and config conv layer.
norm_cfg (dict) – dictionary to construct and config norm layer.
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None
- class mmdet.models.necks.NASFPN(in_channels, out_channels, num_outs, stack_times, start_level=0, end_level=-1, add_extra_convs=False, norm_cfg=None, init_cfg={'layer': 'Conv2d', 'type': 'Caffe2Xavier'})[source]¶
NAS-FPN.
Implementation of NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Detection
- Parameters
in_channels (List[int]) – Number of input channels per scale.
out_channels (int) – Number of output channels (used at each scale)
num_outs (int) – Number of output scales.
stack_times (int) – The number of times the pyramid architecture will be stacked.
start_level (int) – Index of the start input backbone level used to build the feature pyramid. Default: 0.
end_level (int) – Index of the end input backbone level (exclusive) to build the feature pyramid. Default: -1, which means the last level.
add_extra_convs (bool) – It decides whether to add conv layers on top of the original feature maps. Default to False. If True, its actual mode is specified by extra_convs_on_inputs.
init_cfg (dict or list[dict], optional) – Initialization config dict.
- class mmdet.models.necks.PAFPN(in_channels, out_channels, num_outs, start_level=0, end_level=-1, add_extra_convs=False, relu_before_extra_convs=False, no_norm_on_lateral=False, conv_cfg=None, norm_cfg=None, act_cfg=None, init_cfg={'distribution': 'uniform', 'layer': 'Conv2d', 'type': 'Xavier'})[source]¶
Path Aggregation Network for Instance Segmentation.
This is an implementation of the PAFPN in Path Aggregation Network.
- Parameters
in_channels (List[int]) – Number of input channels per scale.
out_channels (int) – Number of output channels (used at each scale)
num_outs (int) – Number of output scales.
start_level (int) – Index of the start input backbone level used to build the feature pyramid. Default: 0.
end_level (int) – Index of the end input backbone level (exclusive) to build the feature pyramid. Default: -1, which means the last level.
add_extra_convs (bool | str) –
If bool, it decides whether to add conv layers on top of the original feature maps. Default to False. If True, it is equivalent to add_extra_convs=’on_input’. If str, it specifies the source feature map of the extra convs. Only the following options are allowed
’on_input’: Last feat map of neck inputs (i.e. backbone feature).
’on_lateral’: Last feature map after lateral convs.
’on_output’: The last output feature map after fpn convs.
relu_before_extra_convs (bool) – Whether to apply relu before the extra conv. Default: False.
no_norm_on_lateral (bool) – Whether to apply norm on lateral. Default: False.
conv_cfg (dict) – Config dict for convolution layer. Default: None.
norm_cfg (dict) – Config dict for normalization layer. Default: None.
act_cfg (str) – Config dict for activation layer in ConvModule. Default: None.
init_cfg (dict or list[dict], optional) – Initialization config dict.
- class mmdet.models.necks.RFP(Recursive Feature Pyramid)[source]¶
This is an implementation of RFP in DetectoRS. Different from standard FPN, the input of RFP should be multi level features along with origin input image of backbone.
- Parameters
rfp_steps (int) – Number of unrolled steps of RFP.
rfp_backbone (dict) – Configuration of the backbone for RFP.
aspp_out_channels (int) – Number of output channels of ASPP module.
aspp_dilations (tuple[int]) – Dilation rates of four branches. Default: (1, 3, 6, 1)
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None
- class mmdet.models.necks.SSDNeck(in_channels, out_channels, level_strides, level_paddings, l2_norm_scale=20.0, last_kernel_size=3, use_depthwise=False, conv_cfg=None, norm_cfg=None, act_cfg={'type': 'ReLU'}, init_cfg=[{'type': 'Xavier', 'distribution': 'uniform', 'layer': 'Conv2d'}, {'type': 'Constant', 'val': 1, 'layer': 'BatchNorm2d'}])[source]¶
Extra layers of SSD backbone to generate multi-scale feature maps.
- Parameters
in_channels (Sequence[int]) – Number of input channels per scale.
out_channels (Sequence[int]) – Number of output channels per scale.
level_strides (Sequence[int]) – Stride of 3x3 conv per level.
level_paddings (Sequence[int]) – Padding size of 3x3 conv per level.
l2_norm_scale (float|None) – L2 normalization layer init scale. If None, not use L2 normalization on the first input feature.
last_kernel_size (int) – Kernel size of the last conv layer. Default: 3.
use_depthwise (bool) – Whether to use DepthwiseSeparableConv. Default: False.
conv_cfg (dict) – Config dict for convolution layer. Default: None.
norm_cfg (dict) – Dictionary to construct and config norm layer. Default: None.
act_cfg (dict) – Config dict for activation layer. Default: dict(type=’ReLU’).
init_cfg (dict or list[dict], optional) – Initialization config dict.
- class mmdet.models.necks.YOLOV3Neck(num_scales, in_channels, out_channels, conv_cfg=None, norm_cfg={'requires_grad': True, 'type': 'BN'}, act_cfg={'negative_slope': 0.1, 'type': 'LeakyReLU'}, init_cfg=None)[source]¶
The neck of YOLOV3.
It can be treated as a simplified version of FPN. It will take the result from Darknet backbone and do some upsampling and concatenation. It will finally output the detection result.
Note
- The input feats should be from top to bottom.
i.e., from high-lvl to low-lvl
- But YOLOV3Neck will process them in reversed order.
i.e., from bottom (high-lvl) to top (low-lvl)
- Parameters
num_scales (int) – The number of scales / stages.
in_channels (List[int]) – The number of input channels per scale.
out_channels (List[int]) – The number of output channels per scale.
conv_cfg (dict, optional) – Config dict for convolution layer. Default: None.
norm_cfg (dict, optional) – Dictionary to construct and config norm layer. Default: dict(type=’BN’, requires_grad=True)
act_cfg (dict, optional) – Config dict for activation layer. Default: dict(type=’LeakyReLU’, negative_slope=0.1).
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None
- forward(feats)[source]¶
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- class mmdet.models.necks.YOLOXPAFPN(in_channels, out_channels, num_csp_blocks=3, use_depthwise=False, upsample_cfg={'mode': 'nearest', 'scale_factor': 2}, conv_cfg=None, norm_cfg={'eps': 0.001, 'momentum': 0.03, 'type': 'BN'}, act_cfg={'type': 'Swish'}, init_cfg={'a': 2.23606797749979, 'distribution': 'uniform', 'layer': 'Conv2d', 'mode': 'fan_in', 'nonlinearity': 'leaky_relu', 'type': 'Kaiming'})[source]¶
Path Aggregation Network used in YOLOX.
- Parameters
in_channels (List[int]) – Number of input channels per scale.
out_channels (int) – Number of output channels (used at each scale)
num_csp_blocks (int) – Number of bottlenecks in CSPLayer. Default: 3
use_depthwise (bool) – Whether to depthwise separable convolution in blocks. Default: False
upsample_cfg (dict) – Config dict for interpolate layer. Default: dict(scale_factor=2, mode=’nearest’)
conv_cfg (dict, optional) – Config dict for convolution layer. Default: None, which means using conv2d.
norm_cfg (dict) – Config dict for normalization layer. Default: dict(type=’BN’)
act_cfg (dict) – Config dict for activation layer. Default: dict(type=’Swish’)
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None.
dense_heads¶
roi_heads¶
losses¶
utils¶
- class mmdet.models.utils.AdaptiveAvgPool2d(output_size: Union[int, None, Tuple[Optional[int], ...]])[source]¶
Handle empty batch dimension to AdaptiveAvgPool2d.
- forward(x)[source]¶
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- class mmdet.models.utils.CSPLayer(in_channels, out_channels, expand_ratio=0.5, num_blocks=1, add_identity=True, use_depthwise=False, conv_cfg=None, norm_cfg={'eps': 0.001, 'momentum': 0.03, 'type': 'BN'}, act_cfg={'type': 'Swish'}, init_cfg=None)[source]¶
Cross Stage Partial Layer.
- Parameters
in_channels (int) – The input channels of the CSP layer.
out_channels (int) – The output channels of the CSP layer.
expand_ratio (float) – Ratio to adjust the number of channels of the hidden layer. Default: 0.5
num_blocks (int) – Number of blocks. Default: 1
add_identity (bool) – Whether to add identity in blocks. Default: True
use_depthwise (bool) – Whether to depthwise separable convolution in blocks. Default: False
conv_cfg (dict, optional) – Config dict for convolution layer. Default: None, which means using conv2d.
norm_cfg (dict) – Config dict for normalization layer. Default: dict(type=’BN’)
act_cfg (dict) – Config dict for activation layer. Default: dict(type=’Swish’)
- forward(x)[source]¶
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- class mmdet.models.utils.ConvUpsample(in_channels, inner_channels, num_layers=1, num_upsample=None, conv_cfg=None, norm_cfg=None, init_cfg=None, **kwargs)[source]¶
ConvUpsample performs 2x upsampling after Conv.
There are several ConvModule layers. In the first few layers, upsampling will be applied after each layer of convolution. The number of upsampling must be no more than the number of ConvModule layers.
- Parameters
in_channels (int) – Number of channels in the input feature map.
inner_channels (int) – Number of channels produced by the convolution.
num_layers (int) – Number of convolution layers.
num_upsample (int | optional) – Number of upsampling layer. Must be no more than num_layers. Upsampling will be applied after the first
num_upsamplelayers of convolution. Default:num_layers.conv_cfg (dict) – Config dict for convolution layer. Default: None, which means using conv2d.
norm_cfg (dict) – Config dict for normalization layer. Default: None.
init_cfg (dict) – Config dict for initialization. Default: None.
kwargs (key word augments) – Other augments used in ConvModule.
- forward(x)[source]¶
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- class mmdet.models.utils.DetrTransformerDecoder(*args, post_norm_cfg={'type': 'LN'}, return_intermediate=False, **kwargs)[source]¶
Implements the decoder in DETR transformer.
- Parameters
return_intermediate (bool) – Whether to return intermediate outputs.
post_norm_cfg (dict) – Config of last normalization layer. Default： LN.
- forward(query, *args, **kwargs)[source]¶
Forward function for TransformerDecoder.
- Parameters
query (Tensor) – Input query with shape (num_query, bs, embed_dims).
- Returns
- Results with shape [1, num_query, bs, embed_dims] when
return_intermediate is False, otherwise it has shape [num_layers, num_query, bs, embed_dims].
- Return type
Tensor
- class mmdet.models.utils.DetrTransformerDecoderLayer(attn_cfgs, feedforward_channels, ffn_dropout=0.0, operation_order=None, act_cfg={'inplace': True, 'type': 'ReLU'}, norm_cfg={'type': 'LN'}, ffn_num_fcs=2, **kwargs)[source]¶
Implements decoder layer in DETR transformer.
- Parameters
attn_cfgs (list[mmcv.ConfigDict] | list[dict] | dict )) – Configs for self_attention or cross_attention, the order should be consistent with it in operation_order. If it is a dict, it would be expand to the number of attention in operation_order.
feedforward_channels (int) – The hidden dimension for FFNs.
ffn_dropout (float) – Probability of an element to be zeroed in ffn. Default 0.0.
operation_order (tuple[str]) – The execution order of operation in transformer. Such as (‘self_attn’, ‘norm’, ‘ffn’, ‘norm’). Default：None
act_cfg (dict) – The activation config for FFNs. Default: LN
norm_cfg (dict) – Config dict for normalization layer. Default: LN.
ffn_num_fcs (int) – The number of fully-connected layers in FFNs. Default：2.
- class mmdet.models.utils.DyReLU(channels, ratio=4, conv_cfg=None, act_cfg=({'type': 'ReLU'}, {'type': 'HSigmoid', 'bias': 3.0, 'divisor': 6.0}), init_cfg=None)[source]¶
Dynamic ReLU (DyReLU) module.
See Dynamic ReLU for details. Current implementation is specialized for task-aware attention in DyHead. HSigmoid arguments in default act_cfg follow DyHead official code. https://github.com/microsoft/DynamicHead/blob/master/dyhead/dyrelu.py
- Parameters
channels (int) – The input (and output) channels of DyReLU module.
ratio (int) – Squeeze ratio in Squeeze-and-Excitation-like module, the intermediate channel will be
int(channels/ratio). Default: 4.conv_cfg (None or dict) – Config dict for convolution layer. Default: None, which means using conv2d.
act_cfg (dict or Sequence[dict]) – Config dict for activation layer. If act_cfg is a dict, two activation layers will be configurated by this dict. If act_cfg is a sequence of dicts, the first activation layer will be configurated by the first dict and the second activation layer will be configurated by the second dict. Default: (dict(type=’ReLU’), dict(type=’HSigmoid’, bias=3.0, divisor=6.0))
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None
- class mmdet.models.utils.DynamicConv(in_channels=256, feat_channels=64, out_channels=None, input_feat_shape=7, with_proj=True, act_cfg={'inplace': True, 'type': 'ReLU'}, norm_cfg={'type': 'LN'}, init_cfg=None)[source]¶
Implements Dynamic Convolution.
This module generate parameters for each sample and use bmm to implement 1*1 convolution. Code is modified from the official github repo .
- Parameters
in_channels (int) – The input feature channel. Defaults to 256.
feat_channels (int) – The inner feature channel. Defaults to 64.
out_channels (int, optional) – The output feature channel. When not specified, it will be set to in_channels by default
input_feat_shape (int) – The shape of input feature. Defaults to 7.
with_proj (bool) – Project two-dimentional feature to one-dimentional feature. Default to True.
act_cfg (dict) – The activation config for DynamicConv.
norm_cfg (dict) – Config dict for normalization layer. Default layer normalization.
(obj (init_cfg) – mmcv.ConfigDict): The Config for initialization. Default: None.
- forward(param_feature, input_feature)[source]¶
Forward function for DynamicConv.
- Parameters
param_feature (Tensor) – The feature can be used to generate the parameter, has shape (num_all_proposals, in_channels).
input_feature (Tensor) – Feature that interact with parameters, has shape (num_all_proposals, in_channels, H, W).
- Returns
The output feature has shape (num_all_proposals, out_channels).
- Return type
Tensor
- class mmdet.models.utils.InvertedResidual(in_channels, out_channels, mid_channels, kernel_size=3, stride=1, se_cfg=None, with_expand_conv=True, conv_cfg=None, norm_cfg={'type': 'BN'}, act_cfg={'type': 'ReLU'}, drop_path_rate=0.0, with_cp=False, init_cfg=None)[source]¶
Inverted Residual Block.
- Parameters
in_channels (int) – The input channels of this Module.
out_channels (int) – The output channels of this Module.
mid_channels (int) – The input channels of the depthwise convolution.
kernel_size (int) – The kernel size of the depthwise convolution. Default: 3.
stride (int) – The stride of the depthwise convolution. Default: 1.
se_cfg (dict) – Config dict for se layer. Default: None, which means no se layer.
with_expand_conv (bool) – Use expand conv or not. If set False, mid_channels must be the same with in_channels. Default: True.
conv_cfg (dict) – Config dict for convolution layer. Default: None, which means using conv2d.
norm_cfg (dict) – Config dict for normalization layer. Default: dict(type=’BN’).
act_cfg (dict) – Config dict for activation layer. Default: dict(type=’ReLU’).
drop_path_rate (float) – stochastic depth rate. Defaults to 0.
with_cp (bool) – Use checkpoint or not. Using checkpoint will save some memory while slowing down the training speed. Default: False.
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None
- Returns
The output tensor.
- Return type
Tensor
- forward(x)[source]¶
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- class mmdet.models.utils.LearnedPositionalEncoding(num_feats, row_num_embed=50, col_num_embed=50, init_cfg={'layer': 'Embedding', 'type': 'Uniform'})[source]¶
Position embedding with learnable embedding weights.
- Parameters
num_feats (int) – The feature dimension for each position along x-axis or y-axis. The final returned dimension for each position is 2 times of this value.
row_num_embed (int, optional) – The dictionary size of row embeddings. Default 50.
col_num_embed (int, optional) – The dictionary size of col embeddings. Default 50.
init_cfg (dict or list[dict], optional) – Initialization config dict.
- forward(mask)[source]¶
Forward function for LearnedPositionalEncoding.
- Parameters
mask (Tensor) – ByteTensor mask. Non-zero values representing ignored positions, while zero values means valid positions for this image. Shape [bs, h, w].
- Returns
- Returned position embedding with shape
[bs, num_feats*2, h, w].
- Return type
pos (Tensor)
- class mmdet.models.utils.NormedConv2d(*args, tempearture=20, power=1.0, eps=1e-06, norm_over_kernel=False, **kwargs)[source]¶
Normalized Conv2d Layer.
- Parameters
tempeature (float, optional) – Tempeature term. Default to 20.
power (int, optional) – Power term. Default to 1.0.
eps (float, optional) – The minimal value of divisor to keep numerical stability. Default to 1e-6.
norm_over_kernel (bool, optional) – Normalize over kernel. Default to False.
- forward(x)[source]¶
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- class mmdet.models.utils.NormedLinear(*args, tempearture=20, power=1.0, eps=1e-06, **kwargs)[source]¶
Normalized Linear Layer.
- Parameters
tempeature (float, optional) – Tempeature term. Default to 20.
power (int, optional) – Power term. Default to 1.0.
eps (float, optional) – The minimal value of divisor to keep numerical stability. Default to 1e-6.
- forward(x)[source]¶
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- class mmdet.models.utils.PatchEmbed(in_channels=3, embed_dims=768, conv_type='Conv2d', kernel_size=16, stride=16, padding='corner', dilation=1, bias=True, norm_cfg=None, input_size=None, init_cfg=None)[source]¶
Image to Patch Embedding.
We use a conv layer to implement PatchEmbed.
- Parameters
in_channels (int) – The num of input channels. Default: 3
embed_dims (int) – The dimensions of embedding. Default: 768
conv_type (str) – The config dict for embedding conv layer type selection. Default: “Conv2d.
kernel_size (int) – The kernel_size of embedding conv. Default: 16.
stride (int) – The slide stride of embedding conv. Default: None (Would be set as kernel_size).
padding (int | tuple | string) – The padding length of embedding conv. When it is a string, it means the mode of adaptive padding, support “same” and “corner” now. Default: “corner”.
dilation (int) – The dilation rate of embedding conv. Default: 1.
bias (bool) – Bias of embed conv. Default: True.
norm_cfg (dict, optional) – Config dict for normalization layer. Default: None.
input_size (int | tuple | None) – The size of input, which will be used to calculate the out size. Only work when dynamic_size is False. Default: None.
init_cfg (mmcv.ConfigDict, optional) – The Config for initialization. Default: None.
- class mmdet.models.utils.ResLayer(block, inplanes, planes, num_blocks, stride=1, avg_down=False, conv_cfg=None, norm_cfg={'type': 'BN'}, downsample_first=True, **kwargs)[source]¶
ResLayer to build ResNet style backbone.
- Parameters
block (nn.Module) – block used to build ResLayer.
inplanes (int) – inplanes of block.
planes (int) – planes of block.
num_blocks (int) – number of blocks.
stride (int) – stride of the first block. Default: 1
avg_down (bool) – Use AvgPool instead of stride conv when downsampling in the bottleneck. Default: False
conv_cfg (dict) – dictionary to construct and config conv layer. Default: None
norm_cfg (dict) – dictionary to construct and config norm layer. Default: dict(type=’BN’)
downsample_first (bool) – Downsample at the first block or last block. False for Hourglass, True for ResNet. Default: True
- class mmdet.models.utils.SELayer(channels, ratio=16, conv_cfg=None, act_cfg=({'type': 'ReLU'}, {'type': 'Sigmoid'}), init_cfg=None)[source]¶
Squeeze-and-Excitation Module.
- Parameters
channels (int) – The input (and output) channels of the SE layer.
ratio (int) – Squeeze ratio in SELayer, the intermediate channel will be
int(channels/ratio). Default: 16.conv_cfg (None or dict) – Config dict for convolution layer. Default: None, which means using conv2d.
act_cfg (dict or Sequence[dict]) – Config dict for activation layer. If act_cfg is a dict, two activation layers will be configurated by this dict. If act_cfg is a sequence of dicts, the first activation layer will be configurated by the first dict and the second activation layer will be configurated by the second dict. Default: (dict(type=’ReLU’), dict(type=’Sigmoid’))
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None
- forward(x)[source]¶
Defines the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.
- class mmdet.models.utils.SimplifiedBasicBlock(inplanes, planes, stride=1, dilation=1, downsample=None, style='pytorch', with_cp=False, conv_cfg=None, norm_cfg={'type': 'BN'}, dcn=None, plugins=None, init_fg=None)[source]¶
Simplified version of original basic residual block. This is used in SCNet.
Norm layer is now optional
Last ReLU in forward function is removed
- property norm1¶
normalization layer after the first convolution layer
- Type
nn.Module
- property norm2¶
normalization layer after the second convolution layer
- Type
nn.Module
- class mmdet.models.utils.SinePositionalEncoding(num_feats, temperature=10000, normalize=False, scale=6.283185307179586, eps=1e-06, offset=0.0, init_cfg=None)[source]¶
Position encoding with sine and cosine functions.
See End-to-End Object Detection with Transformers for details.
- Parameters
num_feats (int) – The feature dimension for each position along x-axis or y-axis. Note the final returned dimension for each position is 2 times of this value.
temperature (int, optional) – The temperature used for scaling the position embedding. Defaults to 10000.
normalize (bool, optional) – Whether to normalize the position embedding. Defaults to False.
scale (float, optional) – A scale factor that scales the position embedding. The scale will be used only when normalize is True. Defaults to 2*pi.
eps (float, optional) – A value added to the denominator for numerical stability. Defaults to 1e-6.
offset (float) – offset add to embed when do the normalization. Defaults to 0.
init_cfg (dict or list[dict], optional) – Initialization config dict. Default: None
- forward(mask)[source]¶
Forward function for SinePositionalEncoding.
- Parameters
mask (Tensor) – ByteTensor mask. Non-zero values representing ignored positions, while zero values means valid positions for this image. Shape [bs, h, w].
- Returns
- Returned position embedding with shape
[bs, num_feats*2, h, w].
- Return type
pos (Tensor)
- class mmdet.models.utils.Transformer(encoder=None, decoder=None, init_cfg=None)[source]¶
Implements the DETR transformer.
Following the official DETR implementation, this module copy-paste from torch.nn.Transformer with modifications:
positional encodings are passed in MultiheadAttention
extra LN at the end of encoder is removed
decoder returns a stack of activations from all decoding layers
See paper: End-to-End Object Detection with Transformers for details.
- Parameters
encoder (mmcv.ConfigDict | Dict) – Config of TransformerEncoder. Defaults to None.
decoder ((mmcv.ConfigDict | Dict)) – Config of TransformerDecoder. Defaults to None
(obj (init_cfg) – mmcv.ConfigDict): The Config for initialization. Defaults to None.
- forward(x, mask, query_embed, pos_embed)[source]¶
Forward function for Transformer.
- Parameters
x (Tensor) – Input query with shape [bs, c, h, w] where c = embed_dims.
mask (Tensor) – The key_padding_mask used for encoder and decoder, with shape [bs, h, w].
query_embed (Tensor) – The query embedding for decoder, with shape [num_query, c].
pos_embed (Tensor) – The positional encoding for encoder and decoder, with the same shape as x.
- Returns
results of decoder containing the following tensor.
- out_dec: Output from decoder. If return_intermediate_dec is True output has shape [num_dec_layers, bs,
num_query, embed_dims], else has shape [1, bs, num_query, embed_dims].
memory: Output results from encoder, with shape [bs, embed_dims, h, w].
- Return type
tuple[Tensor]
- mmdet.models.utils.adaptive_avg_pool2d(input, output_size)[source]¶
Handle empty batch dimension to adaptive_avg_pool2d.
- Parameters
input (tensor) – 4D tensor.
output_size (int, tuple[int,int]) – the target output size.
- mmdet.models.utils.build_linear_layer(cfg, *args, **kwargs)[source]¶
Build linear layer. :param cfg: The linear layer config, which should contain:
type (str): Layer type.
layer args: Args needed to instantiate an linear layer.
- Parameters
args (argument list) – Arguments passed to the __init__ method of the corresponding linear layer.
kwargs (keyword arguments) – Keyword arguments passed to the __init__ method of the corresponding linear layer.
- Returns
Created linear layer.
- Return type
nn.Module
- mmdet.models.utils.gaussian_radius(det_size, min_overlap)[source]¶
Generate 2D gaussian radius.
This function is modified from the official github repo.
Given
min_overlap, radius could computed by a quadratic equation according to Vieta’s formulas.There are 3 cases for computing gaussian radius, details are following:
Explanation of figure:
ltandbrindicates the left-top and bottom-right corner of ground truth box.xindicates the generated corner at the limited position whenradius=r.Case1: one corner is inside the gt box and the other is outside.
|< width >| lt-+----------+ - | | | ^ +--x----------+--+ | | | | | | | | height | | overlap | | | | | | | | | | v +--+---------br--+ - | | | +----------+--x
To ensure IoU of generated box and gt box is larger than
min_overlap:\[\begin{split}\cfrac{(w-r)*(h-r)}{w*h+(w+h)r-r^2} \ge {iou} \quad\Rightarrow\quad {r^2-(w+h)r+\cfrac{1-iou}{1+iou}*w*h} \ge 0 \\ {a} = 1,\quad{b} = {-(w+h)},\quad{c} = {\cfrac{1-iou}{1+iou}*w*h} \\ {r} \le \cfrac{-b-\sqrt{b^2-4*a*c}}{2*a}\end{split}\]Case2: both two corners are inside the gt box.
|< width >| lt-+----------+ - | | | ^ +--x-------+ | | | | | | |overlap| | height | | | | | +-------x--+ | | | v +----------+-br -
To ensure IoU of generated box and gt box is larger than
min_overlap:\[\begin{split}\cfrac{(w-2*r)*(h-2*r)}{w*h} \ge {iou} \quad\Rightarrow\quad {4r^2-2(w+h)r+(1-iou)*w*h} \ge 0 \\ {a} = 4,\quad {b} = {-2(w+h)},\quad {c} = {(1-iou)*w*h} \\ {r} \le \cfrac{-b-\sqrt{b^2-4*a*c}}{2*a}\end{split}\]Case3: both two corners are outside the gt box.
|< width >| x--+----------------+ | | | +-lt-------------+ | - | | | | ^ | | | | | | overlap | | height | | | | | | | | v | +------------br--+ - | | | +----------------+--x
To ensure IoU of generated box and gt box is larger than
min_overlap:\[\begin{split}\cfrac{w*h}{(w+2*r)*(h+2*r)} \ge {iou} \quad\Rightarrow\quad {4*iou*r^2+2*iou*(w+h)r+(iou-1)*w*h} \le 0 \\ {a} = {4*iou},\quad {b} = {2*iou*(w+h)},\quad {c} = {(iou-1)*w*h} \\ {r} \le \cfrac{-b+\sqrt{b^2-4*a*c}}{2*a}\end{split}\]- Parameters
det_size (list[int]) – Shape of object.
min_overlap (float) – Min IoU with ground truth for boxes generated by keypoints inside the gaussian kernel.
- Returns
Radius of gaussian kernel.
- Return type
radius (int)
- mmdet.models.utils.gen_gaussian_target(heatmap, center, radius, k=1)[source]¶
Generate 2D gaussian heatmap.
- Parameters
heatmap (Tensor) – Input heatmap, the gaussian kernel will cover on it and maintain the max value.
center (list[int]) – Coord of gaussian kernel’s center.
radius (int) – Radius of gaussian kernel.
k (int) – Coefficient of gaussian kernel. Default: 1.
- Returns
Updated heatmap covered by gaussian kernel.
- Return type
out_heatmap (Tensor)
- mmdet.models.utils.get_uncertain_point_coords_with_randomness(mask_pred, labels, num_points, oversample_ratio, importance_sample_ratio)[source]¶
Get
num_pointsmost uncertain points with random points during train.Sample points in [0, 1] x [0, 1] coordinate space based on their uncertainty. The uncertainties are calculated for each point using ‘get_uncertainty()’ function that takes point’s logit prediction as input.
- Parameters
mask_pred (Tensor) – A tensor of shape (num_rois, num_classes, mask_height, mask_width) for class-specific or class-agnostic prediction.
labels (list) – The ground truth class for each instance.
num_points (int) – The number of points to sample.
oversample_ratio (int) – Oversampling parameter.
importance_sample_ratio (float) – Ratio of points that are sampled via importnace sampling.
- Returns
- A tensor of shape (num_rois, num_points, 2)
that contains the coordinates sampled points.
- Return type
point_coords (Tensor)
- mmdet.models.utils.get_uncertainty(mask_pred, labels)[source]¶
Estimate uncertainty based on pred logits.
We estimate uncertainty as L1 distance between 0.0 and the logits prediction in ‘mask_pred’ for the foreground class in classes.
- Parameters
mask_pred (Tensor) – mask predication logits, shape (num_rois, num_classes, mask_height, mask_width).
labels (list[Tensor]) – Either predicted or ground truth label for each predicted mask, of length num_rois.
- Returns
- Uncertainty scores with the most uncertain
locations having the highest uncertainty score, shape (num_rois, 1, mask_height, mask_width)
- Return type
scores (Tensor)
- mmdet.models.utils.interpolate_as(source, target, mode='bilinear', align_corners=False)[source]¶
Interpolate the source to the shape of the target.
The source must be a Tensor, but the target can be a Tensor or a np.ndarray with the shape (…, target_h, target_w).
- Parameters
source (Tensor) – A 3D/4D Tensor with the shape (N, H, W) or (N, C, H, W).
target (Tensor | np.ndarray) – The interpolation target with the shape (…, target_h, target_w).
mode (str) – Algorithm used for interpolation. The options are the same as those in F.interpolate(). Default:
'bilinear'.align_corners (bool) – The same as the argument in F.interpolate().
- Returns
The interpolated source Tensor.
- Return type
Tensor
- mmdet.models.utils.make_divisible(value, divisor, min_value=None, min_ratio=0.9)[source]¶
Make divisible function.
This function rounds the channel number to the nearest value that can be divisible by the divisor. It is taken from the original tf repo. It ensures that all layers have a channel number that is divisible by divisor. It can be seen here: https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet.py # noqa
- Parameters
value (int) – The original channel number.
divisor (int) – The divisor to fully divide the channel number.
min_value (int) – The minimum value of the output channel. Default: None, means that the minimum value equal to the divisor.
min_ratio (float) – The minimum ratio of the rounded channel number to the original channel number. Default: 0.9.
- Returns
The modified output channel number.
- Return type
int
- mmdet.models.utils.nchw_to_nlc(x)[source]¶
Flatten [N, C, H, W] shape tensor to [N, L, C] shape tensor.
- Parameters
x (Tensor) – The input tensor of shape [N, C, H, W] before conversion.
- Returns
The output tensor of shape [N, L, C] after conversion.
- Return type
Tensor
- mmdet.models.utils.nlc_to_nchw(x, hw_shape)[source]¶
Convert [N, L, C] shape tensor to [N, C, H, W] shape tensor.
- Parameters
x (Tensor) – The input tensor of shape [N, L, C] before conversion.
hw_shape (Sequence[int]) – The height and width of output feature map.
- Returns
The output tensor of shape [N, C, H, W] after conversion.
- Return type
Tensor
- mmdet.models.utils.preprocess_panoptic_gt(gt_labels, gt_masks, gt_semantic_seg, num_things, num_stuff, img_metas)[source]¶
Preprocess the ground truth for a image.
- Parameters
gt_labels (Tensor) – Ground truth labels of each bbox, with shape (num_gts, ).
gt_masks (BitmapMasks) – Ground truth masks of each instances of a image, shape (num_gts, h, w).
gt_semantic_seg (Tensor | None) – Ground truth of semantic segmentation with the shape (1, h, w). [0, num_thing_class - 1] means things, [num_thing_class, num_class-1] means stuff, 255 means VOID. It’s None when training instance segmentation.
img_metas (dict) – List of image meta information.
- Returns
a tuple containing the following targets.
- labels (Tensor): Ground truth class indices for a
image, with shape (n, ), n is the sum of number of stuff type and number of instance in a image.
- masks (Tensor): Ground truth mask for a image, with
shape (n, h, w). Contains stuff and things when training panoptic segmentation, and things only when training instance segmentation.
- Return type
tuple
mmdet.utils¶
NPU (HUAWEI Ascend)¶
Usage¶
Please refer to the building documentation of MMCV to install MMCV on NPU devices
Here we use 8 NPUs on your computer to train the model with the following command:
bash tools/dist_train.sh configs/ssd/ssd300_coco.py 8
Also, you can use only one NPU to train the model with the following command:
python tools/train.py configs/ssd/ssd300_coco.py
Models Results¶
| Model | box AP | mask AP | Config | Download |
|---|---|---|---|---|
| ssd300 | 25.6 | --- | config | log |
| ssd512 | 29.4 | --- | config | log |
| ssdlite-mbv2* | 20.2 | --- | config | log |
| retinanet-r18 | 31.8 | --- | config | log |
| retinanet-r50 | 36.6 | --- | config | log |
| yolov3-608 | 34.7 | --- | config | log |
| yolox-s** | 39.9 | --- | config | log |
| centernet-r18 | 26.1 | --- | config | log |
| fcos-r50* | 36.1 | --- | config | log |
| solov2-r50 | --- | 34.7 | config | log |
Notes:
If not specially marked, the results on NPU are the same as those on the GPU with FP32.
(*) The results on the NPU of these models are aligned with the results of the mixed-precision training on the GPU, but are lower than the results of the FP32. This situation is mainly related to the phase of the model itself in mixed-precision training, users may need to adjust the hyperparameters to achieve better results.
(**) The accuracy of yolox-s on the GPU in mixed precision is 40.1, with
persister_woker=Truein the data loader config by default. There are currently some bugs on NPUs that prevent the last few epochs from running, but the accuracy is less affected and the difference can be ignored.
High-performance Model Result on Ascend Device¶
Introduction to optimization:
Modify the loop calculation as a whole batch calculation to reduce the number of instructions issued.
Modify the index calculation to mask calculation, because the SIMD architecture is good at processing continuous data calculation.
| Model | Config | v100 iter time | 910A iter time |
|---|---|---|---|
| ascend-ssd300 | config | 0.165s/iter | 0.383s/iter -> 0.13s/iter |
| ascend-retinanet-r18 | config | 0.567s/iter | 0.780s/iter -> 0.420s/iter |
All above models are provided by Huawei Ascend group.