Swin Transformer for Object Detection

This repo contains the supported code and configuration files to reproduce object detection results of Swin Transformer. It is based on mmdetection.

Updates

05/11/2021 Models for MoBY are released

04/12/2021 Initial commits

Results and Models

Mask R-CNN

Backbone	Pretrain	Lr Schd	box mAP	mask mAP	#params	FLOPs	config	log	model
Swin-T	ImageNet-1K	1x	43.7	39.8	48M	267G	config	github/baidu	github/baidu
Swin-T	ImageNet-1K	3x	46.0	41.6	48M	267G	config	github/baidu	github/baidu
Swin-S	ImageNet-1K	3x	48.5	43.3	69M	359G	config	github/baidu	github/baidu

Cascade Mask R-CNN

Backbone	Pretrain	Lr Schd	box mAP	mask mAP	#params	FLOPs	config	log	model
Swin-T	ImageNet-1K	1x	48.1	41.7	86M	745G	config	github/baidu	github/baidu
Swin-T	ImageNet-1K	3x	50.4	43.7	86M	745G	config	github/baidu	github/baidu
Swin-S	ImageNet-1K	3x	51.9	45.0	107M	838G	config	github/baidu	github/baidu
Swin-B	ImageNet-1K	3x	51.9	45.0	145M	982G	config	github/baidu	github/baidu

RepPoints V2

Backbone	Pretrain	Lr Schd	box mAP	mask mAP	#params	FLOPs
Swin-T	ImageNet-1K	3x	50.0	-	45M	283G

Mask RepPoints V2

Backbone	Pretrain	Lr Schd	box mAP	mask mAP	#params	FLOPs
Swin-T	ImageNet-1K	3x	50.3	43.6	47M	292G

Notes:

• Pre-trained models can be downloaded from Swin Transformer for ImageNet Classification.
• Access code for baidu is swin.

Results of MoBY with Swin Transformer

Mask R-CNN

Backbone	Pretrain	Lr Schd	box mAP	mask mAP	#params	FLOPs	config	log	model
Swin-T	ImageNet-1K	1x	43.6	39.6	48M	267G	config	github/baidu	github/baidu
Swin-T	ImageNet-1K	3x	46.0	41.7	48M	267G	config	github/baidu	github/baidu

Cascade Mask R-CNN

Backbone	Pretrain	Lr Schd	box mAP	mask mAP	#params	FLOPs	config	log	model
Swin-T	ImageNet-1K	1x	48.1	41.5	86M	745G	config	github/baidu	github/baidu
Swin-T	ImageNet-1K	3x	50.2	43.5	86M	745G	config	github/baidu	github/baidu

Notes:

• The drop path rate needs to be tuned for best practice.
• MoBY pre-trained models can be downloaded from MoBY with Swin Transformer.

Usage

Installation

Please refer to get_started.md for installation and dataset preparation.

Inference

# single-gpu testing
python tools/test.py <CONFIG_FILE> <DET_CHECKPOINT_FILE> --eval bbox segm

# multi-gpu testing
tools/dist_test.sh <CONFIG_FILE> <DET_CHECKPOINT_FILE> <GPU_NUM> --eval bbox segm

Training

To train a detector with pre-trained models, run:

# single-gpu training
python tools/train.py <CONFIG_FILE> --cfg-options model.pretrained=<PRETRAIN_MODEL> [model.backbone.use_checkpoint=True] [other optional arguments]

# multi-gpu training
tools/dist_train.sh <CONFIG_FILE> <GPU_NUM> --cfg-options model.pretrained=<PRETRAIN_MODEL> [model.backbone.use_checkpoint=True] [other optional arguments] 

For example, to train a Cascade Mask R-CNN model with a Swin-T backbone and 8 gpus, run:

tools/dist_train.sh configs/swin/cascade_mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_giou_4conv1f_adamw_3x_coco.py 8 --cfg-options model.pretrained=<PRETRAIN_MODEL> 

Note: use_checkpoint is used to save GPU memory. Please refer to this page for more details.

Apex (optional):

We use apex for mixed precision training by default. To install apex, run:

git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

If you would like to disable apex, modify the type of runner as EpochBasedRunner and comment out the following code block in the configuration files:

# do not use mmdet version fp16
fp16 = None
optimizer_config = dict(
    type="DistOptimizerHook",
    update_interval=1,
    grad_clip=None,
    coalesce=True,
    bucket_size_mb=-1,
    use_fp16=True,
)

Citing Swin Transformer

@article{liu2021Swin,
  title={Swin Transformer: Hierarchical Vision Transformer using Shifted Windows},
  author={Liu, Ze and Lin, Yutong and Cao, Yue and Hu, Han and Wei, Yixuan and Zhang, Zheng and Lin, Stephen and Guo, Baining},
  journal={arXiv preprint arXiv:2103.14030},
  year={2021}
}

Other Links

Image Classification: See Swin Transformer for Image Classification.

Semantic Segmentation: See Swin Transformer for Semantic Segmentation.

Self-Supervised Learning: See MoBY with Swin Transformer.

Video Recognition, See Video Swin Transformer.