EfficientDet-Pytorch

This project is a kind of implementation of EfficientDet using mmdetection.

It is based on the

• the paper EfficientDet: Scalable and Efficient Object Detection
• official TensorFlow implementation
• Pytorch implementation of EfficientNet

Models

Variant	mAP(val2017)	Params	FLOPs	mAP(val2017) in paper	Params in paper	FLOPs in paper
D0	32.02	3.87M	2.55B	33.5	3.9M	2.5B
D1	37.78	6.62M	6.12B	39.1	6.6M	6.1B
D2	——	8.09M	11B	42.5	8.1M	11B
D3	——	12.02M	24.88B	45.9	12M	25B
D4	——	20.7M	55.13B	49.0	21M	55B
D5	——	33.63M	135.31B	50.5	34M	135B
D6	——	——	——	51.3	52M	226B

Usage

1. Install mmdetection

   This implementation is based on mmdetection(v1.1.0+8732ed9). Please refer to INSTALL.md for installation and dataset preparation.

2. Copy the codes to mmdetection directory

   cp -r mmdet/ ${MMDETECTION_PATH}/
   cp -r configs/ ${MMDETECTION_PATH}/

3. Prepare data

   The directories should be arranged like this:

   >   mmdetection
   >     ├── mmdet
   >     ├── tools
   >     ├── configs
   >     ├── data
   >     │   ├── coco
   >     │   │   ├── annotations
   >     │   │   ├── train2017
   >     │   │   ├── val2017
   >     │   │   ├── test2017
   

4. Train D0 with 4 GPUs

   CONFIG_FILE=configs/efficientdet/efficientdet_d0_4gpu.py
   ./ tools/dist_train.py ${CONFIG_FILE} 4

5. Calculate parameters and flops

    python tools/get_flops.py ${CONFIG_FILE} --shape $SIZE $SIZE

6. Test

   python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} --out  ${OUTPUT_FILE} --eval bbox

More usages can reference mmdetection documentation.

Update log

• [2020-04-27] Update results and add SyncBN in backbone.
• [2020-04-20] Fix some bug in bifpn and use separate BN in head.
• [2020-04-17] Add efficientdet-d0 training config.
• [2020-04-16] Add efficientnet.py and retina_sepconv_head.py.
• [2020-04-06] Create this repository.

Notice

1. For small reason, I can't release the model. But you can reproduce the result easily using the config file that I provide.
2. I find the training procedure of EfficientDet is unstable and there is a small chance that results can be 3% mAP lower.
3. The number of bifpn in the latest version of paper is a little different from the first version, but the parameters and flops are the same. I use the structure in the latest version of paper.
4. Training from scratch is a time-consuming task. For exmaple, it took me 4 days to train D0 from scratch using 4 GTX TiTAN V GPUs.