A simple Detectron2 EfficientDet (working in progress)

update 2020/11/3: There are still some hidden bugs in the code, but due to lack of GPU, I can't continue the experiments

This repository is a simple detectron2 based implementation of EfficientDet

• The backbone part is ported from EfficientNet-PyTorch
• The BiFPN implementation is based on the official implementation
• The detection framework is based on Detectron2

Requirements:

• python>=3.5
• detectron2

Getting Started

1. Install detectron2

2. Download COCO dataset and put annotations, train2017, val2017 (or create symlinks) into DETECTRON2_PATH/datasets/coco

3. Clone this repo:

   git clone https://github.com/zzzxxxttt/simple_detectron2_efficientdet /path/to/efficientdet
   

4. Download the pretrained EfficientNet weights. For example, you downloaded the EfficientNet-B0 weights and name it as b0.pth, run the following codes in python console:

   >>> import torch
   >>> ckpt = torch.load('b0.pth', map_location = 'cpu')
   >>> ckpt = {'model': ckpt, 'matching_heuristics': True}
   >>> torch.save(ckpt, 'b0_detectron2.pth')

5. Start training:

   python train_net.py --config-file configs/EfficientDet_D0.yaml MODEL.WEIGHTS /path/to/checkpoint_file
   

Results

Model	mAP (val, 100 epochs)	mAP (val, 300 epochs)	paper mAP (val, 300 epochs)
EfficientDet-D0	31.9%	32.4%	33.5%