Dynamic DETR: End-to-End Object Detection with Dynamic Attention

This is the CCBDA final project, NYCU, 2022 Fall which is an PyTorch implementation for Dynamic DETR.

For paper's details see Dynamic DETR: End-to-End Object Detection with Dynamic Attention by Xiyang Dai, Yinpeng Chen, Jianwei Yang, Pengchuan Zhang, Lu Yuan, Lei Zhang

Data preparation

Download and extract COCO 2017 train and val images with annotations from http://cocodataset.org. We expect the directory structure to be the following:

path/to/coco/
  annotations/  # annotation json files
  train2017/    # train images
  val2017/      # val images

Training

To train Dynamic DETR on a single gpu with 12 epochs:

python main.py --output_dir [./DIR] --epochs 12 --lr_drop 11 --coco_path /path/to/coco 

We train Dynamic DETR with AdamW setting learning rate in the transformer to 1e-4 and 1e-5 in the backbone. Horizontal flips, scales and crops are used for augmentation. The transformer is trained with dropout of 0.1, and the whole model is trained with grad clip of 0.1.

Evaluation

To evaluate Dynamic DETR:

python main.py --batch_size 2 --no_aux_loss --eval --resume /path/to/checkpoint --coco_path /path/to/coco

Distributed training

python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --coco_path /path/to/coco