Deformable-DETR: Deformable Transformers for End-to-End Object Detection

Deformable-DETR

This an implementation of Deformable-DETR. Codes are based on DETR project. My code is inspired by his/her work. Many thanks.

Preparation

For DETR stuffs, etc. data preparation, evaluation, and others , please refer to DETR.

Training

My machine is equipped with two GTX 2080TIs. Below is the training script for DDP training.

bash train.sh

For single gpu training, try the code below

python main.py
--output_dir my_output \
--coco_path ~/dev/data/coco \
--lr 0.0002 \
--lr_backbone 0.00001 \
--num_queries 300 \
--batch_size 1 \
--enc_layers 6 \
--dec_layers 6 \
--no_aux_loss \
--amp

If you do not need AMP, just remove this flag.

Change logs

• 2020-11-30

  • add focal loss for classification

• 2020-11-29

  • integrate MS-Deformable-Attention into DETR architecture
  • modify transfomer's implementation to be adapted to Deformable-Attention
  • add image mask to MS-Deformable-Attention
  • add automatic mixed precision training
  • use adam for the optimizer
  • change lr for projection layers

• 2020-11-24

  • add scale embedding
  • change remove outer loop for scales
  • add backbone modifications for returning multi-scale feature maps
  • add test code for using Deformable-Attention module

• 2020-11-22

  • add Multi-scale Deformabe Attention Module