Results | Updates | Usage | Todo | Acknowledge

This branch contains the unofficial pytorch implementation of Exploring Plain Vision Transformer Backbones for Object Detection. Thanks for their wonderful work!

The models are trained on 4 A100 machines with 2 images per gpu, which makes a batch size of 64 during training.

[2022-04-18] Explore using small 1K supervised trained models (20M parameters) for ViTDet (45.6 mAP). The results with multi-stage structure is 46.0 mAP for Swin-T and 47.8 mAP for ViTAEv2-S with Mask RCNN on COCO.

[2022-04-17] Release the pretrained weights and logs for ViT-B and ViTAE-B on MS COCO. The models are totally trained with PyTorch on GPU.

[2022-04-16] Release the initial unofficial implementation of ViTDet with ViT-Base model! It obtains 51.1 mAP and 45.5 mAP on detection and segmentation, respectively. The weights and logs will be uploaded soon.

Applications of ViTAE Transformer include: image classification | object detection | semantic segmentation | animal pose segmentation | remote sensing | matting

We use PyTorch 1.9.0 or NGC docker 21.06, and mmcv 1.3.9 for the experiments.

git clone https://github.com/open-mmlab/mmcv.git
cd mmcv
git checkout v1.3.9
MMCV_WITH_OPS=1 pip install -e .
cd ..
git clone https://github.com/ViTAE-Transformer/ViTDet.git
cd ViTDet
pip install -v -e .

After install the two repos, install timm and einops, i.e.,

pip install timm==0.4.9 einops

Download the pretrained models from MAE or ViTAE, and then conduct the experiments by

# for single machine
bash tools/dist_train.sh <Config PATH> <NUM GPUs> --cfg-options model.pretrained=<Pretrained PATH>

# for multiple machines
python -m torch.distributed.launch --nnodes <Num Machines> --node_rank <Rank of Machine> --nproc_per_node <GPUs Per Machine> --master_addr <Master Addr> --master_port <Master Port> tools/train.py <Config PATH> --cfg-options model.pretrained=<Pretrained PATH> --launcher pytorch

This repo current contains modifications including:

• using LN for the convolutions in RPN and heads
• using large scale jittor for augmentation
• using RPE from MViT
• using longer training epochs and 1024 test size
• using global attention layers

There are other things to do:

• Implement the conv blocks for global information communication

• Tune the models for Cascade RCNN

• Train ViT models for the LVIS dataset

• Train ViTAE model with the ViTDet framework

We acknowledge the excellent implementation from mmdetection, MAE, MViT, and BeiT.

@article{Li2022ExploringPV,
  title={Exploring Plain Vision Transformer Backbones for Object Detection},
  author={Yanghao Li and Hanzi Mao and Ross B. Girshick and Kaiming He},
  journal={ArXiv},
  year={2022},
  volume={abs/2203.16527}
}

For ViTAE and ViTAEv2, please refer to:

@article{xu2021vitae,
  title={Vitae: Vision transformer advanced by exploring intrinsic inductive bias},
  author={Xu, Yufei and Zhang, Qiming and Zhang, Jing and Tao, Dacheng},
  journal={Advances in Neural Information Processing Systems},
  volume={34},
  year={2021}
}

@article{zhang2022vitaev2,
  title={ViTAEv2: Vision Transformer Advanced by Exploring Inductive Bias for Image Recognition and Beyond},
  author={Zhang, Qiming and Xu, Yufei and Zhang, Jing and Tao, Dacheng},
  journal={arXiv preprint arXiv:2202.10108},
  year={2022}
}