Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions

This repository contains PyTorch evaluation code, training code and pretrained models for PVT (Pyramid Vision Transformer).

Like ResNet, PVT is a pure transformer backbone that can be easily plugged in most downstream task models.

With a comparable number of parameters, PVT-Small+RetinaNet achieves 40.4 AP on the COCO dataset, surpassing ResNet50+RetinNet (36.3 AP) by 4.1 AP.

Figure 1: Performance of RetinaNet 1x with different backbones.

This repository is developed on the top of pytorch-image-models and deit.

For details see Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions.

If you use this code for a paper please cite:

@misc{wang2021pyramid,
      title={Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions}, 
      author={Wenhai Wang and Enze Xie and Xiang Li and Deng-Ping Fan and Kaitao Song and Ding Liang and Tong Lu and Ping Luo and Ling Shao},
      year={2021},
      eprint={2102.12122},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Todo List

• ImageNet model weights
• PVT + RetinaNet/Mask R-CNN config & models
• PVT + Semantic FPN config & models
• PVT + DETR config & models
• PVT + Trans2Seg config & models

Model Zoo

We provide baseline PVT models pretrained on ImageNet 2012.

name	acc@1	#params (M)	url
PVT-Tiny	75.1	13.2	51 M
PVT-Small	79.8	24.5	93 M
PVT-Medium	81.2	44.2	Todo.
PVT-Large	81.7	61.4	Todo.

Before using it, make sure you have the pytorch-image-models package timm==0.3.2 by Ross Wightman installed. Note that our work relies of the augmentations proposed in this library.

Usage

First, clone the repository locally:

git clone https://github.com/whai362/PVT.git

Then, install PyTorch 1.6.0+ and torchvision 0.7.0+ and pytorch-image-models 0.3.2:

conda install -c pytorch pytorch torchvision
pip install timm==0.3.2

Data preparation

Download and extract ImageNet train and val images from http://image-net.org/. The directory structure is the standard layout for the torchvision datasets.ImageFolder, and the training and validation data is expected to be in the train/ folder and val folder respectively:

/path/to/imagenet/
  train/
    class1/
      img1.jpeg
    class2/
      img2.jpeg
  val/
    class1/
      img3.jpeg
    class/2
      img4.jpeg

Evaluation

To evaluate a pre-trained PVT-Small on ImageNet val with a single GPU run:

sh dist_train.sh pvt_small 1 /path/to/checkpoint_root --data-path /path/to/imagenet --resume /path/to/checkpoint_file --eval

This should give

Todo.

Training

To train PVT-Small on ImageNet on a single node with 8 gpus for 300 epochs run:

sh dist_train.sh pvt_small 8 /path/to/checkpoint_root --data-path /path/to/imagenet

License

This repository is released under the Apache 2.0 license as found in the LICENSE file.