PVT-tensorflow2

A Tensorflow2.x implementation of Pyramid Vision Transformer as described in Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions

Update Log

[2021-06-29]

• Fix bug on saving model

[2021-03-20]

• Add PVT-tiny,PVT-small,PVT-medium,PVT-large.

Installation

1. Clone project

git clone https://github.com/wangermeng2021/PVT-tensorflow2.git
cd PVT-tensorflow2

2. Install environment

• install tesnorflow ( skip this step if it's already installed)

• pip install -r requirements.txt
  

Training

• For training on cifar10 dataset,use:

  python train.py --dataset cifar10  --model PVT-tiny --epochs 200 --batch-size 256 --img-size 32
  

• For training on your custom dataset,use:

  python train.py --dataset dataset/RockPaperScissor  --model PVT-tiny --epochs 200 --batch-size 128 --img-size 128 --init-lr 5e-5
  

Evaluation results:

model	cifar10
PVT-tiny	0.82

References

• https://github.com/whai362/PVT