This project uses ViT to perform image classification tasks on DATA set CIFAR10. The implement of Vit and pretrained weight are from https://github.com/asyml/vision-transformer-pytorch. Different from https://github.com/Kaicheng-Yang0828/Vision-Transformer-ViT, this project use multi-process distributed training and it also use Apex to reduce GPU resource consumption.

pytorch 1.7.1
python 3.7.3

1、 git clone https://github.com/NVIDIA/apex.git
2、 cd apex
3、 python setup.py install

Download the CIFAR10 from http://www.cs.toronto.edu/~kriz/cifar.html or you can get it from https://pan.baidu.com/s/1ogAFopdVzswge2Aaru_lvw (code: k5v8), creat data floder and unzip the cifar-10-python.tar.gz under './data'

You can download the pretrained file from https://pan.baidu.com/s/1CuUj-XIXwecxWMEcLoJzPg (code: ox9n), creat Vit_weights floder and pretrained file under ./Vit_weights

python main.py 

Base on the pretrained weight, after one epoch, I get 98.1 Accuracy (I didn't adjust the parameters carefully, you can get better results by adjusting the parameters)

1、Multi-process parallel training reduces the training time by one-fifth
2、Apex reduce about 30% GPU resources under the premise of ensuring the same accuracy rate