pytorch

Swin Transformer

Model description

The Swin Transformer is a type of Vision Transformer. It builds hierarchical feature maps by merging image patches (shown in gray) in deeper layers and has linear computation complexity to input image size due to computation of self-attention only within each local window (shown in red). It can thus serve as a general-purpose backbone for both image classification and dense recognition tasks.

Step 1: Installing

pip install timm==0.4.12
pip install yacs

Sign up and login in ImageNet official website, then choose 'Download' to download the whole ImageNet dataset. Specify /path/to/imagenet to your ImageNet path in later training process.

The ImageNet dataset path structure should look like:

imagenet
├── train
│   └── n01440764
│       ├── n01440764_10026.JPEG
│       └── ...
├── train_list.txt
├── val
│   └── n01440764
│       ├── ILSVRC2012_val_00000293.JPEG
│       └── ...
└── val_list.txt

Step 2: Training

Multiple GPUs on one machine

python3 -m torch.distributed.launch --nproc_per_node 8 --master_port 12345  main.py \
    --cfg configs/swin/swin_tiny_patch4_window7_224.yaml --data-path /path/to/imagenet --batch-size 128

Reference

https://github.com/microsoft/Swin-Transformer