Swin Transformer PyTorch Hub

This is just a quick way to load Swin Transformers from image classification from PyTorch Hub. This repository makes it possible to load Swin Transformers in 1 line of code.

The official Swin transformer repository can be found here:

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

Dependencies

• torch - PyTorch
• timm - Torchvision Image Models

Load Model

import torch
HUB_URL = "SharanSMenon/swin-transformer-hub"
MODEL_NAME = "swin_tiny_patch4_window7_224"
# check hubconf for more models.
model = torch.hub.load(HUB_URL, MODEL_NAME, pretrained=True) # load from torch hub

Transforms

Transforms for passing in PIL images for inference.

from torchvision import transforms as T
from PIL import Image
import timm

transform = T.Compose([
    T.Resize(256),
    T.CenterCrop(224),
    T.ToTensor(),
    T.Normalize(timm.data.IMAGENET_DEFAULT_MEAN, timm.data.IMAGENET_DEFAULT_STD)
])

Imagenet Labels

Get a list of imagenet labels.

import json
from urllib.request import urlopen

URL = "https://raw.githubusercontent.com/SharanSMenon/swin-transformer-hub/main/imagenet_labels.json"
response = urlopen(URL)
classes = json.loads(response.read())
len(classes) # Should return 1000

TODO

• Add support for more model weights