xrosliang / swin-transformer-v2

PyTorch reimplementation of the paper "Swin Transformer V2: Scaling Up Capacity and Resolution" [arXiv 2021].

Swin Transformer V2: Scaling Up Capacity and Resolution

Unofficial PyTorch reimplementation of the paper Swin Transformer V2: Scaling Up Capacity and Resolution by Ze Liu, Han Hu et al. (Microsoft Research Asia).

This repository includes a pure PyTorch implementation of the Swin Transformer V2 and provides pre-trained weights (CIFAR10 & Places365).

The official Swin Transformer V1 implementation is available here. Currently (25.01.2022), an official implementation of the Swin Transformer V2 is not publicly available.

Installation

You can simply install the Swin Transformer V2 implementation as a Python package by using pip.

pip install git+https://github.com/ChristophReich1996/Swin-Transformer-V2

Alternatively, you can clone the repository and use the implementation in swin_transformer_v2 directly in your project.

Usage

This implementation provides the configurations reported in the paper (SwinV2-T, SwinV2-S, etc.). You can build the model by calling the corresponding function. Please note that the Swin Transformer V2 (SwinTransformerV2 class) implementation returns the feature maps of each stage of the network (List[torch.Tensor]). If you want to use this implementation for image classification simply wrap this model and take the final feature map (a wrapper example can be found here).

from swin_transformer_v2 import SwinTransformerV2

from swin_transformer_v2 import swin_transformer_v2_t, swin_transformer_v2_s, swin_transformer_v2_b, \
    swin_transformer_v2_l, swin_transformer_v2_h, swin_transformer_v2_g

# SwinV2-T
swin_transformer: SwinTransformerV2 = swin_transformer_v2_t(in_channels=3,
                                                            window_size=8,
                                                            input_resolution=(256, 256),
                                                            sequential_self_attention=False,
                                                            use_checkpoint=False)

If you want to change the resolution and/or the window size for fine-tuning or inference please use the update_resolution method.

# Change resolution and window size of the model
swin_transformer.update_resolution(new_window_size=16, new_input_resolution=(512, 512))

In case you want to use a custom configuration you can use the SwinTransformerV2 class. The constructor method takes the following parameters.

Parameter	Description	Type
in_channels	Number of input channels	int
depth	Depth of the stage (number of layers)	int
downscale	If true input is downsampled (see Fig. 3 or V1 paper)	bool
input_resolution	Input resolution	Tuple[int, int]
number_of_heads	Number of attention heads to be utilized	int
window_size	Window size to be utilized	int
shift_size	Shifting size to be used	int
ff_feature_ratio	Ratio of the hidden dimension in the FFN to the input channels	int
dropout	Dropout in input mapping	float
dropout_attention	Dropout rate of attention map	float
dropout_path	Dropout in main path	float
use_checkpoint	If true checkpointing is utilized	bool
sequential_self_attention	If true sequential self-attention is performed	bool
use_deformable_block	If true deformable block is used	bool

This file includes a full example how to use this implementation.

This implementation also includes a deformable version of the Swin Transformer V2 inspired by the paper Vision Transformer with Deformable Attention. Deformable attention can be utilized by setting use_deformable_block=True.

This repository also provides an image classification training script for CIFAR10 and Places365.

Results

Model	Dataset	Accuracy	Weights
Swin Transformer V2 T	CIFAR10	0.8974	backbone weights
Swin Transformer V2 T deformable	CIFAR10	0.8962	backbone weights
Swin Transformer V2 B	Places365 (256 X 256)	0.4456 (after 13 epochs)	backbone weights

Disclaimer

This is a very experimental implementation based on the Swin Transformer V2 paper and the official implementation of the Swin Transformer V1. Especially, the sequential self-attention implementation is currently not really memory efficient, if you have any idea for a more efficient sequential implementation please open a pull request. Since an official implementation of the Swin Transformer V2 is not yet published, it is not possible to say to which extent this implementation might differ from the original one. If you have any issues with this implementation please raise an issue.

Reference

@article{Liu2021,
    title={{Swin Transformer V2: Scaling Up Capacity and Resolution}},
    author={Liu, Ze and Hu, Han and Lin, Yutong and Yao, Zhuliang and Xie, Zhenda and Wei, Yixuan and Ning, Jia and Cao, 
            Yue and Zhang, Zheng and Dong, Li and others},
    journal={arXiv preprint arXiv:2111.09883},
    year={2021}
}