Swin-Transformer-Tensorflow

A direct translation of the official PyTorch implementation of "Swin Transformer: Hierarchical Vision Transformer using Shifted Windows" to TensorFlow 2.

The official Pytorch implementation can be found here.

Introduction:

Swin Transformer (the name Swin stands for Shifted window) is initially described in arxiv, which capably serves as a general-purpose backbone for computer vision. It is basically a hierarchical Transformer whose representation is computed with shifted windows. The shifted windowing scheme brings greater efficiency by limiting self-attention computation to non-overlapping local windows while also allowing for cross-window connection.

Swin Transformer achieves strong performance on COCO object detection (58.7 box AP and 51.1 mask AP on test-dev) and ADE20K semantic segmentation (53.5 mIoU on val), surpassing previous models by a large margin.

Usage:

1. To Run a Pre-trained Swin Transformer

Swin-T:

python main.py --cfg configs/swin_tiny_patch4_window7_224.yaml --include_top 1 --resume 1 --weights_type imagenet_1k

Swin-S:

python main.py --cfg configs/swin_small_patch4_window7_224.yaml --include_top 1 --resume 1 --weights_type imagenet_1k

Swin-B:

python main.py --cfg configs/swin_base_patch4_window7_224.yaml --include_top 1 --resume 1 --weights_type imagenet_1k

The possible options for cfg and weights_type are:

cfg	weights_type	22K model	1K Model
configs/swin_tiny_patch4_window7_224.yaml	imagenet_1k	-	github
configs/swin_small_patch4_window7_224.yaml	imagenet_1k	-	github
configs/swin_base_patch4_window7_224.yaml	imagenet_1k	-	github
configs/swin_base_patch4_window12_384.yaml	imagenet_1k	-	github
configs/swin_base_patch4_window7_224.yaml	imagenet_22kto1k	-	github
configs/swin_base_patch4_window12_384.yaml	imagenet_22kto1k	-	github
configs/swin_large_patch4_window7_224.yaml	imagenet_22kto1k	-	github
configs/swin_large_patch4_window12_384.yaml	imagenet_22kto1k	-	github
configs/swin_base_patch4_window7_224.yaml	imagenet_22k	github	-
configs/swin_base_patch4_window12_384.yaml	imagenet_22k	github	-
configs/swin_large_patch4_window7_224.yaml	imagenet_22k	github	-
configs/swin_large_patch4_window12_384.yaml	imagenet_22k	github	-

2. Create custom models

To create a custom classification model:

import argparse

import tensorflow as tf

from config import get_config
from models.build import build_model

parser = argparse.ArgumentParser('Custom Swin Transformer')

parser.add_argument(
    '--cfg',
    type=str,
    metavar="FILE",
    help='path to config file',
    default="CUSTOM_YAML_FILE_PATH"
)
parser.add_argument(
    '--resume',
    type=int,
    help='Whether or not to resume training from pretrained weights',
    choices={0, 1},
    default=1,
)
parser.add_argument(
    '--weights_type',
    type=str,
    help='Type of pretrained weight file to load including number of classes',
    choices={"imagenet_1k", "imagenet_22k", "imagenet_22kto1k"},
    default="imagenet_1k",
)

args = parser.parse_args()
custom_config = get_config(args, include_top=False)

swin_transformer = tf.keras.Sequential([
    build_model(config=custom_config, load_pretrained=args.resume, weights_type=args.weights_type),
    tf.keras.layers.Dense(CUSTOM_NUM_CLASSES)
)

Model ouputs are logits, so don't forget to include softmax in training/inference!!

You can easily customize the model configs with custom YAML files. Predefined YAML files provided by Microsoft are located in the configs directory.

3. Convert PyTorch pretrained weights into Tensorflow checkpoints

We provide a python script with which we convert official PyTorch weights into Tensorflow checkpoints.

$ python convert_weights.py --cfg config_file --weights the_path_to_pytorch_weights --weights_type type_of_pretrained_weights --output the_path_to_output_tf_weights

TODO:

• Translate model code over to TensorFlow
• Load PyTorch pretrained weights into TensorFlow model
• Write trainer code
• Reproduce results presented in paper
  • Object Detection
• Reproduce training efficiency of official code in TensorFlow

Citations:

@misc{liu2021swin,
      title={Swin Transformer: Hierarchical Vision Transformer using Shifted Windows}, 
      author={Ze Liu and Yutong Lin and Yue Cao and Han Hu and Yixuan Wei and Zheng Zhang and Stephen Lin and Baining Guo},
      year={2021},
      eprint={2103.14030},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}