Global Filter Networks for Image Classification

Created by Yongming Rao, Wenliang Zhao, Zheng Zhu, Jiwen Lu, Jie Zhou

This repository contains PyTorch implementation for GFNet.

Global Filter Networks is a transformer-style architecture that learns long-term spatial dependencies in the frequency domain with log-linear complexity. Our architecture replaces the self-attention layer in vision transformers with three key operations: a 2D discrete Fourier transform, an element-wise multiplication between frequency-domain features and learnable global filters, and a 2D inverse Fourier transform.

Our code is based on pytorch-image-models and DeiT.

[Project Page] [arXiv]

Global Filter Layer

GFNet is a conceptually simple yet computationally efficient architecture, which consists of several stacking Global Filter Layers and Feedforward Networks (FFN). The Global Filter Layer mixes tokens with log-linear complexity benefiting from the highly efficient Fast Fourier Transform (FFT) algorithm. The layer is easy to implement:

import torch
import torch.nn as nn
import torch.fft

class GlobalFilter(nn.Module):
    def __init__(self, dim, h=14, w=8):
        super().__init__()
        self.complex_weight = nn.Parameter(torch.randn(h, w, dim, 2, dtype=torch.float32) * 0.02)
        self.w = w
        self.h = h

    def forward(self, x):
        B, H, W, C = x.shape
        x = torch.fft.rfft2(x, dim=(1, 2), norm='ortho')
        weight = torch.view_as_complex(self.complex_weight)
        x = x * weight
        x = torch.fft.irfft2(x, s=(H, W), dim=(1, 2), norm='ortho')
        return x

Compared to self-attention and spatial MLP, our Global Filter Layer is much more efficient to process high-resolution feature maps:

Model Zoo

We provide our GFNet models pretrained on ImageNet:

Usage

Requirements

• torch>=1.8.1
• torchvision
• timm

Data preparation: download and extract ImageNet images from http://image-net.org/. The directory structure should be

│ILSVRC2012/
├──train/
│  ├── n01440764
│  │   ├── n01440764_10026.JPEG
│  │   ├── n01440764_10027.JPEG
│  │   ├── ......
│  ├── ......
├──val/
│  ├── n01440764
│  │   ├── ILSVRC2012_val_00000293.JPEG
│  │   ├── ILSVRC2012_val_00002138.JPEG
│  │   ├── ......
│  ├── ......

Evaluation

To evaluate a pre-trained GFNet model on the ImageNet validation set with a single GPU, run:

python infer.py --data-path /path/to/ILSVRC2012/ --arch arch_name --path /path/to/model

Training

ImageNet

To train GFNet models on ImageNet from scratch, run:

python -m torch.distributed.launch --nproc_per_node=8 --use_env main_gfnet.py  --output_dir logs/gfnet-xs --arch gfnet-xs --batch-size 128 --data-path /path/to/ILSVRC2012/

To finetune a pre-trained model at higher resolution, run:

python -m torch.distributed.launch --nproc_per_node=8 --use_env main_gfnet.py  --output_dir logs/gfnet-xs-img384 --arch gfnet-xs --input-size 384 --batch-size 64 --data-path /path/to/ILSVRC2012/ --lr 5e-6 --weight-decay 1e-8 --min-lr 5e-6 --epochs 30 --finetune /path/to/model

Transfer Learning Datasets

To finetune a pre-trained model on a transfer learning dataset, run:

python -m torch.distributed.launch --nproc_per_node=8 --use_env main_gfnet_transfer.py  --output_dir logs/gfnet-xs-cars --arch gfnet-xs --batch-size 64 --data-set CARS --data-path /path/to/stanford_cars --epochs 1000 --dist-eval --lr 0.0001 --weight-decay 1e-4 --clip-grad 1 --warmup-epochs 5 --finetune /path/to/model 

License

MIT License

Citation

If you find our work useful in your research, please consider citing:

@article{rao2021global,
  title={Global Filter Networks for Image Classification},
  author={Rao, Yongming and Zhao, Wenliang and Zhu, Zheng and Lu, Jiwen and Zhou, Jie},
  journal={arXiv preprint arXiv:2107.00645},
  year={2021}
}