Attention Branch Network

Writer : Hiroshi Fukui

Maintainer : Tsubasa Hirakawa

Change Log

• 15 Apr 2020: Add Docker environment (torch 0.4.0) [0.4.0 implementation]
• 02 Sep 2022: Update PyTorch version (torch 1.12.0)

Abstract

This repository contains the source code of Attention Branch Network for image classification. The attention branch network is designed to extend the top-down visual explanation model by introducing a branch structure with an attention mechanism. ABN improves CNN’s performance and visual explanation at the same time by gnerating attention maps in the forward pass.

[CVF open access] [arXiv paper]

Citation

If you find this repository is useful. Please cite the following references.

@article{fukui2018cvpr,
    author = {Hiroshi Fukui and Tsubasa Hirakawa and Takayoshi Yamashita and Hironobu Fujiyoshi},
    title = {Attention Branch Network: Learning of Attention Mechanism for Visual Explanation},
    journal = {Computer Vision and Pattern Recognition},
    year = {2019},
    pages = {10705-10714}
}

@article{fukui2018arxiv,
    author = {Hiroshi Fukui and Tsubasa Hirakawa and Takayoshi Yamashita and Hironobu Fujiyoshi},
    title = {Attention Branch Network: Learning of Attention Mechanism for Visual Explanation},
    journal = {arXiv preprint arXiv:1812.10025},
    year = {2018}
}  

Enviroment

Our source code is based on https://github.com/bearpaw/pytorch-classification/ implemented with PyTorch. We are grateful for the author!

Now, we have updated PyTorch version (1.12.0) from the original implementation (0.4.0). Required PyTorch version is as follows:

• PyTorch : 1.12.0
• PyTorch vision : 0.13.0

Docker

We prepared Docker environments for ABN. You can quickly start to use Docker and run scripts. For more details, please see docker/README.md.

Execution

Example of run command is as follows:

Training

# CIFAR-100 dataset
python3 cifar.py -a resnet --dataset cifar100 --depth 110 --epochs 300 --schedule 150 225 --gamma 0.1 --wd 1e-4 --checkpoint checkpoints/cifar100/resnet-110 --gpu-id 0,1

# ImageNet dataset
python3 imagenet.py -a resnet152 --data ../../dataset/imagenet_data/ --epochs 90 --schedule 31 61 --gamma 0.1 -c checkpoints/imagenet/resnet152 --gpu-id 4,5,6,7 --test-batch 100

Evaluation

# CIFAR-100 dataset
python3 cifar.py -a resnet --dataset cifar100 --depth 110 --epochs 300 --schedule 150 225 --gamma 0.1 --wd 1e-4 --checkpoint checkpoints/cifar100/resnet-110 --gpu-id 0,1 --evaluate --resume checkpoints/cifar100/resnet-110/model_best.pth.tar

# ImageNet dataset
python3 imagenet.py -a resnet152 --data ../../../../dataset/imagenet_data/ --epochs 90 --schedule 31 61 --gamma 0.1 -c checkpoints/imagenet/resnet152 --gpu-id 4,5,6 --test-batch 10 --evaluate --resume checkpoints/imagenet/resnet152/model_best.pth.tar

If you try the other models, please see TRAINING.md.

Trained models

We have published the model files of ABN, which are ResNet family models on CIFAR100 and ImageNet2012 dataset.

Models on CIFAR100 Dataset

• ResNet110
• DenseNet
• Wide ResNet
• ResNeXt

Models on ImageNet2012 Dataset

• ResNet50
• ResNet101
• ResNet152

Performances

CIFAR100

	top-1 error (ABN)	top-1 error (original)
ResNet110	22.5	24.1
DenseNet	21.6	22.5
Wide ResNet	18.1	18.9
ResNeXt	17.7	18.3

ImageNet2012

	top-1 error (ABN)	top-1 error (original)
ResNet50	23.1	24.1
ResNet101	21.8	22.5
ResNet152	21.4	22.2

Examples of attention map