luuuyi / cbam.pytorch

Non-official implement of Paper：CBAM: Convolutional Block Attention Module

CBAM.PyTorch

Non-official implement of Paper：CBAM: Convolutional Block Attention Module

Introduction

The codes are PyTorch re-implement version for paper: CBAM: Convolutional Block Attention Module

Woo S, Park J, Lee J Y, et al. CBAM: Convolutional Block Attention Module[J]. 2018. ECCV2018

Structure

The overview of CBAM. The module has two sequential sub-modules: channel and spatial. The intermediate feature map is adaptively refined through our module (CBAM) at every convolutional block of deep networks.

Requirements

• Python3
• PyTorch 0.4.1
• tensorboardX (optional)
• torchnet
• pretrainedmodels (optional)

Results

We just test four models in ImageNet-1K, both train set and val set are scaled to 256(minimal side), only use Mirror and RandomResizeCrop as training data augmentation, during validation, we use center crop to get 224x224 patch.

ImageNet-1K

Models	validation(Top-1)	validation(Top-5)
ResNet50	74.26	91.91
ResNet50-CBAM	75.45	92.55

cbam.pytorch's Issues

Pre-trained model

Hello, may you provide a pre-trained model of cbam, thank you.

Can pytorch 0.4.0 use this code?

怎么运行

博主大大 能给出运行的步骤吗 刚接触python和pytorch ，真的是一脸懵

Global Second Order Pooling

I know you said that you do not have time but this is also interesting: https://arxiv.org/abs/1811.12006

Questions about ChannelAttention

Thank you very much! In the paper, this block meant to be the same as SE block (except for the MaxPool addition). Is there any reason why you used nn.Conv2d rather than nn.Linear in the original SE block? I am working with time series, would you recommend me to use Conv1d or keep using Linear? Finally, is there any reason why bias = False? I obtained better results with bias = True with SE block in my case.

Lastly, have you done any comparison with Gather & Excite? https://github.com/hujie-frank/GENet

about accuracy?

I find your code can't reach the accuary

how to use grad cam to visualize the attention part？

关于空间注意力模块的max_out出错

我将模块加在resnet网络上，训练时出现max_out的print结果为两个tensor，被认定为是list，无法运行，请问大家有遇到此类问题吗

数据集要怎么弄

datasets？

fc or conv in file resnet_cbam.py, line 31 and line 33?

For the file resnet_cbam.py, I think line31 and line 33 are not consistent with the paper. fc1 and fc2 should be nn.Linear because the paper said:

Both descriptors are then forwarded to a shared network
to produce our channel attention map Mc 2 RC�1�1. The shared network is
composed of multi-layer perceptron (MLP) with one hidden layer. To reduce
parameter overhead, the hidden activation size is set to RC=r�1�1, where r is
the reduction ratio.

May I know why you use conv instead of Linear ?

I think there are some mistakes

After applying Channel Attention Module, maybe it would be better to apply a convolution layer in order to modify the channels to the original value (usually 3 channels), instead of applying Spatial Attention Module instantly. Or Spatial Attention Module can't make sense.

Error in testing

hi I ma getting this error while testing
Traceback (most recent call last):
File "test.py", line 213, in
test_and_generate_result_round2('2','resnet50',416, True)
File "test.py", line 133, in test_and_generate_result_round2
my_model = resnet50.MyResNet50(model_ft)
NameError: name 'resnet50' is not defined

what means of in_planes in ChannelAttention？

RT.

Can you share a pretrained model link for resnet50?

Thanks for your contribution! I hope you can share a pretrained model link for resnet50!