CBAM-keras:在Keras上实施CBAM

# CBAM-Keras This is a Keras implementation of ["CBAM: Convolutional Block Attention Module"](https://arxiv.org/pdf/1807.06521). This repository includes the implementation of ["Squeeze-and-Excitation Networks"](https://arxiv.org/pdf/1709.01507) as well, so that you can train and compare among base CNN model, base model with CBAM block and base model with SE block. ## CBAM: Convolutional Block Attention Module **CBAM** proposes an architectural unit called *"Convolutional Block Attention Module" (CBAM)* block to improve representation power by using attention mechanism: focusing on important features and supressing unnecessary ones. This research can be considered as a descendant and an improvement of ["Squeeze-and-Excitation Networks"](https://arxiv.org/pdf/1709.01507). ### Diagram of a CBAM_block <div align="center"> <img src="https://github.com/kobiso/CBAM-keras/blob/master/figures/overview.png"> </div> ### Diagram of each attention sub-module <div align="center"> <img src="https://github.com/kobiso/CBAM-keras/blob/master/figures/submodule.png"> </div> ### Classification results on ImageNet-1K <div align="center"> <img src="https://github.com/kobiso/CBAM-keras/blob/master/figures/exp4.png"> </div> <div align="center"> <img src="https://github.com/kobiso/CBAM-keras/blob/master/figures/exp5.png" width="750"> </div> ## Prerequisites - Python 3.x - Keras ## Prepare Data set This repository use [*Cifar10*](https://www.cs.toronto.edu/~kriz/cifar.html) dataset. When you run the training script, the dataset will be automatically downloaded. (Note that you **can not run Inception series model** with Cifar10 dataset, since the smallest input size available in Inception series model is 139 when Cifar10 is 32. So, try to use Inception series model with other dataset.) ## CBAM_block and SE_block Supportive Models You can train and test base CNN model, base model with CBAM block and base model with SE block. You can run **CBAM_block** or **SE_block** added models in the below list. - Inception V3 + CBAM / + SE - Inception-ResNet-v2 + CBAM / + SE - ResNet_v1 + CBAM / + SE (ResNet20, ResNet32, ResNet44, ResNet56, ResNet110, ResNet164, ResNet1001) - ResNet_v2 + CBAM / + SE (ResNet20, ResNet56, ResNet110, ResNet164, ResNet1001) - ResNeXt + CBAM / + SE - MobileNet + CBAM / + SE - DenseNet + CBAM / + SE (DenseNet121, DenseNet161, DenseNet169, DenseNet201, DenseNet264) ### Change *Reduction ratio* To change *reduction ratio*, you can set `ratio` on `se_block` and `cbam_block` method in `models/attention_module.py` ## Train a Model You can simply train a model with `main.py`. 1. Set a model you want to train. - e.g. `model = resnet_v1.resnet_v1(input_shape=input_shape, depth=depth, attention_module=attention_module)` 2. Set attention_module parameter - e.g. `attention_module = 'cbam_block'` 3. Set other parameter such as *batch_size*, *epochs*, *data_augmentation* and so on. 4. Run the `main.py` file - e.g. `python main.py` ## Related Works - Blog: [CBAM: Convolutional Block Attention Module](https://kobiso.github.io//research/research-CBAM/) - Repository: [CBAM-TensorFlow](https://github.com/kobiso/CBAM-tensorflow) - Repository: [CBAM-TensorFlow-Slim](https://github.com/kobiso/CBAM-tensorflow-slim) - Repository: [SENet-TensorFlow-Slim](https://github.com/kobiso/SENet-tensorflow-slim) ## Reference - Paper: [CBAM: Convolutional Block Attention Module](https://arxiv.org/pdf/1807.06521) - Paper: [Squeeze-and-Excitation Networks](https://arxiv.org/pdf/1709.01507) - Repository: [Keras: Cifar10 ResNet example](https://github.com/keras-team/keras/blob/master/examples/cifar10_resnet.py) - Repository: [keras-squeeze-excite-network](https://github.com/titu1994/keras-squeeze-excite-network) ## Author Byung Soo Ko / kobiso62@gmail.com