Attentions module: SE, BAM and CBAM on CIFAR100 for ResNet34 and ResNet50

Cao Nguyen-Van
[email protected]
20184172

Requirements:

• Python 2.7 (only 2.7)
• pytorch 1.0.1
• tqdm, argparse, yaml
• ...

This work is organized as followings:

Attentions
|       README.md: This file
|       main.py: File to train the models
|       grad_cam.py File for generating Grad-CAM images 
|_______utils: Folder for dataset and other functions 
|       |______datasets.py: Datasets
|       |______utils.py: Other utils
|_______models: Folder for networks and modules
|       |______attention_resnet.py: ResNet with additional arguments for attention modules
|       |______attention_modules.py: attention modules
|       |______resnet.py: Original ResNet implementation from Torchvision
|       |______models.py: wrapper to create models 
|_______configs: Folder for configuration
|_______|______config.yaml: configurations for training and testing network
        

To run train the model:

In order to train the model, the dataset directory should be placed to DATA_DIR in the config.yaml file. Please point to the lowest level directory, and keep every file same as downloaded from CIFAR100 website. In case you changed file name, please go to utils/dataset.py and change the name of data files. The default directory for logging and saving models are checkpoint/logs and checkpoint/<baseline>_<attention_type>. These paths can be changed to any where by changing value of LOG_DIR and CHECKPOINT in config.yaml. Such directories will be automatically created if they don't existed.

python main.py --arch resnet50 --attention CBAM

• --arch: resnet34 or resnet50
• --attention: 'None', SE, BAM, CBAM, BAMspatial, BAMchannel, CBAMspatial, CBAMchannel.

To run the grad-cam code:

python grad_cam.py --arch resnet50 --attention CBAM --output-dir gradCAM_images

• --arch: None, resnet34 or resnet50. If None, the original input images will be produced
• --attention: 'None', SE, BAM, CBAM, BAMspatial, BAMchannel, CBAMspatial, CBAMchannel.
• --output-dir: Output folder to save the images This script will take the first 8 images from the test set and compute Grad-CAM image with the specified model. In case of None in the --arch option, the output will be saved in the format <index>_<class name>.png. In other cases, the output will be saved in the format <index>_<archirecture>_<attention>_<probability of the true label>.png.