CNNs Visualization using CAM, gradCAM or gradCAM++

Using Grad, Grad-CAM or Grad-CAM++ for visualizing feature maps of Deep Convolutional Networks

Reference

[1] CAM: Learning Deep Features for Discriminative Localization
[2] GradCAM: Visual Explanations from Deep Networks via Gradient-based Localization
[3] Grad-CAM++: Improved Visual Explanations for Deep Convolutional Networks

Project Structure

CNNs_visualization_pytorch
                    |
                    ├── models
                    |	    ├── definitions  # including all definition of models
                    |	    └── weights      # including all trained weights for loading into models.
                    |
                    ├── modules
                    |	    ├── CAM
                    |       |     ├── class_activation_map.py
                    |       |  	  └── config.yaml
                    |	    ├── gradCAM
                    |       |     ├── gradCAM.py
                    |       |  	  └── config.yaml
                    |       └── gradCAMpp
                    |             ├── gradCAMpp.py
                    |       	  └── config.yaml
                    ├── run.py
                    └── utils.py

TODO

• Adding Class Activation Map method for visualizing the last convolutional layer in CNNs with AveragePooling.
• Adding Grad-CAM method for visualizing any convolutional layer in CNNs.
• Applying GuidedBackprop method which is proposed in paper[2].
• Applying grad-CAM++ method which is proposed in paper[3].
• Adding .flake8 to check coding style.
• Adding mypy to check my code and find common bugs.
• Adding pre-commit to solve hook issues.

Usage

CAM, Grad-CAM using model which is trained with custom model and cifar 10 dataset (10 classes).

• CAM

python run.py <image_path> --show-image --config-path module/CAM/config.yaml --module-name cifar_10
python run.py <image_dir> --pattern <image_pattern> --show-image --config-path module/CAM/config.yaml --module-name cifar_10

• Grad-CAM

python run.py <image_path> --show-image --config-path module/gradCAM/config.yaml --module-name cifar_10
python run.py <image_dir> --pattern <image_pattern> --show-image --config-path module/gradCAM/config.yaml --module-name cifar_10

CAM, Grad-CAM using torchvision.models.resnet18 with pretrained weight and imagenet dataset (1000 classes).

• CAM

python run.py <image_path/image_dir> --show-image --config-path module/CAM/config.yaml --module-name image_net
python run.py <image_dir> --pattern <image_pattern> --show-image --config-path module/CAM/config.yaml --module-name image_net

• Grad-CAM

python run.py <image_path> --show-image --config-path module/gradCAM/config.yaml --module-name image_net
python run.py <image_dir> --pattern <image_pattern> --show-image --config-path module/gradCAM/config.yaml --module-name image_net

Results

Method	CIFAR 10	IMAGE NET
CAM
grad-CAM
CAM
grad-CAM

Explainations

CAM (Updating)

Grad-CAM (Updating)

Grad-CAM++ (Updating)

Contributor

Xuan-Phung Pham