Grad-CAM wrapper for PyTorch (WIP)

Grad-CAM wrapper is a utility that easily allows you extract and visualize gradient-weighted class activation mappings for your custom model. It works by wrapping around your model, without overwriting any of your models properties/function definitions.

Features

• Computes Grad-CAM using your custom model
• Support for multitarget classification
• Grad-CAM computation is differentiable, allowing you to make it trainable
• Visualization using opencv, or Pillow, depending on your preference
• Licensed under MIT

Installation

Minimum requirements:

python >= 3.6
torch

For visualization:

matplotlib
numpy
python-opencv OR Pillow

Example

+ from gradcam_wrapper import gradcam_wrapper

...

- model = MyCustomModel(params)
+ model = gradcam_wrapper(MyCustomModel)
+ gradcam_layer = 'my_selected_layer'
+ model = model(params, gradcam_layer)

...

for inputs, target in dataloader:
-   logits = model(inputs)
+   logits, gcams = model(inputs)
    # When using the ground truth labels:
+   logits, gcams = model(inputs, labels=targets)

FAQ

• Can I use it with any of the torchvision models?
  Yes, but you cannot use the functions that usually define the model directly, as they are not instances of the torch.nn.Module class. So this requires a little more involvement than usual. An example would be:

model = gradcam_wrapper(torchvision.models.ResNet)
model_params = {'block': torchvision.models.resnet.BasicBlock, 'layers': [2, 2, 2, 2]}
gradcam_layer = 'layer4'
model = model(model_params, gradcam_layer)

Refer to the PyTorch source to do the same for other models

References

The paper that introduces Grad-CAM:

@inproceedings{selvaraju2017grad,
  title={Grad-cam: Visual explanations from deep networks via gradient-based localization},
  author={Selvaraju, Ramprasaath R and Cogswell, Michael and Das, Abhishek and Vedantam, Ramakrishna and Parikh, Devi and Batra, Dhruv},
  booktitle={Proceedings of the IEEE international conference on computer vision},
  pages={618--626},
  year={2017}
}