PyTorch implementation of Grad-CAM (Gradient-weighted Class Activation Mapping) [1]. Grad-CAM localizes and highlights discriminative regions that convolutional neural networks-based models activate to predict visual concepts. This repository contains only implementation for image classification models. Please checkout pytorch_0.3.1 branch for pytorch 0.3.1.

• Python 2.7
• pytorch 0.4.0
• torchvision 0.2.1
• click
• opencv

python main.py --help

• -i, --image-path: a path to an image (required)
• -a, --arch: a model name from torchvision.models, e.g., 'resnet152' (required)
• -k, --topk: the number of classes to generate (default: 3)
• --cuda/--no-cuda: GPU or CPU

The command above generates, for top k classes:

• Gradients by vanilla backpropagation
• Gradients by guided backpropagation [2]
• Gradients by deconvnet [2]
• Grad-CAM [1]
• Guided Grad-CAM [1]

The guided-* do not support F.relu but only nn.ReLU in this codes. For instance, off-the-shelf inception_v3 cannot cut off negative gradients during backward operation (#2).

layer4.2 of torchvision.models.resnet152

	bull mastiff	tiger cat	boxer
Probability	0.54285	0.19302	0.10428
Grad-CAM [1]
Vanilla backpropagation
"Deconvnet" [2]
Guided backpropagation [2]
Guided Grad-CAM [1]

Grad-CAM visualization of bull mastiff

Model	resnet152	vgg19	vgg19_bn	inception_v3	densenet201
Input	224x224	224x224	224x224	299x299	224x224
Grad-CAM [1]

[1] R. R. Selvaraju, A. Das, R. Vedantam, M. Cogswell, D. Parikh, and D. Batra. "Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization". arXiv, 2016
[2] J. T. Springenberg, A. Dosovitskiy, T. Brox, and M. Riedmiller. "Striving for Simplicity: The All Convolutional Net". arXiv, 2014