Grad-CAM with PyTorch

PyTorch implementation of Grad-CAM (Gradient-weighted Class Activation Mapping) [1] in image classification. This repository also contains implementations of vanilla backpropagation, guided backpropagation [2], deconvnet [2], and guided Grad-CAM [1], occlusion sensitivity maps [3].

Requirements

Python 2.7 / 3.+

$ pip install click opencv-python matplotlib tqdm numpy
$ pip install "torch>=0.4.1" torchvision

Basic usage

python main.py [DEMO_ID] [OPTIONS]

Demo ID:

• demo1
• demo2
• demo3

Options:

• -i, --image-paths: image path, which can be provided multiple times (required)
• -a, --arch: a model name from torchvision.models, e.g. "resnet152" (required)
• -t, --target-layer: a module name to be visualized, e.g. "layer4.2" (required)
• -k, --topk: the number of classes to generate (default: 3)
• -o, --output-dir: a directory to store results (default: ./results)
• --cuda/--cpu: 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 (issue #2).

Demo 1

Generate all kinds of visualization maps given a torchvision model, a target layer, and images.

python main.py demo1 -a resnet152 -t layer4 \
                     -i samples/cat_dog.png -i samples/vegetables.jpg # You can add more images

Predicted class	#1 boxer	#2 bull mastiff	#3 tiger cat
Grad-CAM [1]
Vanilla backpropagation
"Deconvnet" [2]
Guided backpropagation [2]
Guided Grad-CAM [1]

Grad-CAM with different models for "bull mastiff" class

Model	resnet152	vgg19	vgg19_bn	densenet201	squeezenet1_1
Layer	layer4	features	features	features	features
Grad-CAM [1]

Demo 2

Generate Grad-CAM maps for "bull mastiff" class, at different layers of ResNet-152 (hardcoded).

python main.py demo2 -i samples/cat_dog.png

Layer	relu	layer1	layer2	layer3	layer4
Grad-CAM [1]

Demo 3

Generate the occlusion sensitivity map [1, 3] based on logit scores. The red and blue regions indicate a relative increase and decrease from non-occluded scores respectively: the blue regions are critical!

python main.py demo3 -a resnet152 -i samples/cat_dog.png

Patch size	10x10	15x15	25x25	35x35	45x45	90x90
"boxer" sensitivity
"bull mastiff" sensitivity
"tiger cat" sensitivity

This demo takes much time to compute per-pixel logits. You can control the resolution by changing sampling stride (--stride), or increasing batch size as to fit on your GPUs (--n-batches). The model is wrapped with torch.nn.DataParallel so that runs on multiple GPUs by default.

References

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. In ICCV, 2017
2. J. T. Springenberg, A. Dosovitskiy, T. Brox, and M. Riedmiller. Striving for Simplicity: The All Convolutional Net. arXiv, 2014
3. M. D. Zeiler, R. Fergus. Visualizing and Understanding Convolutional Networks. In ECCV, 2013