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].
Python 2.7 / 3.+
$ pip install click opencv-python matplotlib tqdm numpy
$ pip install "torch>=0.4.1" torchvisionpython main.py [DEMO_ID] [OPTIONS]Demo ID:
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 CPUThe command above generates, for top k classes:
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).
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] | |
 |
 |
 |
 |
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] |  |
 |
 |
 |
|
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.