gradCAM, guidedBackProp, smoothGrad on pytorch

Pytorch implementation of gradCAM, guidedBackProp, smoothGrad

Resnet50 is used in this implementation (also other models can be used)

• gradCAM https://arxiv.org/pdf/1610.02391.pdf
• guidedBackProp https://arxiv.org/pdf/1412.6806.pdf
• smoothGrad https://arxiv.org/pdf/1706.03825.pdf

dependencies

• Python 3.6
• torch 1.0.0
• torchvison 0.2.1
• cv2
• tqdm

usage

install requirements

pip install -r requirements.txt

execute script

python main.py image_path --cuda --index hoge

• add cuda option to use gpu

• add index option to specify the target imagenet index of gradCAM

  Please see imagenet_class_index.json (e.g. ostrich: 9, tusker: 101)

  If not specified, target index will be estimated by the model

examples

	ostrich.jpg	elephant.jpg
raw
gradCAM
guidedBackProp
guidedGradCAM
smoothGrad