SmoothGrad implementation in PyTorch

PyTorch implementation of SmoothGrad: removing noise by adding noise.

|Vanilla Gradients|SmoothGrad|Guided backpropagation|Guided SmoothGrad| |:-:|:-:|:-:|:-:| |||||

And other techniques such as the following are implemented.

• Vanilla Gradients

• Guided backpropagation

• Grad-CAM

Download

git clone https://github.com/pkdn/pytorch-smoothgrad

Usage

Save the saliency maps using VanillaGrad, GuidedBackpropGrad, SmoothGrad, GuidedBackpropSmoothGrad.

python saliency.py --img <image-path>

Save the Grad-CAM image.

python grad_cam.py --img <image-path>

If you do not specify an image path, read the raccoon's data (scipy.misc.face()).

Requirements

• PyTorch

• torch-vision

• numpy

• scipy

• OpenCV

Environment under Python 3.5.2 is tested.

Acknowledgments

This code is insipired by pytorch-grad-cam.