kazuto1011 / smoothgrad-pytorch

PyTorch re-implementation of SmoothGrad

SmoothGrad with PyTorch

PyTorch implementation of SmoothGrad [1]. WIP, not tested on GPU.

Dependencies

• Python 2.7
• PyTorch
• torchvision
• tqdm

Examples

python main.py --image samples/cat_dog.png [--no-cuda] [--guided]

With the --guided option, you can generate smoothed maps from guided backproped gradients.

Model: ResNet-152 pre-trained on ImageNet
Prediction: bull mastiff - 54.3% @1
#samples: 100

Noise level (σ)	10%	15%	20%
SmoothGrad [1]
Guided Backprop + SmoothGrad

References

[1] D. Smikov, N. Thorat, B. Kim, F. Viégas, M. Wattenberg. "SmoothGrad: removing noise by adding noise". arXiv, 2017