PyTorch implementation of SmoothGrad [1]. WIP, not tested on GPU.
- Python 2.7
- PyTorch
- torchvision
- tqdm
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 |  |
 |
 |
[1] D. Smikov, N. Thorat, B. Kim, F. Viégas, M. Wattenberg. "SmoothGrad: removing noise by adding noise". arXiv, 2017