search

marcoancona / DeepExplain

A unified framework of perturbation and gradient-based attribution methods for Deep Neural Networks interpretability. DeepExplain also includes support for Shapley Values sampling. (ICLR 2018)
https://arxiv.org/abs/1711.06104
MIT License
725 stars 133 forks source link

'Garter Snake' attribution comparison example codes needed #3

Closed yulongwang12 closed 6 years ago

yulongwang12 commented 6 years ago

Would you please include the example codes which can generate the 'Garter Snake' attribution comparison picture in docs/comparison.png? Actually I'm currently working on reproducing the results on PyTorch. The MNIST examples are comparable, but for snake picture, the attribution heatmaps are not salient. I want to know if any configurations have been made (like the initial relevance value r_i scaled ?) when performing on naturalistic images with imagenet-pretrained models. Thank you very much!

marcoancona commented 6 years ago

We have used Inception V3 and filtered the 99th percentile as suggested in https://arxiv.org/abs/1706.03825 Also make sure the colormap range is symmetric, in positive and negative values. I will upload the code for it soon.

marcoancona commented 6 years ago

Closed in 2ed086953c07a0a1652cf0b0d6b9e1359fd03426