DeepLIFT with a multiclass classifier

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

720 stars 133 forks source link

DeepLIFT with a multiclass classifier #52

Closed gregmcinnes closed 4 years ago

gregmcinnes commented 4 years ago

I'm trying to run DeepLIFT on a multiclass classifier. The model is trained in keras and outputs binary classifications for two classes. I've been able to get DeepLIFT to work with DeepExplain to work for a single class, but I can't figure out how to get a multiclass classifier to work in the DeepExplain framework. Can you provide any examples or some guidance on what might be different between the two cases?

Thanks

gregmcinnes commented 4 years ago

I'm an idiot. I had the shape of the y's matrix wrong and didn't realize it. I looked at the shape of the y matrix in the CNN MNIST tutorial and realized my mistake.