pytorch
Add warnings for using GuidedGradCAM and GuidedBackprop
It has been shown that these methods may produce attributions that do not reflect what the model has learnt from training data. Original paper: https://arxiv.org/pdf/1810.03292.pdf Relevant blog post: https://glassboxmedicine.com/2019/10/12/guided-grad-cam-is-broken-sanity-checks-for-saliency-maps/
I think it's necessary to warn users about such pitfalls since these methods can produce pleasing images regardless of the model. Users may believe that their model quality is good because of these attributions which can be wrong.
Hi @m30m!
Thank you for your pull request and welcome to our community. We require contributors to sign our Contributor License Agreement, and we don't seem to have you on file.
In order for us to review and merge your code, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.
If you have received this in error or have any questions, please contact us at [email protected]. Thanks!
Hi @m30m , thanks for the contribution! It is good to note this information regarding the sanity check results. Since different methods have different disadvantages, we think it might be better to add this information to the comparison matrix on our website instead of the docstring. This matrix is located in docs/algorithms_comparison_matrix.md in the repo and on https://captum.ai/docs/algorithms_comparison_matrix .This information can either be in a new column for Limitations or in the Description for now. What do you think? Can you update the PR to move the info there? Thanks! cc: @NarineK
Thank you for signing our Contributor License Agreement. We can now accept your code for this (and any) Facebook open source project. Thanks!