alibi
alibi copied to clipboard
[IntegratedGradients] Pytorch models
Our IntegratedGradients
method is for TensorFlow models only, for feature parity it would be desirable to extend it to PyTorch models as well.
There are a few options and some design decisions to make. The captum library has PyTorch methods for all kinds of gradient-based attribution methods. The easiest thing would be to just wrap this but it has downsides:
- dependency on
torch
- foralibi
we want this to be optional - temptation to wrap all
captum
methods at the cost of abandoning feature parity withtensorflow
So it likely makes sense to implement this from scratch.
This could also be an opportunity for a much larger project, designing the public and private API for all kinds of gradient-based methods not just IG, but the priority should be on getting a PyTorch IG going.
As a side note, kserve supports TorchServe
which offers explainability via captum
. So this issue also affects SC (v2) feature parity with kserve at this level.