SeldonIO / alibi

Algorithms for explaining machine learning models
https://docs.seldon.io/projects/alibi/en/stable/
Other
2.35k stars 249 forks source link

Integrated Gradients serialization with layers that are not accessible via model.layers #401

Open jklaise opened 3 years ago

jklaise commented 3 years ago

Because of https://github.com/SeldonIO/alibi/pull/398 it is not clear how to handle serialization of IG explainers specifying custom layers not accessible via model.layers.

jklaise commented 3 years ago

This is fairly important to resolve as we don't serialize the original model as part of serializing the explainer, so by extension we can't serialize the layer object, but if there is no way to refer to it otherwise than just by having a reference to it then we're stuck...