James Budarz

Hi, has anyone found a workaround for this?

@jklaise Thanks for the response. Anchors also fails when eager mode is disabled. `RuntimeError: Attempting to capture an EagerTensor without building a function.` However, if I don't disable eager mode...

Sorry- to answer your first question: > does onnx runtime work in eager mode only? Making predictions with eager mode disabled results in the same error: `RuntimeError: Attempting to capture...

Thanks for looking into this. > A very simple thing to check first (if you can) is if you can make predictions using the ONNX runtime with a TF2.x model...

I've just replicated this experiment with the original Keras model, and I realize that this might not be an ONNX issue- it might be incompatibility with any keras model trained...

This update will break compatibility with scikit-learn < 0.20, so it would be necessary to update the requirements and setup.py to require at least this version (eli5 currently allows >=0.18).

Out-of-bag scores are produced during the training process, while Permutation Importance relies on being provided a pretrained model. PermutationImportance is asking for a method of the model object to use...

As with #397, changing the requirements and setup to require sklearn >= 0.20 would maintain compatibility.

Additional information: I've now found that `eli5.explain_weights(perm)` results in all weights of zero, and I also realized that this model predicts the negative class for all data points. This would...

The way to think about the baseline is the expected value for the whole training dataset. (i.e. if the mean prediction of your model, based on that dataset, is 38%,...