Possible error in xgboost-census-debugger-rules.ipynb

[jane-delaney]

I think there may be an error in the code for https://github.com/aws/amazon-sagemaker-examples/tree/master/sagemaker-debugger/xgboost_census_explanations)/xgboost-census-debugger-rules.ipynb

For the bar plot in the Global Explanations section, the shap values are taken as trial.tensor("full_shap/f0").value(trial.last_complete_step)

Does this not take the values for the first feature only instead of all features?

Apologies if this is my lack of understanding.

[jane-delaney]

This seems to be related to this discussion: awslabs/sagemaker-debugger#294.
Turns out all of the tensors have all of the shap values for every feature instead of just for the feature in question, so the tensors are actually identical ( I tested this).
This is why it is possible to used only the first tensor and get all shap values for all features.
Very confusing though, suggest it be fixed as per the suggestion in the above link and an explanation added to the example.