Luca Massaron

icon indicating a website link https://www.kaggle.com/lucamassaron

icon company @google Developers Expert icon location Verona, Italy icon location Data scientist, author of books on AI, machine learning, deep learning & Kaggle. 3x Kaggle Grandmaster, 7th worldwide rank achieved on Kaggle competitions.

Results 3 comments of Luca Massaron

SparseCategoricalFocalLoss cannot compile with TPUs

Yes, if I replace it with a tf.keras.losses.SparseCategoricalCrossentropy() the code just works fine. Here is the complete error reported by the code: ```--------------------------------------------------------------------------- InternalError Traceback (most recent call last) /tmp/ipykernel_42/3395278551.py...

SparseCategoricalFocalLoss cannot compile with TPUs

So probably is it necessary just to change the resizing operations?

Feature importance

Both LIME (https://github.com/marcotcr/lime) and SHAP (https://github.com/slundberg/shap) can provide you with feature importance of a DNN model. They require some extra-work, though. I will write some articles on that in the...