Subscores in neural ranking GAM

[lauragalera]

Hello there,

Following announcement #202:

I've successfully trained a neural GAM model following this tutorial, but using GAMScorer instead.

I've been able to predict the ranking of a test query using the saved model:

tf_example_predictor = loaded_model.signatures[tf.saved_model.REGRESS_METHOD_NAME]
scores = tf_example_predictor(tf.convert_to_tensor(instances))[tf.saved_model.REGRESS_OUTPUTS]

Nevertheless, I'd like to understand the logic behind these scores, which should be possible since GAM is interpretable. The API mentions that GAM layers retrieve the subscore of each feature per each document.

Could you provide some support?

Thank you