[FEA] Computing Training/Eval metrics after each n steps

[sararb]

🚀 Feature request

Create a custom Callback for computing metrics after each n-steps instead of computing them for each batch. This callback can be then provided to the fit method.

Motivation

• Evaluating the RetrievalModels trained with negative sampling strategies requires comparing the positive scores against a set of implicit negative ones using ranking metrics. One required step to compute those metrics is the extraction of the top-k negatives. This process can be time-consuming (when the catalog of items get large) and the goal of this feature is to gather positive scores and labels for n steps before calling the top-k extraction and ranking metrics calculation.

Your contribution