[FEA] Feature to extract attention values from transformer heads

[vivpra89]

🚀 Feature request

Ability to extract attention weights from various heads of transformer

Motivation

Plotting attention provides insights into the inner-workings and user behaviors that business teams can relate with. This is easily available with pytorch / Tensorflow.

Is there a way to convert the trained model to PT / TF models to capture the attention values?