would be nice to add an input mask, so we can use arbitrary length input during the training.

[shamanez]

The forward function of the TransformerEncoderLayer can have src_key_padding_mask. Maybe we can update it too.