Closed JunweiLiang closed 2 years ago
Hi, you can follow cal_rel_pos_spatial
to add a separate temporal rel pos embedding.
Our official spatial-temporal rel pos embedding will be released in PySlowFast (our video codebase) soon together with the MViTv2 video models.
Any suggestions on how to implement spatial-temporal relative positional embeddings? I'm trying to extend based on the
cal_rel_pos_spatial
function inattention.py
Thanks, Junwei