Understanding TokenLearner

[AshStuff]

I am trying to understand the tokenlearner and I guess the current implementation is different from the original TokenLearner proposed right ? I guess the original Tokenlearner performs the token learner for each temporal dimension and then reshape back.

[920232796]

In this work, token learner only controls the sequence length. It doesn't need to reshape back because the network has a segmentation decoder.