About Temporal Sampling on Features (TSF)

Thanks for the wonderful work, Haoning! There's one question about the Temporal Sampling on Features (TSF) in the paper, since the original token set has been sampled in this stage, how to get the **Tpe** whose shape is _B*N*C_ (as far as I'm concerned, the dimension N should be reduced?).

Ps. It would be great if it was convenient for you to open source the codes. 