search

Atze00 / MoViNet-pytorch

MoViNets PyTorch implementation: Mobile Video Networks for Efficient Video Recognition;
MIT License
258 stars 50 forks source link

There seems no implementation of positional_encoding #32

Closed zhouruirui1998 closed 2 years ago

zhouruirui1998 commented 2 years ago

Hello, thanks for the PyTorch implementation! There seems no implementation of positional_encoding,does that influence in the stream model?

zhouruirui1998 commented 2 years ago

sorry,I just realized A0-A3 don't need positional_encoding

gabrielegoletto commented 2 years ago

Hi @zhouruirui1998, could you please explain why you mentioned that they are not needed with A0-A3?

sorry,I just realized A0-A3 don't need positional_encoding

mkmohangb commented 2 years ago

@ezius07 From here:

Note: A3, A4, and A5 models use a positional encoding in the squeeze-excitation blocks, while A0, A1, and A2 do not. For the smaller models, accuracy is unaffected without positional encoding, while for the larger models accuracy is significantly worse without positional encoding.

gabrielegoletto commented 2 years ago

@usernameandme I had not seen it! Many thanks!