search

zczcwh / POTTER

The project is an official implementation of our paper "POTTER: Pooling Attention Transformer for Efficient Human Mesh Recovery".
44 stars 1 forks source link

Selection of embed-wise pooling attention dimension #6

Closed jinxuzh closed 8 months ago

jinxuzh commented 1 year ago

Hi, appreciate the amazing works! I have one questions about embed-wise pooling attention: how do you determine the embedding dimensions D_h and D_w? In the code D_h is hardcoded as 32, is there any reasons or intuitions why you choose this value? Also, have you ever tried using other dimensions (e.g. D_h=8, which results in a squared embedded feature if D=64) and see if the performance would be improved?

zczcwh commented 1 year ago

Hi, you may try different num of dimensions. I chose 32 just because it's suitable for most cases.