question about 'wh_offset_base'

[xieenze]

Hi, thanks for your great work. I am confused that 'wh_offset_base' =16 in ttf_head.py, but I don't know why you need to set this parameter.

Why do you need to multiply it in the regression head? Thanks.

[liuzili97]

Hi. We use it to enlarge the prediction since directly predicting large values is hard. After using this parameter, we just need to predict the 1/16 of the values, which can be easier.

[xieenze]

@liuzili97 Thanks for quick reply. But the thing is that I do not find you use this parameter during inference. I think 'wh_offset_base' should also be used during inference. =-=

[liuzili97]

This parameter is used in both training and inference since the forward will be called anyway.

[xieenze]

Thanks. I get it.