raoyongming / HorNet

[NeurIPS 2022] HorNet: Efficient High-Order Spatial Interactions with Recursive Gated Convolutions

Home Page:https://hornet.ivg-research.xyz/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

i have some problem with using GlobalLocalFilterFPN

870231652 opened this issue · comments

commented

when i use 7x7 filter,there is no question。but if i use GlobalLocalFilter,There will be a lot of alarms

image
and it is related to the val
image

Hi, thanks for your interest in our work. It seems the input tensor and the complex weights in GFNet have different dtypes (float vs. half). I didn't find the line self.pws[i](x).to(torch.float32) in our released code. I think the problem may be caused by the data type convert.

commented

Thank you for your answer.self.pwsi.to(torch.float32),this sentence is writing in this code by myself.With or without, The warning has not changed。The reason I write this sentence is that I want to address this warning。To ensure consistency with your code,I re-entered the gnconv code, which was copied from github。The mistake is repeated。
image
image
Maybe it has something to do with this piece of code I wrote?I really appreciate your kindness!

commented

This problem is caused by my code,sorry to bother you。issue was settled。