Question about always_pw in IRFBlock
andyhahaha opened this issue · comments
andyhahaha commented
According to the block architecture in the paper, seems that FBNet uses always_pw=True.
But in this repo, the release FBNet model use always_pw=False.
Why don't use the same architecture as paper?
Does always_pw=True or False affect the accuracy a lot?
Using always_pw=True can't reproduce the results on the paper?
Thanks.
Stzpz commented
We use always_pw = True for the models in the paper, which are released as the caffe2 models. Later when we reimplemented the models using pytorch, we noticed always_pw = False produces almost the the same results but with slightly less Flops, so we use that as default in pytorch models.
Bests,
… 在 2020年7月15日,下午10:57,andyhahaha ***@***.***> 写道:
According to the block architecture in the paper, seems that FBNet uses always_pw=True.
But in this repo, the release FBNet model use always_pw=False.
Why don't use the same architecture as paper?
Does always_pw=True or False affect the accuracy a lot?
Using always_pw=True can't reproduce the results on the paper?
Thanks.
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub, or unsubscribe.