Open mjack3 opened 2 years ago
Thank you for your comment and sorry for late reply.
I saw your repo working on FastFlow reproduction and using AllInOneBlock
as backend.
There are mainly two reasons why we choose GLOW design. First, according to the paper, FastFlow block has an activation normalization layer followed by channel permute layer, and consequent two "coupling layers”. We assume this is the very architecture of GLOW. Second, GLOW could stabilize learning phase in our environment.
We would be happy to hear your views on this.
Hello, I agree with you. I used GLOW after having studied this in deep with 1.2 clamping. and results turned up very good.
Hello. Is good to know that there's people trying to reproduce this work. I'm also working on it.