CFNet icon indicating copy to clipboard operation
CFNet copied to clipboard

Gamma, Beta in the model weight

Open WANG-KX opened this issue 4 years ago • 6 comments

Why the value of gamma_s3, gamma_s2, beta_s3, beta_s2 are all zeros in your provided model weights? If they are all zeros, meaning that they are not functional, right?

WANG-KX avatar Sep 07 '21 06:09 WANG-KX

Hello, gamma and beta are normalization factors, which is initialized as 0 and gradually learns a weight.

gallenszl avatar Sep 07 '21 06:09 gallenszl

I mean in your provided trained weights, they are still zeros.

WANG-KX avatar Sep 07 '21 06:09 WANG-KX

This is a little strange and I will check it later.

gallenszl avatar Sep 07 '21 06:09 gallenszl

The gradients of gamma_s3, gamma_s2, beta_s3, beta_s2 are None. Is that expected?

callmeray avatar Apr 19 '22 10:04 callmeray

The gradients of gamma_s3, gamma_s2, beta_s3, beta_s2 are None. Is that expected?

I think these parameters are not updated during training, and the parameters from the provided trained weights are all zeros.

wangpinzhi avatar May 19 '23 07:05 wangpinzhi

I also found this problem :the gradients of gamma_s3, gamma_s2, beta_s3, beta_s2 are None ,and they are not updated during training.it will lead to poor training results

august779188 avatar Dec 19 '23 03:12 august779188