how to gradient descent

[shz1314]

hi @ZZUTK
i have a problem about your model's training.first i remove the GAN loss,just use autoencorder to generate blurred image.Then set other weight of loss from 0 to 0.0001 ,but the EG_err=0.2870 at 50th epoch.

thanks
Song

[susanqq]

Can you try to set the stddev for weight init as 0.02?

[shz1314]

Thanks @susanqq
can you tell me where should i add it, the stddev for weight.
so thanks,
Song

[shz1314]

OH i find ,do you mean the weight of 'fc' 'conv''deconv'

[susanqq]

yes, please change it in the ops.py

[shz1314]

hi,
shall you send me the pretrained model to me, i want to do the contrast experiment with your model.My email is haoze_song_jnu@163.com .
Yours Haoze