You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
After resizing from 96 to 64 then the images are not belong to 0-255 any more, then doing the "img/127.5 + 1" process not exactly obey the thick transform img to -1~1, don't know weather it matters
The 'x + eps * (G(z) - x)' step, the dimention of eps is '[batch_size, h, w, c]' or
'[batchsize, 1, 1, 1] and then broadcast to [batch_size, h, w, c]' ?
2. ACGAN in mnist with WGAN-GP loss
What is learned
BN 跟在 RuLU后边是坠吼的
网络结构这一块是试出来的,对 G 不加 z 和 label 的 embedding 会崩掉。也可能实际上是每个model对应的lr等超参数没调好。