#This generator refuses to learn

6 messages · Page 1 of 1 (latest)

undone rune
#

So i was following the zero to gans course, completed it. Wanted to adjust the code to generate the mnist integers instead of anime faces. I wrote this generator

    # in: latent_size x 1 x 1

    nn.ConvTranspose2d(latent_size, 512, kernel_size=3, stride=1, padding=0, bias=False),
    nn.BatchNorm2d(512),
    nn.ReLU(True),
    # out: 512 x 3 x 3

    nn.ConvTranspose2d(512, 256, kernel_size=3, stride=2, padding=0, bias=False),
    nn.BatchNorm2d(256),
    nn.ReLU(True),
    # out: 256 x 7 x 7

    nn.ConvTranspose2d(256, 128, kernel_size=4, stride=2, padding=1, bias=False),
    nn.BatchNorm2d(128),
    nn.ReLU(True),
    # out: 128 x 14 x 14

    # Self attention layer
    nn.Conv2d(128, 128, kernel_size=1, stride=1, padding=0, bias=False),
    nn.BatchNorm2d(128),
    nn.ReLU(True),
    # out: 128 x 14 x 14

    nn.ConvTranspose2d(128, 1, kernel_size=4, stride=2, padding=1, bias=False),    
    nn.Tanh()
    # out: 1 x 28 x 28
)```
However it just refuses to train?

If i use this one (and resize the mnist images to 64x64) it works
```generator = nn.Sequential(
    # in: latent_size x 1 x 1

    nn.ConvTranspose2d(latent_size, 512, kernel_size=4, stride=1, padding=0, bias=False),
    nn.BatchNorm2d(512),
    nn.ReLU(True),
    # out: 512 x 4 x 4

    nn.ConvTranspose2d(512, 256, kernel_size=4, stride=2, padding=1, bias=False),
    nn.BatchNorm2d(256),
    nn.ReLU(True),
    # out: 256 x 8 x 8

    nn.ConvTranspose2d(256, 128, kernel_size=4, stride=2, padding=1, bias=False),
    nn.BatchNorm2d(128),
    nn.ReLU(True),
    # out: 128 x 16 x 16

    nn.ConvTranspose2d(128, 64, kernel_size=4, stride=2, padding=1, bias=False),
    nn.BatchNorm2d(64),
    nn.ReLU(True),
    # out: 64 x 32 x 32

    nn.ConvTranspose2d(64, 1, kernel_size=4, stride=2, padding=1, bias=False),
    nn.Tanh()
    # out: 1 x 64 x 64
)
generated-images-0010.png
#

This is what it looks like after (1-10) epochs with the first code snippet

#

With the second snippet it works fine (1 epoch)

generated-images-0001.png
daring storm
#

I don't really see anything wrong with the first snippet, are you sure this is the only thing that is changed between the two experiments?

Maybe you need to adjust the learning rate if the total number of parameters is different.

undone rune
#

the discriminator is also changed ofcourse corressponding to the new size