WGAN outputs only noises

(Mirodil) #1

Hi,

I am doing a small experiment. I am trying to use WGAN example from lesson 7. I created my own dataset for this purpose. I am using the following code snippet, but the outputs for every epoch are the same, noises.

def get_data(bs, size):
    return (GANItemList.from_folder(path, noise_sz=100)
               .no_split()
               .label_from_func(noop)
               .transform(tfms=[[crop_pad(size=size, row_pct=(0,1), col_pct=(0,1))], []], size=size, tfm_y=True)
               .databunch(bs=bs)
               .normalize(stats = [torch.tensor([0.5,0.5,0.5]), torch.tensor([0.5,0.5,0.5])], do_x=False, do_y=True))

data = get_data(4, 64)

generator = basic_generator(in_size=64, n_channels=3, n_extra_layers=1)
critic    = basic_critic(in_size=64, n_channels=3, n_extra_layers=1)

learn = GANLearner.wgan(data, generator, critic, switch_eval=False, opt_func = partial(optim.Adam, betas = (0.,0.99)), wd=0.)

learn.fit(30,2e-4)

output:

|epoch|train_loss|gen_loss|disc_loss|time|
|---|---|---|---|---|
|1|-0.490096|0.378111|-0.666829|00:13|
|2|-0.835214|0.540004|-1.113979|00:15|
|3|-1.000573|0.620919|-1.334438|00:17|
....
|27|-1.182753|0.798620|-1.604142|00:17|
|28|-1.182750|0.799152|-1.604292|00:19|
|29|-1.182985|0.799610|-1.604540|00:18|
|30|-1.182571|0.800078|-1.604233|00:18|

What do you think what could be the cause of these noises? Do I need to tweak some parameters? or create bigger dataset?

Any help would be greatly appreciated.

0 Likes

(Mirodil) #2

I solved by setting None to transformations:

.transform(tfms=None, size=size, tfm_y=True)
0 Likes

(p) #3

I am having this same issue. How large of a dataset did you build? @Mirodil

0 Likes

(Mirodil) #4

Dataset was not big around 1000(128x128) images.

0 Likes

(stephen oni) #5

not working for me, i dont know why

0 Likes

(p) #6

How many epochs did it take you/what batch size did you use? I’m having the same issue but only 125 images.

My code looks like:

def get_data(bs, size):
return (GANItemList.from_folder(path, noise_sz=1000)
.split_none()
.label_from_func(noop)
.transform(tfms=None, size=size)
.databunch(bs=bs)
.normalize(stats = [torch.tensor([0.5,0.5,0.5]), torch.tensor([0.5,0.5,0.5])], do_x=False, do_y=False))

data = get_data(20, 512)

0 Likes

(Mirodil) #7

As I ran it in CPU I had to use small batch size 4 images of 32x32 size, and I ran 10 epoch to see something else then noises.

0 Likes

(p) #8

It seems like this is a classic issue with GANs related to the generator loss not decreasing. I increased the learning rate and it’s working a bit better now.

1 Like