alexiajm / deep-learning-with-cats Goto Github PK

source code:
https://github.com/bioinf-jku/TTUR/blob/ca9b6572d08f81d0725de8558400fb17585266d3/WGAN_GP/gan_64x64_FID.py#L45-L49

I am confusing about the code in training Discriminator phase:

errD_real = D(x)
errD_real.backward(one)
...
errD_fake = D(x_fake)
errD_fake.backward(one_neg)

But D is $maxV(G,D)=E(D(real))-E(D(fake))$ , i think in “loss form” is inverse

errD_real = D(x)
errD_real.backward(one_neg)
...
errD_fake = D(x_fake)
errD_fake.backward(one)

Maybe you could do a simple unweighted average pixel difference over the dataset for the example cats and see if it is not just overfitting on the training data?
I would almost say they don't look distorted enough to be really generated :)

When we generate the same number of images as batch_size during training, it performs very well. But when we only want to generate one picture, there is a problem with its brightness. We also set the batch_size to 64 when generating the picture, but this problem still exists when saving a picture. We think it is the normalization of the following line of code.
vutils.save_image(fake_test.data[0:1,:,:,:], './output/%01d.png' % i , normalize=True)
How can we store one image at a time?
We are looking forward to your reply.

the link to cat dataset returns "503 service unavailable"

the Cat Dataset (https://web.archive.org/web/20150703060412/http://137.189.35.203/WebUI/CatDatabase/catData.html)
unable to download.

I'm interested in your research，But unfortunately, the link of the downloaded data set you provided is invalid. I can only go to the web page and cannot download the data. Could you please provide other downloads?

for the wgan-gp?
Thanks!

My mac os python3.6. How to solve this?

Traceback (most recent call last):
File "Meow_DCGAN.py", line 267, in
for i, data_batch in enumerate(dataset, 0):
File "/usr/local/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 212, in next
return self._process_next_batch(batch)
File "/usr/local/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 239, in _process_next_batch
raise batch.exc_type(batch.exc_msg)
TypeError: Traceback (most recent call last):
File "/usr/local/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 41, in _worker_loop
samples = collate_fn([dataset[i] for i in batch_indices])
File "/usr/local/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 41, in
samples = collate_fn([dataset[i] for i in batch_indices])
File "/usr/local/lib/python3.6/site-packages/torchvision/datasets/folder.py", line 67, in getitem
img = self.transform(img)
File "/usr/local/lib/python3.6/site-packages/torchvision/transforms.py", line 29, in call
img = t(img)
File "/usr/local/lib/python3.6/site-packages/torchvision/transforms.py", line 139, in call
ow = int(self.size * w / h)
TypeError: unsupported operand type(s) for /: 'tuple' and 'int'

Thanks for the good code!!!

Can anybody explain the difference between training the real and fake date separately and training them simultaneously. The code backwards the real first then the fake.

Very nice cats! Have you thought of trying BEGAN https://arxiv.org/abs/1703.10717 ? Their results on faces look great.

Not sure what I'm missing but I keep getting this:

File "DCGAN.py", line 43
base_dir = f"{param.output_folder}/run-{run}"

I really like your post and the code. In your post, you mentioned that WGAN-GP may cause the image to be blurry. Have you solved the problem or it is due to the Wasserstein loss? Thanks!

first, i want to say thanks so much to Alexia. Now, t need make a folder contain subfolders save images, which be generated by DCGAN. I tried but not success. Help me.

Shouldn't the weight initialization for SELU be something like:

def selu_weights_init(m):
    classname = m.__class__.__name__
    if classname.find('Conv') != -1:
        m.weight.data.normal_(0.0, 0.5 / math.sqrt(m.weight.numel()))

    elif classname.find('BatchNorm') != -1:
        size = m.weight.size()
        fan_out = size[0] # number of rows
        fan_in = size[1] # number of columns

        m.weight.data.normal_(0.0, 1.0 / math.sqrt(fan_in))
        # Estimated mean, must be around 0
        m.bias.data.fill_(0)

(The 0.5 factor for the conv. coming from reading the PyTorch forums about what worked for someone, in other places 1.0 is used)