eriklindernoren / pytorch-gan Goto Github PK

Sorry this is not an issue. I have some questions regarding the implementation of ACGAN and training ACGAN.

1. In your implementation the encoded label vector is multiplied with the noise vector and given as the input to the G. But shouldn't it be concatenated?
2. The CrossEntropy loss in PyTorch already includes a softmax function. Therefore I am unclear whether a softmax function should be included in the Discriminator or not.
3. I am unclear about when to stop training. It seems (https://pytorch.org/tutorials/beginner/dcgan_faces_tutorial.html) the accuracy of the D for real vs fake images will be initially high for the real images, but low for the fake images. However, when I run ACGAN both are high initially (over 95%) and then reduces to ~70%. This accuracy does not go below 60%. So do you think my training is correct? Also I am unclear about when to stop training. May be this accuracy close to 50%, or the loss of the D converges, or loss of the G converges..?

Hi, I have a question about the cgan implementation.
In your code, you use nn.embedding to embed the prior labels. The problem is, when the learnable weights are not specified, the vocabulary will be randomly initialized.

In both generator and discriminator, you use two different nn.embedding, and they are initialized differently. However, when we generate a fake image, we use one embedding, but when we use discriminator to distinguish the fake image, we use another embedding. Will this have effect on the final performance?

I am not very familiar with GAN. But I just think this is strange. It's true that we still use the same labels, but the actual embeddings are different. I think using the same embedding for the discriminator and generator will be more reasonable?

Thank you very much for sharing such a great code!Will it be better to use if __name__ == '__main__':?

I am having this error using horse2zebra, I have checked all the input images and they all have the size [3, 265, 265], so, most probable that the error is caused by G_AB(real_B). I have, however, tried cyclegan with cifar-100 and monet2photo and everything went fine. I am using PyTorch 0.4 and Python 3.6.

line 157, in <module>
    loss_id_B = criterion_identity(G_AB(real_B), real_B)
...
Given groups=1, weight of size [64, 3, 7, 7], expected input[1, 1, 262, 262] to have 3 channels, but got 1 channels instead

Running wgan.py:

Traceback (most recent call last):
  File "wgan.py", line 179, in <module>
    gen_validity.backward(valid)
  File "/home/alcaster/.pyenv/versions/ml/lib/python3.6/site-packages/torch/tensor.py", line 93, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph)
  File "/home/alcaster/.pyenv/versions/ml/lib/python3.6/site-packages/torch/autograd/__init__.py", line 89, in backward
    allow_unreachable=True)  # allow_unreachable flag
RuntimeError: Trying to backward through the graph a second time, but the buffers have already been freed. Specify retain_graph=True when calling backward the first time.

My versions of packages:
torch==0.4.0
torchvision==0.2.1

In your WGAN implementation, you have your n-critique loop around the Generator learning when it should actually be around the Discriminator's. Critique stands for Discriminator.

When evaluating the ESRGAN and SRGAN I can see that the class FeatureExtractor() is not defined anywhere. I can see latest commit is 13 days ago, so I assume you are currently working on implementing these models?

The dataset in the Drobox can't be downloaded.

I'm a little confused about why use mse_loss in GAN loss and use L1 loss in Cycle loss and Identity loss,and i didn't find this in the paper.
And the second question is why use the fake_A_buffer.push_and_pop() ,it seems to do something like if the len<50 do nothing,and when len>50,the part of >50 do the random choice the sample? i really confused about this

The source code of Identity loss is shown below:
loss_id_A = criterion_identity(G_BA(real_A), real_A)
loss_id_B = criterion_identity(G_AB(real_B), real_B)

This seems a little bit weird to me, maybe it should be:
loss_id_A = criterion_identity(G_AB(real_A), real_A)
loss_id_B = criterion_identity(G_BA(real_B), real_B)

Hi, thanks for every contributor!
I was wondering if this repo could add some GAN implementation in NLP field?
Is it possible to add DPGAN?

and TextGAN

I wish i could contribute but i am new to pyotch. I would be happy to see pytorch implementation of these GANs!

It raises RuntimeError: cannot join current thread when the whole training process ends.
Nothing was changed in the code.

What exactly is it? Here can I learn about it?

Where have you told pytorch that you are going to use channel first image? Previously I used channel last image with pytorch.

@eriklindernoren According to the Paper the coupled discriminators should share the parameters of the last layers, but you implemented it to share the parameters of the first layers (just like the generator does).

The L2 norm of the gradient penalty term in WGAN-GP should be calculated across all dimensions of an image, but the current implementation calculates it across each dimension of an image separately (i.e., the absolute value of each pixel in an image is calculated).

Indeed, in the following line, gradients is a tensor of size (batch_size, nb_channels, img_width, img_height):

gradient_penalty = ((gradients.norm(2, dim=1) - 1) ** 2).mean()

To solve the issue, the 4-dimensional tensor containing the gradients should be flattened across the last 3 dimensions:

gradients = gradients.view(real_samples.size(0), -1)
gradient_penalty = ((gradients.norm(2, dim=1) - 1) ** 2).mean()

Hi~

I found an error while using InfoGAN. The error is as follows. Hope to repair.

/usr/local/lib/python3.7/site-packages/torch/nn/modules/upsampling.py:129: UserWarning: nn.Upsample is deprecated. Use nn.functional.interpolate instead.
  warnings.warn("nn.{} is deprecated. Use nn.functional.interpolate instead.".format(self.name))
/usr/local/lib/python3.7/site-packages/torch/nn/modules/container.py:92: UserWarning: Implicit dimension choice for softmax has been deprecated. Change the call to include dim=X as an argument.
  input = module(input)

Thank you for writing the code, gave me a lot of inspiration!

Hi there, thanks very much for the wonderful repo.
When I want to download the img_align_celeba.zip from Dropbox, I found the link is turned off. So can you update the link or share me a private link for downloading the dataset?

Thanks very much.

I run your code about cgan and the generated image looks quite perfect. However, when I tried to classified the generated image by another network (Resnet18), the predicted label is always 'eight'. Is this a common feature of cgan?

Hello, I noticed that within your implementation of pix2pix, Datasets will return image in form of like this

But when reading it again in training loop, it was written like

This happened multiple times within pix2pix.py when loading the image. Is this switching intentional?

Hello and my thanks for this great repo. I like how your code is simple and effective.

I would like to point out that you are using MSE for your Discriminator Loss instead of Binary Cross Entropy. If you have a specific reason for why you are doing that, could you share it?

I can't git clone https://github.com/eriklindernoren/PyTorch-GAN
Error message is below.
fatal: could not create work tree dir 'PyTorch-GAN'

Why does this error happen?

when I run python3 wgan.py, appears print ("[Epoch %d/%d] [Batch %d/%d] [D loss: %f] [G loss: %f]" % (epoch, opt.n_epochs,batches_done % len(dataloader), len(dataloader),d_loss.item(), gen_validity.item()))
ValueError: only one element tensors can be converted to Python scalars
how can i fix it?

In line 265 the code does not look right :

code_input = Variable(FloatTensor(np.random.normal(-1, 1, (batch_size, opt.code_dim))))

instead of 'normal' shouldn't be 'uniform'?

• Mirtha

In the line 161 of the 'began' implementation, should that be...?

g_loss = torch.mean(torch.abs(discriminator(gen_imgs) - real_imgs))

instead of

g_loss = torch.mean(torch.abs(discriminator(gen_imgs) - gen_imgs))

Please correct me if I miss something. Many thanks!

in https://github.com/eriklindernoren/PyTorch-GAN/blob/master/implementations/relativistic_gan/relativistic_gan.py

        if opt.rel_avg_gan:
            g_loss = adversarial_loss(fake_pred - real_pred.mean(0, keepdim=True), valid)
        else:
            g_loss = adversarial_loss(fake_pred - real_pred, valid)

        # Loss measures generator's ability to fool the discriminator
        g_loss = adversarial_loss(discriminator(gen_imgs), valid)

        g_loss.backward()
        optimizer_G.step()

Is this expected? Does it look like g_loss is getting overwritten?

While calling Discriminator() from models.py in cycleGan and pix2pix, I get a syntax error response as ;

Traceback (most recent call last):
  File "cyclegan.py", line 15, in <module>
    from models import *
  File "./PyTorch-GAN/implementations/cyclegan/models.py", line 167
    *discriminator_block(64, 128, 2, True),
    ^
SyntaxError: invalid syntax

This happens with both python3 and python2.7.
Looks like dereferencing does not work and I could not find a way to make it work.

Any bits of advice?

Thanks

An attempt has been made to start a new process before the
current process has finished its bootstrapping phase.

    This probably means that you are not using fork to start your
    child processes and you have forgotten to use the proper idiom
    in the main module:

        if __name__ == '__main__':
            freeze_support()
            ...

    The "freeze_support()" line can be omitted if the program
    is not going to be frozen to produce an executable.
ForkingPickler(file, protocol).dump(obj)

BrokenPipeError: [Errno 32] Broken pipe

Namespace(b1=0.5, b2=0.999, batch_size=1, channels=3, checkpoint_interval=-1, dataset_name='facades', decay_epoch=100, epoch=0, img_height=256, img_width=256, lr=0.0002, n_cpu=8, n_epochs=200, sample_interval=500)

Namespace(b1=0.5, b2=0.999, batch_size=1, channels=3, checkpoint_interval=-1, dataset_name='facades', decay_epoch=100, epoch=0, img_height=256, img_width=256, lr=0.0002, n_cpu=8, n_epochs=200, sample_interval=500)

Abnormal to me, this line was printed twice

Traceback (most recent call last):
File "gan/gan.py", line 154, in
d_loss.item(), g_loss.item()))
File "/usr/local/lib/python3.5/dist-packages/torch/autograd/variable.py", line 67, in getattr
return object.getattribute(self, name)
AttributeError: 'Variable' object has no attribute 'item'

I notice that the generated images have higher brightness and more colors than the original image, or the resulting images of other approaches. What causes it?

@eriklindernoren I think you should put .detach() to real_samples and fake_samples. Isn't it?

Hi
Thank you for your wonderful effort in implementing so many papers.
I have a query regarding your EBGAN implementation.
https://github.com/eriklindernoren/PyTorch-GAN/blob/master/implementations/ebgan/ebgan.py

In line 175 when you are optimizing the generator G why is the pixelwise_loss computed using gen_imgs.detach() and why not simply gen_imgs() ? If we do .detach() while updating G then the pixelwise_loss will not contribute any gradients towards optimizing the generator weights. Is it the right way to do that ?

Please clarify my doubt.
Thank You in Advance !

hi,your code is so amazing.I want to use your WGAN_gp code to generate my images

Original paper used a noise with the input of the generator. why did you not use it?

Hi,
I was wondering which GPUs did you use for cycleGAN?

Thanks!

acgan.py --- sgan.py

DCGAN fails learning the mnist dataset. Is there a problem in implementation.

hi, i believe the implementation of wgan-gp is buggy. the interpolation in

i believe the line should be replaced by
alpha = Tensor(np.random.random(size=(real_samples.shape[0], 1, 1, 1)))

hi, i am wonder in cgan. the auxiliary_loss is not use in optimizer_G.step(). But the CGAN can training normal and get correct result. I think, Maybe i overlook some significant details. So who can give me some tips. thanks !

I'm running the srgan.py implementation and receive the following error:

Namespace(b1=0.5, b2=0.999, batch_size=1, channels=3, checkpoint_interval=-1, dataset_name='img_align_celeba', decay_epoch=100, epoch=0, hr_height=256, hr_width=256, lr=0.0002, n_cpu=8, n_epochs=200, sample_interval=100)
Downloading: "https://download.pytorch.org/models/vgg19-dcbb9e9d.pth" to /Users/Username/.torch/models/vgg19-dcbb9e9d.pth
100%|███████████████████████████████████████████| 574673361/574673361 [00:34<00:00, 16895858.86it/s]
Traceback (most recent call last):
  File "srgan.py", line 104, in <module>
    lr_transforms = [   transforms.Resize((opt.hr_height//4, opt.hr_height//4), Image.BICUBIC),
AttributeError: module 'torchvision.transforms' has no attribute 'Resize'

If you need any additional information let me know...

1. It can be summarized in the chronological order of the papers rather than the first letter of the model so that it is easier to study the related development of GAN more clearly.

2. It would be better if you can briefly summarize the connections and differences between the models in papers.

I run CycleGAN following all commands the author gave. But there is empty in images and saved_models. I tested that the cyclegan.py didn't enter the 140 line "for i, batch in enumerate(dataloader):".
Any advices is appreciated.

Hi:
Thanks for sharing your impressive repo, I found the loss_D in wan.py is :
loss_D = -torch.mean(discriminator(real_imgs)) + torch.mean(discriminator(fake_imgs)) ,but the wgan's paper described the loss_D as follow:

So is it a bug or other reasons？

Hi @eriklindernoren and all,

Thanks to all contributor for the awesome repository.

@eriklindernoren the loss of the discriminator within the current implementation seems not consistent with the Algorithm 1 of WGAN-div.

tensor.sub_(mean[:, None, None]).div_(std[:, None, None])
RuntimeError: output with shape [1, 32, 32] doesn't match the broadcast shape [3, 32, 32]

When I was runing the DCGANS example, it gives me this error.