arnab39 / cyclegan-pytorch

A clean and lucid implementation of cycleGAN using PyTorch

CycleGAN using PyTorch

CycleGAN is one of the most interesting works I have read. Although the idea behind cycleGAN looks quite intuitive after you read the paper: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks, the official PyTorch implementation by junyanz is difficult to understand for beginners (The code is really well written but it's just that it has multiple things implemented together). As I am writing a simpler version of the code for some other work, I thought of making my version of cycleGAN public for those who are looking for an easier implementation of the paper.

All the credit goes to the authors of the paper. This code is inspired by the actual implementation by junyanz which can be found here.

Requirements

• The code has been written in Python (3.5.2) and PyTorch (0.4.1)

How to run

• To download datasets (eg. horse2zebra)

$ sh ./download_dataset.sh horse2zebra

• To run training

$ python main.py --training True

• To run testing

$ python main.py --testing True

• Try tweaking the arguments to get best performance according to the dataset.

Results

• For horse to zebra dataset. ( Real - Generated - Reconstructed)

cyclegan-pytorch's Issues

RuntimeError: Given groups=1, weight of size [64, 1, 7, 7], expected input[5, 64, 70, 70] to have 1 channels, but got 64 channels instead

Hello,

I am getting this error.
I have implemented your code ,but written a custom dataset loader for my image patches.

Can you tell me how can I solve this error

About setting grad to false in train iteration.

In training, discriminator gradient is set to false while generator is updating.
But I tried with no discriminator lock, and it it seems the result was same.
I wonder the lock code needs in PyTorch?

OSError: [Errno 95] Operation not supported

Whenever i try to run it on colab it shows:
File "main.py", line 40, in main
create_link(args.dataset_dir)
File "/content/drive/My Drive/cycleGAN-PyTorch/utils.py", line 37, in create_link
os.path.join(dirs[key], 'Link'))

Operation not supported: '/content/drive/My Drive/cycleGAN-PyTorch/datasets/horse2zebra/trainA' -> '/content/drive/My Drive/cycleGAN-PyTorch/datasets/horse2zebra/ltrainA/Link'

what can i do?

Is this mode collapse?

Do you think this is mode collapse? (Training: 50% epoch: 100/200)
horse2zebra

I want to know how to Improve this problem, thanks.

OSError: image file is truncated

Hi , I ma getting an error for OSError: image file is truncated. for my dataset.
Could you please advice. Thanks

how to train the model on CPU ??

Currently the code use GPU by default. Can we train it by using CPU??

Confusing naming

Hi, your code is very clean thanks!

But I think that naming can be confusing. At https://github.com/arnab39/cycleGAN-PyTorch/blob/master/model.py#L107 you named a_fake = self.Gab(b_real), the generator of a to b taking a b example should not generate the identity (a_idt)? In all examples I show, the followed nomenclature for generators is this Gxy, generator from x to y...

Assuming that, all generators naming are reversed. What do you think?

Need to add B-->A in testing?

Hello, great code.
I think you should add one more option to map from B to A in testing

cycleGAN-PyTorch/test.py

Line 24 in ecdc973

a_test_data = dsets.ImageFolder(dataset_dirs['testA'], transform=transform)

Besides, we do not need to shuffle in testing.

    a_test_data = dsets.ImageFolder(dataset_dirs['testA'], transform=transform)
    b_test_data = dsets.ImageFolder(dataset_dirs['testB'], transform=transform)


    a_test_loader = torch.utils.data.DataLoader(a_test_data, batch_size=args.batch_size, shuffle=True, num_workers=4)
    b_test_loader = torch.utils.data.DataLoader(b_test_data, batch_size=args.batch_size, shuffle=True, num_workers=4)

You save my life~

What does offset mean in 'LambdaLR' class?

You are setting '0′ to the offset parameter at 'utils.LambdaLR'.
I don't know how to use it. Is it always '0' even if training was continued?

Running on Google Colab, create_link error occurs.

I run main.py on Google Colab , and I got the message bellow.

File "/content/grive/My Drive/CycleGAN/cycleGAN-PyTorch/utils.py", line 37, in create_link os.path.join(dirs[key], 'Link')) OSError: [Errno 95] Operation not supported: '/content/grive/My Drive/CycleGAN/cycleGAN-PyTorch/datasets/horse2zebra/trainA' -> './datasets/horse2zebra/ltrainA/Link'

I think it's better to copy trainA to ltrainA without using the link.