Greyscale modifications? about pro_gan_pytorch

[akanimax]

@msmacpherson,
Sorry, you are right. Currently, there is no way for changing the number of channels of the output generated images. In case of MNIST, I did the same things (replicating the channels thrice). You could either do this (works pretty well), or change a little bit of code to include single channel images as well. Please feel free to open a PR here if you do implement it.

In my case, I was going to do it, but procrastination got in the way ... you know 😆. Besides, colour images are more fun you know 😄.

Glad that this repo is helpful for you.

Cheers 🍻!
@akanimax

[msmacpherson]

Thanks a lot! I'm spending my whole life trying to make tensors match up at the moment. If you don't mind I also found an error when i run the ProGan (unconditional mode) with cifar10 (I just removed classes from the command line to run using your code otherwise):

pro_gan = pg.ProGAN(depth=depth,
latent_size=latent_size, device=device)

Which gives me this error:

Files already downloaded and verified
Starting the training process ...

Currently working on Depth: 0
Current resolution: 4 x 4

Epoch: 1
Traceback (most recent call last):

File "", line 62, in
batch_sizes=batch_sizes

File "/home/matt/MSC/Progan/pro_gan_pytorch/PRO_GAN.py", line 607, in train
images = batch.to(self.device)

AttributeError: 'list' object has no attribute 'to'

Is there an additional preprocessing required for unconditional training please? For my project I'm looking at single class x-ray images initially so I only have a single image class.