PyTorch-GAN
PyTorch-GAN copied to clipboard
eriklindernoren
Problem in DCGAN
DCGAN fails learning the mnist dataset. Is there a problem in implementation.
Very vague issue. What is happening instead of it learning the dataset? I ran the DCGAN code [as is] and had legible results in only 10 epochs.
while discriminator's error decreases to 0 rapidly, generator's error increases to ~10. generated images are so poor. I tried lowering the learning rate of the discriminator, or using nesterov instead of adam but still generated images are so bad.
an example 90 epoch with nesterov, 3e-4 discriminator error, 8.5 generator error.
When I run gan.py, it's ok. But when I run dcgan.py, something is wrong. RuntimeError: cuda runtime error (2) : out of memory at/pytorch/aten/src/THC/generic/THCStorage.cu:58 I reduce the batch size, but the mistake still exists. Waht should I do? thank you for giving me some advice.
You can refer to the implementation details for DCGAN in my repository(https://github.com/Lornatang/Deep-Convolutional-Generative-Adversarial-Network)。 @githubcyy
