PyTorch-CycleGAN icon indicating copy to clipboard operation
PyTorch-CycleGAN copied to clipboard
verified publisher icon aitorzip

Minor fixes

Open tanvach opened this issue 7 years ago • 1 comments

Fixes two small issues I came across before I could get the script to run:

  • Pytorch now treats true scalar (dimension [1]) differently to a singleton (dimension [1x1]). This broke the loss computation at loss_GAN_A2B = criterion_GAN(pred_fake, target_real) step. Fixed by making target_real and target_fake tensors.
  • Method to start Visdom is now updated

@aitorzip

tanvach avatar May 20 '18 05:05 tanvach

Thank you @tanvach ! Your patch also does the job on my setup.

Pandinosaurus avatar Jul 04 '18 13:07 Pandinosaurus