Taking too much training time for CycleGAN

I am trying to reimplement CycleGAN from this GitHub: (https://github.com/eriklindernoren/PyTorch-GAN#cyclegan) but it’s taking showing estimated time in days. I am using GPU, is there any way to optimize this or can get pretrained weights.

Based on this repository and this Colab notebook one epoch should take approx. 90 seconds on 4x V100s.
Assuming your GPUs are slower and you are using a single GPU, the estimation of days is still quite large.

You could try to run the notebook on Colab and compare the speed with your local system.
If training on your system is indeed that slow, I would recommend to narrow down the bottleneck and check if it’s might be the data loading (are you storing the data on a local SSD or network drive etc.) or the model itself.

1 Like