wgan-gp
wgan-gp copied to clipboard
Results not reproducible
Thanks for sharing this repo.
Running the code (main.py
, 200 epochs, parameters as mentioned in readme, only fixing the the issue mentioned in #1), I couldn't reproduce the results shown in https://github.com/EmilienDupont/wgan-gp/blob/master/gifs/mnist_200_epochs.gif. Instead, the generated samples look as follows:
Any hints on what would need to be changed to make the training successful?
One step towards reproducibility would be to fix random seeds in the beginning of main.py
, e.g. with the following code:
import random
import numpy as np
import torch
random_seed = 42
random.seed(random_seed)
np.random.seed(random_seed)
torch.manual_seed(random_seed)
torch.cuda.manual_seed_all(random_seed)
Hello ,I met the same problem as you. And I find it that this gif what my model generated 6.60Mb, But in this code example gif folder,it is 8.22 Mb. What make this difference? I hope author to anser this question,thank you very much!
Hello ,I met the same problem as you. And I find it that this gif what my model generated 6.60Mb, But in this code example gif folder,it is 8.22 Mb. What make this difference? I hope author to anser this question,thank you very much!
guess it's the learning rate and trainning epoch causing that problem, i changed it to 5x larger: lr = 5e-4, and the performance will be better:
after 50 epochs trainning:
(you can train it longer and check your result)