代码拉取完成，页面将自动刷新
Pytorch implementation of Improved Training of Wasserstein GANs by Gulrajani et al.
Parameters used were lr=1e-4, betas=(.9, .99), dim=16, latent_dim=100. Note that the images were resized from (28, 28) to (32, 32).
Gif [work in progress]
Samples [work in progress]
Set up a generator and discriminator model
from models import Generator, Discriminator
generator = Generator(img_size=(32, 32, 1), latent_dim=100, dim=16)
discriminator = Discriminator(img_size=(32, 32, 1), dim=16)
The generator and discriminator are built to automatically scale with image sizes, so you can easily use images from your own dataset.
Train the generator and discriminator with the WGAN-GP loss
import torch
# Initialize optimizers
G_optimizer = torch.optim.Adam(generator.parameters(), lr=1e-4, betas=(.9, .99))
D_optimizer = torch.optim.Adam(discriminator.parameters(), lr=1e-4, betas=(.9, .99))
# Set up trainer
from training import Trainer
trainer = Trainer(generator, discriminator, G_optimizer, D_optimizer,
use_cuda=torch.cuda.is_available())
# Train model for 200 epochs
trainer.train(data_loader, epochs=200, save_training_gif=True)
This will train the models and generate a gif of the training progress.
Note that WGAN-GPs take a long time to converge. Even on MNIST it takes about 50 epochs to start seeing decent results. For more information and a full example on MNIST, check out main.py.
此处可能存在不合适展示的内容，页面不予展示。您可通过相关编辑功能自查并修改。
如您确认内容无涉及 不当用语 / 纯广告导流 / 暴力 / 低俗色情 / 侵权 / 盗版 / 虚假 / 无价值内容或违法国家有关法律法规的内容，可点击提交进行申诉，我们将尽快为您处理。