You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
# Set motion module learnable
for name, module in denoising_unet.named_modules():
if "motion_modules" in name:
for params in module.parameters():
params.requires_grad = True
感谢分享~
1、参考hallo2和[Moore-AnimateAnyone进行第2阶段代码复现,denoise_unet部分的第一个参数latent是跟2个开源项目一致吗?
(1)self.denoising_unet第1个参数:noisy_latents = train_noise_scheduler.add_noise(latents, noise, timesteps)
(2)权重冻结:
vae.requires_grad_(False)
denoising_unet.requires_grad_(False)
reference_unet.requires_grad_(False)
face_locator.requires_grad_(False)
2、使用跟hallo2相同的原图加噪与denoise_unet预测的值算mse loss,第2批数据后就出现loss为nan的情况;
mse_loss: tensor(0.2483, device='cuda:0', dtype=torch.float16, grad_fn=)
{'global_step:1, train_loss: 0.248291015625'}
mse_loss: tensor(nan, device='cuda:0', dtype=torch.float16, grad_fn=)
{'global_step:2, train_loss: nan'}
3、请问能参考哪个开源进行实现,再次感谢!
The text was updated successfully, but these errors were encountered: