Open-Sora: Democratizing Efficient Video Production for All
Python
debug
Shen-Chenhuifb0f59171c38
diff --git a/configs/vae_magvit_v2/train/17x128x128.py b/configs/vae_magvit_v2/train/17x128x128.pyindex 6e16503..5236fc0 100644--- a/configs/vae_magvit_v2/train/17x128x128.py+++ b/configs/vae_magvit_v2/train/17x128x128.py@@ -59,7 +59,7 @@ kl_loss_weight = 0.000001perceptual_loss_weight = 0.1 # use vgg is not None and more than 0discriminator_factor = 1.0discriminator_loss="hinge"-discriminator_start = 50001+discriminator_start = 1 # 50001 TODO: change to correct val, debug use 1 for now# Othersseed = 42diff --git a/opensora/models/vae/vae_3d_v2.py b/opensora/models/vae/vae_3d_v2.pyindex 0be32da..f03f057 100644--- a/opensora/models/vae/vae_3d_v2.py+++ b/opensora/models/vae/vae_3d_v2.py@@ -784,7 +784,8 @@ class VAE_3D_V2(nn.Module):return videodef get_last_layer(self):- return self.conv_out.weight+ # CausalConv3d wraps the conv+ return self.conv_out.conv.weight# def parameters(self):diff --git a/scripts/train-vae-v2.py b/scripts/train-vae-v2.pyindex cf5edcc..b7dc6de 100644--- a/scripts/train-vae-v2.py+++ b/scripts/train-vae-v2.py@@ -292,11 +292,11 @@ def main():vae_loss = nll_loss# adversarial lossif global_step > cfg.discriminator_start:- fake_logits = discriminator(fake_video.contiguous)+ fake_logits = discriminator(fake_video.contiguous())adversarial_loss = adversarial_loss_fn(fake_logits,nll_loss,- vae.get_last_layer(),+ vae.module.get_last_layer(),global_step,is_training = vae.training,)@@ -313,8 +313,8 @@ def main():disc_optimizer.zero_grad()# if video_contains_first_frame:# Since we don't have enough T frames, pad anyways- real_logits = discriminator(real_video.contiguous.detach())- fake_logits = discriminator(fake_video.contiguous.detach())+ real_logits = discriminator(real_video.contiguous().detach())+ fake_logits = discriminator(fake_video.contiguous().detach())disc_loss = disc_loss_fn(real_logits, fake_logits, global_step)# Backward & updatebooster.backward(loss=disc_loss, optimizer=disc_optimizer)