Skip to content

Open-Sora: Democratizing Efficient Video Production for All

Python

debug

shenchenhui5935c5b10b55
diff --git a/opensora/models/vae/vae_3d.py b/opensora/models/vae/vae_3d.py
index e9a315f..8bdce80 100644
--- a/opensora/models/vae/vae_3d.py
+++ b/opensora/models/vae/vae_3d.py
@@ -400,7 +400,7 @@ class VAE_3D(nn.Module):
dtype=dtype,
)
- self.loss = model_utils.VEA3DLoss(kl_weight=kl_weight)
+ # self.loss = model_utils.VEA3DLoss(kl_weight=kl_weight)
self.quant_conv = nn.Conv3d(2*latent_embed_dim, 2*kl_embed_dim, 1)
self.post_quant_conv = nn.Conv3d(kl_embed_dim, latent_embed_dim, 1)
@@ -435,13 +435,12 @@ class VAE_3D(nn.Module):
dec = self.decode(z)
return dec, posterior
-
- def get_loss(
- self,
- batch, # [B, C, T, H, W]
- optimizer_idx=None, # NOTE: to add GAN later
- ):
- reconstructions, posterior = self(batch)
- loss = self.loss(batch, reconstructions, posterior)
-
- return loss
+ # def get_loss(
+ # self,
+ # batch, # [B, C, T, H, W]
+ # optimizer_idx=None, # NOTE: to add GAN later
+ # ):
+ # reconstructions, posterior = self(batch)
+ # loss = self.loss(batch, reconstructions, posterior)
+
+ # return loss
diff --git a/scripts/train-vae.py b/scripts/train-vae.py
index ebec5bc..bfe9f82 100644
--- a/scripts/train-vae.py
+++ b/scripts/train-vae.py
@@ -33,7 +33,7 @@ from opensora.utils.config_utils import (
)
from opensora.utils.misc import all_reduce_mean, format_numel_str, get_model_numel, requires_grad, to_torch_dtype
from opensora.utils.train_utils import update_ema, MaskGenerator
-
+from opensora.models.vae.model_utils import VEA3DLoss
def main():
@@ -205,7 +205,11 @@ def main():
batch = next(dataloader_iter)
x = batch["video"].to(device, dtype) # [B, C, T, H, W]
- loss = vae.get_loss(x)
+ # loss = vae.get_loss(x)
+ loss_function = VEA3DLoss(kl_weight=cfg.kl_weight)
+ reconstructions, posterior = vae(x)
+ loss = loss_function(batch, reconstructions, posterior)
+
# Backward & update
booster.backward(loss=loss, optimizer=optimizer)
optimizer.step()