Skip to content

Open-Sora: Democratizing Efficient Video Production for All

Python

debug

Shen-Chenhuid1e57cbfd575
diff --git a/configs/vae_3d/train/16x256x256.py b/configs/vae_3d/train/16x256x256.py
index f5e4e21..c9aacaa 100644
--- a/configs/vae_3d/train/16x256x256.py
+++ b/configs/vae_3d/train/16x256x256.py
@@ -14,6 +14,9 @@ grad_checkpoint = True
plugin = "zero2"
sp_size = 1
+# Define Loss
+kl_weight = 0.000001
+
# Define model
model = dict(
@@ -30,7 +33,6 @@ model = dict(
kl_embed_dim = 64,
custom_conv_padding = None,
activation_fn = 'swish',
- kl_weight = 0.000001,
)
# TODO: check all these settings
diff --git a/opensora/models/vae/vae_3d.py b/opensora/models/vae/vae_3d.py
index 993d22d..822b038 100644
--- a/opensora/models/vae/vae_3d.py
+++ b/opensora/models/vae/vae_3d.py
@@ -442,14 +442,4 @@ class VAE_3D(nn.Module):
else:
z = posterior.mode()
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
+ return dec, posterior
\ No newline at end of file
diff --git a/scripts/train-vae.py b/scripts/train-vae.py
index 6772f40..17407e6 100644
--- a/scripts/train-vae.py
+++ b/scripts/train-vae.py
@@ -206,7 +206,7 @@ def main():
x = batch["video"].to(device, dtype) # [B, C, T, H, W]
# loss = vae.get_loss(x)
- loss_function = VEA3DLoss(kl_weight=cfg.model.kl_weight).to(device, dtype)
+ loss_function = VEA3DLoss(kl_weight=cfg.kl_weight).to(device, dtype)
reconstructions, posterior = vae(x)
loss = loss_function(x, reconstructions, posterior)