Skip to content

Open-Sora: Democratizing Efficient Video Production for All

Python

updated

zhengzangwf7a4fcfc3f91
diff --git a/configs/vae/train/image.py b/configs/vae/train/image.py
index 77ef011..d20602c 100644
--- a/configs/vae/train/image.py
+++ b/configs/vae/train/image.py
@@ -34,7 +34,7 @@ model = dict(
)
# loss weights
-perceptual_loss_weight = 0.1 # use vgg is not None and more than 0
+perceptual_loss_weight = 0.0 # use vgg is not None and more than 0
kl_loss_weight = 1e-6
mixed_image_ratio = 0.1
diff --git a/configs/vae/train/video.py b/configs/vae/train/video.py
index b330059..259945e 100644
--- a/configs/vae/train/video.py
+++ b/configs/vae/train/video.py
@@ -19,7 +19,7 @@ plugin = "zero2"
# Define model
model = dict(
type="VideoAutoencoderPipeline",
- freeze_vae_2d=True,
+ freeze_vae_2d=False,
from_pretrained=None,
vae_2d=dict(
type="VideoAutoencoderKL",
diff --git a/opensora/models/vae/vae.py b/opensora/models/vae/vae.py
index e7912e7..17520eb 100644
--- a/opensora/models/vae/vae.py
+++ b/opensora/models/vae/vae.py
@@ -122,6 +122,7 @@ class VideoAutoencoderPipeline(nn.Module):
super().__init__()
self.spatial_vae = build_module(vae_2d, MODELS)
self.temporal_vae = build_module(vae_temporal, MODELS)
+
if from_pretrained is not None:
load_checkpoint(self, from_pretrained)
if freeze_vae_2d:
diff --git a/opensora/utils/config_utils.py b/opensora/utils/config_utils.py
index 8b9bf65..1da9fd4 100644
--- a/opensora/utils/config_utils.py
+++ b/opensora/utils/config_utils.py
@@ -62,6 +62,7 @@ def parse_args(training=False):
# Training
# ======================================================
else:
+ parser.add_argument("--lr", default=None, type=float, help="learning rate")
parser.add_argument("--wandb", default=None, type=bool, help="enable wandb")
parser.add_argument("--load", default=None, type=str, help="path to continue training")
parser.add_argument("--data-path", default=None, type=str, help="path to data csv")