Skip to content

Open-Sora: Democratizing Efficient Video Production for All

Python

config

Shen-Chenhui96a42e08db0b
diff --git a/configs/vae_magvit_v2/train/pipeline_16x128x128.py b/configs/vae_magvit_v2/train/pipeline_16x128x128.py
index 6da9ee1..8b2dd77 100644
--- a/configs/vae_magvit_v2/train/pipeline_16x128x128.py
+++ b/configs/vae_magvit_v2/train/pipeline_16x128x128.py
@@ -21,6 +21,7 @@ sp_size = 1
vae_2d = dict(
type="VideoAutoencoderKL",
from_pretrained="stabilityai/sd-vae-ft-ema",
+ # SDXL
)
model = dict(
@@ -82,7 +83,7 @@ magvit uses about # samples (K) * epochs ~ 2-5 K, num_frames = 4, reso = 128
3-6 epochs for pexel, from pexel observation its correct
'''
-epochs = 200
+epochs = 1000
log_every = 1
ckpt_every = 200
load = None
diff --git a/opensora/models/vae/vae_3d_v2.py b/opensora/models/vae/vae_3d_v2.py
index fe97dda..66f8f3e 100644
--- a/opensora/models/vae/vae_3d_v2.py
+++ b/opensora/models/vae/vae_3d_v2.py
@@ -1094,6 +1094,7 @@ class AdversarialLoss(nn.Module):
):
# NOTE: following MAGVIT to allow non_saturating
assert self.generator_loss_type in ["hinge", "vanilla", "non-saturating"]
+
if self.generator_loss_type == "hinge":
gen_loss = -torch.mean(fake_logits)
elif self.generator_loss_type == "non-saturating":
diff --git a/scripts/train-vae-v2.py b/scripts/train-vae-v2.py
index d9900ef..e007878 100644
--- a/scripts/train-vae-v2.py
+++ b/scripts/train-vae-v2.py
@@ -425,6 +425,8 @@ def main():
"num_samples": global_step * total_batch_size,
"epoch": epoch,
"loss": vae_loss.item(),
+ "kl_loss": weighted_kl_loss.item(),
+ "gen_adv_loss": adversarial_loss.item(),
"disc_loss": disc_loss.item(),
"avg_loss": avg_loss,
},