Skip to content

Open-Sora: Democratizing Efficient Video Production for All

Python

debug

Shen-Chenhui1b66919091f9
diff --git a/configs/vae_3d/inference/16x256x256.py b/configs/vae_3d/inference/16x256x256.py
index 8ae73e2..1728661 100644
--- a/configs/vae_3d/inference/16x256x256.py
+++ b/configs/vae_3d/inference/16x256x256.py
@@ -11,7 +11,7 @@ use_image_transform = False
num_workers = 4
# Define acceleration
-dtype = "fp16"
+dtype = "bf16"
grad_checkpoint = True
plugin = "zero2"
sp_size = 1
diff --git a/configs/vae_3d/train/16x256x256.py b/configs/vae_3d/train/16x256x256.py
index 5d5927a..4759f4e 100644
--- a/configs/vae_3d/train/16x256x256.py
+++ b/configs/vae_3d/train/16x256x256.py
@@ -16,7 +16,7 @@ sp_size = 1
# Define Loss
kl_weight = 0.000001
-perceptual_weight = 1.0 # TODO: need to change this to 0.1 !!! according to MAGVIT paper
+perceptual_weight = 0.1 # TODO: need to change this to 0.1 !!! according to MAGVIT paper
# Define model
@@ -47,7 +47,7 @@ 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 = 3
+epochs = 5
log_every = 1
ckpt_every = 1000
load = None
diff --git a/scripts/inference-vae.py b/scripts/inference-vae.py
index 8aeddee..b4c873e 100644
--- a/scripts/inference-vae.py
+++ b/scripts/inference-vae.py
@@ -121,7 +121,7 @@ def main():
with tqdm(
range(total_steps),
- desc=f"Avg Loss: {running_loss}",
+ # desc=f"Avg Loss: {running_loss}",
disable=not coordinator.is_master(),
total=total_steps,
initial=0,