Skip to content

Open-Sora: Democratizing Efficient Video Production for All

Python

[hotfix] fix ema model init (#43)

Hongxin Liu0afa188eeb5c
diff --git a/train.py b/train.py
index 7ddd738..22a372f 100644
--- a/train.py
+++ b/train.py
@@ -9,7 +9,6 @@ A minimal training script for DiT using PyTorch DDP.
"""
import argparse
import os
-from copy import deepcopy
from functools import partial
import torch
@@ -119,7 +118,8 @@ def main(args):
# Step 4: Create DiT and EMA
model = DiT_models[args.model](**model_kwargs).to(get_current_device())
patch_size = model.patch_size
- ema = deepcopy(model)
+ ema = DiT_models[args.model](**model_kwargs).to(get_current_device())
+ update_ema(ema, model, decay=0)
requires_grad(ema, False)
model.train() # important! This enables embedding dropout for classifier-free guidance