Skip to content

Open-Sora: Democratizing Efficient Video Production for All

Python

debug

Shen-Chenhuiccb92cd1ab77
diff --git a/configs/vae_magvit_v2/train/17x128x128.py b/configs/vae_magvit_v2/train/17x128x128.py
index 546a5a2..5b8880d 100644
--- a/configs/vae_magvit_v2/train/17x128x128.py
+++ b/configs/vae_magvit_v2/train/17x128x128.py
@@ -1,3 +1,4 @@
+num_frames = 17
frame_interval = 3
image_size = (128, 128)
@@ -18,6 +19,7 @@ sp_size = 1
model = dict(
type="VAE_MAGVIT_V2",
+ num_frames = num_frames,
in_out_channels = 3,
latent_embed_dim = 256,
filters = 128,
@@ -32,7 +34,6 @@ model = dict(
separate_first_frame_encoding = False,
kl_loss_weight = 0.000001,
perceptual_loss_weight = 0.1, # use vgg is not None and more than 0
- num_frames = 17,
adversarial_loss_weight = 1.0,
discriminator_in_channels = 3,
discriminator_filters = 128,
diff --git a/opensora/models/vae/README.md b/opensora/models/vae/README.md
index 8184257..b71f63a 100644
--- a/opensora/models/vae/README.md
+++ b/opensora/models/vae/README.md
@@ -41,4 +41,9 @@ CUDA_VISIBLE_DEVICES=5 torchrun --master_port=29530 --nnodes=1 --nproc_per_node=
'x-transformers'
```
+### 2.2 Train
+
+```yaml
+CUDA_VISIBLE_DEVICES7 torchrun --master_port=29510 --nnodes=1 --nproc_per_node=1 scripts/train-vae-v2.py configs/vae_magvit_v2/train/17x128x128.py --data-path /home/shenchenhui/data/pexels/train.csv
+```