Skip to content

Open-Sora: Democratizing Efficient Video Production for All

Python

add disc initialization code

Shen-Chenhui34be856fdc74
diff --git a/configs/vae_magvit_v2/train/pipeline_16x128x128.py b/configs/vae_magvit_v2/train/pipeline_16x128x128.py
index 1e3b157..2d4f1a9 100644
--- a/configs/vae_magvit_v2/train/pipeline_16x128x128.py
+++ b/configs/vae_magvit_v2/train/pipeline_16x128x128.py
@@ -46,6 +46,7 @@ discriminator = dict(
num_frames = num_frames,
in_channels = 4,
filters = 128,
+ use_pretrained=False, # NOTE: set to False only if we want to fresh train using a different discriminator!
# channel_multipliers = (2,4,4,4,4), # (2,4,4,4) for 64x64 resolution
channel_multipliers= (2,4,4) # since on intermediate layer, 16 x 16 x 16 dimension z
)
diff --git a/opensora/models/vae/vae_3d_v2.py b/opensora/models/vae/vae_3d_v2.py
index a0ac258..d81af23 100644
--- a/opensora/models/vae/vae_3d_v2.py
+++ b/opensora/models/vae/vae_3d_v2.py
@@ -1209,15 +1209,19 @@ def VAE_MAGVIT_V2(from_pretrained=None, **kwargs):
return model
@MODELS.register_module("DISCRIMINATOR_3D")
-def DISCRIMINATOR_3D(from_pretrained=None, inflate_from_2d=False, **kwargs):
+def DISCRIMINATOR_3D(from_pretrained=None, inflate_from_2d=False, use_pretrained=True, **kwargs):
model = StyleGANDiscriminatorBlur(**kwargs).apply(xavier_uniform_weight_init)
# model = StyleGANDiscriminator(**kwargs).apply(xavier_uniform_weight_init) # SCH: DEBUG: to change back
# model = NLayerDiscriminator3D(input_nc=3, n_layers=3,).apply(n_layer_disc_weights_init)
if from_pretrained is not None:
- if inflate_from_2d:
- load_checkpoint_with_inflation(model, from_pretrained)
+ if use_pretrained:
+ if inflate_from_2d:
+ load_checkpoint_with_inflation(model, from_pretrained)
+ else:
+ load_checkpoint(model, from_pretrained)
else:
- load_checkpoint(model, from_pretrained)
+ print(f"discriminator use_pretrained={use_pretrained}, initializing new discriminator")
+
return model