Skip to content

Open-Sora: Democratizing Efficient Video Production for All

Python

debug

Shen-Chenhui3aa248d7cf3c
diff --git a/opensora/models/vae/model_utils.py b/opensora/models/vae/model_utils.py
index 2737fec..95cb456 100644
--- a/opensora/models/vae/model_utils.py
+++ b/opensora/models/vae/model_utils.py
@@ -91,10 +91,10 @@ class DiagonalGaussianDistribution(object):
self.std = torch.exp(0.5 * self.logvar)
self.var = torch.exp(self.logvar)
if self.deterministic:
- self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device, dtype=dtype)
+ self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device, dtype=self.mean.dtype)
def sample(self):
- x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device, dtype=self.dtype)
+ x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device, dtype=self.mean.dtype)
return x
def kl(self, other=None):
diff --git a/opensora/models/vae/vae_3d.py b/opensora/models/vae/vae_3d.py
index cec4011..45e13ea 100644
--- a/opensora/models/vae/vae_3d.py
+++ b/opensora/models/vae/vae_3d.py
@@ -413,8 +413,11 @@ class VAE_3D(nn.Module):
# self.loss = model_utils.VEA3DLoss(kl_weight=kl_weight)
- self.quant_conv = nn.Conv3d(latent_embed_dim, 2*kl_embed_dim, 1, dtype=dtype, device=device)
- self.post_quant_conv = nn.Conv3d(kl_embed_dim, latent_embed_dim, 1, dtype=dtype, device=device)
+ # self.quant_conv = nn.Conv3d(latent_embed_dim, 2*kl_embed_dim, 1, dtype=dtype, device=device)
+ self.quant_conv = nn.Conv3d(latent_embed_dim, 2*kl_embed_dim, 1, device=device)
+
+ # self.post_quant_conv = nn.Conv3d(kl_embed_dim, latent_embed_dim, 1, dtype=dtype, device=device)
+ self.post_quant_conv = nn.Conv3d(kl_embed_dim, latent_embed_dim, 1, device=device)
self.dtype = dtype
def encode(
@@ -422,7 +425,7 @@ class VAE_3D(nn.Module):
x,
):
encoded_feature = self.encoder(x)
- moments = self.quant_conv(encoded_feature)
+ moments = self.quant_conv(encoded_feature).to(x.dtype)
posterior = model_utils.DiagonalGaussianDistribution(moments, dtype=self.dtype)
return posterior
@@ -430,7 +433,8 @@ class VAE_3D(nn.Module):
self,
z,
):
- z = self.post_quant_conv(z)
+ dtype = z.dtype
+ z = self.post_quant_conv(z).to(dtype)
dec = self.decoder(z)
return dec