Open-Sora: Democratizing Efficient Video Production for All
Python
fix padding
Shen-Chenhui1522f683b866
diff --git a/opensora/models/vae/vae_temporal.py b/opensora/models/vae/vae_temporal.pyindex d4e10ba..fd5c7e8 100644--- a/opensora/models/vae/vae_temporal.py+++ b/opensora/models/vae/vae_temporal.py@@ -374,8 +374,7 @@ class VAE_Temporal(nn.Module):return input_sizedef encode(self, x):- # time_padding = 0 if (x.shape[2] % self.time_downsample_factor == 0) else self.time_downsample_factor - x.shape[2] % self.time_downsample_factor- time_padding = self.time_downsample_factor - x.shape[2] % self.time_downsample_factor+ time_padding = 0 if (x.shape[2] % self.time_downsample_factor == 0) else self.time_downsample_factor - x.shape[2] % self.time_downsample_factorx = pad_at_dim(x, (time_padding, 0), dim=2)encoded_feature = self.encoder(x)moments = self.quant_conv(encoded_feature).to(x.dtype)@@ -383,8 +382,7 @@ class VAE_Temporal(nn.Module):return posteriordef decode(self, z, num_frames=None):- # time_padding = 0 if (num_frames % self.time_downsample_factor == 0) else self.time_downsample_factor - num_frames % self.time_downsample_factor- time_padding = self.time_downsample_factor - num_frames % self.time_downsample_factor+ time_padding = 0 if (num_frames % self.time_downsample_factor == 0) else self.time_downsample_factor - num_frames % self.time_downsample_factorz = self.post_quant_conv(z)x = self.decoder(z)x = x[:, :, time_padding:]