Open-Sora: Democratizing Efficient Video Production for All
Python
add z nll loss
Shen-Chenhuic631358bbf41
diff --git a/opensora/models/vae/losses.py b/opensora/models/vae/losses.pyindex 0561e70..4adc85c 100644--- a/opensora/models/vae/losses.py+++ b/opensora/models/vae/losses.py@@ -79,7 +79,6 @@ class VEALoss(nn.Module):recon_video,posterior,nll_weights=None,- split="train",):video = rearrange(video, "b c t h w -> (b t) c h w").contiguous()recon_video = rearrange(recon_video, "b c t h w -> (b t) c h w").contiguous()diff --git a/scripts/inference-vae.py b/scripts/inference-vae.pyindex e566a24..bb36d91 100644--- a/scripts/inference-vae.py+++ b/scripts/inference-vae.py@@ -171,7 +171,7 @@ def main():if cfg.calc_loss:# ====== Calc Loss ======# simple nll loss- nll_loss, weighted_nll_loss, weighted_kl_loss = vae_loss_fn(video, recon_video, posterior, split="eval")+ nll_loss, weighted_nll_loss, weighted_kl_loss = vae_loss_fn(video, recon_video, posterior)fake_video = pad_at_dim(recon_video, (disc_time_padding, 0), value=0.0, dim=2)fake_logits = discriminator(fake_video.contiguous())diff --git a/scripts/train-vae.py b/scripts/train-vae.pyindex 387e28a..aa8f07f 100644--- a/scripts/train-vae.py+++ b/scripts/train-vae.py@@ -340,7 +340,11 @@ def main():# ====== Generator Loss ======# simple nll lossnll_loss, weighted_nll_loss, weighted_kl_loss = vae_loss_fn(- video, recon_video, posterior, split="train"+ video, recon_video, posterior+ )++ _, weighted_z_nll_loss, _ = vae_loss_fn(+ t_video, t_recon_video, posterior)adversarial_loss = torch.tensor(0.0)@@ -357,7 +361,7 @@ def main():is_training=vae.training,)- vae_loss = weighted_nll_loss + weighted_kl_loss + adversarial_loss+ vae_loss = weighted_nll_loss + weighted_kl_loss + adversarial_loss + weighted_z_nll_lossoptimizer.zero_grad()# Backward & update