Correct SNR weighted loss in v-prediction case by only adding 1 to SNR on the denominator (#6307)

* fix minsnr implementation for v-prediction case

* format code

* always compute snr when snr_gamma is specified

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

examples/controlnet/train_controlnet_flax.py

@@ -907,10 +907,12 @@ def main():
 
             if args.snr_gamma is not None:
                 snr = jnp.array(compute_snr(timesteps))
-                if noise_scheduler.config.prediction_type == "v_prediction":
-                    # Velocity objective requires that we add one to SNR values before we divide by them.
-                    snr = snr + 1
-                snr_loss_weights = jnp.where(snr < args.snr_gamma, snr, jnp.ones_like(snr) * args.snr_gamma) / snr
+                snr_loss_weights = jnp.where(snr < args.snr_gamma, snr, jnp.ones_like(snr) * args.snr_gamma)
+                if noise_scheduler.config.prediction_type == "epsilon":
+                    snr_loss_weights = snr_loss_weights / snr
+                elif noise_scheduler.config.prediction_type == "v_prediction":
+                    snr_loss_weights = snr_loss_weights / (snr + 1)
+
                 loss = loss * snr_loss_weights
 
             loss = loss.mean()

examples/kandinsky2_2/text_to_image/train_text_to_image_decoder.py

@@ -781,12 +781,13 @@ def main():
                     # Since we predict the noise instead of x_0, the original formulation is slightly changed.
                     # This is discussed in Section 4.2 of the same paper.
                     snr = compute_snr(noise_scheduler, timesteps)
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective requires that we add one to SNR values before we divide by them.
-                        snr = snr + 1
-                    mse_loss_weights = (
-                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
-                    )
+                    mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(
+                        dim=1
+                    )[0]
+                    if noise_scheduler.config.prediction_type == "epsilon":
+                        mse_loss_weights = mse_loss_weights / snr
+                    elif noise_scheduler.config.prediction_type == "v_prediction":
+                        mse_loss_weights = mse_loss_weights / (snr + 1)
 
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights

examples/kandinsky2_2/text_to_image/train_text_to_image_lora_decoder.py

@@ -631,12 +631,13 @@ def main():
                     # Since we predict the noise instead of x_0, the original formulation is slightly changed.
                     # This is discussed in Section 4.2 of the same paper.
                     snr = compute_snr(noise_scheduler, timesteps)
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective requires that we add one to SNR values before we divide by them.
-                        snr = snr + 1
-                    mse_loss_weights = (
-                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
-                    )
+                    mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(
+                        dim=1
+                    )[0]
+                    if noise_scheduler.config.prediction_type == "epsilon":
+                        mse_loss_weights = mse_loss_weights / snr
+                    elif noise_scheduler.config.prediction_type == "v_prediction":
+                        mse_loss_weights = mse_loss_weights / (snr + 1)
 
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights

examples/kandinsky2_2/text_to_image/train_text_to_image_lora_prior.py

@@ -664,12 +664,13 @@ def main():
                     # Since we predict the noise instead of x_0, the original formulation is slightly changed.
                     # This is discussed in Section 4.2 of the same paper.
                     snr = compute_snr(noise_scheduler, timesteps)
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective requires that we add one to SNR values before we divide by them.
-                        snr = snr + 1
-                    mse_loss_weights = (
-                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
-                    )
+                    mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(
+                        dim=1
+                    )[0]
+                    if noise_scheduler.config.prediction_type == "epsilon":
+                        mse_loss_weights = mse_loss_weights / snr
+                    elif noise_scheduler.config.prediction_type == "v_prediction":
+                        mse_loss_weights = mse_loss_weights / (snr + 1)
 
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights

examples/kandinsky2_2/text_to_image/train_text_to_image_prior.py

@@ -811,12 +811,13 @@ def main():
                     # Since we predict the noise instead of x_0, the original formulation is slightly changed.
                     # This is discussed in Section 4.2 of the same paper.
                     snr = compute_snr(noise_scheduler, timesteps)
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective requires that we add one to SNR values before we divide by them.
-                        snr = snr + 1
-                    mse_loss_weights = (
-                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
-                    )
+                    mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(
+                        dim=1
+                    )[0]
+                    if noise_scheduler.config.prediction_type == "epsilon":
+                        mse_loss_weights = mse_loss_weights / snr
+                    elif noise_scheduler.config.prediction_type == "v_prediction":
+                        mse_loss_weights = mse_loss_weights / (snr + 1)
 
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights

examples/research_projects/onnxruntime/text_to_image/train_text_to_image.py

@@ -848,12 +848,13 @@ def main():
                     # Since we predict the noise instead of x_0, the original formulation is slightly changed.
                     # This is discussed in Section 4.2 of the same paper.
                     snr = compute_snr(noise_scheduler, timesteps)
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective requires that we add one to SNR values before we divide by them.
-                        snr = snr + 1
-                    mse_loss_weights = (
-                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
-                    )
+                    mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(
+                        dim=1
+                    )[0]
+                    if noise_scheduler.config.prediction_type == "epsilon":
+                        mse_loss_weights = mse_loss_weights / snr
+                    elif noise_scheduler.config.prediction_type == "v_prediction":
+                        mse_loss_weights = mse_loss_weights / (snr + 1)
 
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights

examples/text_to_image/train_text_to_image.py

@@ -943,12 +943,13 @@ def main():
                     # Since we predict the noise instead of x_0, the original formulation is slightly changed.
                     # This is discussed in Section 4.2 of the same paper.
                     snr = compute_snr(noise_scheduler, timesteps)
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective requires that we add one to SNR values before we divide by them.
-                        snr = snr + 1
-                    mse_loss_weights = (
-                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
-                    )
+                    mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(
+                        dim=1
+                    )[0]
+                    if noise_scheduler.config.prediction_type == "epsilon":
+                        mse_loss_weights = mse_loss_weights / snr
+                    elif noise_scheduler.config.prediction_type == "v_prediction":
+                        mse_loss_weights = mse_loss_weights / (snr + 1)
 
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights

examples/text_to_image/train_text_to_image_lora.py

@@ -759,12 +759,13 @@ def main():
                     # Since we predict the noise instead of x_0, the original formulation is slightly changed.
                     # This is discussed in Section 4.2 of the same paper.
                     snr = compute_snr(noise_scheduler, timesteps)
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective requires that we add one to SNR values before we divide by them.
-                        snr = snr + 1
-                    mse_loss_weights = (
-                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
-                    )
+                    mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(
+                        dim=1
+                    )[0]
+                    if noise_scheduler.config.prediction_type == "epsilon":
+                        mse_loss_weights = mse_loss_weights / snr
+                    elif noise_scheduler.config.prediction_type == "v_prediction":
+                        mse_loss_weights = mse_loss_weights / (snr + 1)
 
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights

examples/text_to_image/train_text_to_image_lora_sdxl.py

@@ -1062,12 +1062,13 @@ def main(args):
                     # Since we predict the noise instead of x_0, the original formulation is slightly changed.
                     # This is discussed in Section 4.2 of the same paper.
                     snr = compute_snr(noise_scheduler, timesteps)
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective requires that we add one to SNR values before we divide by them.
-                        snr = snr + 1
-                    mse_loss_weights = (
-                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
-                    )
+                    mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(
+                        dim=1
+                    )[0]
+                    if noise_scheduler.config.prediction_type == "epsilon":
+                        mse_loss_weights = mse_loss_weights / snr
+                    elif noise_scheduler.config.prediction_type == "v_prediction":
+                        mse_loss_weights = mse_loss_weights / (snr + 1)
 
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights

examples/text_to_image/train_text_to_image_sdxl.py

@@ -1087,12 +1087,13 @@ def main(args):
                     # Since we predict the noise instead of x_0, the original formulation is slightly changed.
                     # This is discussed in Section 4.2 of the same paper.
                     snr = compute_snr(noise_scheduler, timesteps)
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective requires that we add one to SNR values before we divide by them.
-                        snr = snr + 1
-                    mse_loss_weights = (
-                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
-                    )
+                    mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(
+                        dim=1
+                    )[0]
+                    if noise_scheduler.config.prediction_type == "epsilon":
+                        mse_loss_weights = mse_loss_weights / snr
+                    elif noise_scheduler.config.prediction_type == "v_prediction":
+                        mse_loss_weights = mse_loss_weights / (snr + 1)
 
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights