Min-SNR Gamma: correct the fix for SNR weighted loss in v-prediction … (huggingface#5238)

Min-SNR Gamma: correct the fix for SNR weighted loss in v-prediction by adding 1 to SNR rather than the resulting loss weights

Co-authored-by: bghira <[email protected]>
Co-authored-by: Sayak Paul <[email protected]>

Author: 3 people

examples/controlnet/train_controlnet_flax.py

@@ -907,17 +907,10 @@ def compute_loss(params, minibatch, sample_rng):
 
             if args.snr_gamma is not None:
                 snr = jnp.array(compute_snr(timesteps))
-                base_weights = jnp.where(snr < args.snr_gamma, snr, jnp.ones_like(snr) * args.snr_gamma) / snr
                 if noise_scheduler.config.prediction_type == "v_prediction":
-                    snr_loss_weights = base_weights + 1
-                else:
-                    # Epsilon and sample prediction use the base weights.
-                    snr_loss_weights = base_weights
-                # For zero-terminal SNR, we have to handle the case where a sigma of Zero results in a Inf value.
-                # When we run this, the MSE loss weights for this timestep is set unconditionally to 1.
-                # If we do not run this, the loss value will go to NaN almost immediately, usually within one step.
-                snr_loss_weights[snr == 0] = 1.0
-
+                    # 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
                 loss = loss * snr_loss_weights
 
             loss = loss.mean()

examples/kandinsky2_2/text_to_image/train_text_to_image_decoder.py

@@ -781,25 +781,13 @@ def collate_fn(examples):
                     # 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)
-                    base_weight = (
+                    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
                     )
 
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective needs to be floored to an SNR weight of one.
-                        mse_loss_weights = base_weight + 1
-                    else:
-                        # Epsilon and sample both use the same loss weights.
-                        mse_loss_weights = base_weight
-
-                    # For zero-terminal SNR, we have to handle the case where a sigma of Zero results in a Inf value.
-                    # When we run this, the MSE loss weights for this timestep is set unconditionally to 1.
-                    # If we do not run this, the loss value will go to NaN almost immediately, usually within one step.
-                    mse_loss_weights[snr == 0] = 1.0
-
-                    # We first calculate the original loss. Then we mean over the non-batch dimensions and
-                    # rebalance the sample-wise losses with their respective loss weights.
-                    # Finally, we take the mean of the rebalanced loss.
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                     loss = loss.mean()

examples/kandinsky2_2/text_to_image/train_text_to_image_lora_decoder.py

@@ -631,25 +631,13 @@ def collate_fn(examples):
                     # 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)
-                    base_weight = (
+                    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
                     )
 
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective needs to be floored to an SNR weight of one.
-                        mse_loss_weights = base_weight + 1
-                    else:
-                        # Epsilon and sample both use the same loss weights.
-                        mse_loss_weights = base_weight
-
-                    # For zero-terminal SNR, we have to handle the case where a sigma of Zero results in a Inf value.
-                    # When we run this, the MSE loss weights for this timestep is set unconditionally to 1.
-                    # If we do not run this, the loss value will go to NaN almost immediately, usually within one step.
-                    mse_loss_weights[snr == 0] = 1.0
-
-                    # We first calculate the original loss. Then we mean over the non-batch dimensions and
-                    # rebalance the sample-wise losses with their respective loss weights.
-                    # Finally, we take the mean of the rebalanced loss.
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                     loss = loss.mean()

examples/kandinsky2_2/text_to_image/train_text_to_image_lora_prior.py

@@ -664,25 +664,13 @@ def collate_fn(examples):
                     # 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)
-                    base_weight = (
+                    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
                     )
 
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective needs to be floored to an SNR weight of one.
-                        mse_loss_weights = base_weight + 1
-                    else:
-                        # Epsilon and sample both use the same loss weights.
-                        mse_loss_weights = base_weight
-
-                    # For zero-terminal SNR, we have to handle the case where a sigma of Zero results in a Inf value.
-                    # When we run this, the MSE loss weights for this timestep is set unconditionally to 1.
-                    # If we do not run this, the loss value will go to NaN almost immediately, usually within one step.
-                    mse_loss_weights[snr == 0] = 1.0
-
-                    # We first calculate the original loss. Then we mean over the non-batch dimensions and
-                    # rebalance the sample-wise losses with their respective loss weights.
-                    # Finally, we take the mean of the rebalanced loss.
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                     loss = loss.mean()

examples/kandinsky2_2/text_to_image/train_text_to_image_prior.py

@@ -811,25 +811,13 @@ def collate_fn(examples):
                     # 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)
-                    base_weight = (
+                    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
                     )
 
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective needs to be floored to an SNR weight of one.
-                        mse_loss_weights = base_weight + 1
-                    else:
-                        # Epsilon and sample both use the same loss weights.
-                        mse_loss_weights = base_weight
-
-                    # For zero-terminal SNR, we have to handle the case where a sigma of Zero results in a Inf value.
-                    # When we run this, the MSE loss weights for this timestep is set unconditionally to 1.
-                    # If we do not run this, the loss value will go to NaN almost immediately, usually within one step.
-                    mse_loss_weights[snr == 0] = 1.0
-
-                    # We first calculate the original loss. Then we mean over the non-batch dimensions and
-                    # rebalance the sample-wise losses with their respective loss weights.
-                    # Finally, we take the mean of the rebalanced loss.
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                     loss = loss.mean()

examples/research_projects/onnxruntime/text_to_image/train_text_to_image.py

@@ -848,24 +848,13 @@ def collate_fn(examples):
                     # 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)
-                    base_weight = (
+                    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
                     )
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # velocity objective prediction requires SNR weights to be floored to a min value of 1.
-                        mse_loss_weights = base_weight + 1
-                    else:
-                        # Epsilon and sample prediction use the base weights.
-                        mse_loss_weights = base_weight
-
-                    # For zero-terminal SNR, we have to handle the case where a sigma of Zero results in a Inf value.
-                    # When we run this, the MSE loss weights for this timestep is set unconditionally to 1.
-                    # If we do not run this, the loss value will go to NaN almost immediately, usually within one step.
-                    mse_loss_weights[snr == 0] = 1.0
-
-                    # We first calculate the original loss. Then we mean over the non-batch dimensions and
-                    # rebalance the sample-wise losses with their respective loss weights.
-                    # Finally, we take the mean of the rebalanced loss.
+
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                     loss = loss.mean()

examples/text_to_image/train_text_to_image.py

@@ -929,25 +929,13 @@ def collate_fn(examples):
                     # 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)
-                    base_weight = (
+                    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
                     )
 
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective needs to be floored to an SNR weight of one.
-                        mse_loss_weights = base_weight + 1
-                    else:
-                        # Epsilon and sample both use the same loss weights.
-                        mse_loss_weights = base_weight
-
-                    # For zero-terminal SNR, we have to handle the case where a sigma of Zero results in a Inf value.
-                    # When we run this, the MSE loss weights for this timestep is set unconditionally to 1.
-                    # If we do not run this, the loss value will go to NaN almost immediately, usually within one step.
-                    mse_loss_weights[snr == 0] = 1.0
-
-                    # We first calculate the original loss. Then we mean over the non-batch dimensions and
-                    # rebalance the sample-wise losses with their respective loss weights.
-                    # Finally, we take the mean of the rebalanced loss.
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                     loss = loss.mean()

examples/text_to_image/train_text_to_image_lora.py

@@ -759,25 +759,13 @@ def collate_fn(examples):
                     # 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)
-                    base_weight = (
+                    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
                     )
 
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective needs to be floored to an SNR weight of one.
-                        mse_loss_weights = base_weight + 1
-                    else:
-                        # Epsilon and sample both use the same loss weights.
-                        mse_loss_weights = base_weight
-
-                    # For zero-terminal SNR, we have to handle the case where a sigma of Zero results in a Inf value.
-                    # When we run this, the MSE loss weights for this timestep is set unconditionally to 1.
-                    # If we do not run this, the loss value will go to NaN almost immediately, usually within one step.
-                    mse_loss_weights[snr == 0] = 1.0
-
-                    # We first calculate the original loss. Then we mean over the non-batch dimensions and
-                    # rebalance the sample-wise losses with their respective loss weights.
-                    # Finally, we take the mean of the rebalanced loss.
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                     loss = loss.mean()

examples/text_to_image/train_text_to_image_lora_sdxl.py

@@ -1050,25 +1050,13 @@ def compute_time_ids(original_size, crops_coords_top_left):
                     # 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)
-                    base_weight = (
+                    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
                     )
 
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective needs to be floored to an SNR weight of one.
-                        mse_loss_weights = base_weight + 1
-                    else:
-                        # Epsilon and sample both use the same loss weights.
-                        mse_loss_weights = base_weight
-
-                    # For zero-terminal SNR, we have to handle the case where a sigma of Zero results in a Inf value.
-                    # When we run this, the MSE loss weights for this timestep is set unconditionally to 1.
-                    # If we do not run this, the loss value will go to NaN almost immediately, usually within one step.
-                    mse_loss_weights[snr == 0] = 1.0
-
-                    # We first calculate the original loss. Then we mean over the non-batch dimensions and
-                    # rebalance the sample-wise losses with their respective loss weights.
-                    # Finally, we take the mean of the rebalanced loss.
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                     loss = loss.mean()

examples/text_to_image/train_text_to_image_sdxl.py

@@ -1067,25 +1067,13 @@ def compute_time_ids(original_size, crops_coords_top_left):
                     # 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)
-                    base_weight = (
+                    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
                     )
 
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective needs to be floored to an SNR weight of one.
-                        mse_loss_weights = base_weight + 1
-                    else:
-                        # Epsilon and sample both use the same loss weights.
-                        mse_loss_weights = base_weight
-
-                    # For zero-terminal SNR, we have to handle the case where a sigma of Zero results in a Inf value.
-                    # When we run this, the MSE loss weights for this timestep is set unconditionally to 1.
-                    # If we do not run this, the loss value will go to NaN almost immediately, usually within one step.
-                    mse_loss_weights[snr == 0] = 1.0
-
-                    # We first calculate the original loss. Then we mean over the non-batch dimensions and
-                    # rebalance the sample-wise losses with their respective loss weights.
-                    # Finally, we take the mean of the rebalanced loss.
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                     loss = loss.mean()