Support training SD V2 with Flax (huggingface#1783)

* Support training SD V2 with Flax

Mostly involves supporting a v_prediction scheduler.

The implementation in huggingface#1777 doesn't take into account a recent refactor of `scheduling_utils_flax`, so this should be used instead.

* Add to other top-level files.

Author: yasyf

examples/dreambooth/train_dreambooth_flax.py

@@ -525,28 +525,35 @@ def compute_loss(params):
                 )[0]
 
             # Predict the noise residual
-            unet_outputs = unet.apply(
+            model_pred = unet.apply(
                 {"params": params["unet"]}, noisy_latents, timesteps, encoder_hidden_states, train=True
-            )
-            noise_pred = unet_outputs.sample
+            ).sample
+
+            # Get the target for loss depending on the prediction type
+            if noise_scheduler.config.prediction_type == "epsilon":
+                target = noise
+            elif noise_scheduler.config.prediction_type == "v_prediction":
+                target = noise_scheduler.get_velocity(noise_scheduler_state, latents, noise, timesteps)
+            else:
+                raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
 
             if args.with_prior_preservation:
                 # Chunk the noise and noise_pred into two parts and compute the loss on each part separately.
-                noise_pred, noise_pred_prior = jnp.split(noise_pred, 2, axis=0)
-                noise, noise_prior = jnp.split(noise, 2, axis=0)
+                model_pred, model_pred_prior = jnp.split(model_pred, 2, axis=0)
+                target, target_prior = jnp.split(target, 2, axis=0)
 
                 # Compute instance loss
-                loss = (noise - noise_pred) ** 2
+                loss = (target - model_pred) ** 2
                 loss = loss.mean()
 
                 # Compute prior loss
-                prior_loss = (noise_prior - noise_pred_prior) ** 2
+                prior_loss = (target_prior - model_pred_prior) ** 2
                 prior_loss = prior_loss.mean()
 
                 # Add the prior loss to the instance loss.
                 loss = loss + args.prior_loss_weight * prior_loss
             else:
-                loss = (noise - noise_pred) ** 2
+                loss = (target - model_pred) ** 2
                 loss = loss.mean()
 
             return loss

examples/text_to_image/train_text_to_image_flax.py

@@ -459,9 +459,19 @@ def compute_loss(params):
             )[0]
 
             # Predict the noise residual and compute loss
-            unet_outputs = unet.apply({"params": params}, noisy_latents, timesteps, encoder_hidden_states, train=True)
-            noise_pred = unet_outputs.sample
-            loss = (noise - noise_pred) ** 2
+            model_pred = unet.apply(
+                {"params": params}, noisy_latents, timesteps, encoder_hidden_states, train=True
+            ).sample
+
+            # Get the target for loss depending on the prediction type
+            if noise_scheduler.config.prediction_type == "epsilon":
+                target = noise
+            elif noise_scheduler.config.prediction_type == "v_prediction":
+                target = noise_scheduler.get_velocity(noise_scheduler_state, latents, noise, timesteps)
+            else:
+                raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
+
+            loss = (target - model_pred) ** 2
             loss = loss.mean()
 
             return loss

examples/textual_inversion/textual_inversion_flax.py

@@ -536,11 +536,20 @@ def compute_loss(params):
             encoder_hidden_states = state.apply_fn(
                 batch["input_ids"], params=params, dropout_rng=dropout_rng, train=True
             )[0]
-            unet_outputs = unet.apply(
+            # Predict the noise residual and compute loss
+            model_pred = unet.apply(
                 {"params": unet_params}, noisy_latents, timesteps, encoder_hidden_states, train=False
-            )
-            noise_pred = unet_outputs.sample
-            loss = (noise - noise_pred) ** 2
+            ).sample
+
+            # Get the target for loss depending on the prediction type
+            if noise_scheduler.config.prediction_type == "epsilon":
+                target = noise
+            elif noise_scheduler.config.prediction_type == "v_prediction":
+                target = noise_scheduler.get_velocity(noise_scheduler_state, latents, noise, timesteps)
+            else:
+                raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
+
+            loss = (target - model_pred) ** 2
             loss = loss.mean()
 
             return loss

src/diffusers/schedulers/scheduling_ddim_flax.py

@@ -29,6 +29,7 @@
     FlaxSchedulerMixin,
     FlaxSchedulerOutput,
     add_noise_common,
+    get_velocity_common,
 )
 
 
@@ -301,5 +302,14 @@ def add_noise(
     ) -> jnp.ndarray:
         return add_noise_common(state.common, original_samples, noise, timesteps)
 
+    def get_velocity(
+        self,
+        state: DDIMSchedulerState,
+        sample: jnp.ndarray,
+        noise: jnp.ndarray,
+        timesteps: jnp.ndarray,
+    ) -> jnp.ndarray:
+        return get_velocity_common(state.common, sample, noise, timesteps)
+
     def __len__(self):
         return self.config.num_train_timesteps

src/diffusers/schedulers/scheduling_ddpm_flax.py

@@ -29,6 +29,7 @@
     FlaxSchedulerMixin,
     FlaxSchedulerOutput,
     add_noise_common,
+    get_velocity_common,
 )
 
 
@@ -293,5 +294,14 @@ def add_noise(
     ) -> jnp.ndarray:
         return add_noise_common(state.common, original_samples, noise, timesteps)
 
+    def get_velocity(
+        self,
+        state: DDPMSchedulerState,
+        sample: jnp.ndarray,
+        noise: jnp.ndarray,
+        timesteps: jnp.ndarray,
+    ) -> jnp.ndarray:
+        return get_velocity_common(state.common, sample, noise, timesteps)
+
     def __len__(self):
         return self.config.num_train_timesteps

src/diffusers/schedulers/scheduling_utils_flax.py

@@ -242,7 +242,7 @@ def create(cls, scheduler):
         )
 
 
-def add_noise_common(
+def get_sqrt_alpha_prod(
     state: CommonSchedulerState, original_samples: jnp.ndarray, noise: jnp.ndarray, timesteps: jnp.ndarray
 ):
     alphas_cumprod = state.alphas_cumprod
@@ -255,5 +255,18 @@ def add_noise_common(
     sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
     sqrt_one_minus_alpha_prod = broadcast_to_shape_from_left(sqrt_one_minus_alpha_prod, original_samples.shape)
 
+    return sqrt_alpha_prod, sqrt_one_minus_alpha_prod
+
+
+def add_noise_common(
+    state: CommonSchedulerState, original_samples: jnp.ndarray, noise: jnp.ndarray, timesteps: jnp.ndarray
+):
+    sqrt_alpha_prod, sqrt_one_minus_alpha_prod = get_sqrt_alpha_prod(state, original_samples, noise, timesteps)
     noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
     return noisy_samples
+
+
+def get_velocity_common(state: CommonSchedulerState, sample: jnp.ndarray, noise: jnp.ndarray, timesteps: jnp.ndarray):
+    sqrt_alpha_prod, sqrt_one_minus_alpha_prod = get_sqrt_alpha_prod(state, sample, noise, timesteps)
+    velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
+    return velocity