diffusers 0.27.1__py3-none-any.whl → 0.32.2__py3-none-any.whl

diffusers/schedulers/scheduling_dpmsolver_sde.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 
 import math
+from dataclasses import dataclass
 from typing import List, Optional, Tuple, Union
 
 import numpy as np
@@ -20,7 +21,31 @@ import torch
 import torchsde
 
 from ..configuration_utils import ConfigMixin, register_to_config
-from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
+from ..utils import BaseOutput, is_scipy_available
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
+
+
+if is_scipy_available():
+    import scipy.stats
+
+
+@dataclass
+# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DPMSolverSDE
+class DPMSolverSDESchedulerOutput(BaseOutput):
+    """
+    Output class for the scheduler's `step` function output.
+
+    Args:
+        prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images):
+            Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+            denoising loop.
+        pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images):
+            The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+            `pred_original_sample` can be used to preview progress or for guidance.
+    """
+
+    prev_sample: torch.Tensor
+    pred_original_sample: Optional[torch.Tensor] = None
 
 
 class BatchedBrownianTree:
@@ -38,7 +63,20 @@ class BatchedBrownianTree:
         except TypeError:
             seed = [seed]
             self.batched = False
-        self.trees = [torchsde.BrownianTree(t0, w0, t1, entropy=s, **kwargs) for s in seed]
+        self.trees = [
+            torchsde.BrownianInterval(
+                t0=t0,
+                t1=t1,
+                size=w0.shape,
+                dtype=w0.dtype,
+                device=w0.device,
+                entropy=s,
+                tol=1e-6,
+                pool_size=24,
+                halfway_tree=True,
+            )
+            for s in seed
+        ]
 
     @staticmethod
     def sort(a, b):
@@ -110,7 +148,7 @@ def betas_for_alpha_bar(
             return math.exp(t * -12.0)
 
     else:
-        raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+        raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}")
 
     betas = []
     for i in range(num_diffusion_timesteps):
@@ -147,6 +185,11 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
         use_karras_sigmas (`bool`, *optional*, defaults to `False`):
             Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`,
             the sigmas are determined according to a sequence of noise levels {σi}.
+        use_exponential_sigmas (`bool`, *optional*, defaults to `False`):
+            Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process.
+        use_beta_sigmas (`bool`, *optional*, defaults to `False`):
+            Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta
+            Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information.
         noise_sampler_seed (`int`, *optional*, defaults to `None`):
             The random seed to use for the noise sampler. If `None`, a random seed is generated.
         timestep_spacing (`str`, defaults to `"linspace"`):
@@ -169,10 +212,18 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
         trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
         prediction_type: str = "epsilon",
         use_karras_sigmas: Optional[bool] = False,
+        use_exponential_sigmas: Optional[bool] = False,
+        use_beta_sigmas: Optional[bool] = False,
         noise_sampler_seed: Optional[int] = None,
         timestep_spacing: str = "linspace",
         steps_offset: int = 0,
     ):
+        if self.config.use_beta_sigmas and not is_scipy_available():
+            raise ImportError("Make sure to install scipy if you want to use beta sigmas.")
+        if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1:
+            raise ValueError(
+                "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used."
+            )
         if trained_betas is not None:
             self.betas = torch.tensor(trained_betas, dtype=torch.float32)
         elif beta_schedule == "linear":
@@ -184,7 +235,7 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
             # Glide cosine schedule
             self.betas = betas_for_alpha_bar(num_train_timesteps)
         else:
-            raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+            raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}")
 
         self.alphas = 1.0 - self.betas
         self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
@@ -233,7 +284,7 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
     @property
     def step_index(self):
         """
-        The index counter for current timestep. It will increae 1 after each scheduler step.
+        The index counter for current timestep. It will increase 1 after each scheduler step.
         """
         return self._step_index
 
@@ -257,21 +308,21 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
 
     def scale_model_input(
         self,
-        sample: torch.FloatTensor,
-        timestep: Union[float, torch.FloatTensor],
-    ) -> torch.FloatTensor:
+        sample: torch.Tensor,
+        timestep: Union[float, torch.Tensor],
+    ) -> torch.Tensor:
         """
         Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
         current timestep.
 
         Args:
-            sample (`torch.FloatTensor`):
+            sample (`torch.Tensor`):
                 The input sample.
             timestep (`int`, *optional*):
                 The current timestep in the diffusion chain.
 
         Returns:
-            `torch.FloatTensor`:
+            `torch.Tensor`:
                 A scaled input sample.
         """
         if self.step_index is None:
@@ -325,9 +376,15 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
         log_sigmas = np.log(sigmas)
         sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
 
-        if self.use_karras_sigmas:
+        if self.config.use_karras_sigmas:
             sigmas = self._convert_to_karras(in_sigmas=sigmas)
             timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
+        elif self.config.use_exponential_sigmas:
+            sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
+            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
+        elif self.config.use_beta_sigmas:
+            sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
+            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
 
         second_order_timesteps = self._second_order_timesteps(sigmas, log_sigmas)
 
@@ -370,7 +427,7 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
         timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sig_proposed])
         return timesteps
 
-    # copied from diffusers.schedulers.scheduling_euler_discrete._sigma_to_t
+    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
     def _sigma_to_t(self, sigma, log_sigmas):
         # get log sigma
         log_sigma = np.log(np.maximum(sigma, 1e-10))
@@ -394,8 +451,8 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
         t = t.reshape(sigma.shape)
         return t
 
-    # copied from diffusers.schedulers.scheduling_euler_discrete._convert_to_karras
-    def _convert_to_karras(self, in_sigmas: torch.FloatTensor) -> torch.FloatTensor:
+    # copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
+    def _convert_to_karras(self, in_sigmas: torch.Tensor) -> torch.Tensor:
         """Constructs the noise schedule of Karras et al. (2022)."""
 
         sigma_min: float = in_sigmas[-1].item()
@@ -408,38 +465,93 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
         sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
         return sigmas
 
+    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential
+    def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor:
+        """Constructs an exponential noise schedule."""
+
+        # Hack to make sure that other schedulers which copy this function don't break
+        # TODO: Add this logic to the other schedulers
+        if hasattr(self.config, "sigma_min"):
+            sigma_min = self.config.sigma_min
+        else:
+            sigma_min = None
+
+        if hasattr(self.config, "sigma_max"):
+            sigma_max = self.config.sigma_max
+        else:
+            sigma_max = None
+
+        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
+
+        sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps))
+        return sigmas
+
+    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta
+    def _convert_to_beta(
+        self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6
+    ) -> torch.Tensor:
+        """From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)"""
+
+        # Hack to make sure that other schedulers which copy this function don't break
+        # TODO: Add this logic to the other schedulers
+        if hasattr(self.config, "sigma_min"):
+            sigma_min = self.config.sigma_min
+        else:
+            sigma_min = None
+
+        if hasattr(self.config, "sigma_max"):
+            sigma_max = self.config.sigma_max
+        else:
+            sigma_max = None
+
+        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
+
+        sigmas = np.array(
+            [
+                sigma_min + (ppf * (sigma_max - sigma_min))
+                for ppf in [
+                    scipy.stats.beta.ppf(timestep, alpha, beta)
+                    for timestep in 1 - np.linspace(0, 1, num_inference_steps)
+                ]
+            ]
+        )
+        return sigmas
+
     @property
     def state_in_first_order(self):
         return self.sample is None
 
     def step(
         self,
-        model_output: Union[torch.FloatTensor, np.ndarray],
-        timestep: Union[float, torch.FloatTensor],
-        sample: Union[torch.FloatTensor, np.ndarray],
+        model_output: Union[torch.Tensor, np.ndarray],
+        timestep: Union[float, torch.Tensor],
+        sample: Union[torch.Tensor, np.ndarray],
         return_dict: bool = True,
         s_noise: float = 1.0,
-    ) -> Union[SchedulerOutput, Tuple]:
+    ) -> Union[DPMSolverSDESchedulerOutput, Tuple]:
         """
         Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
         process from the learned model outputs (most often the predicted noise).
 
         Args:
-            model_output (`torch.FloatTensor` or `np.ndarray`):
+            model_output (`torch.Tensor` or `np.ndarray`):
                 The direct output from learned diffusion model.
-            timestep (`float` or `torch.FloatTensor`):
+            timestep (`float` or `torch.Tensor`):
                 The current discrete timestep in the diffusion chain.
-            sample (`torch.FloatTensor` or `np.ndarray`):
+            sample (`torch.Tensor` or `np.ndarray`):
                 A current instance of a sample created by the diffusion process.
-            return_dict (`bool`, *optional*, defaults to `True`):
-                Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple.
+            return_dict (`bool`):
+                Whether or not to return a [`~schedulers.scheduling_dpmsolver_sde.DPMSolverSDESchedulerOutput`] or
+                tuple.
             s_noise (`float`, *optional*, defaults to 1.0):
                 Scaling factor for noise added to the sample.
 
         Returns:
-            [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
-                If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
-                tuple is returned where the first element is the sample tensor.
+            [`~schedulers.scheduling_dpmsolver_sde.DPMSolverSDESchedulerOutput`] or `tuple`:
+                If return_dict is `True`, [`~schedulers.scheduling_dpmsolver_sde.DPMSolverSDESchedulerOutput`] is
+                returned, otherwise a tuple is returned where the first element is the sample tensor.
         """
         if self.step_index is None:
             self._init_step_index(timestep)
@@ -450,10 +562,10 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
             self.noise_sampler = BrownianTreeNoiseSampler(sample, min_sigma, max_sigma, self.noise_sampler_seed)
 
         # Define functions to compute sigma and t from each other
-        def sigma_fn(_t: torch.FloatTensor) -> torch.FloatTensor:
+        def sigma_fn(_t: torch.Tensor) -> torch.Tensor:
             return _t.neg().exp()
 
-        def t_fn(_sigma: torch.FloatTensor) -> torch.FloatTensor:
+        def t_fn(_sigma: torch.Tensor) -> torch.Tensor:
             return _sigma.log().neg()
 
         if self.state_in_first_order:
@@ -519,17 +631,20 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
         self._step_index += 1
 
         if not return_dict:
-            return (prev_sample,)
+            return (
+                prev_sample,
+                pred_original_sample,
+            )
 
-        return SchedulerOutput(prev_sample=prev_sample)
+        return DPMSolverSDESchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample)
 
     # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise
     def add_noise(
         self,
-        original_samples: torch.FloatTensor,
-        noise: torch.FloatTensor,
-        timesteps: torch.FloatTensor,
-    ) -> torch.FloatTensor:
+        original_samples: torch.Tensor,
+        noise: torch.Tensor,
+        timesteps: torch.Tensor,
+    ) -> torch.Tensor:
         # Make sure sigmas and timesteps have the same device and dtype as original_samples
         sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
         if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
@@ -543,7 +658,11 @@ class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
         # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index
         if self.begin_index is None:
             step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
+        elif self.step_index is not None:
+            # add_noise is called after first denoising step (for inpainting)
+            step_indices = [self.step_index] * timesteps.shape[0]
         else:
+            # add noise is called before first denoising step to create initial latent(img2img)
             step_indices = [self.begin_index] * timesteps.shape[0]
 
         sigma = sigmas[step_indices].flatten()