scheduling_dpmsolver_singlestep.py 52.2 KB
Newer Older
1
# Copyright 2024 TSAIL Team and The HuggingFace Team. All rights reserved.
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver

import math
from typing import List, Optional, Tuple, Union

import numpy as np
import torch

from ..configuration_utils import ConfigMixin, register_to_config
24
from ..utils import deprecate, is_scipy_available, logging
25
from ..utils.torch_utils import randn_tensor
Kashif Rasul's avatar
Kashif Rasul committed
26
from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
27
28


29
30
31
if is_scipy_available():
    import scipy.stats

Patrick von Platen's avatar
Patrick von Platen committed
32
33
34
logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


35
# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
YiYi Xu's avatar
YiYi Xu committed
36
37
38
39
40
def betas_for_alpha_bar(
    num_diffusion_timesteps,
    max_beta=0.999,
    alpha_transform_type="cosine",
):
41
42
43
44
45
46
47
48
49
50
51
52
    """
    Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
    (1-beta) over time from t = [0,1].

    Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
    to that part of the diffusion process.


    Args:
        num_diffusion_timesteps (`int`): the number of betas to produce.
        max_beta (`float`): the maximum beta to use; use values lower than 1 to
                     prevent singularities.
YiYi Xu's avatar
YiYi Xu committed
53
54
        alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
                     Choose from `cosine` or `exp`
55
56
57
58

    Returns:
        betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
    """
YiYi Xu's avatar
YiYi Xu committed
59
    if alpha_transform_type == "cosine":
60

YiYi Xu's avatar
YiYi Xu committed
61
62
63
64
65
66
67
68
69
        def alpha_bar_fn(t):
            return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2

    elif alpha_transform_type == "exp":

        def alpha_bar_fn(t):
            return math.exp(t * -12.0)

    else:
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
70
        raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}")
71
72
73
74
75

    betas = []
    for i in range(num_diffusion_timesteps):
        t1 = i / num_diffusion_timesteps
        t2 = (i + 1) / num_diffusion_timesteps
YiYi Xu's avatar
YiYi Xu committed
76
        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
77
78
79
80
81
    return torch.tensor(betas, dtype=torch.float32)


class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
    """
82
    `DPMSolverSinglestepScheduler` is a fast dedicated high-order solver for diffusion ODEs.
83

84
85
    This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
    methods the library implements for all schedulers such as loading and saving.
86
87

    Args:
88
89
90
91
92
93
94
95
        num_train_timesteps (`int`, defaults to 1000):
            The number of diffusion steps to train the model.
        beta_start (`float`, defaults to 0.0001):
            The starting `beta` value of inference.
        beta_end (`float`, defaults to 0.02):
            The final `beta` value.
        beta_schedule (`str`, defaults to `"linear"`):
            The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
96
            `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
97
98
99
100
        trained_betas (`np.ndarray`, *optional*):
            Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
        solver_order (`int`, defaults to 2):
            The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided
101
            sampling, and `solver_order=3` for unconditional sampling.
102
103
104
105
106
107
108
109
110
111
112
113
114
        prediction_type (`str`, defaults to `epsilon`, *optional*):
            Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
            `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
            Video](https://imagen.research.google/video/paper.pdf) paper).
        thresholding (`bool`, defaults to `False`):
            Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
            as Stable Diffusion.
        dynamic_thresholding_ratio (`float`, defaults to 0.995):
            The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
        sample_max_value (`float`, defaults to 1.0):
            The threshold value for dynamic thresholding. Valid only when `thresholding=True` and
            `algorithm_type="dpmsolver++"`.
        algorithm_type (`str`, defaults to `dpmsolver++`):
115
116
117
118
119
            Algorithm type for the solver; can be `dpmsolver` or `dpmsolver++` or `sde-dpmsolver++`. The `dpmsolver`
            type implements the algorithms in the [DPMSolver](https://huggingface.co/papers/2206.00927) paper, and the
            `dpmsolver++` type implements the algorithms in the [DPMSolver++](https://huggingface.co/papers/2211.01095)
            paper. It is recommended to use `dpmsolver++` or `sde-dpmsolver++` with `solver_order=2` for guided
            sampling like in Stable Diffusion.
120
121
122
123
124
125
        solver_type (`str`, defaults to `midpoint`):
            Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the
            sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers.
        lower_order_final (`bool`, defaults to `True`):
            Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can
            stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10.
126
        use_karras_sigmas (`bool`, *optional*, defaults to `False`):
127
128
            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}.
129
130
        use_exponential_sigmas (`bool`, *optional*, defaults to `False`):
            Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process.
131
132
133
        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.
134
        final_sigmas_type (`str`, *optional*, defaults to `"zero"`):
135
136
            The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final
            sigma is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0.
137
138
139
        lambda_min_clipped (`float`, defaults to `-inf`):
            Clipping threshold for the minimum value of `lambda(t)` for numerical stability. This is critical for the
            cosine (`squaredcos_cap_v2`) noise schedule.
140
        variance_type (`str`, *optional*):
141
142
            Set to "learned" or "learned_range" for diffusion models that predict variance. If set, the model's output
            contains the predicted Gaussian variance.
143
144
    """

Kashif Rasul's avatar
Kashif Rasul committed
145
    _compatibles = [e.name for e in KarrasDiffusionSchedulers]
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
    order = 1

    @register_to_config
    def __init__(
        self,
        num_train_timesteps: int = 1000,
        beta_start: float = 0.0001,
        beta_end: float = 0.02,
        beta_schedule: str = "linear",
        trained_betas: Optional[np.ndarray] = None,
        solver_order: int = 2,
        prediction_type: str = "epsilon",
        thresholding: bool = False,
        dynamic_thresholding_ratio: float = 0.995,
        sample_max_value: float = 1.0,
        algorithm_type: str = "dpmsolver++",
        solver_type: str = "midpoint",
163
        lower_order_final: bool = False,
164
        use_karras_sigmas: Optional[bool] = False,
165
        use_exponential_sigmas: Optional[bool] = False,
166
        use_beta_sigmas: Optional[bool] = False,
167
        final_sigmas_type: Optional[str] = "zero",  # "zero", "sigma_min"
168
169
        lambda_min_clipped: float = -float("inf"),
        variance_type: Optional[str] = None,
170
    ):
171
172
173
174
175
176
        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."
            )
177
178
179
180
        if algorithm_type == "dpmsolver":
            deprecation_message = "algorithm_type `dpmsolver` is deprecated and will be removed in a future version. Choose from `dpmsolver++` or `sde-dpmsolver++` instead"
            deprecate("algorithm_types=dpmsolver", "1.0.0", deprecation_message)

181
182
183
184
185
186
        if trained_betas is not None:
            self.betas = torch.tensor(trained_betas, dtype=torch.float32)
        elif beta_schedule == "linear":
            self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
        elif beta_schedule == "scaled_linear":
            # this schedule is very specific to the latent diffusion model.
187
            self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
188
189
190
191
        elif beta_schedule == "squaredcos_cap_v2":
            # Glide cosine schedule
            self.betas = betas_for_alpha_bar(num_train_timesteps)
        else:
192
            raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}")
193
194
195
196
197
198
199

        self.alphas = 1.0 - self.betas
        self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
        # Currently we only support VP-type noise schedule
        self.alpha_t = torch.sqrt(self.alphas_cumprod)
        self.sigma_t = torch.sqrt(1 - self.alphas_cumprod)
        self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t)
200
        self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5
201
202
203
204
205

        # standard deviation of the initial noise distribution
        self.init_noise_sigma = 1.0

        # settings for DPM-Solver
206
        if algorithm_type not in ["dpmsolver", "dpmsolver++", "sde-dpmsolver++"]:
207
            if algorithm_type == "deis":
208
                self.register_to_config(algorithm_type="dpmsolver++")
209
            else:
210
                raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}")
211
        if solver_type not in ["midpoint", "heun"]:
212
            if solver_type in ["logrho", "bh1", "bh2"]:
213
                self.register_to_config(solver_type="midpoint")
214
            else:
215
                raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}")
216

217
        if algorithm_type not in ["dpmsolver++", "sde-dpmsolver++"] and final_sigmas_type == "zero":
218
219
220
221
            raise ValueError(
                f"`final_sigmas_type` {final_sigmas_type} is not supported for `algorithm_type` {algorithm_type}. Please chooose `sigma_min` instead."
            )

222
223
224
225
226
227
228
        # setable values
        self.num_inference_steps = None
        timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy()
        self.timesteps = torch.from_numpy(timesteps)
        self.model_outputs = [None] * solver_order
        self.sample = None
        self.order_list = self.get_order_list(num_train_timesteps)
229
        self._step_index = None
230
        self._begin_index = None
231
        self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
232
233
234
235
236
237
238

    def get_order_list(self, num_inference_steps: int) -> List[int]:
        """
        Computes the solver order at each time step.

        Args:
            num_inference_steps (`int`):
239
                The number of diffusion steps used when generating samples with a pre-trained model.
240
241
        """
        steps = num_inference_steps
242
        order = self.config.solver_order
243
244
        if order > 3:
            raise ValueError("Order > 3 is not supported by this scheduler")
245
        if self.config.lower_order_final:
246
247
248
249
250
251
252
253
254
            if order == 3:
                if steps % 3 == 0:
                    orders = [1, 2, 3] * (steps // 3 - 1) + [1, 2] + [1]
                elif steps % 3 == 1:
                    orders = [1, 2, 3] * (steps // 3) + [1]
                else:
                    orders = [1, 2, 3] * (steps // 3) + [1, 2]
            elif order == 2:
                if steps % 2 == 0:
255
                    orders = [1, 2] * (steps // 2 - 1) + [1, 1]
256
257
258
259
260
261
262
263
264
265
266
                else:
                    orders = [1, 2] * (steps // 2) + [1]
            elif order == 1:
                orders = [1] * steps
        else:
            if order == 3:
                orders = [1, 2, 3] * (steps // 3)
            elif order == 2:
                orders = [1, 2] * (steps // 2)
            elif order == 1:
                orders = [1] * steps
StAlKeR7779's avatar
StAlKeR7779 committed
267
268
269
270

        if self.config.final_sigmas_type == "zero":
            orders[-1] = 1

271
272
        return orders

273
274
275
    @property
    def step_index(self):
        """
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
276
        The index counter for current timestep. It will increase 1 after each scheduler step.
277
278
279
        """
        return self._step_index

280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
    @property
    def begin_index(self):
        """
        The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
        """
        return self._begin_index

    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
    def set_begin_index(self, begin_index: int = 0):
        """
        Sets the begin index for the scheduler. This function should be run from pipeline before the inference.

        Args:
            begin_index (`int`):
                The begin index for the scheduler.
        """
        self._begin_index = begin_index

298
299
300
301
302
303
    def set_timesteps(
        self,
        num_inference_steps: int = None,
        device: Union[str, torch.device] = None,
        timesteps: Optional[List[int]] = None,
    ):
304
        """
305
        Sets the discrete timesteps used for the diffusion chain (to be run before inference).
306
307
308

        Args:
            num_inference_steps (`int`):
309
310
311
                The number of diffusion steps used when generating samples with a pre-trained model.
            device (`str` or `torch.device`, *optional*):
                The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
312
313
314
315
            timesteps (`List[int]`, *optional*):
                Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
                timestep spacing strategy of equal spacing between timesteps schedule is used. If `timesteps` is
                passed, `num_inference_steps` must be `None`.
316
        """
317
318
319
320
321
322
        if num_inference_steps is None and timesteps is None:
            raise ValueError("Must pass exactly one of  `num_inference_steps` or `timesteps`.")
        if num_inference_steps is not None and timesteps is not None:
            raise ValueError("Must pass exactly one of  `num_inference_steps` or `timesteps`.")
        if timesteps is not None and self.config.use_karras_sigmas:
            raise ValueError("Cannot use `timesteps` when `config.use_karras_sigmas=True`.")
323
324
        if timesteps is not None and self.config.use_exponential_sigmas:
            raise ValueError("Cannot set `timesteps` with `config.use_exponential_sigmas = True`.")
325
326
        if timesteps is not None and self.config.use_beta_sigmas:
            raise ValueError("Cannot set `timesteps` with `config.use_beta_sigmas = True`.")
327
328

        num_inference_steps = num_inference_steps or len(timesteps)
329
        self.num_inference_steps = num_inference_steps
330
331
332
333
334
335
336

        if timesteps is not None:
            timesteps = np.array(timesteps).astype(np.int64)
        else:
            # Clipping the minimum of all lambda(t) for numerical stability.
            # This is critical for cosine (squaredcos_cap_v2) noise schedule.
            clipped_idx = torch.searchsorted(torch.flip(self.lambda_t, [0]), self.config.lambda_min_clipped)
YiYi Xu's avatar
YiYi Xu committed
337
            clipped_idx = clipped_idx.item()
338
339
340
341
342
343
            timesteps = (
                np.linspace(0, self.config.num_train_timesteps - 1 - clipped_idx, num_inference_steps + 1)
                .round()[::-1][:-1]
                .copy()
                .astype(np.int64)
            )
344

345
        sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
346
        log_sigmas = np.log(sigmas)
347
        if self.config.use_karras_sigmas:
348
            sigmas = np.flip(sigmas).copy()
349
350
            sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round()
351
        elif self.config.use_exponential_sigmas:
352
353
            sigmas = np.flip(sigmas).copy()
            sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
354
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
355
        elif self.config.use_beta_sigmas:
356
357
            sigmas = np.flip(sigmas).copy()
            sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
358
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
359
360
        else:
            sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
361
362

        if self.config.final_sigmas_type == "sigma_min":
363
            sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5
364
365
366
367
368
369
370
        elif self.config.final_sigmas_type == "zero":
            sigma_last = 0
        else:
            raise ValueError(
                f" `final_sigmas_type` must be one of `sigma_min` or `zero`, but got {self.config.final_sigmas_type}"
            )
        sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32)
371

372
        self.sigmas = torch.from_numpy(sigmas).to(device=device)
373

374
        self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64)
375
376
        self.model_outputs = [None] * self.config.solver_order
        self.sample = None
Patrick von Platen's avatar
Patrick von Platen committed
377
378

        if not self.config.lower_order_final and num_inference_steps % self.config.solver_order != 0:
379
            logger.warning(
380
                "Changing scheduler {self.config} to have `lower_order_final` set to True to handle uneven amount of inference steps. Please make sure to always use an even number of `num_inference steps when using `lower_order_final=False`."
Patrick von Platen's avatar
Patrick von Platen committed
381
382
383
            )
            self.register_to_config(lower_order_final=True)

384
        if not self.config.lower_order_final and self.config.final_sigmas_type == "zero":
385
            logger.warning(
386
387
388
389
                " `last_sigmas_type='zero'` is not supported for `lower_order_final=False`. Changing scheduler {self.config} to have `lower_order_final` set to True."
            )
            self.register_to_config(lower_order_final=True)

Patrick von Platen's avatar
Patrick von Platen committed
390
        self.order_list = self.get_order_list(num_inference_steps)
391

392
393
        # add an index counter for schedulers that allow duplicated timesteps
        self._step_index = None
394
        self._begin_index = None
395
        self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
396

397
    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
398
    def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor:
399
400
401
402
403
404
405
406
407
408
        """
        "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
        prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
        s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
        pixels from saturation at each step. We find that dynamic thresholding results in significantly better
        photorealism as well as better image-text alignment, especially when using very large guidance weights."

        https://arxiv.org/abs/2205.11487
        """
        dtype = sample.dtype
409
        batch_size, channels, *remaining_dims = sample.shape
410
411
412
413
414

        if dtype not in (torch.float32, torch.float64):
            sample = sample.float()  # upcast for quantile calculation, and clamp not implemented for cpu half

        # Flatten sample for doing quantile calculation along each image
415
        sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
416
417
418
419
420
421
422
423
424
425

        abs_sample = sample.abs()  # "a certain percentile absolute pixel value"

        s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
        s = torch.clamp(
            s, min=1, max=self.config.sample_max_value
        )  # When clamped to min=1, equivalent to standard clipping to [-1, 1]
        s = s.unsqueeze(1)  # (batch_size, 1) because clamp will broadcast along dim=0
        sample = torch.clamp(sample, -s, s) / s  # "we threshold xt0 to the range [-s, s] and then divide by s"

426
        sample = sample.reshape(batch_size, channels, *remaining_dims)
427
428
429
        sample = sample.to(dtype)

        return sample
430

431
432
433
    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
    def _sigma_to_t(self, sigma, log_sigmas):
        # get log sigma
434
        log_sigma = np.log(np.maximum(sigma, 1e-10))
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454

        # get distribution
        dists = log_sigma - log_sigmas[:, np.newaxis]

        # get sigmas range
        low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
        high_idx = low_idx + 1

        low = log_sigmas[low_idx]
        high = log_sigmas[high_idx]

        # interpolate sigmas
        w = (low - log_sigma) / (low - high)
        w = np.clip(w, 0, 1)

        # transform interpolation to time range
        t = (1 - w) * low_idx + w * high_idx
        t = t.reshape(sigma.shape)
        return t

455
456
457
458
459
460
461
    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t
    def _sigma_to_alpha_sigma_t(self, sigma):
        alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
        sigma_t = sigma * alpha_t

        return alpha_t, sigma_t

462
    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
463
    def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor:
464
465
        """Constructs the noise schedule of Karras et al. (2022)."""

Suraj Patil's avatar
Suraj Patil committed
466
467
468
469
470
471
472
473
474
475
476
477
478
479
        # 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()
480
481
482
483
484
485
486
487

        rho = 7.0  # 7.0 is the value used in the paper
        ramp = np.linspace(0, 1, num_inference_steps)
        min_inv_rho = sigma_min ** (1 / rho)
        max_inv_rho = sigma_max ** (1 / rho)
        sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
        return sigmas

488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
    # 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()

507
        sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps))
508
509
        return sigmas

510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
    # 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()

531
        sigmas = np.array(
532
533
534
535
536
537
538
539
540
541
            [
                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

542
    def convert_model_output(
543
        self,
544
        model_output: torch.Tensor,
545
        *args,
546
        sample: torch.Tensor = None,
547
        **kwargs,
548
    ) -> torch.Tensor:
549
        """
550
551
552
553
554
        Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is
        designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an
        integral of the data prediction model.

        <Tip>
555

556
557
        The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both noise
        prediction and data prediction models.
558

559
        </Tip>
560
561

        Args:
562
            model_output (`torch.Tensor`):
563
                The direct output from the learned diffusion model.
564
            sample (`torch.Tensor`):
565
                A current instance of a sample created by the diffusion process.
566
567

        Returns:
568
            `torch.Tensor`:
569
                The converted model output.
570
        """
571
572
573
574
575
576
577
578
579
580
581
582
        timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
        if sample is None:
            if len(args) > 1:
                sample = args[1]
            else:
                raise ValueError("missing `sample` as a required keyward argument")
        if timestep is not None:
            deprecate(
                "timesteps",
                "1.0.0",
                "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )
583
        # DPM-Solver++ needs to solve an integral of the data prediction model.
584
        if self.config.algorithm_type in ["dpmsolver++", "sde-dpmsolver++"]:
585
            if self.config.prediction_type == "epsilon":
586
                # DPM-Solver and DPM-Solver++ only need the "mean" output.
587
                if self.config.variance_type in ["learned", "learned_range"]:
588
                    model_output = model_output[:, :3]
589
590
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
591
592
593
594
                x0_pred = (sample - sigma_t * model_output) / alpha_t
            elif self.config.prediction_type == "sample":
                x0_pred = model_output
            elif self.config.prediction_type == "v_prediction":
595
596
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
597
598
599
600
601
602
603
604
                x0_pred = alpha_t * sample - sigma_t * model_output
            else:
                raise ValueError(
                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
                    " `v_prediction` for the DPMSolverSinglestepScheduler."
                )

            if self.config.thresholding:
605
606
                x0_pred = self._threshold_sample(x0_pred)

607
            return x0_pred
608

609
610
611
        # DPM-Solver needs to solve an integral of the noise prediction model.
        elif self.config.algorithm_type == "dpmsolver":
            if self.config.prediction_type == "epsilon":
612
                # DPM-Solver and DPM-Solver++ only need the "mean" output.
613
614
615
616
                if self.config.variance_type in ["learned", "learned_range"]:
                    epsilon = model_output[:, :3]
                else:
                    epsilon = model_output
617
            elif self.config.prediction_type == "sample":
618
619
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
620
621
                epsilon = (sample - alpha_t * model_output) / sigma_t
            elif self.config.prediction_type == "v_prediction":
622
623
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
624
625
626
627
628
629
630
                epsilon = alpha_t * model_output + sigma_t * sample
            else:
                raise ValueError(
                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
                    " `v_prediction` for the DPMSolverSinglestepScheduler."
                )

631
632
633
634
635
636
637
638
            if self.config.thresholding:
                alpha_t, sigma_t = self.alpha_t[timestep], self.sigma_t[timestep]
                x0_pred = (sample - sigma_t * epsilon) / alpha_t
                x0_pred = self._threshold_sample(x0_pred)
                epsilon = (sample - alpha_t * x0_pred) / sigma_t

            return epsilon

639
640
    def dpm_solver_first_order_update(
        self,
641
        model_output: torch.Tensor,
642
        *args,
643
        sample: torch.Tensor = None,
644
        noise: Optional[torch.Tensor] = None,
645
        **kwargs,
646
    ) -> torch.Tensor:
647
        """
648
        One step for the first-order DPMSolver (equivalent to DDIM).
649
650

        Args:
651
            model_output (`torch.Tensor`):
652
653
654
655
656
                The direct output from the learned diffusion model.
            timestep (`int`):
                The current discrete timestep in the diffusion chain.
            prev_timestep (`int`):
                The previous discrete timestep in the diffusion chain.
657
            sample (`torch.Tensor`):
658
                A current instance of a sample created by the diffusion process.
659
660

        Returns:
661
            `torch.Tensor`:
662
                The sample tensor at the previous timestep.
663
        """
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
        timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
        prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
        if sample is None:
            if len(args) > 2:
                sample = args[2]
            else:
                raise ValueError(" missing `sample` as a required keyward argument")
        if timestep is not None:
            deprecate(
                "timesteps",
                "1.0.0",
                "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        if prev_timestep is not None:
            deprecate(
                "prev_timestep",
                "1.0.0",
                "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )
        sigma_t, sigma_s = self.sigmas[self.step_index + 1], self.sigmas[self.step_index]
        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
        alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s)
        lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
        lambda_s = torch.log(alpha_s) - torch.log(sigma_s)
689
690
691
692
693
        h = lambda_t - lambda_s
        if self.config.algorithm_type == "dpmsolver++":
            x_t = (sigma_t / sigma_s) * sample - (alpha_t * (torch.exp(-h) - 1.0)) * model_output
        elif self.config.algorithm_type == "dpmsolver":
            x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output
694
695
696
697
698
699
700
        elif self.config.algorithm_type == "sde-dpmsolver++":
            assert noise is not None
            x_t = (
                (sigma_t / sigma_s * torch.exp(-h)) * sample
                + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output
                + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
            )
701
702
703
704
        return x_t

    def singlestep_dpm_solver_second_order_update(
        self,
705
        model_output_list: List[torch.Tensor],
706
        *args,
707
        sample: torch.Tensor = None,
708
        noise: Optional[torch.Tensor] = None,
709
        **kwargs,
710
    ) -> torch.Tensor:
711
        """
712
713
        One step for the second-order singlestep DPMSolver that computes the solution at time `prev_timestep` from the
        time `timestep_list[-2]`.
714
715

        Args:
716
            model_output_list (`List[torch.Tensor]`):
717
718
719
720
721
                The direct outputs from learned diffusion model at current and latter timesteps.
            timestep (`int`):
                The current and latter discrete timestep in the diffusion chain.
            prev_timestep (`int`):
                The previous discrete timestep in the diffusion chain.
722
            sample (`torch.Tensor`):
723
                A current instance of a sample created by the diffusion process.
724
725

        Returns:
726
            `torch.Tensor`:
727
                The sample tensor at the previous timestep.
728
        """
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
        timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
        prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
        if sample is None:
            if len(args) > 2:
                sample = args[2]
            else:
                raise ValueError(" missing `sample` as a required keyward argument")
        if timestep_list is not None:
            deprecate(
                "timestep_list",
                "1.0.0",
                "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        if prev_timestep is not None:
            deprecate(
                "prev_timestep",
                "1.0.0",
                "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )
        sigma_t, sigma_s0, sigma_s1 = (
            self.sigmas[self.step_index + 1],
            self.sigmas[self.step_index],
            self.sigmas[self.step_index - 1],
        )

        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
        alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
        alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)

        lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
        lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
        lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)

763
        m0, m1 = model_output_list[-1], model_output_list[-2]
764

765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
        h, h_0 = lambda_t - lambda_s1, lambda_s0 - lambda_s1
        r0 = h_0 / h
        D0, D1 = m1, (1.0 / r0) * (m0 - m1)
        if self.config.algorithm_type == "dpmsolver++":
            # See https://arxiv.org/abs/2211.01095 for detailed derivations
            if self.config.solver_type == "midpoint":
                x_t = (
                    (sigma_t / sigma_s1) * sample
                    - (alpha_t * (torch.exp(-h) - 1.0)) * D0
                    - 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (sigma_t / sigma_s1) * sample
                    - (alpha_t * (torch.exp(-h) - 1.0)) * D0
                    + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
                )
        elif self.config.algorithm_type == "dpmsolver":
            # See https://arxiv.org/abs/2206.00927 for detailed derivations
            if self.config.solver_type == "midpoint":
                x_t = (
                    (alpha_t / alpha_s1) * sample
                    - (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - 0.5 * (sigma_t * (torch.exp(h) - 1.0)) * D1
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (alpha_t / alpha_s1) * sample
                    - (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
                )
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
        elif self.config.algorithm_type == "sde-dpmsolver++":
            assert noise is not None
            if self.config.solver_type == "midpoint":
                x_t = (
                    (sigma_t / sigma_s1 * torch.exp(-h)) * sample
                    + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0
                    + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1
                    + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (sigma_t / sigma_s1 * torch.exp(-h)) * sample
                    + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0
                    + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1
                    + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
                )
812
813
814
815
        return x_t

    def singlestep_dpm_solver_third_order_update(
        self,
816
        model_output_list: List[torch.Tensor],
817
        *args,
818
        sample: torch.Tensor = None,
StAlKeR7779's avatar
StAlKeR7779 committed
819
        noise: Optional[torch.Tensor] = None,
820
        **kwargs,
821
    ) -> torch.Tensor:
822
        """
823
824
        One step for the third-order singlestep DPMSolver that computes the solution at time `prev_timestep` from the
        time `timestep_list[-3]`.
825
826

        Args:
827
            model_output_list (`List[torch.Tensor]`):
828
829
830
831
832
                The direct outputs from learned diffusion model at current and latter timesteps.
            timestep (`int`):
                The current and latter discrete timestep in the diffusion chain.
            prev_timestep (`int`):
                The previous discrete timestep in the diffusion chain.
833
            sample (`torch.Tensor`):
834
                A current instance of a sample created by diffusion process.
835
836

        Returns:
837
            `torch.Tensor`:
838
                The sample tensor at the previous timestep.
839
        """
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866

        timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
        prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
        if sample is None:
            if len(args) > 2:
                sample = args[2]
            else:
                raise ValueError(" missing`sample` as a required keyward argument")
        if timestep_list is not None:
            deprecate(
                "timestep_list",
                "1.0.0",
                "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        if prev_timestep is not None:
            deprecate(
                "prev_timestep",
                "1.0.0",
                "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        sigma_t, sigma_s0, sigma_s1, sigma_s2 = (
            self.sigmas[self.step_index + 1],
            self.sigmas[self.step_index],
            self.sigmas[self.step_index - 1],
            self.sigmas[self.step_index - 2],
867
        )
868
869
870
871
872
873
874
875
876
877
878
879
880

        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
        alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
        alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)
        alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2)

        lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
        lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
        lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)
        lambda_s2 = torch.log(alpha_s2) - torch.log(sigma_s2)

        m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3]

881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
        h, h_0, h_1 = lambda_t - lambda_s2, lambda_s0 - lambda_s2, lambda_s1 - lambda_s2
        r0, r1 = h_0 / h, h_1 / h
        D0 = m2
        D1_0, D1_1 = (1.0 / r1) * (m1 - m2), (1.0 / r0) * (m0 - m2)
        D1 = (r0 * D1_0 - r1 * D1_1) / (r0 - r1)
        D2 = 2.0 * (D1_1 - D1_0) / (r0 - r1)
        if self.config.algorithm_type == "dpmsolver++":
            # See https://arxiv.org/abs/2206.00927 for detailed derivations
            if self.config.solver_type == "midpoint":
                x_t = (
                    (sigma_t / sigma_s2) * sample
                    - (alpha_t * (torch.exp(-h) - 1.0)) * D0
                    + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1_1
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (sigma_t / sigma_s2) * sample
                    - (alpha_t * (torch.exp(-h) - 1.0)) * D0
                    + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
                    - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2
                )
        elif self.config.algorithm_type == "dpmsolver":
            # See https://arxiv.org/abs/2206.00927 for detailed derivations
            if self.config.solver_type == "midpoint":
                x_t = (
                    (alpha_t / alpha_s2) * sample
                    - (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1_1
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (alpha_t / alpha_s2) * sample
                    - (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
                    - (sigma_t * ((torch.exp(h) - 1.0 - h) / h**2 - 0.5)) * D2
                )
StAlKeR7779's avatar
StAlKeR7779 committed
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
        elif self.config.algorithm_type == "sde-dpmsolver++":
            assert noise is not None
            if self.config.solver_type == "midpoint":
                x_t = (
                    (sigma_t / sigma_s2 * torch.exp(-h)) * sample
                    + (alpha_t * (1.0 - torch.exp(-2.0 * h))) * D0
                    + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1_1
                    + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (sigma_t / sigma_s2 * torch.exp(-h)) * sample
                    + (alpha_t * (1.0 - torch.exp(-2.0 * h))) * D0
                    + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1
                    + (alpha_t * ((1.0 - torch.exp(-2.0 * h) + (-2.0 * h)) / (-2.0 * h) ** 2 - 0.5)) * D2
                    + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
                )
934
935
936
937
        return x_t

    def singlestep_dpm_solver_update(
        self,
938
        model_output_list: List[torch.Tensor],
939
        *args,
940
        sample: torch.Tensor = None,
941
        order: int = None,
942
        noise: Optional[torch.Tensor] = None,
943
        **kwargs,
944
    ) -> torch.Tensor:
945
        """
946
        One step for the singlestep DPMSolver.
947
948

        Args:
949
            model_output_list (`List[torch.Tensor]`):
950
951
952
953
954
                The direct outputs from learned diffusion model at current and latter timesteps.
            timestep (`int`):
                The current and latter discrete timestep in the diffusion chain.
            prev_timestep (`int`):
                The previous discrete timestep in the diffusion chain.
955
            sample (`torch.Tensor`):
956
                A current instance of a sample created by diffusion process.
957
            order (`int`):
958
                The solver order at this step.
959
960

        Returns:
961
            `torch.Tensor`:
962
                The sample tensor at the previous timestep.
963
        """
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
        timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
        prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
        if sample is None:
            if len(args) > 2:
                sample = args[2]
            else:
                raise ValueError(" missing`sample` as a required keyward argument")
        if order is None:
            if len(args) > 3:
                order = args[3]
            else:
                raise ValueError(" missing `order` as a required keyward argument")
        if timestep_list is not None:
            deprecate(
                "timestep_list",
                "1.0.0",
                "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        if prev_timestep is not None:
            deprecate(
                "prev_timestep",
                "1.0.0",
                "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

990
        if order == 1:
991
            return self.dpm_solver_first_order_update(model_output_list[-1], sample=sample, noise=noise)
992
        elif order == 2:
993
            return self.singlestep_dpm_solver_second_order_update(model_output_list, sample=sample, noise=noise)
994
        elif order == 3:
StAlKeR7779's avatar
StAlKeR7779 committed
995
            return self.singlestep_dpm_solver_third_order_update(model_output_list, sample=sample, noise=noise)
996
997
998
        else:
            raise ValueError(f"Order must be 1, 2, 3, got {order}")

999
1000
1001
1002
    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep
    def index_for_timestep(self, timestep, schedule_timesteps=None):
        if schedule_timesteps is None:
            schedule_timesteps = self.timesteps
1003

1004
        index_candidates = (schedule_timesteps == timestep).nonzero()
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016

        if len(index_candidates) == 0:
            step_index = len(self.timesteps) - 1
        # The sigma index that is taken for the **very** first `step`
        # is always the second index (or the last index if there is only 1)
        # This way we can ensure we don't accidentally skip a sigma in
        # case we start in the middle of the denoising schedule (e.g. for image-to-image)
        elif len(index_candidates) > 1:
            step_index = index_candidates[1].item()
        else:
            step_index = index_candidates[0].item()

1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
        return step_index

    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index
    def _init_step_index(self, timestep):
        """
        Initialize the step_index counter for the scheduler.
        """

        if self.begin_index is None:
            if isinstance(timestep, torch.Tensor):
                timestep = timestep.to(self.timesteps.device)
            self._step_index = self.index_for_timestep(timestep)
        else:
            self._step_index = self._begin_index
1031

1032
1033
    def step(
        self,
1034
        model_output: torch.Tensor,
1035
        timestep: Union[int, torch.Tensor],
1036
        sample: torch.Tensor,
1037
        generator=None,
1038
1039
1040
        return_dict: bool = True,
    ) -> Union[SchedulerOutput, Tuple]:
        """
1041
1042
        Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
        the singlestep DPMSolver.
1043
1044

        Args:
1045
            model_output (`torch.Tensor`):
1046
1047
1048
                The direct output from learned diffusion model.
            timestep (`int`):
                The current discrete timestep in the diffusion chain.
1049
            sample (`torch.Tensor`):
1050
1051
1052
                A current instance of a sample created by the diffusion process.
            return_dict (`bool`):
                Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`.
1053
1054

        Returns:
1055
1056
1057
            [`~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.
1058
1059
1060
1061
1062
1063
1064

        """
        if self.num_inference_steps is None:
            raise ValueError(
                "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
            )

1065
1066
        if self.step_index is None:
            self._init_step_index(timestep)
1067

1068
        model_output = self.convert_model_output(model_output, sample=sample)
1069
1070
1071
1072
        for i in range(self.config.solver_order - 1):
            self.model_outputs[i] = self.model_outputs[i + 1]
        self.model_outputs[-1] = model_output

1073
1074
1075
1076
1077
1078
1079
        if self.config.algorithm_type == "sde-dpmsolver++":
            noise = randn_tensor(
                model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype
            )
        else:
            noise = None

1080
        order = self.order_list[self.step_index]
Patrick von Platen's avatar
Patrick von Platen committed
1081
1082
1083
1084
1085
1086

        #  For img2img denoising might start with order>1 which is not possible
        #  In this case make sure that the first two steps are both order=1
        while self.model_outputs[-order] is None:
            order -= 1

1087
1088
1089
1090
        # For single-step solvers, we use the initial value at each time with order = 1.
        if order == 1:
            self.sample = sample

1091
1092
1093
        prev_sample = self.singlestep_dpm_solver_update(
            self.model_outputs, sample=self.sample, order=order, noise=noise
        )
1094

1095
        # upon completion increase step index by one, noise=noise
1096
        self._step_index += 1
1097
1098
1099
1100
1101
1102

        if not return_dict:
            return (prev_sample,)

        return SchedulerOutput(prev_sample=prev_sample)

1103
    def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor:
1104
1105
1106
1107
1108
        """
        Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
        current timestep.

        Args:
1109
            sample (`torch.Tensor`):
1110
                The input sample.
1111
1112

        Returns:
1113
            `torch.Tensor`:
1114
                A scaled input sample.
1115
1116
1117
        """
        return sample

1118
    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.add_noise
1119
1120
    def add_noise(
        self,
1121
1122
        original_samples: torch.Tensor,
        noise: torch.Tensor,
1123
        timesteps: torch.IntTensor,
1124
    ) -> torch.Tensor:
1125
1126
1127
1128
1129
1130
1131
1132
1133
        # 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):
            # mps does not support float64
            schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
            timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
        else:
            schedule_timesteps = self.timesteps.to(original_samples.device)
            timesteps = timesteps.to(original_samples.device)
1134

1135
        # begin_index is None when the scheduler is used for training or pipeline does not implement set_begin_index
1136
1137
        if self.begin_index is None:
            step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
1138
1139
1140
        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]
1141
        else:
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
1142
            # add noise is called before first denoising step to create initial latent(img2img)
1143
            step_indices = [self.begin_index] * timesteps.shape[0]
1144

1145
1146
1147
        sigma = sigmas[step_indices].flatten()
        while len(sigma.shape) < len(original_samples.shape):
            sigma = sigma.unsqueeze(-1)
1148

1149
1150
        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
        noisy_samples = alpha_t * original_samples + sigma_t * noise
1151
1152
1153
1154
        return noisy_samples

    def __len__(self):
        return self.config.num_train_timesteps