scheduling_deis_multistep.py 39.1 KB
Newer Older
Aryan's avatar
Aryan committed
1
# Copyright 2025 FLAIR Lab and The HuggingFace Team. All rights reserved.
2
3
4
5
6
7
8
9
10
11
12
13
14
#
# 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.

Quentin Gallouédec's avatar
Quentin Gallouédec committed
15
# DISCLAIMER: check https://huggingface.co/papers/2204.13902 and https://github.com/qsh-zh/deis for more info
16
17
18
19
20
21
22
23
24
# The codebase is modified based on https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_dpmsolver_multistep.py

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

import numpy as np
import torch

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


29
30
31
32
if is_scipy_available():
    import scipy.stats


33
# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
YiYi Xu's avatar
YiYi Xu committed
34
35
36
37
38
def betas_for_alpha_bar(
    num_diffusion_timesteps,
    max_beta=0.999,
    alpha_transform_type="cosine",
):
39
40
41
42
43
44
45
46
47
48
49
50
    """
    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
51
52
        alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
                     Choose from `cosine` or `exp`
53
54
55
56

    Returns:
        betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
    """
YiYi Xu's avatar
YiYi Xu committed
57
    if alpha_transform_type == "cosine":
58

YiYi Xu's avatar
YiYi Xu committed
59
60
61
62
63
64
65
66
67
        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
68
        raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}")
69
70
71
72
73

    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
74
        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
75
76
77
78
79
    return torch.tensor(betas, dtype=torch.float32)


class DEISMultistepScheduler(SchedulerMixin, ConfigMixin):
    """
80
    `DEISMultistepScheduler` is a fast high order solver for diffusion ordinary differential equations (ODEs).
81

82
83
    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.
84
85

    Args:
86
87
88
89
90
91
92
93
        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
94
            `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
        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 DEIS order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided
            sampling, and `solver_order=3` for unconditional sampling.
        prediction_type (`str`, defaults to `epsilon`):
            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`.
        algorithm_type (`str`, defaults to `deis`):
            The algorithm type for the solver.
        lower_order_final (`bool`, defaults to `True`):
            Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps.
115
        use_karras_sigmas (`bool`, *optional*, defaults to `False`):
116
117
             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}.
118
119
        use_exponential_sigmas (`bool`, *optional*, defaults to `False`):
            Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process.
120
121
122
        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.
123
124
125
126
        timestep_spacing (`str`, defaults to `"linspace"`):
            The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
            Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
        steps_offset (`int`, defaults to 0):
127
            An offset added to the inference steps, as required by some model families.
128
129
    """

Kashif Rasul's avatar
Kashif Rasul committed
130
    _compatibles = [e.name for e in KarrasDiffusionSchedulers]
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
    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 = "deis",
        solver_type: str = "logrho",
        lower_order_final: bool = True,
149
        use_karras_sigmas: Optional[bool] = False,
150
        use_exponential_sigmas: Optional[bool] = False,
151
        use_beta_sigmas: Optional[bool] = False,
152
153
        use_flow_sigmas: Optional[bool] = False,
        flow_shift: Optional[float] = 1.0,
154
155
        timestep_spacing: str = "linspace",
        steps_offset: int = 0,
156
157
        use_dynamic_shifting: bool = False,
        time_shift_type: str = "exponential",
158
    ):
159
160
161
162
163
164
        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."
            )
165
166
167
168
169
170
        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.
171
            self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
172
173
174
175
        elif beta_schedule == "squaredcos_cap_v2":
            # Glide cosine schedule
            self.betas = betas_for_alpha_bar(num_train_timesteps)
        else:
176
            raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}")
177
178
179
180
181
182
183

        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)
184
        self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5
185
186
187
188
189
190
191

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

        # settings for DEIS
        if algorithm_type not in ["deis"]:
            if algorithm_type in ["dpmsolver", "dpmsolver++"]:
192
                self.register_to_config(algorithm_type="deis")
193
            else:
194
                raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}")
195
196

        if solver_type not in ["logrho"]:
197
            if solver_type in ["midpoint", "heun", "bh1", "bh2"]:
198
                self.register_to_config(solver_type="logrho")
199
            else:
200
                raise NotImplementedError(f"solver type {solver_type} is not implemented for {self.__class__}")
201
202
203
204
205
206
207

        # 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.lower_order_nums = 0
208
        self._step_index = None
209
        self._begin_index = None
210
        self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
211
212
213
214

    @property
    def step_index(self):
        """
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
215
        The index counter for current timestep. It will increase 1 after each scheduler step.
216
217
        """
        return self._step_index
218

219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
    @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

237
238
239
    def set_timesteps(
        self, num_inference_steps: int, device: Union[str, torch.device] = None, mu: Optional[float] = None
    ):
240
        """
241
        Sets the discrete timesteps used for the diffusion chain (to be run before inference).
242
243
244

        Args:
            num_inference_steps (`int`):
245
246
247
                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.
248
        """
249
250
251
        if mu is not None:
            assert self.config.use_dynamic_shifting and self.config.time_shift_type == "exponential"
            self.config.flow_shift = np.exp(mu)
Quentin Gallouédec's avatar
Quentin Gallouédec committed
252
        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
        if self.config.timestep_spacing == "linspace":
            timesteps = (
                np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps + 1)
                .round()[::-1][:-1]
                .copy()
                .astype(np.int64)
            )
        elif self.config.timestep_spacing == "leading":
            step_ratio = self.config.num_train_timesteps // (num_inference_steps + 1)
            # creates integer timesteps by multiplying by ratio
            # casting to int to avoid issues when num_inference_step is power of 3
            timesteps = (np.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(np.int64)
            timesteps += self.config.steps_offset
        elif self.config.timestep_spacing == "trailing":
            step_ratio = self.config.num_train_timesteps / num_inference_steps
            # creates integer timesteps by multiplying by ratio
            # casting to int to avoid issues when num_inference_step is power of 3
            timesteps = np.arange(self.config.num_train_timesteps, 0, -step_ratio).round().copy().astype(np.int64)
            timesteps -= 1
        else:
            raise ValueError(
                f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
            )
276

277
        sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
278
        log_sigmas = np.log(sigmas)
279
        if self.config.use_karras_sigmas:
280
            sigmas = np.flip(sigmas).copy()
281
282
            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()
283
            sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32)
284
        elif self.config.use_exponential_sigmas:
285
286
            sigmas = np.flip(sigmas).copy()
            sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
287
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
288
            sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32)
289
        elif self.config.use_beta_sigmas:
290
291
            sigmas = np.flip(sigmas).copy()
            sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
292
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
293
            sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32)
294
295
296
        elif self.config.use_flow_sigmas:
            alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1)
            sigmas = 1.0 - alphas
hlky's avatar
hlky committed
297
            sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1].copy()
298
            timesteps = (sigmas * self.config.num_train_timesteps).copy()
299
            sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32)
300
301
302
303
        else:
            sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
            sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5
            sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32)
304

305
306
        self.sigmas = torch.from_numpy(sigmas)
        self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64)
307
308
309

        self.num_inference_steps = len(timesteps)

310
311
312
313
314
        self.model_outputs = [
            None,
        ] * self.config.solver_order
        self.lower_order_nums = 0

315
316
        # add an index counter for schedulers that allow duplicated timesteps
        self._step_index = None
317
        self._begin_index = None
318
        self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
319

320
    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
321
    def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor:
322
323
324
325
326
327
328
        """
        "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."

Quentin Gallouédec's avatar
Quentin Gallouédec committed
329
        https://huggingface.co/papers/2205.11487
330
331
        """
        dtype = sample.dtype
332
        batch_size, channels, *remaining_dims = sample.shape
333
334
335
336
337

        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
338
        sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
339
340
341
342
343
344
345
346
347
348

        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"

349
        sample = sample.reshape(batch_size, channels, *remaining_dims)
350
351
352
        sample = sample.to(dtype)

        return sample
353

354
355
356
    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
    def _sigma_to_t(self, sigma, log_sigmas):
        # get log sigma
357
        log_sigma = np.log(np.maximum(sigma, 1e-10))
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379

        # 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

    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t
    def _sigma_to_alpha_sigma_t(self, sigma):
380
381
382
383
384
385
        if self.config.use_flow_sigmas:
            alpha_t = 1 - sigma
            sigma_t = sigma
        else:
            alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
            sigma_t = sigma * alpha_t
386
387
388
389

        return alpha_t, sigma_t

    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
390
    def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor:
391
392
        """Constructs the noise schedule of Karras et al. (2022)."""

Suraj Patil's avatar
Suraj Patil committed
393
394
395
396
397
398
399
400
401
402
403
404
405
406
        # 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()
407
408
409
410
411
412
413
414

        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

415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
    # 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()

434
        sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps))
435
436
        return sigmas

437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
    # 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()

458
        sigmas = np.array(
459
460
461
462
463
464
465
466
467
468
            [
                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

469
    def convert_model_output(
470
        self,
471
        model_output: torch.Tensor,
472
        *args,
473
        sample: torch.Tensor = None,
474
        **kwargs,
475
    ) -> torch.Tensor:
476
        """
477
        Convert the model output to the corresponding type the DEIS algorithm needs.
478
479

        Args:
480
            model_output (`torch.Tensor`):
481
482
483
                The direct output from the learned diffusion model.
            timestep (`int`):
                The current discrete timestep in the diffusion chain.
484
            sample (`torch.Tensor`):
485
                A current instance of a sample created by the diffusion process.
486
487

        Returns:
488
            `torch.Tensor`:
489
                The converted model output.
490
        """
491
492
493
494
495
        timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
        if sample is None:
            if len(args) > 1:
                sample = args[1]
            else:
496
                raise ValueError("missing `sample` as a required keyword argument")
497
498
499
500
501
502
503
504
505
        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`",
            )

        sigma = self.sigmas[self.step_index]
        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
506
507
508
509
510
511
        if self.config.prediction_type == "epsilon":
            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":
            x0_pred = alpha_t * sample - sigma_t * model_output
512
513
514
        elif self.config.prediction_type == "flow_prediction":
            sigma_t = self.sigmas[self.step_index]
            x0_pred = sample - sigma_t * model_output
515
516
        else:
            raise ValueError(
517
518
                f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, "
                "`v_prediction`, or `flow_prediction` for the DEISMultistepScheduler."
519
520
521
            )

        if self.config.thresholding:
522
            x0_pred = self._threshold_sample(x0_pred)
523
524
525
526
527
528
529
530

        if self.config.algorithm_type == "deis":
            return (sample - alpha_t * x0_pred) / sigma_t
        else:
            raise NotImplementedError("only support log-rho multistep deis now")

    def deis_first_order_update(
        self,
531
        model_output: torch.Tensor,
532
        *args,
533
        sample: torch.Tensor = None,
534
        **kwargs,
535
    ) -> torch.Tensor:
536
537
538
539
        """
        One step for the first-order DEIS (equivalent to DDIM).

        Args:
540
            model_output (`torch.Tensor`):
541
542
543
544
545
                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.
546
            sample (`torch.Tensor`):
547
                A current instance of a sample created by the diffusion process.
548
549

        Returns:
550
            `torch.Tensor`:
551
                The sample tensor at the previous timestep.
552
        """
553
554
555
556
557
558
        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:
559
                raise ValueError("missing `sample` as a required keyword argument")
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
        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)

580
581
582
583
584
585
586
587
588
        h = lambda_t - lambda_s
        if self.config.algorithm_type == "deis":
            x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output
        else:
            raise NotImplementedError("only support log-rho multistep deis now")
        return x_t

    def multistep_deis_second_order_update(
        self,
589
        model_output_list: List[torch.Tensor],
590
        *args,
591
        sample: torch.Tensor = None,
592
        **kwargs,
593
    ) -> torch.Tensor:
594
595
596
597
        """
        One step for the second-order multistep DEIS.

        Args:
598
            model_output_list (`List[torch.Tensor]`):
599
                The direct outputs from learned diffusion model at current and latter timesteps.
600
            sample (`torch.Tensor`):
601
                A current instance of a sample created by the diffusion process.
602
603

        Returns:
604
            `torch.Tensor`:
605
                The sample tensor at the previous timestep.
606
        """
607
608
609
610
611
612
        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:
613
                raise ValueError("missing `sample` as a required keyword argument")
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
        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)

638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
        m0, m1 = model_output_list[-1], model_output_list[-2]

        rho_t, rho_s0, rho_s1 = sigma_t / alpha_t, sigma_s0 / alpha_s0, sigma_s1 / alpha_s1

        if self.config.algorithm_type == "deis":

            def ind_fn(t, b, c):
                # Integrate[(log(t) - log(c)) / (log(b) - log(c)), {t}]
                return t * (-np.log(c) + np.log(t) - 1) / (np.log(b) - np.log(c))

            coef1 = ind_fn(rho_t, rho_s0, rho_s1) - ind_fn(rho_s0, rho_s0, rho_s1)
            coef2 = ind_fn(rho_t, rho_s1, rho_s0) - ind_fn(rho_s0, rho_s1, rho_s0)

            x_t = alpha_t * (sample / alpha_s0 + coef1 * m0 + coef2 * m1)
            return x_t
        else:
            raise NotImplementedError("only support log-rho multistep deis now")

    def multistep_deis_third_order_update(
        self,
658
        model_output_list: List[torch.Tensor],
659
        *args,
660
        sample: torch.Tensor = None,
661
        **kwargs,
662
    ) -> torch.Tensor:
663
664
665
666
        """
        One step for the third-order multistep DEIS.

        Args:
667
            model_output_list (`List[torch.Tensor]`):
668
                The direct outputs from learned diffusion model at current and latter timesteps.
669
            sample (`torch.Tensor`):
670
                A current instance of a sample created by diffusion process.
671
672

        Returns:
673
            `torch.Tensor`:
674
                The sample tensor at the previous timestep.
675
        """
676
677
678
679
680
681
682

        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:
683
                raise ValueError("missing `sample` as a required keyword argument")
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
        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],
        )

        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)

710
        m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3]
711

712
713
714
715
        rho_t, rho_s0, rho_s1, rho_s2 = (
            sigma_t / alpha_t,
            sigma_s0 / alpha_s0,
            sigma_s1 / alpha_s1,
716
            sigma_s2 / alpha_s2,
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
        )

        if self.config.algorithm_type == "deis":

            def ind_fn(t, b, c, d):
                # Integrate[(log(t) - log(c))(log(t) - log(d)) / (log(b) - log(c))(log(b) - log(d)), {t}]
                numerator = t * (
                    np.log(c) * (np.log(d) - np.log(t) + 1)
                    - np.log(d) * np.log(t)
                    + np.log(d)
                    + np.log(t) ** 2
                    - 2 * np.log(t)
                    + 2
                )
                denominator = (np.log(b) - np.log(c)) * (np.log(b) - np.log(d))
                return numerator / denominator

            coef1 = ind_fn(rho_t, rho_s0, rho_s1, rho_s2) - ind_fn(rho_s0, rho_s0, rho_s1, rho_s2)
            coef2 = ind_fn(rho_t, rho_s1, rho_s2, rho_s0) - ind_fn(rho_s0, rho_s1, rho_s2, rho_s0)
            coef3 = ind_fn(rho_t, rho_s2, rho_s0, rho_s1) - ind_fn(rho_s0, rho_s2, rho_s0, rho_s1)

            x_t = alpha_t * (sample / alpha_s0 + coef1 * m0 + coef2 * m1 + coef3 * m2)

            return x_t
        else:
            raise NotImplementedError("only support log-rho multistep deis now")

744
745
746
747
    # 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
748

749
        index_candidates = (schedule_timesteps == timestep).nonzero()
750
751
752
753
754
755
756
757
758
759
760
761

        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()

762
763
764
765
766
767
768
769
770
771
772
773
774
775
        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
776

777
778
    def step(
        self,
779
        model_output: torch.Tensor,
780
        timestep: Union[int, torch.Tensor],
781
        sample: torch.Tensor,
782
783
784
        return_dict: bool = True,
    ) -> Union[SchedulerOutput, Tuple]:
        """
785
786
        Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
        the multistep DEIS.
787
788

        Args:
789
            model_output (`torch.Tensor`):
790
                The direct output from learned diffusion model.
791
            timestep (`int`):
792
                The current discrete timestep in the diffusion chain.
793
            sample (`torch.Tensor`):
794
795
796
                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`.
797
798

        Returns:
799
800
801
            [`~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.
802
803
804
805
806
807
808

        """
        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"
            )

809
810
811
        if self.step_index is None:
            self._init_step_index(timestep)

812
        lower_order_final = (
813
            (self.step_index == len(self.timesteps) - 1) and self.config.lower_order_final and len(self.timesteps) < 15
814
815
        )
        lower_order_second = (
816
            (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15
817
818
        )

819
        model_output = self.convert_model_output(model_output, sample=sample)
820
821
822
823
824
        for i in range(self.config.solver_order - 1):
            self.model_outputs[i] = self.model_outputs[i + 1]
        self.model_outputs[-1] = model_output

        if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final:
825
            prev_sample = self.deis_first_order_update(model_output, sample=sample)
826
        elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second:
827
            prev_sample = self.multistep_deis_second_order_update(self.model_outputs, sample=sample)
828
        else:
829
            prev_sample = self.multistep_deis_third_order_update(self.model_outputs, sample=sample)
830
831
832
833

        if self.lower_order_nums < self.config.solver_order:
            self.lower_order_nums += 1

834
835
836
        # upon completion increase step index by one
        self._step_index += 1

837
838
839
840
841
        if not return_dict:
            return (prev_sample,)

        return SchedulerOutput(prev_sample=prev_sample)

842
    def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor:
843
844
845
846
847
        """
        Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
        current timestep.

        Args:
848
            sample (`torch.Tensor`):
849
                The input sample.
850
851

        Returns:
852
            `torch.Tensor`:
853
                A scaled input sample.
854
855
856
        """
        return sample

857
    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.add_noise
858
859
    def add_noise(
        self,
860
861
        original_samples: torch.Tensor,
        noise: torch.Tensor,
862
        timesteps: torch.IntTensor,
863
    ) -> torch.Tensor:
864
865
866
867
868
869
870
871
872
        # 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)
873

874
        # begin_index is None when the scheduler is used for training or pipeline does not implement set_begin_index
875
876
        if self.begin_index is None:
            step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
877
878
879
        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]
880
        else:
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
881
            # add noise is called before first denoising step to create initial latent(img2img)
882
            step_indices = [self.begin_index] * timesteps.shape[0]
883

884
885
886
        sigma = sigmas[step_indices].flatten()
        while len(sigma.shape) < len(original_samples.shape):
            sigma = sigma.unsqueeze(-1)
887

888
889
        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
        noisy_samples = alpha_t * original_samples + sigma_t * noise
890
891
892
893
        return noisy_samples

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