scheduling_dpmsolver_multistep.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
Dhruv Nair's avatar
Dhruv Nair committed
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
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
    return torch.tensor(betas, dtype=torch.float32)


78
79
80
81
82
83
84
# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
def rescale_zero_terminal_snr(betas):
    """
    Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)


    Args:
85
        betas (`torch.Tensor`):
86
87
88
            the betas that the scheduler is being initialized with.

    Returns:
89
        `torch.Tensor`: rescaled betas with zero terminal SNR
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
    """
    # Convert betas to alphas_bar_sqrt
    alphas = 1.0 - betas
    alphas_cumprod = torch.cumprod(alphas, dim=0)
    alphas_bar_sqrt = alphas_cumprod.sqrt()

    # Store old values.
    alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
    alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()

    # Shift so the last timestep is zero.
    alphas_bar_sqrt -= alphas_bar_sqrt_T

    # Scale so the first timestep is back to the old value.
    alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)

    # Convert alphas_bar_sqrt to betas
    alphas_bar = alphas_bar_sqrt**2  # Revert sqrt
    alphas = alphas_bar[1:] / alphas_bar[:-1]  # Revert cumprod
    alphas = torch.cat([alphas_bar[0:1], alphas])
    betas = 1 - alphas

    return betas


115
116
class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
    """
117
    `DPMSolverMultistepScheduler` is a fast dedicated high-order solver for diffusion ODEs.
118

119
120
    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.
121
122

    Args:
123
124
125
126
127
128
129
130
        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
131
            `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
132
133
134
135
        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
136
            sampling, and `solver_order=3` for unconditional sampling.
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
        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++`):
            Algorithm type for the solver; can be `dpmsolver`, `dpmsolver++`, `sde-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.
        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.
161
162
163
164
        euler_at_final (`bool`, defaults to `False`):
            Whether to use Euler's method in the final step. It is a trade-off between numerical stability and detail
            richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference
            steps, but sometimes may result in blurring.
165
        use_karras_sigmas (`bool`, *optional*, defaults to `False`):
166
167
            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}.
168
169
        use_exponential_sigmas (`bool`, *optional*, defaults to `False`):
            Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process.
170
171
172
        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.
173
174
175
176
        use_lu_lambdas (`bool`, *optional*, defaults to `False`):
            Whether to use the uniform-logSNR for step sizes proposed by Lu's DPM-Solver in the noise schedule during
            the sampling process. If `True`, the sigmas and time steps are determined according to a sequence of
            `lambda(t)`.
177
        final_sigmas_type (`str`, defaults to `"zero"`):
178
179
            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.
180
181
182
        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.
183
        variance_type (`str`, *optional*):
184
185
186
187
188
189
            Set to "learned" or "learned_range" for diffusion models that predict variance. If set, the model's output
            contains the predicted Gaussian variance.
        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):
190
            An offset added to the inference steps, as required by some model families.
191
192
193
194
        rescale_betas_zero_snr (`bool`, defaults to `False`):
            Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
            dark samples instead of limiting it to samples with medium brightness. Loosely related to
            [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
195
196
    """

Kashif Rasul's avatar
Kashif Rasul committed
197
    _compatibles = [e.name for e in KarrasDiffusionSchedulers]
198
    order = 1
199
200
201
202
203
204
205
206

    @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",
207
        trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
208
        solver_order: int = 2,
209
        prediction_type: str = "epsilon",
210
211
212
213
214
215
        thresholding: bool = False,
        dynamic_thresholding_ratio: float = 0.995,
        sample_max_value: float = 1.0,
        algorithm_type: str = "dpmsolver++",
        solver_type: str = "midpoint",
        lower_order_final: bool = True,
216
        euler_at_final: bool = False,
217
        use_karras_sigmas: Optional[bool] = False,
218
        use_exponential_sigmas: Optional[bool] = False,
219
        use_beta_sigmas: Optional[bool] = False,
220
        use_lu_lambdas: Optional[bool] = False,
221
        final_sigmas_type: Optional[str] = "zero",  # "zero", "sigma_min"
222
223
        lambda_min_clipped: float = -float("inf"),
        variance_type: Optional[str] = None,
224
225
        timestep_spacing: str = "linspace",
        steps_offset: int = 0,
226
        rescale_betas_zero_snr: bool = False,
227
    ):
228
229
230
231
232
233
        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."
            )
234
235
236
237
        if algorithm_type in ["dpmsolver", "sde-dpmsolver"]:
            deprecation_message = f"algorithm_type {algorithm_type} is deprecated and will be removed in a future version. Choose from `dpmsolver++` or `sde-dpmsolver++` instead"
            deprecate("algorithm_types dpmsolver and sde-dpmsolver", "1.0.0", deprecation_message)

238
        if trained_betas is not None:
239
            self.betas = torch.tensor(trained_betas, dtype=torch.float32)
240
241
242
243
        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.
244
            self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
245
246
247
248
        elif beta_schedule == "squaredcos_cap_v2":
            # Glide cosine schedule
            self.betas = betas_for_alpha_bar(num_train_timesteps)
        else:
249
            raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}")
250

251
252
253
        if rescale_betas_zero_snr:
            self.betas = rescale_zero_terminal_snr(self.betas)

254
255
        self.alphas = 1.0 - self.betas
        self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
256
257
258
259
260
261

        if rescale_betas_zero_snr:
            # Close to 0 without being 0 so first sigma is not inf
            # FP16 smallest positive subnormal works well here
            self.alphas_cumprod[-1] = 2**-24

262
263
264
265
        # 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)
266
        self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5
267
268
269
270
271

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

        # settings for DPM-Solver
272
        if algorithm_type not in ["dpmsolver", "dpmsolver++", "sde-dpmsolver", "sde-dpmsolver++"]:
273
            if algorithm_type == "deis":
274
                self.register_to_config(algorithm_type="dpmsolver++")
275
            else:
276
                raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}")
277

278
        if solver_type not in ["midpoint", "heun"]:
279
            if solver_type in ["logrho", "bh1", "bh2"]:
280
                self.register_to_config(solver_type="midpoint")
281
            else:
282
                raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}")
283

284
285
286
287
288
        if algorithm_type not in ["dpmsolver++", "sde-dpmsolver++"] and final_sigmas_type == "zero":
            raise ValueError(
                f"`final_sigmas_type` {final_sigmas_type} is not supported for `algorithm_type` {algorithm_type}. Please choose `sigma_min` instead."
            )

289
290
291
292
293
294
        # 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
295
        self._step_index = None
296
        self._begin_index = None
297
        self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
298
299
300
301

    @property
    def step_index(self):
        """
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
302
        The index counter for current timestep. It will increase 1 after each scheduler step.
303
304
        """
        return self._step_index
305

306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
    @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

    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

323
324
325
326
327
328
    def set_timesteps(
        self,
        num_inference_steps: int = None,
        device: Union[str, torch.device] = None,
        timesteps: Optional[List[int]] = None,
    ):
329
        """
330
        Sets the discrete timesteps used for the diffusion chain (to be run before inference).
331
332
333

        Args:
            num_inference_steps (`int`):
334
335
336
                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.
337
338
339
340
            timesteps (`List[int]`, *optional*):
                Custom timesteps used to support arbitrary timesteps schedule. If `None`, timesteps will be generated
                based on the `timestep_spacing` attribute. If `timesteps` is passed, `num_inference_steps` and `sigmas`
                must be `None`, and `timestep_spacing` attribute will be ignored.
341
        """
342
343
344
345
346
347
348
349
        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("Can only pass one of `num_inference_steps` or `custom_timesteps`.")
        if timesteps is not None and self.config.use_karras_sigmas:
            raise ValueError("Cannot use `timesteps` with `config.use_karras_sigmas = True`")
        if timesteps is not None and self.config.use_lu_lambdas:
            raise ValueError("Cannot use `timesteps` with `config.use_lu_lambdas = True`")
350
351
        if timesteps is not None and self.config.use_exponential_sigmas:
            raise ValueError("Cannot set `timesteps` with `config.use_exponential_sigmas = True`.")
352
353
        if timesteps is not None and self.config.use_beta_sigmas:
            raise ValueError("Cannot set `timesteps` with `config.use_beta_sigmas = True`.")
354
355
356

        if timesteps is not None:
            timesteps = np.array(timesteps).astype(np.int64)
357
        else:
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
            # 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)
            last_timestep = ((self.config.num_train_timesteps - clipped_idx).numpy()).item()

            # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
            if self.config.timestep_spacing == "linspace":
                timesteps = (
                    np.linspace(0, last_timestep - 1, num_inference_steps + 1)
                    .round()[::-1][:-1]
                    .copy()
                    .astype(np.int64)
                )
            elif self.config.timestep_spacing == "leading":
                step_ratio = last_timestep // (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(last_timestep, 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'."
                )
389

390
        sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
391
392
        log_sigmas = np.log(sigmas)

393
        if self.config.use_karras_sigmas:
394
            sigmas = np.flip(sigmas).copy()
395
396
            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()
397
398
399
400
401
        elif self.config.use_lu_lambdas:
            lambdas = np.flip(log_sigmas.copy())
            lambdas = self._convert_to_lu(in_lambdas=lambdas, num_inference_steps=num_inference_steps)
            sigmas = np.exp(lambdas)
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round()
402
        elif self.config.use_exponential_sigmas:
403
404
            sigmas = np.flip(sigmas).copy()
            sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
405
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
406
        elif self.config.use_beta_sigmas:
407
408
            sigmas = np.flip(sigmas).copy()
            sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
409
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
410
411
        else:
            sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
412
413

        if self.config.final_sigmas_type == "sigma_min":
414
            sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5
415
416
417
418
419
420
421
422
        elif self.config.final_sigmas_type == "zero":
            sigma_last = 0
        else:
            raise ValueError(
                f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}"
            )

        sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32)
423

424
425
        self.sigmas = torch.from_numpy(sigmas)
        self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64)
426
427
428

        self.num_inference_steps = len(timesteps)

429
430
431
432
433
        self.model_outputs = [
            None,
        ] * self.config.solver_order
        self.lower_order_nums = 0

434
435
        # add an index counter for schedulers that allow duplicated timesteps
        self._step_index = None
436
        self._begin_index = None
437
        self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
438

439
    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
440
    def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor:
441
442
443
444
445
446
447
448
449
450
        """
        "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
451
        batch_size, channels, *remaining_dims = sample.shape
452
453
454
455
456

        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
457
        sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
458
459
460
461
462
463
464
465
466
467

        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"

468
        sample = sample.reshape(batch_size, channels, *remaining_dims)
469
470
471
        sample = sample.to(dtype)

        return sample
472

473
474
475
    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
    def _sigma_to_t(self, sigma, log_sigmas):
        # get log sigma
476
        log_sigma = np.log(np.maximum(sigma, 1e-10))
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496

        # 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

497
498
499
500
501
502
    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

503
    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
504
    def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor:
505
506
        """Constructs the noise schedule of Karras et al. (2022)."""

Suraj Patil's avatar
Suraj Patil committed
507
508
509
510
511
512
513
514
515
516
517
518
519
520
        # 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()
521
522
523
524
525
526
527
528

        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

529
    def _convert_to_lu(self, in_lambdas: torch.Tensor, num_inference_steps) -> torch.Tensor:
530
531
532
533
534
535
536
537
538
539
540
541
        """Constructs the noise schedule of Lu et al. (2022)."""

        lambda_min: float = in_lambdas[-1].item()
        lambda_max: float = in_lambdas[0].item()

        rho = 1.0  # 1.0 is the value used in the paper
        ramp = np.linspace(0, 1, num_inference_steps)
        min_inv_rho = lambda_min ** (1 / rho)
        max_inv_rho = lambda_max ** (1 / rho)
        lambdas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
        return lambdas

542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
    # 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()

561
        sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps))
562
563
        return sigmas

564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
    # 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()

585
        sigmas = np.array(
586
587
588
589
590
591
592
593
594
595
            [
                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

596
    def convert_model_output(
597
        self,
598
        model_output: torch.Tensor,
599
        *args,
600
        sample: torch.Tensor = None,
601
        **kwargs,
602
    ) -> torch.Tensor:
603
        """
604
605
606
        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.
607

608
        <Tip>
609

610
611
612
613
        The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both noise
        prediction and data prediction models.

        </Tip>
614
615

        Args:
616
            model_output (`torch.Tensor`):
617
                The direct output from the learned diffusion model.
618
            sample (`torch.Tensor`):
619
                A current instance of a sample created by the diffusion process.
620
621

        Returns:
622
            `torch.Tensor`:
623
                The converted model output.
624
        """
625
626
627
628
629
630
631
632
633
634
635
636
        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`",
            )
637

638
        # DPM-Solver++ needs to solve an integral of the data prediction model.
639
        if self.config.algorithm_type in ["dpmsolver++", "sde-dpmsolver++"]:
640
            if self.config.prediction_type == "epsilon":
641
                # DPM-Solver and DPM-Solver++ only need the "mean" output.
642
                if self.config.variance_type in ["learned", "learned_range"]:
643
                    model_output = model_output[:, :3]
644
645
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
646
                x0_pred = (sample - sigma_t * model_output) / alpha_t
647
            elif self.config.prediction_type == "sample":
648
                x0_pred = model_output
649
            elif self.config.prediction_type == "v_prediction":
650
651
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
652
                x0_pred = alpha_t * sample - sigma_t * model_output
653
654
            else:
                raise ValueError(
655
656
                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
                    " `v_prediction` for the DPMSolverMultistepScheduler."
657
658
                )

659
            if self.config.thresholding:
660
661
                x0_pred = self._threshold_sample(x0_pred)

662
            return x0_pred
663

664
        # DPM-Solver needs to solve an integral of the noise prediction model.
665
        elif self.config.algorithm_type in ["dpmsolver", "sde-dpmsolver"]:
666
            if self.config.prediction_type == "epsilon":
667
                # DPM-Solver and DPM-Solver++ only need the "mean" output.
668
669
670
671
                if self.config.variance_type in ["learned", "learned_range"]:
                    epsilon = model_output[:, :3]
                else:
                    epsilon = model_output
672
            elif self.config.prediction_type == "sample":
673
674
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
675
                epsilon = (sample - alpha_t * model_output) / sigma_t
676
            elif self.config.prediction_type == "v_prediction":
677
678
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
679
                epsilon = alpha_t * model_output + sigma_t * sample
680
681
            else:
                raise ValueError(
682
683
                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
                    " `v_prediction` for the DPMSolverMultistepScheduler."
684
                )
685

686
            if self.config.thresholding:
687
688
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
689
690
691
692
693
694
                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

695
696
    def dpm_solver_first_order_update(
        self,
697
        model_output: torch.Tensor,
698
        *args,
699
700
        sample: torch.Tensor = None,
        noise: Optional[torch.Tensor] = None,
701
        **kwargs,
702
    ) -> torch.Tensor:
703
        """
704
        One step for the first-order DPMSolver (equivalent to DDIM).
705
706

        Args:
707
            model_output (`torch.Tensor`):
708
                The direct output from the learned diffusion model.
709
            sample (`torch.Tensor`):
710
                A current instance of a sample created by the diffusion process.
711
712

        Returns:
713
            `torch.Tensor`:
714
                The sample tensor at the previous timestep.
715
        """
716
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
        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)

743
744
745
746
747
        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
748
749
750
751
752
753
754
755
756
757
758
759
760
761
        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
            )
        elif self.config.algorithm_type == "sde-dpmsolver":
            assert noise is not None
            x_t = (
                (alpha_t / alpha_s) * sample
                - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * model_output
                + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
            )
762
763
764
765
        return x_t

    def multistep_dpm_solver_second_order_update(
        self,
766
        model_output_list: List[torch.Tensor],
767
        *args,
768
769
        sample: torch.Tensor = None,
        noise: Optional[torch.Tensor] = None,
770
        **kwargs,
771
    ) -> torch.Tensor:
772
        """
773
        One step for the second-order multistep DPMSolver.
774
775

        Args:
776
            model_output_list (`List[torch.Tensor]`):
777
                The direct outputs from learned diffusion model at current and latter timesteps.
778
            sample (`torch.Tensor`):
779
                A current instance of a sample created by the diffusion process.
780
781

        Returns:
782
            `torch.Tensor`:
783
                The sample tensor at the previous timestep.
784
        """
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
        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)

820
        m0, m1 = model_output_list[-1], model_output_list[-2]
821

822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
        h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1
        r0 = h_0 / h
        D0, D1 = m0, (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_s0) * 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_s0) * 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_s0) * 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_s0) * sample
                    - (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
                )
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
        elif self.config.algorithm_type == "sde-dpmsolver++":
            assert noise is not None
            if self.config.solver_type == "midpoint":
                x_t = (
                    (sigma_t / sigma_s0 * 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_s0 * 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
                )
        elif self.config.algorithm_type == "sde-dpmsolver":
            assert noise is not None
            if self.config.solver_type == "midpoint":
                x_t = (
                    (alpha_t / alpha_s0) * sample
                    - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - (sigma_t * (torch.exp(h) - 1.0)) * D1
                    + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (alpha_t / alpha_s0) * sample
                    - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - 2.0 * (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
                    + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
                )
885
886
887
888
        return x_t

    def multistep_dpm_solver_third_order_update(
        self,
889
        model_output_list: List[torch.Tensor],
890
        *args,
891
        sample: torch.Tensor = None,
892
        **kwargs,
893
    ) -> torch.Tensor:
894
        """
895
        One step for the third-order multistep DPMSolver.
896
897

        Args:
898
            model_output_list (`List[torch.Tensor]`):
899
                The direct outputs from learned diffusion model at current and latter timesteps.
900
            sample (`torch.Tensor`):
901
                A current instance of a sample created by diffusion process.
902
903

        Returns:
904
            `torch.Tensor`:
905
                The sample tensor at the previous timestep.
906
        """
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933

        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],
934
        )
935
936
937
938
939
940
941
942
943
944
945
946
947

        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]

948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
        h, h_0, h_1 = lambda_t - lambda_s0, lambda_s0 - lambda_s1, lambda_s1 - lambda_s2
        r0, r1 = h_0 / h, h_1 / h
        D0 = m0
        D1_0, D1_1 = (1.0 / r0) * (m0 - m1), (1.0 / r1) * (m1 - m2)
        D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1)
        D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1)
        if self.config.algorithm_type == "dpmsolver++":
            # See https://arxiv.org/abs/2206.00927 for detailed derivations
            x_t = (
                (sigma_t / sigma_s0) * 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
            x_t = (
                (alpha_t / alpha_s0) * 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
            )
        return x_t

972
973
974
    def index_for_timestep(self, timestep, schedule_timesteps=None):
        if schedule_timesteps is None:
            schedule_timesteps = self.timesteps
975

976
        index_candidates = (schedule_timesteps == timestep).nonzero()
977
978
979
980
981
982
983
984
985
986
987
988

        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()

989
990
991
992
993
994
995
996
997
998
999
1000
1001
        return 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
1002

1003
1004
    def step(
        self,
1005
        model_output: torch.Tensor,
1006
        timestep: Union[int, torch.Tensor],
1007
        sample: torch.Tensor,
1008
        generator=None,
1009
        variance_noise: Optional[torch.Tensor] = None,
1010
1011
1012
        return_dict: bool = True,
    ) -> Union[SchedulerOutput, Tuple]:
        """
1013
1014
        Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
        the multistep DPMSolver.
1015
1016

        Args:
1017
            model_output (`torch.Tensor`):
1018
1019
1020
                The direct output from learned diffusion model.
            timestep (`int`):
                The current discrete timestep in the diffusion chain.
1021
            sample (`torch.Tensor`):
1022
1023
1024
                A current instance of a sample created by the diffusion process.
            generator (`torch.Generator`, *optional*):
                A random number generator.
1025
            variance_noise (`torch.Tensor`):
1026
1027
                Alternative to generating noise with `generator` by directly providing the noise for the variance
                itself. Useful for methods such as [`LEdits++`].
1028
1029
            return_dict (`bool`):
                Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`.
1030
1031

        Returns:
1032
1033
1034
            [`~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.
1035
1036
1037
1038
1039
1040
1041

        """
        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"
            )

1042
1043
1044
        if self.step_index is None:
            self._init_step_index(timestep)

1045
1046
        # Improve numerical stability for small number of steps
        lower_order_final = (self.step_index == len(self.timesteps) - 1) and (
1047
1048
1049
            self.config.euler_at_final
            or (self.config.lower_order_final and len(self.timesteps) < 15)
            or self.config.final_sigmas_type == "zero"
1050
1051
        )
        lower_order_second = (
1052
            (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15
1053
1054
        )

1055
        model_output = self.convert_model_output(model_output, sample=sample)
1056
1057
1058
1059
        for i in range(self.config.solver_order - 1):
            self.model_outputs[i] = self.model_outputs[i + 1]
        self.model_outputs[-1] = model_output

1060
1061
        # Upcast to avoid precision issues when computing prev_sample
        sample = sample.to(torch.float32)
1062
        if self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"] and variance_noise is None:
1063
            noise = randn_tensor(
1064
                model_output.shape, generator=generator, device=model_output.device, dtype=torch.float32
1065
            )
1066
1067
        elif self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"]:
            noise = variance_noise.to(device=model_output.device, dtype=torch.float32)
1068
1069
1070
        else:
            noise = None

1071
        if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final:
1072
            prev_sample = self.dpm_solver_first_order_update(model_output, sample=sample, noise=noise)
1073
        elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second:
1074
            prev_sample = self.multistep_dpm_solver_second_order_update(self.model_outputs, sample=sample, noise=noise)
1075
        else:
1076
            prev_sample = self.multistep_dpm_solver_third_order_update(self.model_outputs, sample=sample)
1077
1078
1079
1080

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

1081
1082
1083
        # Cast sample back to expected dtype
        prev_sample = prev_sample.to(model_output.dtype)

1084
1085
1086
        # upon completion increase step index by one
        self._step_index += 1

1087
1088
1089
1090
1091
        if not return_dict:
            return (prev_sample,)

        return SchedulerOutput(prev_sample=prev_sample)

1092
    def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor:
1093
1094
1095
1096
1097
        """
        Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
        current timestep.

        Args:
1098
            sample (`torch.Tensor`):
1099
                The input sample.
1100
1101

        Returns:
1102
            `torch.Tensor`:
1103
                A scaled input sample.
1104
1105
1106
1107
1108
        """
        return sample

    def add_noise(
        self,
1109
1110
        original_samples: torch.Tensor,
        noise: torch.Tensor,
1111
        timesteps: torch.IntTensor,
1112
    ) -> torch.Tensor:
1113
1114
1115
1116
1117
1118
1119
1120
1121
        # 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)
1122

1123
        # begin_index is None when the scheduler is used for training or pipeline does not implement set_begin_index
1124
1125
        if self.begin_index is None:
            step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
1126
1127
1128
        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]
1129
        else:
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
1130
            # add noise is called before first denoising step to create initial latent(img2img)
1131
            step_indices = [self.begin_index] * timesteps.shape[0]
1132

1133
1134
1135
        sigma = sigmas[step_indices].flatten()
        while len(sigma.shape) < len(original_samples.shape):
            sigma = sigma.unsqueeze(-1)
1136

1137
1138
        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
        noisy_samples = alpha_t * original_samples + sigma_t * noise
1139
1140
1141
1142
        return noisy_samples

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