scheduling_euler_discrete.py 31.8 KB
Newer Older
1
# Copyright 2024 Katherine Crowson and The HuggingFace Team. All rights reserved.
hlky's avatar
hlky committed
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.

15
import math
hlky's avatar
hlky committed
16
from dataclasses import dataclass
17
from typing import List, Optional, Tuple, Union
hlky's avatar
hlky committed
18
19
20
21
22

import numpy as np
import torch

from ..configuration_utils import ConfigMixin, register_to_config
Dhruv Nair's avatar
Dhruv Nair committed
23
24
from ..utils import BaseOutput, logging
from ..utils.torch_utils import randn_tensor
Kashif Rasul's avatar
Kashif Rasul committed
25
from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
hlky's avatar
hlky committed
26
27
28
29
30
31
32
33
34


logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


@dataclass
# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->EulerDiscrete
class EulerDiscreteSchedulerOutput(BaseOutput):
    """
35
    Output class for the scheduler's `step` function output.
hlky's avatar
hlky committed
36
37

    Args:
38
        prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images):
39
            Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
hlky's avatar
hlky committed
40
            denoising loop.
41
        pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images):
42
            The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
hlky's avatar
hlky committed
43
44
45
            `pred_original_sample` can be used to preview progress or for guidance.
    """

46
47
    prev_sample: torch.Tensor
    pred_original_sample: Optional[torch.Tensor] = None
hlky's avatar
hlky committed
48
49


50
# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
YiYi Xu's avatar
YiYi Xu committed
51
52
53
54
55
def betas_for_alpha_bar(
    num_diffusion_timesteps,
    max_beta=0.999,
    alpha_transform_type="cosine",
):
56
57
58
59
60
61
62
63
64
65
66
67
    """
    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
68
69
        alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
                     Choose from `cosine` or `exp`
70
71
72
73

    Returns:
        betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
    """
YiYi Xu's avatar
YiYi Xu committed
74
    if alpha_transform_type == "cosine":
75

YiYi Xu's avatar
YiYi Xu committed
76
77
78
79
80
81
82
83
84
        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
85
        raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}")
86
87
88
89
90

    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
91
        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
92
93
94
    return torch.tensor(betas, dtype=torch.float32)


95
96
97
98
99
100
101
# 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:
102
        betas (`torch.Tensor`):
103
104
105
            the betas that the scheduler is being initialized with.

    Returns:
106
        `torch.Tensor`: rescaled betas with zero terminal SNR
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
    """
    # 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


hlky's avatar
hlky committed
132
133
class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
    """
134
    Euler scheduler.
hlky's avatar
hlky committed
135

136
137
    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.
hlky's avatar
hlky committed
138
139

    Args:
140
141
142
143
144
145
146
147
        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
hlky's avatar
hlky committed
148
            `linear` or `scaled_linear`.
149
150
151
152
153
154
155
156
157
        trained_betas (`np.ndarray`, *optional*):
            Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
        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).
        interpolation_type(`str`, defaults to `"linear"`, *optional*):
            The interpolation type to compute intermediate sigmas for the scheduler denoising steps. Should be on of
            `"linear"` or `"log_linear"`.
158
        use_karras_sigmas (`bool`, *optional*, defaults to `False`):
159
160
            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}.
161
162
        use_exponential_sigmas (`bool`, *optional*, defaults to `False`):
            Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process.
163
164
165
166
        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):
167
            An offset added to the inference steps, as required by some model families.
168
169
170
171
        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).
172
173
174
        final_sigmas_type (`str`, defaults to `"zero"`):
            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.
hlky's avatar
hlky committed
175
176
    """

Kashif Rasul's avatar
Kashif Rasul committed
177
    _compatibles = [e.name for e in KarrasDiffusionSchedulers]
178
    order = 1
179

hlky's avatar
hlky committed
180
181
182
183
184
185
186
    @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",
187
        trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
Suraj Patil's avatar
Suraj Patil committed
188
        prediction_type: str = "epsilon",
189
        interpolation_type: str = "linear",
190
        use_karras_sigmas: Optional[bool] = False,
191
        use_exponential_sigmas: Optional[bool] = False,
Suraj Patil's avatar
Suraj Patil committed
192
193
        sigma_min: Optional[float] = None,
        sigma_max: Optional[float] = None,
194
        timestep_spacing: str = "linspace",
Suraj Patil's avatar
Suraj Patil committed
195
        timestep_type: str = "discrete",  # can be "discrete" or "continuous"
196
        steps_offset: int = 0,
197
        rescale_betas_zero_snr: bool = False,
198
        final_sigmas_type: str = "zero",  # can be "zero" or "sigma_min"
hlky's avatar
hlky committed
199
    ):
200
201
        if sum([self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1:
            raise ValueError("Only one of `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used.")
hlky's avatar
hlky committed
202
        if trained_betas is not None:
203
            self.betas = torch.tensor(trained_betas, dtype=torch.float32)
hlky's avatar
hlky committed
204
205
206
207
        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.
208
            self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
209
210
211
        elif beta_schedule == "squaredcos_cap_v2":
            # Glide cosine schedule
            self.betas = betas_for_alpha_bar(num_train_timesteps)
hlky's avatar
hlky committed
212
        else:
213
            raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}")
hlky's avatar
hlky committed
214

215
216
217
        if rescale_betas_zero_snr:
            self.betas = rescale_zero_terminal_snr(self.betas)

hlky's avatar
hlky committed
218
219
220
        self.alphas = 1.0 - self.betas
        self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)

221
222
223
224
225
        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

226
        sigmas = (((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5).flip(0)
Suraj Patil's avatar
Suraj Patil committed
227
228
        timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=float)[::-1].copy()
        timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32)
hlky's avatar
hlky committed
229
230
231

        # setable values
        self.num_inference_steps = None
Suraj Patil's avatar
Suraj Patil committed
232
233
234
235
236
237
238
239
240

        # TODO: Support the full EDM scalings for all prediction types and timestep types
        if timestep_type == "continuous" and prediction_type == "v_prediction":
            self.timesteps = torch.Tensor([0.25 * sigma.log() for sigma in sigmas])
        else:
            self.timesteps = timesteps

        self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])

hlky's avatar
hlky committed
241
        self.is_scale_input_called = False
242
        self.use_karras_sigmas = use_karras_sigmas
243
        self.use_exponential_sigmas = use_exponential_sigmas
hlky's avatar
hlky committed
244

YiYi Xu's avatar
YiYi Xu committed
245
        self._step_index = None
246
        self._begin_index = None
247
        self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
YiYi Xu's avatar
YiYi Xu committed
248

249
250
251
    @property
    def init_noise_sigma(self):
        # standard deviation of the initial noise distribution
252
        max_sigma = max(self.sigmas) if isinstance(self.sigmas, list) else self.sigmas.max()
253
        if self.config.timestep_spacing in ["linspace", "trailing"]:
254
            return max_sigma
255

256
        return (max_sigma**2 + 1) ** 0.5
257

YiYi Xu's avatar
YiYi Xu committed
258
259
260
    @property
    def step_index(self):
        """
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
261
        The index counter for current timestep. It will increase 1 after each scheduler step.
YiYi Xu's avatar
YiYi Xu committed
262
263
264
        """
        return self._step_index

265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
    @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

283
    def scale_model_input(self, sample: torch.Tensor, timestep: Union[float, torch.Tensor]) -> torch.Tensor:
hlky's avatar
hlky committed
284
        """
285
286
        Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
        current timestep. Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm.
hlky's avatar
hlky committed
287
288

        Args:
289
            sample (`torch.Tensor`):
290
291
292
                The input sample.
            timestep (`int`, *optional*):
                The current timestep in the diffusion chain.
hlky's avatar
hlky committed
293
294

        Returns:
295
            `torch.Tensor`:
296
                A scaled input sample.
hlky's avatar
hlky committed
297
        """
YiYi Xu's avatar
YiYi Xu committed
298
299
        if self.step_index is None:
            self._init_step_index(timestep)
300

YiYi Xu's avatar
YiYi Xu committed
301
        sigma = self.sigmas[self.step_index]
hlky's avatar
hlky committed
302
        sample = sample / ((sigma**2 + 1) ** 0.5)
303

hlky's avatar
hlky committed
304
305
306
        self.is_scale_input_called = True
        return sample

307
308
309
310
311
312
313
    def set_timesteps(
        self,
        num_inference_steps: int = None,
        device: Union[str, torch.device] = None,
        timesteps: Optional[List[int]] = None,
        sigmas: Optional[List[float]] = None,
    ):
hlky's avatar
hlky committed
314
        """
315
        Sets the discrete timesteps used for the diffusion chain (to be run before inference).
hlky's avatar
hlky committed
316
317
318

        Args:
            num_inference_steps (`int`):
319
320
321
                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.
322
323
324
325
326
327
328
329
330
            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.
            sigmas (`List[float]`, *optional*):
                Custom sigmas used to support arbitrary timesteps schedule schedule. If `None`, timesteps and sigmas
                will be generated based on the relevant scheduler attributes. If `sigmas` is passed,
                `num_inference_steps` and `timesteps` must be `None`, and the timesteps will be generated based on the
                custom sigmas schedule.
hlky's avatar
hlky committed
331
332
        """

333
334
335
336
337
338
339
340
        if timesteps is not None and sigmas is not None:
            raise ValueError("Only one of `timesteps` or `sigmas` should be set.")
        if num_inference_steps is None and timesteps is None and sigmas is None:
            raise ValueError("Must pass exactly one of `num_inference_steps` or `timesteps` or `sigmas.")
        if num_inference_steps is not None and (timesteps is not None or sigmas is not None):
            raise ValueError("Can only pass one of `num_inference_steps` or `timesteps` or `sigmas`.")
        if timesteps is not None and self.config.use_karras_sigmas:
            raise ValueError("Cannot set `timesteps` with `config.use_karras_sigmas = True`.")
341
342
        if timesteps is not None and self.config.use_exponential_sigmas:
            raise ValueError("Cannot set `timesteps` with `config.use_exponential_sigmas = True`.")
343
344
345
346
347
        if (
            timesteps is not None
            and self.config.timestep_type == "continuous"
            and self.config.prediction_type == "v_prediction"
        ):
348
            raise ValueError(
349
                "Cannot set `timesteps` with `config.timestep_type = 'continuous'` and `config.prediction_type = 'v_prediction'`."
350
351
            )

352
353
354
        if num_inference_steps is None:
            num_inference_steps = len(timesteps) if timesteps is not None else len(sigmas) - 1
        self.num_inference_steps = num_inference_steps
355

356
357
358
359
        if sigmas is not None:
            log_sigmas = np.log(np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5))
            sigmas = np.array(sigmas).astype(np.float32)
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas[:-1]])
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
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
        else:
            if timesteps is not None:
                timesteps = np.array(timesteps).astype(np.float32)
            else:
                # "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, self.config.num_train_timesteps - 1, num_inference_steps, dtype=np.float32
                    )[::-1].copy()
                elif self.config.timestep_spacing == "leading":
                    step_ratio = self.config.num_train_timesteps // self.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(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32)
                    )
                    timesteps += self.config.steps_offset
                elif self.config.timestep_spacing == "trailing":
                    step_ratio = self.config.num_train_timesteps / self.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.float32)
                    )
                    timesteps -= 1
                else:
                    raise ValueError(
                        f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
                    )

            sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
            log_sigmas = np.log(sigmas)
            if self.config.interpolation_type == "linear":
                sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
            elif self.config.interpolation_type == "log_linear":
                sigmas = torch.linspace(np.log(sigmas[-1]), np.log(sigmas[0]), num_inference_steps + 1).exp().numpy()
            else:
                raise ValueError(
                    f"{self.config.interpolation_type} is not implemented. Please specify interpolation_type to either"
                    " 'linear' or 'log_linear'"
                )

            if self.config.use_karras_sigmas:
                sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=self.num_inference_steps)
                timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])

407
408
409
410
            elif self.config.use_exponential_sigmas:
                sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=self.num_inference_steps)
                timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])

411
412
413
414
415
416
417
418
419
420
            if self.config.final_sigmas_type == "sigma_min":
                sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5
            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)
421

Suraj Patil's avatar
Suraj Patil committed
422
        sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device)
YiYi Xu's avatar
YiYi Xu committed
423

Suraj Patil's avatar
Suraj Patil committed
424
425
        # TODO: Support the full EDM scalings for all prediction types and timestep types
        if self.config.timestep_type == "continuous" and self.config.prediction_type == "v_prediction":
426
            self.timesteps = torch.Tensor([0.25 * sigma.log() for sigma in sigmas[:-1]]).to(device=device)
Suraj Patil's avatar
Suraj Patil committed
427
428
429
        else:
            self.timesteps = torch.from_numpy(timesteps.astype(np.float32)).to(device=device)

YiYi Xu's avatar
YiYi Xu committed
430
        self._step_index = None
431
        self._begin_index = None
432
        self.sigmas = sigmas.to("cpu")  # to avoid too much CPU/GPU communication
hlky's avatar
hlky committed
433

434
435
    def _sigma_to_t(self, sigma, log_sigmas):
        # get log sigma
436
        log_sigma = np.log(np.maximum(sigma, 1e-10))
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457

        # 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 https://github.com/crowsonkb/k-diffusion/blob/686dbad0f39640ea25c8a8c6a6e56bb40eacefa2/k_diffusion/sampling.py#L17
458
    def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor:
459
460
        """Constructs the noise schedule of Karras et al. (2022)."""

Suraj Patil's avatar
Suraj Patil committed
461
462
463
464
465
466
467
468
469
470
471
472
473
474
        # 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()
475
476

        rho = 7.0  # 7.0 is the value used in the paper
477
        ramp = np.linspace(0, 1, num_inference_steps)
478
479
480
481
482
        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

483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
    # Copied from https://github.com/crowsonkb/k-diffusion/blob/686dbad0f39640ea25c8a8c6a6e56bb40eacefa2/k_diffusion/sampling.py#L26
    def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor:
        """Constructs an exponential noise schedule."""

        # Hack to make sure that other schedulers which copy this function don't break
        # TODO: Add this logic to the other schedulers
        if hasattr(self.config, "sigma_min"):
            sigma_min = self.config.sigma_min
        else:
            sigma_min = None

        if hasattr(self.config, "sigma_max"):
            sigma_max = self.config.sigma_max
        else:
            sigma_max = None

        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()

        sigmas = torch.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps).exp()
        return sigmas

505
506
507
    def index_for_timestep(self, timestep, schedule_timesteps=None):
        if schedule_timesteps is None:
            schedule_timesteps = self.timesteps
YiYi Xu's avatar
YiYi Xu committed
508

509
        indices = (schedule_timesteps == timestep).nonzero()
YiYi Xu's avatar
YiYi Xu committed
510
511
512
513
514

        # 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)
515
        pos = 1 if len(indices) > 1 else 0
YiYi Xu's avatar
YiYi Xu committed
516

517
518
519
520
521
522
523
524
525
        return indices[pos].item()

    def _init_step_index(self, timestep):
        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
YiYi Xu's avatar
YiYi Xu committed
526

hlky's avatar
hlky committed
527
528
    def step(
        self,
529
530
531
        model_output: torch.Tensor,
        timestep: Union[float, torch.Tensor],
        sample: torch.Tensor,
hlky's avatar
hlky committed
532
533
534
535
536
537
538
539
        s_churn: float = 0.0,
        s_tmin: float = 0.0,
        s_tmax: float = float("inf"),
        s_noise: float = 1.0,
        generator: Optional[torch.Generator] = None,
        return_dict: bool = True,
    ) -> Union[EulerDiscreteSchedulerOutput, Tuple]:
        """
540
        Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
hlky's avatar
hlky committed
541
542
543
        process from the learned model outputs (most often the predicted noise).

        Args:
544
            model_output (`torch.Tensor`):
545
546
547
                The direct output from learned diffusion model.
            timestep (`float`):
                The current discrete timestep in the diffusion chain.
548
            sample (`torch.Tensor`):
549
550
551
552
553
554
555
556
557
558
559
                A current instance of a sample created by the diffusion process.
            s_churn (`float`):
            s_tmin  (`float`):
            s_tmax  (`float`):
            s_noise (`float`, defaults to 1.0):
                Scaling factor for noise added to the sample.
            generator (`torch.Generator`, *optional*):
                A random number generator.
            return_dict (`bool`):
                Whether or not to return a [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or
                tuple.
hlky's avatar
hlky committed
560
561

        Returns:
562
563
564
            [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or `tuple`:
                If return_dict is `True`, [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] is
                returned, otherwise a tuple is returned where the first element is the sample tensor.
hlky's avatar
hlky committed
565
566
        """

567
        if isinstance(timestep, (int, torch.IntTensor, torch.LongTensor)):
hlky's avatar
hlky committed
568
            raise ValueError(
Patrick von Platen's avatar
Patrick von Platen committed
569
570
571
572
573
                (
                    "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to"
                    " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass"
                    " one of the `scheduler.timesteps` as a timestep."
                ),
hlky's avatar
hlky committed
574
575
576
            )

        if not self.is_scale_input_called:
577
            logger.warning(
hlky's avatar
hlky committed
578
579
580
581
                "The `scale_model_input` function should be called before `step` to ensure correct denoising. "
                "See `StableDiffusionPipeline` for a usage example."
            )

YiYi Xu's avatar
YiYi Xu committed
582
583
        if self.step_index is None:
            self._init_step_index(timestep)
hlky's avatar
hlky committed
584

585
586
587
        # Upcast to avoid precision issues when computing prev_sample
        sample = sample.to(torch.float32)

YiYi Xu's avatar
YiYi Xu committed
588
        sigma = self.sigmas[self.step_index]
hlky's avatar
hlky committed
589
590
591

        gamma = min(s_churn / (len(self.sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigma <= s_tmax else 0.0

592
593
594
        noise = randn_tensor(
            model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator
        )
595

hlky's avatar
hlky committed
596
597
598
599
600
601
602
        eps = noise * s_noise
        sigma_hat = sigma * (gamma + 1)

        if gamma > 0:
            sample = sample + eps * (sigma_hat**2 - sigma**2) ** 0.5

        # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
603
604
605
        # NOTE: "original_sample" should not be an expected prediction_type but is left in for
        # backwards compatibility
        if self.config.prediction_type == "original_sample" or self.config.prediction_type == "sample":
606
607
            pred_original_sample = model_output
        elif self.config.prediction_type == "epsilon":
Suraj Patil's avatar
Suraj Patil committed
608
            pred_original_sample = sample - sigma_hat * model_output
609
        elif self.config.prediction_type == "v_prediction":
Suraj Patil's avatar
Suraj Patil committed
610
            # denoised = model_output * c_out + input * c_skip
Suraj Patil's avatar
Suraj Patil committed
611
612
613
            pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1))
        else:
            raise ValueError(
614
                f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`"
Suraj Patil's avatar
Suraj Patil committed
615
            )
hlky's avatar
hlky committed
616
617
618
619

        # 2. Convert to an ODE derivative
        derivative = (sample - pred_original_sample) / sigma_hat

YiYi Xu's avatar
YiYi Xu committed
620
        dt = self.sigmas[self.step_index + 1] - sigma_hat
hlky's avatar
hlky committed
621
622
623

        prev_sample = sample + derivative * dt

624
625
626
        # Cast sample back to model compatible dtype
        prev_sample = prev_sample.to(model_output.dtype)

YiYi Xu's avatar
YiYi Xu committed
627
628
629
        # upon completion increase step index by one
        self._step_index += 1

hlky's avatar
hlky committed
630
631
632
633
634
635
636
        if not return_dict:
            return (prev_sample,)

        return EulerDiscreteSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample)

    def add_noise(
        self,
637
638
639
640
        original_samples: torch.Tensor,
        noise: torch.Tensor,
        timesteps: torch.Tensor,
    ) -> torch.Tensor:
hlky's avatar
hlky committed
641
        # Make sure sigmas and timesteps have the same device and dtype as original_samples
642
        sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
hlky's avatar
hlky committed
643
644
        if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
            # mps does not support float64
645
            schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
hlky's avatar
hlky committed
646
647
            timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
        else:
648
            schedule_timesteps = self.timesteps.to(original_samples.device)
hlky's avatar
hlky committed
649
650
            timesteps = timesteps.to(original_samples.device)

651
652
653
        # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index
        if self.begin_index is None:
            step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
654
655
656
        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]
657
        else:
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
658
            # add noise is called before first denoising step to create initial latent(img2img)
659
            step_indices = [self.begin_index] * timesteps.shape[0]
hlky's avatar
hlky committed
660

661
        sigma = sigmas[step_indices].flatten()
hlky's avatar
hlky committed
662
663
664
665
666
667
        while len(sigma.shape) < len(original_samples.shape):
            sigma = sigma.unsqueeze(-1)

        noisy_samples = original_samples + noise * sigma
        return noisy_samples

668
    def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.Tensor) -> torch.Tensor:
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
        if (
            isinstance(timesteps, int)
            or isinstance(timesteps, torch.IntTensor)
            or isinstance(timesteps, torch.LongTensor)
        ):
            raise ValueError(
                (
                    "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to"
                    " `EulerDiscreteScheduler.get_velocity()` is not supported. Make sure to pass"
                    " one of the `scheduler.timesteps` as a timestep."
                ),
            )

        if sample.device.type == "mps" and torch.is_floating_point(timesteps):
            # mps does not support float64
            schedule_timesteps = self.timesteps.to(sample.device, dtype=torch.float32)
            timesteps = timesteps.to(sample.device, dtype=torch.float32)
        else:
            schedule_timesteps = self.timesteps.to(sample.device)
            timesteps = timesteps.to(sample.device)

        step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
        alphas_cumprod = self.alphas_cumprod.to(sample)
        sqrt_alpha_prod = alphas_cumprod[step_indices] ** 0.5
        sqrt_alpha_prod = sqrt_alpha_prod.flatten()
        while len(sqrt_alpha_prod.shape) < len(sample.shape):
            sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)

        sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[step_indices]) ** 0.5
        sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
        while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape):
            sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)

        velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
        return velocity

hlky's avatar
hlky committed
705
706
    def __len__(self):
        return self.config.num_train_timesteps