-
Notifications
You must be signed in to change notification settings - Fork 4.9k
/
scheduling_dpmsolver_singlestep.py
630 lines (551 loc) 路 29.3 KB
/
scheduling_dpmsolver_singlestep.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
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
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
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
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
# Copyright 2023 TSAIL Team and The HuggingFace Team. All rights reserved.
#
# 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
from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999):
"""
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.
Returns:
betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
"""
def alpha_bar(time_step):
return math.cos((time_step + 0.008) / 1.008 * math.pi / 2) ** 2
betas = []
for i in range(num_diffusion_timesteps):
t1 = i / num_diffusion_timesteps
t2 = (i + 1) / num_diffusion_timesteps
betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
return torch.tensor(betas, dtype=torch.float32)
class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
"""
DPM-Solver (and the improved version DPM-Solver++) is a fast dedicated high-order solver for diffusion ODEs with
the convergence order guarantee. Empirically, sampling by DPM-Solver with only 20 steps can generate high-quality
samples, and it can generate quite good samples even in only 10 steps.
For more details, see the original paper: https://arxiv.org/abs/2206.00927 and https://arxiv.org/abs/2211.01095
Currently, we support the singlestep DPM-Solver for both noise prediction models and data prediction models. We
recommend to use `solver_order=2` for guided sampling, and `solver_order=3` for unconditional sampling.
We also support the "dynamic thresholding" method in Imagen (https://arxiv.org/abs/2205.11487). For pixel-space
diffusion models, you can set both `algorithm_type="dpmsolver++"` and `thresholding=True` to use the dynamic
thresholding. Note that the thresholding method is unsuitable for latent-space diffusion models (such as
stable-diffusion).
[`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__`
function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`.
[`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
[`~SchedulerMixin.from_pretrained`] functions.
Args:
num_train_timesteps (`int`): number of diffusion steps used to train the model.
beta_start (`float`): the starting `beta` value of inference.
beta_end (`float`): the final `beta` value.
beta_schedule (`str`):
the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
`linear`, `scaled_linear`, or `squaredcos_cap_v2`.
trained_betas (`np.ndarray`, optional):
option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc.
solver_order (`int`, default `2`):
the order of DPM-Solver; can be `1` or `2` or `3`. We recommend to use `solver_order=2` for guided
sampling, and `solver_order=3` for unconditional sampling.
prediction_type (`str`, default `epsilon`):
indicates whether the model predicts the noise (epsilon), or the data / `x0`. One of `epsilon`, `sample`,
or `v-prediction`.
thresholding (`bool`, default `False`):
whether to use the "dynamic thresholding" method (introduced by Imagen, https://arxiv.org/abs/2205.11487).
For pixel-space diffusion models, you can set both `algorithm_type=dpmsolver++` and `thresholding=True` to
use the dynamic thresholding. Note that the thresholding method is unsuitable for latent-space diffusion
models (such as stable-diffusion).
dynamic_thresholding_ratio (`float`, default `0.995`):
the ratio for the dynamic thresholding method. Default is `0.995`, the same as Imagen
(https://arxiv.org/abs/2205.11487).
sample_max_value (`float`, default `1.0`):
the threshold value for dynamic thresholding. Valid only when `thresholding=True` and
`algorithm_type="dpmsolver++`.
algorithm_type (`str`, default `dpmsolver++`):
the algorithm type for the solver. Either `dpmsolver` or `dpmsolver++`. The `dpmsolver` type implements the
algorithms in https://arxiv.org/abs/2206.00927, and the `dpmsolver++` type implements the algorithms in
https://arxiv.org/abs/2211.01095. We recommend to use `dpmsolver++` with `solver_order=2` for guided
sampling (e.g. stable-diffusion).
solver_type (`str`, default `midpoint`):
the solver type for the second-order solver. Either `midpoint` or `heun`. The solver type slightly affects
the sample quality, especially for small number of steps. We empirically find that `midpoint` solvers are
slightly better, so we recommend to use the `midpoint` type.
lower_order_final (`bool`, default `True`):
whether to use lower-order solvers in the final steps. For singlestep schedulers, we recommend to enable
this to use up all the function evaluations.
"""
_compatibles = [e.name for e in KarrasDiffusionSchedulers]
order = 1
@register_to_config
def __init__(
self,
num_train_timesteps: int = 1000,
beta_start: float = 0.0001,
beta_end: float = 0.02,
beta_schedule: str = "linear",
trained_betas: Optional[np.ndarray] = None,
solver_order: int = 2,
prediction_type: str = "epsilon",
thresholding: bool = False,
dynamic_thresholding_ratio: float = 0.995,
sample_max_value: float = 1.0,
algorithm_type: str = "dpmsolver++",
solver_type: str = "midpoint",
lower_order_final: bool = True,
):
if trained_betas is not None:
self.betas = torch.tensor(trained_betas, dtype=torch.float32)
elif beta_schedule == "linear":
self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
elif beta_schedule == "scaled_linear":
# this schedule is very specific to the latent diffusion model.
self.betas = (
torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
)
elif beta_schedule == "squaredcos_cap_v2":
# Glide cosine schedule
self.betas = betas_for_alpha_bar(num_train_timesteps)
else:
raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
self.alphas = 1.0 - self.betas
self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
# Currently we only support VP-type noise schedule
self.alpha_t = torch.sqrt(self.alphas_cumprod)
self.sigma_t = torch.sqrt(1 - self.alphas_cumprod)
self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t)
# standard deviation of the initial noise distribution
self.init_noise_sigma = 1.0
# settings for DPM-Solver
if algorithm_type not in ["dpmsolver", "dpmsolver++"]:
if algorithm_type == "deis":
self.register_to_config(algorithm_type="dpmsolver++")
else:
raise NotImplementedError(f"{algorithm_type} does is not implemented for {self.__class__}")
if solver_type not in ["midpoint", "heun"]:
if solver_type in ["logrho", "bh1", "bh2"]:
self.register_to_config(solver_type="midpoint")
else:
raise NotImplementedError(f"{solver_type} does is not implemented for {self.__class__}")
# setable values
self.num_inference_steps = None
timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy()
self.timesteps = torch.from_numpy(timesteps)
self.model_outputs = [None] * solver_order
self.sample = None
self.order_list = self.get_order_list(num_train_timesteps)
def get_order_list(self, num_inference_steps: int) -> List[int]:
"""
Computes the solver order at each time step.
Args:
num_inference_steps (`int`):
the number of diffusion steps used when generating samples with a pre-trained model.
"""
steps = num_inference_steps
order = self.config.solver_order
if self.config.lower_order_final:
if order == 3:
if steps % 3 == 0:
orders = [1, 2, 3] * (steps // 3 - 1) + [1, 2] + [1]
elif steps % 3 == 1:
orders = [1, 2, 3] * (steps // 3) + [1]
else:
orders = [1, 2, 3] * (steps // 3) + [1, 2]
elif order == 2:
if steps % 2 == 0:
orders = [1, 2] * (steps // 2)
else:
orders = [1, 2] * (steps // 2) + [1]
elif order == 1:
orders = [1] * steps
else:
if order == 3:
orders = [1, 2, 3] * (steps // 3)
elif order == 2:
orders = [1, 2] * (steps // 2)
elif order == 1:
orders = [1] * steps
return orders
def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
"""
Sets the timesteps used for the diffusion chain. Supporting function to be run before inference.
Args:
num_inference_steps (`int`):
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.
"""
self.num_inference_steps = num_inference_steps
timesteps = (
np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps + 1)
.round()[::-1][:-1]
.copy()
.astype(np.int64)
)
self.timesteps = torch.from_numpy(timesteps).to(device)
self.model_outputs = [None] * self.config.solver_order
self.sample = None
self.orders = self.get_order_list(num_inference_steps)
# Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
"""
"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
batch_size, channels, height, width = sample.shape
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
sample = sample.reshape(batch_size, channels * height * width)
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"
sample = sample.reshape(batch_size, channels, height, width)
sample = sample.to(dtype)
return sample
def convert_model_output(
self, model_output: torch.FloatTensor, timestep: int, sample: torch.FloatTensor
) -> torch.FloatTensor:
"""
Convert the model output to the corresponding type that the algorithm (DPM-Solver / DPM-Solver++) 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. So we need to first convert the model output to the
corresponding type to match the algorithm.
Note that the algorithm type and the model type is decoupled. That is to say, we can use either DPM-Solver or
DPM-Solver++ for both noise prediction model and data prediction model.
Args:
model_output (`torch.FloatTensor`): direct output from learned diffusion model.
timestep (`int`): current discrete timestep in the diffusion chain.
sample (`torch.FloatTensor`):
current instance of sample being created by diffusion process.
Returns:
`torch.FloatTensor`: the converted model output.
"""
# DPM-Solver++ needs to solve an integral of the data prediction model.
if self.config.algorithm_type == "dpmsolver++":
if self.config.prediction_type == "epsilon":
alpha_t, sigma_t = self.alpha_t[timestep], self.sigma_t[timestep]
x0_pred = (sample - sigma_t * model_output) / alpha_t
elif self.config.prediction_type == "sample":
x0_pred = model_output
elif self.config.prediction_type == "v_prediction":
alpha_t, sigma_t = self.alpha_t[timestep], self.sigma_t[timestep]
x0_pred = alpha_t * sample - sigma_t * model_output
else:
raise ValueError(
f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
" `v_prediction` for the DPMSolverSinglestepScheduler."
)
if self.config.thresholding:
x0_pred = self._threshold_sample(x0_pred)
return x0_pred
# DPM-Solver needs to solve an integral of the noise prediction model.
elif self.config.algorithm_type == "dpmsolver":
if self.config.prediction_type == "epsilon":
return model_output
elif self.config.prediction_type == "sample":
alpha_t, sigma_t = self.alpha_t[timestep], self.sigma_t[timestep]
epsilon = (sample - alpha_t * model_output) / sigma_t
return epsilon
elif self.config.prediction_type == "v_prediction":
alpha_t, sigma_t = self.alpha_t[timestep], self.sigma_t[timestep]
epsilon = alpha_t * model_output + sigma_t * sample
return epsilon
else:
raise ValueError(
f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
" `v_prediction` for the DPMSolverSinglestepScheduler."
)
def dpm_solver_first_order_update(
self,
model_output: torch.FloatTensor,
timestep: int,
prev_timestep: int,
sample: torch.FloatTensor,
) -> torch.FloatTensor:
"""
One step for the first-order DPM-Solver (equivalent to DDIM).
See https://arxiv.org/abs/2206.00927 for the detailed derivation.
Args:
model_output (`torch.FloatTensor`): direct output from learned diffusion model.
timestep (`int`): current discrete timestep in the diffusion chain.
prev_timestep (`int`): previous discrete timestep in the diffusion chain.
sample (`torch.FloatTensor`):
current instance of sample being created by diffusion process.
Returns:
`torch.FloatTensor`: the sample tensor at the previous timestep.
"""
lambda_t, lambda_s = self.lambda_t[prev_timestep], self.lambda_t[timestep]
alpha_t, alpha_s = self.alpha_t[prev_timestep], self.alpha_t[timestep]
sigma_t, sigma_s = self.sigma_t[prev_timestep], self.sigma_t[timestep]
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
return x_t
def singlestep_dpm_solver_second_order_update(
self,
model_output_list: List[torch.FloatTensor],
timestep_list: List[int],
prev_timestep: int,
sample: torch.FloatTensor,
) -> torch.FloatTensor:
"""
One step for the second-order singlestep DPM-Solver.
It computes the solution at time `prev_timestep` from the time `timestep_list[-2]`.
Args:
model_output_list (`List[torch.FloatTensor]`):
direct outputs from learned diffusion model at current and latter timesteps.
timestep (`int`): current and latter discrete timestep in the diffusion chain.
prev_timestep (`int`): previous discrete timestep in the diffusion chain.
sample (`torch.FloatTensor`):
current instance of sample being created by diffusion process.
Returns:
`torch.FloatTensor`: the sample tensor at the previous timestep.
"""
t, s0, s1 = prev_timestep, timestep_list[-1], timestep_list[-2]
m0, m1 = model_output_list[-1], model_output_list[-2]
lambda_t, lambda_s0, lambda_s1 = self.lambda_t[t], self.lambda_t[s0], self.lambda_t[s1]
alpha_t, alpha_s1 = self.alpha_t[t], self.alpha_t[s1]
sigma_t, sigma_s1 = self.sigma_t[t], self.sigma_t[s1]
h, h_0 = lambda_t - lambda_s1, lambda_s0 - lambda_s1
r0 = h_0 / h
D0, D1 = m1, (1.0 / r0) * (m0 - m1)
if self.config.algorithm_type == "dpmsolver++":
# See https://arxiv.org/abs/2211.01095 for detailed derivations
if self.config.solver_type == "midpoint":
x_t = (
(sigma_t / sigma_s1) * sample
- (alpha_t * (torch.exp(-h) - 1.0)) * D0
- 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1
)
elif self.config.solver_type == "heun":
x_t = (
(sigma_t / sigma_s1) * sample
- (alpha_t * (torch.exp(-h) - 1.0)) * D0
+ (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
)
elif self.config.algorithm_type == "dpmsolver":
# See https://arxiv.org/abs/2206.00927 for detailed derivations
if self.config.solver_type == "midpoint":
x_t = (
(alpha_t / alpha_s1) * sample
- (sigma_t * (torch.exp(h) - 1.0)) * D0
- 0.5 * (sigma_t * (torch.exp(h) - 1.0)) * D1
)
elif self.config.solver_type == "heun":
x_t = (
(alpha_t / alpha_s1) * sample
- (sigma_t * (torch.exp(h) - 1.0)) * D0
- (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
)
return x_t
def singlestep_dpm_solver_third_order_update(
self,
model_output_list: List[torch.FloatTensor],
timestep_list: List[int],
prev_timestep: int,
sample: torch.FloatTensor,
) -> torch.FloatTensor:
"""
One step for the third-order singlestep DPM-Solver.
It computes the solution at time `prev_timestep` from the time `timestep_list[-3]`.
Args:
model_output_list (`List[torch.FloatTensor]`):
direct outputs from learned diffusion model at current and latter timesteps.
timestep (`int`): current and latter discrete timestep in the diffusion chain.
prev_timestep (`int`): previous discrete timestep in the diffusion chain.
sample (`torch.FloatTensor`):
current instance of sample being created by diffusion process.
Returns:
`torch.FloatTensor`: the sample tensor at the previous timestep.
"""
t, s0, s1, s2 = prev_timestep, timestep_list[-1], timestep_list[-2], timestep_list[-3]
m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3]
lambda_t, lambda_s0, lambda_s1, lambda_s2 = (
self.lambda_t[t],
self.lambda_t[s0],
self.lambda_t[s1],
self.lambda_t[s2],
)
alpha_t, alpha_s2 = self.alpha_t[t], self.alpha_t[s2]
sigma_t, sigma_s2 = self.sigma_t[t], self.sigma_t[s2]
h, h_0, h_1 = lambda_t - lambda_s2, lambda_s0 - lambda_s2, lambda_s1 - lambda_s2
r0, r1 = h_0 / h, h_1 / h
D0 = m2
D1_0, D1_1 = (1.0 / r1) * (m1 - m2), (1.0 / r0) * (m0 - m2)
D1 = (r0 * D1_0 - r1 * D1_1) / (r0 - r1)
D2 = 2.0 * (D1_1 - D1_0) / (r0 - r1)
if self.config.algorithm_type == "dpmsolver++":
# See https://arxiv.org/abs/2206.00927 for detailed derivations
if self.config.solver_type == "midpoint":
x_t = (
(sigma_t / sigma_s2) * sample
- (alpha_t * (torch.exp(-h) - 1.0)) * D0
+ (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1_1
)
elif self.config.solver_type == "heun":
x_t = (
(sigma_t / sigma_s2) * sample
- (alpha_t * (torch.exp(-h) - 1.0)) * D0
+ (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
- (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2
)
elif self.config.algorithm_type == "dpmsolver":
# See https://arxiv.org/abs/2206.00927 for detailed derivations
if self.config.solver_type == "midpoint":
x_t = (
(alpha_t / alpha_s2) * sample
- (sigma_t * (torch.exp(h) - 1.0)) * D0
- (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1_1
)
elif self.config.solver_type == "heun":
x_t = (
(alpha_t / alpha_s2) * sample
- (sigma_t * (torch.exp(h) - 1.0)) * D0
- (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
- (sigma_t * ((torch.exp(h) - 1.0 - h) / h**2 - 0.5)) * D2
)
return x_t
def singlestep_dpm_solver_update(
self,
model_output_list: List[torch.FloatTensor],
timestep_list: List[int],
prev_timestep: int,
sample: torch.FloatTensor,
order: int,
) -> torch.FloatTensor:
"""
One step for the singlestep DPM-Solver.
Args:
model_output_list (`List[torch.FloatTensor]`):
direct outputs from learned diffusion model at current and latter timesteps.
timestep (`int`): current and latter discrete timestep in the diffusion chain.
prev_timestep (`int`): previous discrete timestep in the diffusion chain.
sample (`torch.FloatTensor`):
current instance of sample being created by diffusion process.
order (`int`):
the solver order at this step.
Returns:
`torch.FloatTensor`: the sample tensor at the previous timestep.
"""
if order == 1:
return self.dpm_solver_first_order_update(model_output_list[-1], timestep_list[-1], prev_timestep, sample)
elif order == 2:
return self.singlestep_dpm_solver_second_order_update(
model_output_list, timestep_list, prev_timestep, sample
)
elif order == 3:
return self.singlestep_dpm_solver_third_order_update(
model_output_list, timestep_list, prev_timestep, sample
)
else:
raise ValueError(f"Order must be 1, 2, 3, got {order}")
def step(
self,
model_output: torch.FloatTensor,
timestep: int,
sample: torch.FloatTensor,
return_dict: bool = True,
) -> Union[SchedulerOutput, Tuple]:
"""
Step function propagating the sample with the singlestep DPM-Solver.
Args:
model_output (`torch.FloatTensor`): direct output from learned diffusion model.
timestep (`int`): current discrete timestep in the diffusion chain.
sample (`torch.FloatTensor`):
current instance of sample being created by diffusion process.
return_dict (`bool`): option for returning tuple rather than SchedulerOutput class
Returns:
[`~scheduling_utils.SchedulerOutput`] or `tuple`: [`~scheduling_utils.SchedulerOutput`] if `return_dict` is
True, otherwise a `tuple`. When returning a tuple, the first element is the sample tensor.
"""
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"
)
if isinstance(timestep, torch.Tensor):
timestep = timestep.to(self.timesteps.device)
step_index = (self.timesteps == timestep).nonzero()
if len(step_index) == 0:
step_index = len(self.timesteps) - 1
else:
step_index = step_index.item()
prev_timestep = 0 if step_index == len(self.timesteps) - 1 else self.timesteps[step_index + 1]
model_output = self.convert_model_output(model_output, timestep, sample)
for i in range(self.config.solver_order - 1):
self.model_outputs[i] = self.model_outputs[i + 1]
self.model_outputs[-1] = model_output
order = self.order_list[step_index]
# For single-step solvers, we use the initial value at each time with order = 1.
if order == 1:
self.sample = sample
timestep_list = [self.timesteps[step_index - i] for i in range(order - 1, 0, -1)] + [timestep]
prev_sample = self.singlestep_dpm_solver_update(
self.model_outputs, timestep_list, prev_timestep, self.sample, order
)
if not return_dict:
return (prev_sample,)
return SchedulerOutput(prev_sample=prev_sample)
def scale_model_input(self, sample: torch.FloatTensor, *args, **kwargs) -> torch.FloatTensor:
"""
Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
current timestep.
Args:
sample (`torch.FloatTensor`): input sample
Returns:
`torch.FloatTensor`: scaled input sample
"""
return sample
# Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise
def add_noise(
self,
original_samples: torch.FloatTensor,
noise: torch.FloatTensor,
timesteps: torch.IntTensor,
) -> torch.FloatTensor:
# Make sure alphas_cumprod and timestep have same device and dtype as original_samples
alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device, dtype=original_samples.dtype)
timesteps = timesteps.to(original_samples.device)
sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
sqrt_alpha_prod = sqrt_alpha_prod.flatten()
while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
return noisy_samples
def __len__(self):
return self.config.num_train_timesteps