kd5678 commited on
Commit
e4c5319
1 Parent(s): 34dc268

Create ddim_with_prob.py

Browse files
Files changed (1) hide show
  1. ddim_with_prob.py +397 -0
ddim_with_prob.py ADDED
@@ -0,0 +1,397 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2022 Stanford University Team and The HuggingFace Team. All rights reserved.
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ # DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion
16
+ # and https://github.com/hojonathanho/diffusion
17
+
18
+ import math
19
+ from dataclasses import dataclass
20
+ from typing import List, Optional, Tuple, Union
21
+ import numpy as np
22
+ import torch
23
+ from diffusers.configuration_utils import ConfigMixin, register_to_config
24
+ from diffusers.utils import BaseOutput
25
+ from diffusers.utils.torch_utils import randn_tensor
26
+ from diffusers.schedulers.scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
27
+
28
+
29
+ @dataclass
30
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DDIM
31
+ class DDIMSchedulerOutput(BaseOutput):
32
+ """
33
+ Output class for the scheduler's step function output.
34
+
35
+ Args:
36
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
37
+ Computed sample (x_{t-1}) of previous timestep. `prev_sample` should be used as next model input in the
38
+ denoising loop.
39
+ pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
40
+ The predicted denoised sample (x_{0}) based on the model output from the current timestep.
41
+ `pred_original_sample` can be used to preview progress or for guidance.
42
+ """
43
+
44
+ prev_sample: torch.FloatTensor
45
+ pred_original_sample: Optional[torch.FloatTensor] = None
46
+ log_prob: Optional[torch.FloatTensor] = None
47
+
48
+
49
+
50
+ def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999) -> torch.Tensor:
51
+ """
52
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
53
+ (1-beta) over time from t = [0,1].
54
+
55
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
56
+ to that part of the diffusion process.
57
+
58
+
59
+ Args:
60
+ num_diffusion_timesteps (`int`): the number of betas to produce.
61
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
62
+ prevent singularities.
63
+
64
+ Returns:
65
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
66
+ """
67
+
68
+ def alpha_bar(time_step):
69
+ return math.cos((time_step + 0.008) / 1.008 * math.pi / 2) ** 2
70
+
71
+ betas = []
72
+ for i in range(num_diffusion_timesteps):
73
+ t1 = i / num_diffusion_timesteps
74
+ t2 = (i + 1) / num_diffusion_timesteps
75
+ betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
76
+ return torch.tensor(betas)
77
+
78
+
79
+ class DDIMSchedulerCustom(SchedulerMixin, ConfigMixin):
80
+ """
81
+ Denoising diffusion implicit models is a scheduler that extends the denoising procedure introduced in denoising
82
+ diffusion probabilistic models (DDPMs) with non-Markovian guidance.
83
+
84
+ [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__`
85
+ function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`.
86
+ [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
87
+ [`~SchedulerMixin.from_pretrained`] functions.
88
+
89
+ For more details, see the original paper: https://arxiv.org/abs/2010.02502
90
+
91
+ Args:
92
+ num_train_timesteps (`int`): number of diffusion steps used to train the model.
93
+ beta_start (`float`): the starting `beta` value of inference.
94
+ beta_end (`float`): the final `beta` value.
95
+ beta_schedule (`str`):
96
+ the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
97
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
98
+ trained_betas (`np.ndarray`, optional):
99
+ option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc.
100
+ clip_sample (`bool`, default `True`):
101
+ option to clip predicted sample between -1 and 1 for numerical stability.
102
+ set_alpha_to_one (`bool`, default `True`):
103
+ each diffusion step uses the value of alphas product at that step and at the previous one. For the final
104
+ step there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`,
105
+ otherwise it uses the value of alpha at step 0.
106
+ steps_offset (`int`, default `0`):
107
+ an offset added to the inference steps. You can use a combination of `offset=1` and
108
+ `set_alpha_to_one=False`, to make the last step use step 0 for the previous alpha product, as done in
109
+ stable diffusion.
110
+ prediction_type (`str`, default `epsilon`, optional):
111
+ prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion
112
+ process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4
113
+ https://imagen.research.google/video/paper.pdf)
114
+ """
115
+
116
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
117
+ order = 1
118
+
119
+ @register_to_config
120
+ def __init__(
121
+ self,
122
+ num_train_timesteps: int = 1000,
123
+ beta_start: float = 0.0001,
124
+ beta_end: float = 0.02,
125
+ beta_schedule: str = "linear",
126
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
127
+ clip_sample: bool = True,
128
+ set_alpha_to_one: bool = True,
129
+ steps_offset: int = 0,
130
+ prediction_type: str = "epsilon",
131
+ ):
132
+ if trained_betas is not None:
133
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
134
+ elif beta_schedule == "linear":
135
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
136
+ elif beta_schedule == "scaled_linear":
137
+ # this schedule is very specific to the latent diffusion model.
138
+ self.betas = (
139
+ torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
140
+ )
141
+ elif beta_schedule == "squaredcos_cap_v2":
142
+ # Glide cosine schedule
143
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
144
+ else:
145
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
146
+
147
+ self.alphas = 1.0 - self.betas
148
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
149
+
150
+ # At every step in ddim, we are looking into the previous alphas_cumprod
151
+ # For the final step, there is no previous alphas_cumprod because we are already at 0
152
+ # `set_alpha_to_one` decides whether we set this parameter simply to one or
153
+ # whether we use the final alpha of the "non-previous" one.
154
+ self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0]
155
+
156
+ # standard deviation of the initial noise distribution
157
+ self.init_noise_sigma = 1.0
158
+
159
+ # setable values
160
+ self.num_inference_steps = None
161
+ self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64))
162
+
163
+ def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
164
+ """
165
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
166
+ current timestep.
167
+
168
+ Args:
169
+ sample (`torch.FloatTensor`): input sample
170
+ timestep (`int`, optional): current timestep
171
+
172
+ Returns:
173
+ `torch.FloatTensor`: scaled input sample
174
+ """
175
+ return sample
176
+
177
+ def _get_variance(self, timestep, prev_timestep):
178
+ alpha_prod_t = self.alphas_cumprod[timestep]
179
+ alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod
180
+ beta_prod_t = 1 - alpha_prod_t
181
+ beta_prod_t_prev = 1 - alpha_prod_t_prev
182
+
183
+ variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev)
184
+
185
+ return variance
186
+
187
+ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
188
+ """
189
+ Sets the discrete timesteps used for the diffusion chain. Supporting function to be run before inference.
190
+
191
+ Args:
192
+ num_inference_steps (`int`):
193
+ the number of diffusion steps used when generating samples with a pre-trained model.
194
+ """
195
+
196
+ if num_inference_steps > self.config.num_train_timesteps:
197
+ raise ValueError(
198
+ f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:"
199
+ f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
200
+ f" maximal {self.config.num_train_timesteps} timesteps."
201
+ )
202
+
203
+ self.num_inference_steps = num_inference_steps
204
+ step_ratio = self.config.num_train_timesteps // self.num_inference_steps
205
+ # creates integer timesteps by multiplying by ratio
206
+ # casting to int to avoid issues when num_inference_step is power of 3
207
+ timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64)
208
+ self.timesteps = torch.from_numpy(timesteps).to(device)
209
+ self.timesteps += self.config.steps_offset
210
+
211
+ def step(
212
+ self,
213
+ model_output: torch.FloatTensor,
214
+ timestep: int,
215
+ sample: torch.FloatTensor,
216
+ eta: float = 0.0,
217
+ use_clipped_model_output: bool = False,
218
+ generator=None,
219
+ variance_noise: Optional[torch.FloatTensor] = None,
220
+ return_dict: bool = True,
221
+ prev_sample: Optional[torch.FloatTensor] = None,
222
+ ) -> Union[DDIMSchedulerOutput, Tuple]:
223
+ """
224
+
225
+ Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
226
+ process from the learned model outputs (most often the predicted noise).
227
+
228
+ First, the model_output is used to calculate the prev_sample_mean. If
229
+ key is not None, some noise is added to produce prev_sample (with
230
+ variance depending on eta). If prev_sample is not None, this function
231
+ essentially just calculates the log_prob of prev_sample given
232
+ prev_sample_mean, and prev_sample is returned unmodified.
233
+
234
+
235
+ Args:
236
+ model_output (`torch.FloatTensor`): direct output from learned diffusion model.
237
+ timestep (`int`): current discrete timestep in the diffusion chain.
238
+ sample (`torch.FloatTensor`):
239
+ current instance of sample being created by diffusion process.
240
+ eta (`float`): weight of noise for added noise in diffusion step.
241
+ use_clipped_model_output (`bool`): if `True`, compute "corrected" `model_output` from the clipped
242
+ predicted original sample. Necessary because predicted original sample is clipped to [-1, 1] when
243
+ `self.config.clip_sample` is `True`. If no clipping has happened, "corrected" `model_output` would
244
+ coincide with the one provided as input and `use_clipped_model_output` will have not effect.
245
+ generator: random number generator.
246
+ variance_noise (`torch.FloatTensor`): instead of generating noise for the variance using `generator`, we
247
+ can directly provide the noise for the variance itself. This is useful for methods such as
248
+ CycleDiffusion. (https://arxiv.org/abs/2210.05559)
249
+ return_dict (`bool`): option for returning tuple rather than DDIMSchedulerOutput class
250
+
251
+ Returns:
252
+ [`~schedulers.scheduling_utils.DDIMSchedulerOutput`] or `tuple`:
253
+ [`~schedulers.scheduling_utils.DDIMSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`. When
254
+ returning a tuple, the first element is the sample tensor.
255
+
256
+ """
257
+ # eta = 1.0
258
+ if self.num_inference_steps is None:
259
+ raise ValueError(
260
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
261
+ )
262
+
263
+ # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
264
+ # Ideally, read DDIM paper in-detail understanding
265
+
266
+ # Notation (<variable name> -> <name in paper>
267
+ # - pred_noise_t -> e_theta(x_t, t)
268
+ # - pred_original_sample -> f_theta(x_t, t) or x_0
269
+ # - std_dev_t -> sigma_t
270
+ # - eta -> η
271
+ # - pred_sample_direction -> "direction pointing to x_t"
272
+ # - pred_prev_sample -> "x_t-1"
273
+
274
+ # 1. get previous step value (=t-1)
275
+ prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps
276
+
277
+
278
+ # 2. compute alphas, betas
279
+ alpha_prod_t = self.alphas_cumprod[timestep]
280
+ alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod
281
+
282
+ beta_prod_t = 1 - alpha_prod_t
283
+
284
+ # 3. compute predicted original sample from predicted noise also called
285
+ # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
286
+ if self.config.prediction_type == "epsilon":
287
+ pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
288
+ elif self.config.prediction_type == "sample":
289
+ pred_original_sample = model_output
290
+ elif self.config.prediction_type == "v_prediction":
291
+ pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output
292
+ # predict V
293
+ model_output = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
294
+ else:
295
+ raise ValueError(
296
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
297
+ " `v_prediction`"
298
+ )
299
+
300
+ # 4. Clip "predicted x_0"
301
+ if self.config.clip_sample:
302
+ pred_original_sample = torch.clamp(pred_original_sample, -1, 1)
303
+
304
+
305
+ # 5. compute variance: "sigma_t(η)" -> see formula (16)
306
+ # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)
307
+ variance = self._get_variance(timestep, prev_timestep)
308
+ std_dev_t = eta * variance ** (0.5)
309
+
310
+
311
+ if use_clipped_model_output:
312
+ # the model_output is always re-derived from the clipped x_0 in Glide
313
+ model_output = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
314
+
315
+ # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
316
+ pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * model_output
317
+
318
+ # 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
319
+ prev_sample_mean = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
320
+
321
+
322
+ if prev_sample is None and eta > 0:
323
+ device = model_output.device
324
+ if variance_noise is not None and generator is not None:
325
+ raise ValueError(
326
+ "Cannot pass both generator and variance_noise. Please make sure that either `generator` or"
327
+ " `variance_noise` stays `None`."
328
+ )
329
+
330
+ if variance_noise is None:
331
+ variance_noise = randn_tensor(
332
+ model_output.shape, generator=generator, device=device, dtype=model_output.dtype
333
+ )
334
+
335
+ prev_sample = prev_sample_mean + std_dev_t * variance_noise
336
+
337
+ # std_dev_t = torch.clip(std_dev_t, min=1e-6)
338
+ log_prob = (
339
+ -((prev_sample - prev_sample_mean) ** 2) / (2 * (std_dev_t**2))
340
+ - math.log(std_dev_t)
341
+ - math.log(math.sqrt(2 * math.pi))
342
+ )
343
+
344
+ log_prob_mean = torch.mean(log_prob, axis=tuple(range(1, log_prob.ndim)))
345
+
346
+
347
+
348
+ if not return_dict:
349
+ return (prev_sample, pred_original_sample, log_prob, log_prob_mean)
350
+
351
+ return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample, log_prob=log_prob_mean)
352
+
353
+ def add_noise(
354
+ self,
355
+ original_samples: torch.FloatTensor,
356
+ noise: torch.FloatTensor,
357
+ timesteps: torch.IntTensor,
358
+ ) -> torch.FloatTensor:
359
+ # Make sure alphas_cumprod and timestep have same device and dtype as original_samples
360
+ self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device, dtype=original_samples.dtype)
361
+ timesteps = timesteps.to(original_samples.device)
362
+
363
+ sqrt_alpha_prod = self.alphas_cumprod[timesteps] ** 0.5
364
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
365
+ while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
366
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
367
+
368
+ sqrt_one_minus_alpha_prod = (1 - self.alphas_cumprod[timesteps]) ** 0.5
369
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
370
+ while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
371
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
372
+
373
+ noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
374
+ return noisy_samples
375
+
376
+ def get_velocity(
377
+ self, sample: torch.FloatTensor, noise: torch.FloatTensor, timesteps: torch.IntTensor
378
+ ) -> torch.FloatTensor:
379
+ # Make sure alphas_cumprod and timestep have same device and dtype as sample
380
+ self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device, dtype=sample.dtype)
381
+ timesteps = timesteps.to(sample.device)
382
+
383
+ sqrt_alpha_prod = self.alphas_cumprod[timesteps] ** 0.5
384
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
385
+ while len(sqrt_alpha_prod.shape) < len(sample.shape):
386
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
387
+
388
+ sqrt_one_minus_alpha_prod = (1 - self.alphas_cumprod[timesteps]) ** 0.5
389
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
390
+ while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape):
391
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
392
+
393
+ velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
394
+ return velocity
395
+
396
+ def __len__(self):
397
+ return self.config.num_train_timesteps