scheduling_ddpm.py 12.3 KB
Newer Older
Ryan Russell's avatar
Ryan Russell committed
1
# Copyright 2022 UC Berkeley Team and The HuggingFace Team. All rights reserved.
Patrick von Platen's avatar
improve  
Patrick von Platen committed
2
3
4
5
6
7
8
9
10
11
12
13
#
# 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.
14
15
16

# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim

anton-l's avatar
anton-l committed
17
import math
18
from dataclasses import dataclass
19
from typing import Optional, Tuple, Union
Patrick von Platen's avatar
Patrick von Platen committed
20

Patrick von Platen's avatar
Patrick von Platen committed
21
import numpy as np
22
import torch
Patrick von Platen's avatar
improve  
Patrick von Platen committed
23

24
from ..configuration_utils import ConfigMixin, register_to_config
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
from ..utils import BaseOutput
from .scheduling_utils import SchedulerMixin


@dataclass
class DDPMSchedulerOutput(BaseOutput):
    """
    Output class for the scheduler's step function output.

    Args:
        prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
            Computed sample (x_{t-1}) of previous timestep. `prev_sample` should be used as next model input in the
            denoising loop.
        pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
            The predicted denoised sample (x_{0}) based on the model output from the current timestep.
            `pred_original_sample` can be used to preview progress or for guidance.
    """

    prev_sample: torch.FloatTensor
    pred_original_sample: Optional[torch.FloatTensor] = None
45
46
47
48


def betas_for_alpha_bar(num_diffusion_timesteps, max_beta=0.999):
    """
Patrick von Platen's avatar
Patrick von Platen committed
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].
51

52
53
54
55
56
57
58
    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
59
                     prevent singularities.
60
61
62

    Returns:
        betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
63
    """
64

65
66
67
68
69
70
71
72
    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))
73
    return torch.tensor(betas, dtype=torch.float32)
Patrick von Platen's avatar
improve  
Patrick von Platen committed
74
75


Patrick von Platen's avatar
Patrick von Platen committed
76
class DDPMScheduler(SchedulerMixin, ConfigMixin):
77
78
79
80
    """
    Denoising diffusion probabilistic models (DDPMs) explores the connections between denoising score matching and
    Langevin dynamics sampling.

81
82
83
    [`~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`.
    [`~ConfigMixin`] also provides general loading and saving functionality via the [`~ConfigMixin.save_config`] and
Nathan Lambert's avatar
Nathan Lambert committed
84
    [`~ConfigMixin.from_config`] functions.
85

86
87
88
89
90
91
92
93
94
    For more details, see the original paper: https://arxiv.org/abs/2006.11239

    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`.
Nathan Lambert's avatar
Nathan Lambert committed
95
96
        trained_betas (`np.ndarray`, optional):
            option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc.
97
98
99
100
101
102
103
104
        variance_type (`str`):
            options to clip the variance used when adding noise to the denoised sample. Choose from `fixed_small`,
            `fixed_small_log`, `fixed_large`, `fixed_large_log`, `learned` or `learned_range`.
        clip_sample (`bool`, default `True`):
            option to clip predicted sample between -1 and 1 for numerical stability.

    """

105
    @register_to_config
Patrick von Platen's avatar
improve  
Patrick von Platen committed
106
107
    def __init__(
        self,
Partho's avatar
Partho committed
108
109
110
111
112
113
114
        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,
        variance_type: str = "fixed_small",
        clip_sample: bool = True,
Patrick von Platen's avatar
improve  
Patrick von Platen committed
115
    ):
116
        if trained_betas is not None:
117
            self.betas = torch.from_numpy(trained_betas)
118
        elif beta_schedule == "linear":
119
            self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
120
121
        elif beta_schedule == "scaled_linear":
            # this schedule is very specific to the latent diffusion model.
122
123
124
            self.betas = (
                torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
            )
anton-l's avatar
anton-l committed
125
        elif beta_schedule == "squaredcos_cap_v2":
Patrick von Platen's avatar
Patrick von Platen committed
126
            # Glide cosine schedule
Nathan Lambert's avatar
Nathan Lambert committed
127
            self.betas = betas_for_alpha_bar(num_train_timesteps)
Patrick von Platen's avatar
improve  
Patrick von Platen committed
128
129
130
        else:
            raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")

Patrick von Platen's avatar
Patrick von Platen committed
131
        self.alphas = 1.0 - self.betas
132
133
        self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
        self.one = torch.tensor(1.0)
Patrick von Platen's avatar
Patrick von Platen committed
134

135
136
        # setable values
        self.num_inference_steps = None
137
        self.timesteps = np.arange(0, num_train_timesteps)[::-1]
Patrick von Platen's avatar
Patrick von Platen committed
138

139
140
        self.variance_type = variance_type

Partho's avatar
Partho committed
141
    def set_timesteps(self, num_inference_steps: int):
142
143
144
145
146
147
148
        """
        Sets the discrete 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.
        """
Patrick von Platen's avatar
Patrick von Platen committed
149
        num_inference_steps = min(self.config.num_train_timesteps, num_inference_steps)
150
151
152
        self.num_inference_steps = num_inference_steps
        self.timesteps = np.arange(
            0, self.config.num_train_timesteps, self.config.num_train_timesteps // self.num_inference_steps
153
        )[::-1]
154

155
    def _get_variance(self, t, predicted_variance=None, variance_type=None):
156
157
        alpha_prod_t = self.alphas_cumprod[t]
        alpha_prod_t_prev = self.alphas_cumprod[t - 1] if t > 0 else self.one
Patrick von Platen's avatar
Patrick von Platen committed
158

Kashif Rasul's avatar
Kashif Rasul committed
159
        # For t > 0, compute predicted variance βt (see formula (6) and (7) from https://arxiv.org/pdf/2006.11239.pdf)
160
        # and sample from it to get previous sample
Kashif Rasul's avatar
Kashif Rasul committed
161
        # x_{t-1} ~ N(pred_prev_sample, variance) == add variance to pred_sample
162
        variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * self.betas[t]
Patrick von Platen's avatar
Patrick von Platen committed
163

164
165
166
        if variance_type is None:
            variance_type = self.config.variance_type

167
        # hacks - were probably added for training stability
168
        if variance_type == "fixed_small":
169
            variance = torch.clamp(variance, min=1e-20)
170
        # for rl-diffuser https://arxiv.org/abs/2205.09991
171
        elif variance_type == "fixed_small_log":
172
            variance = torch.log(torch.clamp(variance, min=1e-20))
173
        elif variance_type == "fixed_large":
174
            variance = self.betas[t]
175
        elif variance_type == "fixed_large_log":
Patrick von Platen's avatar
Patrick von Platen committed
176
            # Glide max_log
177
            variance = torch.log(self.betas[t])
178
179
180
181
182
183
184
        elif variance_type == "learned":
            return predicted_variance
        elif variance_type == "learned_range":
            min_log = variance
            max_log = self.betas[t]
            frac = (predicted_variance + 1) / 2
            variance = frac * max_log + (1 - frac) * min_log
Patrick von Platen's avatar
Patrick von Platen committed
185
186
187

        return variance

188
189
    def step(
        self,
190
        model_output: torch.FloatTensor,
191
        timestep: int,
192
        sample: torch.FloatTensor,
193
        predict_epsilon=True,
Patrick von Platen's avatar
Patrick von Platen committed
194
        generator=None,
195
        return_dict: bool = True,
196
    ) -> Union[DDPMSchedulerOutput, Tuple]:
197
198
199
200
201
        """
        Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
        process from the learned model outputs (most often the predicted noise).

        Args:
202
            model_output (`torch.FloatTensor`): direct output from learned diffusion model.
203
            timestep (`int`): current discrete timestep in the diffusion chain.
204
            sample (`torch.FloatTensor`):
205
206
207
208
                current instance of sample being created by diffusion process.
            predict_epsilon (`bool`):
                optional flag to use when model predicts the samples directly instead of the noise, epsilon.
            generator: random number generator.
209
            return_dict (`bool`): option for returning tuple rather than DDPMSchedulerOutput class
210
211

        Returns:
212
213
            [`~schedulers.scheduling_utils.DDPMSchedulerOutput`] or `tuple`:
            [`~schedulers.scheduling_utils.DDPMSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`. When
214
            returning a tuple, the first element is the sample tensor.
215
216

        """
217
        t = timestep
218

219
220
221
222
223
        if model_output.shape[1] == sample.shape[1] * 2 and self.variance_type in ["learned", "learned_range"]:
            model_output, predicted_variance = torch.split(model_output, sample.shape[1], dim=1)
        else:
            predicted_variance = None

Patrick von Platen's avatar
Patrick von Platen committed
224
        # 1. compute alphas, betas
225
226
        alpha_prod_t = self.alphas_cumprod[t]
        alpha_prod_t_prev = self.alphas_cumprod[t - 1] if t > 0 else self.one
Patrick von Platen's avatar
Patrick von Platen committed
227
228
229
        beta_prod_t = 1 - alpha_prod_t
        beta_prod_t_prev = 1 - alpha_prod_t_prev

230
        # 2. compute predicted original sample from predicted noise also called
Patrick von Platen's avatar
Patrick von Platen committed
231
        # "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
232
        if predict_epsilon:
Patrick von Platen's avatar
Patrick von Platen committed
233
            pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
234
        else:
Patrick von Platen's avatar
Patrick von Platen committed
235
            pred_original_sample = model_output
Patrick von Platen's avatar
Patrick von Platen committed
236
237

        # 3. Clip "predicted x_0"
238
        if self.config.clip_sample:
239
            pred_original_sample = torch.clamp(pred_original_sample, -1, 1)
Patrick von Platen's avatar
Patrick von Platen committed
240

241
        # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t
Patrick von Platen's avatar
Patrick von Platen committed
242
        # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
243
244
        pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * self.betas[t]) / beta_prod_t
        current_sample_coeff = self.alphas[t] ** (0.5) * beta_prod_t_prev / beta_prod_t
Patrick von Platen's avatar
Patrick von Platen committed
245

246
        # 5. Compute predicted previous sample µ_t
Patrick von Platen's avatar
Patrick von Platen committed
247
        # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
248
        pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample
Patrick von Platen's avatar
Patrick von Platen committed
249

Patrick von Platen's avatar
Patrick von Platen committed
250
251
252
        # 6. Add noise
        variance = 0
        if t > 0:
253
254
255
            noise = torch.randn(
                model_output.size(), dtype=model_output.dtype, layout=model_output.layout, generator=generator
            ).to(model_output.device)
256
            variance = (self._get_variance(t, predicted_variance=predicted_variance) ** 0.5) * noise
Patrick von Platen's avatar
Patrick von Platen committed
257
258
259

        pred_prev_sample = pred_prev_sample + variance

260
261
262
        if not return_dict:
            return (pred_prev_sample,)

263
        return DDPMSchedulerOutput(prev_sample=pred_prev_sample, pred_original_sample=pred_original_sample)
Patrick von Platen's avatar
Patrick von Platen committed
264

Partho's avatar
Partho committed
265
266
    def add_noise(
        self,
267
268
269
270
271
272
        original_samples: torch.FloatTensor,
        noise: torch.FloatTensor,
        timesteps: torch.IntTensor,
    ) -> torch.FloatTensor:
        timesteps = timesteps.to(self.alphas_cumprod.device)

anton-l's avatar
anton-l committed
273
        sqrt_alpha_prod = self.alphas_cumprod[timesteps] ** 0.5
274
275
276
277
        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)

anton-l's avatar
anton-l committed
278
        sqrt_one_minus_alpha_prod = (1 - self.alphas_cumprod[timesteps]) ** 0.5
279
280
281
        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)
282
283

        noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
anton-l's avatar
anton-l committed
284
        return noisy_samples
anton-l's avatar
anton-l committed
285

Patrick von Platen's avatar
improve  
Patrick von Platen committed
286
    def __len__(self):
Nathan Lambert's avatar
Nathan Lambert committed
287
        return self.config.num_train_timesteps