File size: 12,140 Bytes
522606a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright 2022 Google Brain 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/yang-song/score_sde_pytorch

import math
from dataclasses import dataclass
from typing import Optional, Tuple, Union

import torch

from ..configuration_utils import ConfigMixin, register_to_config
from ..utils import BaseOutput
from .scheduling_utils import SchedulerMixin, SchedulerOutput


@dataclass
class SdeVeOutput(BaseOutput):
    """
    Output class for the ScoreSdeVeScheduler'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.
        prev_sample_mean (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
            Mean averaged `prev_sample`. Same as `prev_sample`, only mean-averaged over previous timesteps.
    """

    prev_sample: torch.FloatTensor
    prev_sample_mean: torch.FloatTensor


class ScoreSdeVeScheduler(SchedulerMixin, ConfigMixin):
    """
    The variance exploding stochastic differential equation (SDE) scheduler.

    For more information, see the original paper: https://arxiv.org/abs/2011.13456

    [`~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.
        snr (`float`):
            coefficient weighting the step from the model_output sample (from the network) to the random noise.
        sigma_min (`float`):
                initial noise scale for sigma sequence in sampling procedure. The minimum sigma should mirror the
                distribution of the data.
        sigma_max (`float`): maximum value used for the range of continuous timesteps passed into the model.
        sampling_eps (`float`): the end value of sampling, where timesteps decrease progressively from 1 to
        epsilon.
        correct_steps (`int`): number of correction steps performed on a produced sample.
    """

    order = 1

    @register_to_config
    def __init__(
        self,
        num_train_timesteps: int = 2000,
        snr: float = 0.15,
        sigma_min: float = 0.01,
        sigma_max: float = 1348.0,
        sampling_eps: float = 1e-5,
        correct_steps: int = 1,
    ):
        # standard deviation of the initial noise distribution
        self.init_noise_sigma = sigma_max

        # setable values
        self.timesteps = None

        self.set_sigmas(num_train_timesteps, sigma_min, sigma_max, sampling_eps)

    def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> 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
            timestep (`int`, optional): current timestep

        Returns:
            `torch.FloatTensor`: scaled input sample
        """
        return sample

    def set_timesteps(
        self, num_inference_steps: int, sampling_eps: float = None, device: Union[str, torch.device] = None
    ):
        """
        Sets the continuous 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.
            sampling_eps (`float`, optional): final timestep value (overrides value given at Scheduler instantiation).

        """
        sampling_eps = sampling_eps if sampling_eps is not None else self.config.sampling_eps

        self.timesteps = torch.linspace(1, sampling_eps, num_inference_steps, device=device)

    def set_sigmas(
        self, num_inference_steps: int, sigma_min: float = None, sigma_max: float = None, sampling_eps: float = None
    ):
        """
        Sets the noise scales used for the diffusion chain. Supporting function to be run before inference.

        The sigmas control the weight of the `drift` and `diffusion` components of sample update.

        Args:
            num_inference_steps (`int`):
                the number of diffusion steps used when generating samples with a pre-trained model.
            sigma_min (`float`, optional):
                initial noise scale value (overrides value given at Scheduler instantiation).
            sigma_max (`float`, optional): final noise scale value (overrides value given at Scheduler instantiation).
            sampling_eps (`float`, optional): final timestep value (overrides value given at Scheduler instantiation).

        """
        sigma_min = sigma_min if sigma_min is not None else self.config.sigma_min
        sigma_max = sigma_max if sigma_max is not None else self.config.sigma_max
        sampling_eps = sampling_eps if sampling_eps is not None else self.config.sampling_eps
        if self.timesteps is None:
            self.set_timesteps(num_inference_steps, sampling_eps)

        self.sigmas = sigma_min * (sigma_max / sigma_min) ** (self.timesteps / sampling_eps)
        self.discrete_sigmas = torch.exp(torch.linspace(math.log(sigma_min), math.log(sigma_max), num_inference_steps))
        self.sigmas = torch.tensor([sigma_min * (sigma_max / sigma_min) ** t for t in self.timesteps])

    def get_adjacent_sigma(self, timesteps, t):
        return torch.where(
            timesteps == 0,
            torch.zeros_like(t.to(timesteps.device)),
            self.discrete_sigmas[timesteps - 1].to(timesteps.device),
        )

    def step_pred(
        self,
        model_output: torch.FloatTensor,
        timestep: int,
        sample: torch.FloatTensor,
        generator: Optional[torch.Generator] = None,
        return_dict: bool = True,
    ) -> Union[SdeVeOutput, Tuple]:
        """
        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:
            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.
            generator: random number generator.
            return_dict (`bool`): option for returning tuple rather than SchedulerOutput class

        Returns:
            [`~schedulers.scheduling_sde_ve.SdeVeOutput`] or `tuple`: [`~schedulers.scheduling_sde_ve.SdeVeOutput`] if
            `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is the sample tensor.

        """
        if self.timesteps is None:
            raise ValueError(
                "`self.timesteps` is not set, you need to run 'set_timesteps' after creating the scheduler"
            )

        timestep = timestep * torch.ones(
            sample.shape[0], device=sample.device
        )  # torch.repeat_interleave(timestep, sample.shape[0])
        timesteps = (timestep * (len(self.timesteps) - 1)).long()

        # mps requires indices to be in the same device, so we use cpu as is the default with cuda
        timesteps = timesteps.to(self.discrete_sigmas.device)

        sigma = self.discrete_sigmas[timesteps].to(sample.device)
        adjacent_sigma = self.get_adjacent_sigma(timesteps, timestep).to(sample.device)
        drift = torch.zeros_like(sample)
        diffusion = (sigma**2 - adjacent_sigma**2) ** 0.5

        # equation 6 in the paper: the model_output modeled by the network is grad_x log pt(x)
        # also equation 47 shows the analog from SDE models to ancestral sampling methods
        diffusion = diffusion.flatten()
        while len(diffusion.shape) < len(sample.shape):
            diffusion = diffusion.unsqueeze(-1)
        drift = drift - diffusion**2 * model_output

        #  equation 6: sample noise for the diffusion term of
        noise = torch.randn(sample.shape, layout=sample.layout, generator=generator).to(sample.device)
        prev_sample_mean = sample - drift  # subtract because `dt` is a small negative timestep
        # TODO is the variable diffusion the correct scaling term for the noise?
        prev_sample = prev_sample_mean + diffusion * noise  # add impact of diffusion field g

        if not return_dict:
            return (prev_sample, prev_sample_mean)

        return SdeVeOutput(prev_sample=prev_sample, prev_sample_mean=prev_sample_mean)

    def step_correct(
        self,
        model_output: torch.FloatTensor,
        sample: torch.FloatTensor,
        generator: Optional[torch.Generator] = None,
        return_dict: bool = True,
    ) -> Union[SchedulerOutput, Tuple]:
        """
        Correct the predicted sample based on the output model_output of the network. This is often run repeatedly
        after making the prediction for the previous timestep.

        Args:
            model_output (`torch.FloatTensor`): direct output from learned diffusion model.
            sample (`torch.FloatTensor`):
                current instance of sample being created by diffusion process.
            generator: random number generator.
            return_dict (`bool`): option for returning tuple rather than SchedulerOutput class

        Returns:
            [`~schedulers.scheduling_sde_ve.SdeVeOutput`] or `tuple`: [`~schedulers.scheduling_sde_ve.SdeVeOutput`] if
            `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is the sample tensor.

        """
        if self.timesteps is None:
            raise ValueError(
                "`self.timesteps` is not set, you need to run 'set_timesteps' after creating the scheduler"
            )

        # For small batch sizes, the paper "suggest replacing norm(z) with sqrt(d), where d is the dim. of z"
        # sample noise for correction
        noise = torch.randn(sample.shape, layout=sample.layout, generator=generator).to(sample.device)

        # compute step size from the model_output, the noise, and the snr
        grad_norm = torch.norm(model_output.reshape(model_output.shape[0], -1), dim=-1).mean()
        noise_norm = torch.norm(noise.reshape(noise.shape[0], -1), dim=-1).mean()
        step_size = (self.config.snr * noise_norm / grad_norm) ** 2 * 2
        step_size = step_size * torch.ones(sample.shape[0]).to(sample.device)
        # self.repeat_scalar(step_size, sample.shape[0])

        # compute corrected sample: model_output term and noise term
        step_size = step_size.flatten()
        while len(step_size.shape) < len(sample.shape):
            step_size = step_size.unsqueeze(-1)
        prev_sample_mean = sample + step_size * model_output
        prev_sample = prev_sample_mean + ((step_size * 2) ** 0.5) * noise

        if not return_dict:
            return (prev_sample,)

        return SchedulerOutput(prev_sample=prev_sample)

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