"vscode:/vscode.git/clone" did not exist on "b036725166ffc2803c0d4156f7abd859b45e92b3"
scheduling_sde_vp.py 3.31 KB
Newer Older
Patrick von Platen's avatar
Patrick von Platen committed
1
# Copyright 2023 Google Brain and The HuggingFace Team. All rights reserved.
Patrick von Platen's avatar
Patrick von Platen committed
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
#
# 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

17
import math
18
from typing import Union
19

Patrick von Platen's avatar
Patrick von Platen committed
20
21
import torch

22
from ..configuration_utils import ConfigMixin, register_to_config
23
from ..utils import randn_tensor
Patrick von Platen's avatar
Patrick von Platen committed
24
25
26
27
from .scheduling_utils import SchedulerMixin


class ScoreSdeVpScheduler(SchedulerMixin, ConfigMixin):
28
29
30
    """
    The variance preserving stochastic differential equation (SDE) scheduler.

31
32
    [`~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`.
33
34
    [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
    [`~SchedulerMixin.from_pretrained`] functions.
35

36
37
38
39
40
41
    For more information, see the original paper: https://arxiv.org/abs/2011.13456

    UNDER CONSTRUCTION

    """

42
43
    order = 1

44
    @register_to_config
Anton Lozhkov's avatar
Anton Lozhkov committed
45
    def __init__(self, num_train_timesteps=2000, beta_min=0.1, beta_max=20, sampling_eps=1e-3):
Patrick von Platen's avatar
Patrick von Platen committed
46
47
48
49
        self.sigmas = None
        self.discrete_sigmas = None
        self.timesteps = None

50
51
    def set_timesteps(self, num_inference_steps, device: Union[str, torch.device] = None):
        self.timesteps = torch.linspace(1, self.config.sampling_eps, num_inference_steps, device=device)
Patrick von Platen's avatar
Patrick von Platen committed
52

53
    def step_pred(self, score, x, t, generator=None):
54
55
56
57
58
        if self.timesteps is None:
            raise ValueError(
                "`self.timesteps` is not set, you need to run 'set_timesteps' after creating the scheduler"
            )

Patrick von Platen's avatar
Patrick von Platen committed
59
        # TODO(Patrick) better comments + non-PyTorch
Nathan Lambert's avatar
Nathan Lambert committed
60
        # postprocess model score
Patrick von Platen's avatar
Patrick von Platen committed
61
62
63
64
        log_mean_coeff = (
            -0.25 * t**2 * (self.config.beta_max - self.config.beta_min) - 0.5 * t * self.config.beta_min
        )
        std = torch.sqrt(1.0 - torch.exp(2.0 * log_mean_coeff))
65
66
67
68
        std = std.flatten()
        while len(std.shape) < len(score.shape):
            std = std.unsqueeze(-1)
        score = -score / std
Patrick von Platen's avatar
Patrick von Platen committed
69

Patrick von Platen's avatar
Patrick von Platen committed
70
71
72
73
        # compute
        dt = -1.0 / len(self.timesteps)

        beta_t = self.config.beta_min + t * (self.config.beta_max - self.config.beta_min)
74
75
76
77
78
        beta_t = beta_t.flatten()
        while len(beta_t.shape) < len(x.shape):
            beta_t = beta_t.unsqueeze(-1)
        drift = -0.5 * beta_t * x

Patrick von Platen's avatar
Patrick von Platen committed
79
        diffusion = torch.sqrt(beta_t)
80
        drift = drift - diffusion**2 * score
Patrick von Platen's avatar
Patrick von Platen committed
81
        x_mean = x + drift * dt
Patrick von Platen's avatar
Patrick von Platen committed
82
83

        # add noise
84
        noise = randn_tensor(x.shape, layout=x.layout, generator=generator, device=x.device, dtype=x.dtype)
85
        x = x_mean + diffusion * math.sqrt(-dt) * noise
Patrick von Platen's avatar
Patrick von Platen committed
86
87

        return x, x_mean
Nathan Lambert's avatar
Nathan Lambert committed
88
89
90

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