File size: 5,458 Bytes
95ba5bc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn.functional as F
import math
import numpy as np


def clip_noise_schedule(alphas2, clip_value=0.001):
    """
    For a noise schedule given by alpha^2, this clips alpha_t / alpha_t-1. This may help improve stability during
    sampling.
    """
    alphas2 = np.concatenate([np.ones(1), alphas2], axis=0)

    alphas_step = (alphas2[1:] / alphas2[:-1])

    alphas_step = np.clip(alphas_step, a_min=clip_value, a_max=1.)
    alphas2 = np.cumprod(alphas_step, axis=0)

    return alphas2


def polynomial_schedule(timesteps: int, s=1e-4, power=3.):
    """
    A noise schedule based on a simple polynomial equation: 1 - x^power.
    """
    steps = timesteps + 1
    x = np.linspace(0, steps, steps)
    alphas2 = (1 - np.power(x / steps, power)) ** 2

    alphas2 = clip_noise_schedule(alphas2, clip_value=0.001)

    precision = 1 - 2 * s

    alphas2 = precision * alphas2 + s

    return alphas2


def cosine_beta_schedule(timesteps, s=0.008, raise_to_power: float = 1):
    """
    cosine schedule
    as proposed in https://openreview.net/forum?id=-NEXDKk8gZ
    """
    steps = timesteps + 2
    x = np.linspace(0, steps, steps)
    alphas_cumprod = np.cos(((x / steps) + s) / (1 + s) * np.pi * 0.5) ** 2
    alphas_cumprod = alphas_cumprod / alphas_cumprod[0]
    betas = 1 - (alphas_cumprod[1:] / alphas_cumprod[:-1])
    betas = np.clip(betas, a_min=0, a_max=0.999)
    alphas = 1. - betas
    alphas_cumprod = np.cumprod(alphas, axis=0)

    if raise_to_power != 1:
        alphas_cumprod = np.power(alphas_cumprod, raise_to_power)

    return alphas_cumprod


class PositiveLinear(torch.nn.Module):
    """Linear layer with weights forced to be positive."""

    def __init__(self, in_features: int, out_features: int, bias: bool = True,
                 weight_init_offset: int = -2):
        super(PositiveLinear, self).__init__()
        self.in_features = in_features
        self.out_features = out_features
        self.weight = torch.nn.Parameter(
            torch.empty((out_features, in_features)))
        if bias:
            self.bias = torch.nn.Parameter(torch.empty(out_features))
        else:
            self.register_parameter('bias', None)
        self.weight_init_offset = weight_init_offset
        self.reset_parameters()

    def reset_parameters(self) -> None:
        torch.nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5))

        with torch.no_grad():
            self.weight.add_(self.weight_init_offset)

        if self.bias is not None:
            fan_in, _ = torch.nn.init._calculate_fan_in_and_fan_out(self.weight)
            bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0
            torch.nn.init.uniform_(self.bias, -bound, bound)

    def forward(self, x):
        positive_weight = F.softplus(self.weight)
        return F.linear(x, positive_weight, self.bias)


class PredefinedNoiseSchedule(torch.nn.Module):
    """
    Predefined noise schedule. Essentially creates a lookup array for predefined (non-learned) noise schedules.
    """

    def __init__(self, noise_schedule, timesteps, precision):
        super(PredefinedNoiseSchedule, self).__init__()
        self.timesteps = timesteps

        if noise_schedule == 'cosine':
            alphas2 = cosine_beta_schedule(timesteps)
        elif 'polynomial' in noise_schedule:
            splits = noise_schedule.split('_')
            assert len(splits) == 2
            power = float(splits[1])
            alphas2 = polynomial_schedule(timesteps, s=precision, power=power)
        else:
            raise ValueError(noise_schedule)

        # print('alphas2', alphas2)

        sigmas2 = 1 - alphas2

        log_alphas2 = np.log(alphas2)
        log_sigmas2 = np.log(sigmas2)

        log_alphas2_to_sigmas2 = log_alphas2 - log_sigmas2

        # print('gamma', -log_alphas2_to_sigmas2)

        self.gamma = torch.nn.Parameter(
            torch.from_numpy(-log_alphas2_to_sigmas2).float(),
            requires_grad=False)

    def forward(self, t):
        t_int = torch.round(t * self.timesteps).long()
        return self.gamma[t_int]


class GammaNetwork(torch.nn.Module):
    """The gamma network models a monotonic increasing function. Construction as in the VDM paper."""

    def __init__(self):
        super().__init__()

        self.l1 = PositiveLinear(1, 1)
        self.l2 = PositiveLinear(1, 1024)
        self.l3 = PositiveLinear(1024, 1)

        self.gamma_0 = torch.nn.Parameter(torch.tensor([-5.]))
        self.gamma_1 = torch.nn.Parameter(torch.tensor([10.]))
        self.show_schedule()

    def show_schedule(self, num_steps=50):
        t = torch.linspace(0, 1, num_steps).view(num_steps, 1)
        gamma = self.forward(t)
        print('Gamma schedule:')
        print(gamma.detach().cpu().numpy().reshape(num_steps))

    def gamma_tilde(self, t):
        l1_t = self.l1(t)
        return l1_t + self.l3(torch.sigmoid(self.l2(l1_t)))

    def forward(self, t):
        zeros, ones = torch.zeros_like(t), torch.ones_like(t)
        # Not super efficient.
        gamma_tilde_0 = self.gamma_tilde(zeros)
        gamma_tilde_1 = self.gamma_tilde(ones)
        gamma_tilde_t = self.gamma_tilde(t)

        # Normalize to [0, 1]
        normalized_gamma = (gamma_tilde_t - gamma_tilde_0) / (
                gamma_tilde_1 - gamma_tilde_0)

        # Rescale to [gamma_0, gamma_1]
        gamma = self.gamma_0 + (self.gamma_1 - self.gamma_0) * normalized_gamma

        return gamma