File size: 6,036 Bytes
5b3329b
986404c
 
5b3329b
986404c
 
 
 
 
 
 
 
5b3329b
986404c
 
5b3329b
986404c
 
5b3329b
986404c
 
 
 
 
 
 
 
 
5b3329b
986404c
5b3329b
 
986404c
5b3329b
 
 
986404c
 
5b3329b
 
986404c
 
5b3329b
 
 
 
 
 
986404c
 
5b3329b
 
986404c
5b3329b
 
986404c
5b3329b
 
986404c
 
5b3329b
 
986404c
 
 
 
 
5b3329b
 
 
 
 
 
 
 
 
986404c
 
 
5b3329b
 
986404c
5b3329b
 
 
 
 
 
986404c
 
 
 
5b3329b
 
 
986404c
5b3329b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
986404c
5b3329b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Compact but faithful conditional-regression and residual-EDM CorrDiff."""

import math

import torch
from torch import nn
from torch.nn import functional as F


class ConvBlock(nn.Module):
    def __init__(self, in_channels, out_channels):
        super().__init__()
        groups = min(8, out_channels)
        self.block = nn.Sequential(
            nn.Conv2d(in_channels, out_channels, 3, padding=1),
            nn.GroupNorm(groups, out_channels),
            nn.SiLU(),
            nn.Conv2d(out_channels, out_channels, 3, padding=1),
            nn.GroupNorm(groups, out_channels),
            nn.SiLU(),
        )
        self.skip = nn.Conv2d(in_channels, out_channels, 1)

    def forward(self, x):
        return self.block(x) + self.skip(x)


class RegressionUNet(nn.Module):
    """Deterministic conditional mean, computed cheaply before 448 px recovery."""

    def __init__(self, in_channels=12, out_channels=4, base_channels=16, feature_size=56,
                 output_size=448):
        super().__init__()
        self.feature_size = feature_size
        self.output_size = output_size
        self.net = nn.Sequential(
            ConvBlock(in_channels, base_channels),
            ConvBlock(base_channels, base_channels * 2),
            ConvBlock(base_channels * 2, base_channels),
            nn.Conv2d(base_channels, out_channels, 1),
        )

    def forward(self, coarse):
        x = F.interpolate(coarse, (self.feature_size, self.feature_size), mode="bilinear",
                          align_corners=False)
        x = self.net(x)
        return F.interpolate(x, (self.output_size, self.output_size), mode="bilinear",
                             align_corners=False)


class ResidualDenoiser(nn.Module):
    """Noise-conditional network used inside EDM preconditioning."""

    def __init__(self, condition_channels=16, out_channels=4, base_channels=16,
                 feature_size=56):
        super().__init__()
        self.feature_size = feature_size
        self.noise_mlp = nn.Sequential(
            nn.Linear(1, base_channels), nn.SiLU(), nn.Linear(base_channels, base_channels)
        )
        self.input = ConvBlock(condition_channels, base_channels)
        self.body = nn.Sequential(
            ConvBlock(base_channels, base_channels * 2),
            ConvBlock(base_channels * 2, base_channels),
            nn.Conv2d(base_channels, out_channels, 1),
        )

    def forward(self, noisy, condition, c_noise):
        size = noisy.shape[-2:]
        x = torch.cat((noisy, condition), dim=1)
        x = F.interpolate(x, (self.feature_size, self.feature_size), mode="bilinear",
                          align_corners=False)
        x = self.input(x)
        x = x + self.noise_mlp(c_noise[:, None].float())[:, :, None, None].to(x.dtype)
        x = self.body(x)
        return F.interpolate(x, size, mode="bilinear", align_corners=False)


class CorrDiff(nn.Module):
    def __init__(self, in_channels=12, out_channels=4, base_channels=16, feature_size=56,
                 output_size=448, sigma_data=0.5):
        super().__init__()
        self.sigma_data = sigma_data
        self.output_size = output_size
        self.regression = RegressionUNet(in_channels, out_channels, base_channels,
                                         feature_size, output_size)
        self.diffusion = ResidualDenoiser(in_channels + 2 * out_channels, out_channels,
                                          base_channels, feature_size)

    def mean(self, coarse):
        return self.regression(coarse)

    def condition(self, coarse, mean):
        coarse = F.interpolate(coarse, mean.shape[-2:], mode="bilinear", align_corners=False)
        return torch.cat((coarse, mean), dim=1)

    def denoise(self, noisy_residual, coarse, mean, sigma):
        sigma = sigma.reshape(-1, 1, 1, 1).to(noisy_residual.dtype)
        sigma_data = self.sigma_data
        c_skip = sigma_data**2 / (sigma.square() + sigma_data**2)
        c_out = sigma * sigma_data / (sigma.square() + sigma_data**2).sqrt()
        c_in = (sigma.square() + sigma_data**2).rsqrt()
        c_noise = sigma.flatten().log() / 4
        network = self.diffusion(c_in * noisy_residual, self.condition(coarse, mean), c_noise)
        return c_skip * noisy_residual + c_out * network

    @staticmethod
    def karras_schedule(steps, sigma_min, sigma_max, rho, device):
        ramp = torch.linspace(0, 1, steps, device=device)
        maximum = sigma_max ** (1 / rho)
        minimum = sigma_min ** (1 / rho)
        sigmas = (maximum + ramp * (minimum - maximum)) ** rho
        return torch.cat((sigmas, sigmas.new_zeros(1)))

    def sample(self, coarse, steps=4, sigma_min=0.002, sigma_max=5.0, rho=7.0,
               solver="heun"):
        mean = self.mean(coarse)
        sigmas = self.karras_schedule(steps, sigma_min, sigma_max, rho, coarse.device)
        x = torch.randn_like(mean) * sigmas[0]
        for index, (current, following) in enumerate(zip(sigmas[:-1], sigmas[1:])):
            sigma = current.expand(coarse.shape[0])
            denoised = self.denoise(x, coarse, mean, sigma)
            derivative = (x - denoised) / current
            proposal = x + (following - current) * derivative
            if solver == "heun" and index < len(sigmas) - 2:
                next_sigma = following.expand(coarse.shape[0])
                next_denoised = self.denoise(proposal, coarse, mean, next_sigma)
                next_derivative = (proposal - next_denoised) / following
                x = x + (following - current) * (derivative + next_derivative) / 2
            else:
                x = proposal
        return mean + x

    def forward(self, coarse, mode="sample", mean=None, noisy=None, sigma=None,
                **sample_options):
        if mode == "mean":
            return self.mean(coarse)
        if mode == "denoise":
            return self.denoise(noisy, coarse, mean, sigma)
        return self.sample(coarse, **sample_options)


__all__ = ["CorrDiff", "RegressionUNet", "ResidualDenoiser"]