DiffMIC/diffusion_utils.py at main · scott-yjyang/DiffMIC · GitHub

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
import math
import torch
import numpy as np

def make_beta_schedule(schedule="linear", num_timesteps=1000, start=1e-5, end=1e-2):
    if schedule == "linear":
        betas = torch.linspace(start, end, num_timesteps)
    elif schedule == "const":
        betas = end * torch.ones(num_timesteps)
    elif schedule == "quad":
        betas = torch.linspace(start ** 0.5, end ** 0.5, num_timesteps) ** 2
    elif schedule == "jsd":
        betas = 1.0 / torch.linspace(num_timesteps, 1, num_timesteps)
    elif schedule == "sigmoid":
        betas = torch.linspace(-6, 6, num_timesteps)
        betas = torch.sigmoid(betas) * (end - start) + start
    elif schedule == "cosine" or schedule == "cosine_reverse":
        max_beta = 0.999
        cosine_s = 0.008
        betas = torch.tensor(
            [min(1 - (math.cos(((i + 1) / num_timesteps + cosine_s) / (1 + cosine_s) * math.pi / 2) ** 2) / (
                    math.cos((i / num_timesteps + cosine_s) / (1 + cosine_s) * math.pi / 2) ** 2), max_beta) for i in
             range(num_timesteps)])
    elif schedule == "cosine_anneal":
        betas = torch.tensor(
            [start + 0.5 * (end - start) * (1 - math.cos(t / (num_timesteps - 1) * math.pi)) for t in
             range(num_timesteps)])
    return betas


def extract(input, t, x):
    shape = x.shape
    out = torch.gather(input, 0, t.to(input.device))
    reshape = [t.shape[0]] + [1] * (len(shape) - 1)
    return out.reshape(*reshape)


# Forward functions
def q_sample(y, y_0_hat, alphas_bar_sqrt, one_minus_alphas_bar_sqrt, t, noise=None):
    """
    y_0_hat: prediction of pre-trained guidance classifier; can be extended to represent
        any prior mean setting at timestep T.
    """
    if noise is None:
        noise = torch.randn_like(y).to(y.device)
    sqrt_alpha_bar_t = extract(alphas_bar_sqrt, t, y)
    sqrt_one_minus_alpha_bar_t = extract(one_minus_alphas_bar_sqrt, t, y)
    # q(y_t | y_0, x)
    y_t = sqrt_alpha_bar_t * y + (1 - sqrt_alpha_bar_t) * y_0_hat + sqrt_one_minus_alpha_bar_t * noise

    return y_t


# Reverse function -- sample y_{t-1} given y_t
def p_sample(model, x, y, y_0_hat, y_T_mean, t, alphas, one_minus_alphas_bar_sqrt):
    """
    Reverse diffusion process sampling -- one time step.

    y: sampled y at time step t, y_t.
    y_0_hat: prediction of pre-trained guidance model.
    y_T_mean: mean of prior distribution at timestep T.
    We replace y_0_hat with y_T_mean in the forward process posterior mean computation, emphasizing that
        guidance model prediction y_0_hat = f_phi(x) is part of the input to eps_theta network, while
        in paper we also choose to set the prior mean at timestep T y_T_mean = f_phi(x).
    """
    device = next(model.parameters()).device
    z = torch.randn_like(y)
    t = torch.tensor([t]).to(device)
    alpha_t = extract(alphas, t, y)
    sqrt_one_minus_alpha_bar_t = extract(one_minus_alphas_bar_sqrt, t, y)
    sqrt_one_minus_alpha_bar_t_m_1 = extract(one_minus_alphas_bar_sqrt, t - 1, y)
    sqrt_alpha_bar_t = (1 - sqrt_one_minus_alpha_bar_t.square()).sqrt()
    sqrt_alpha_bar_t_m_1 = (1 - sqrt_one_minus_alpha_bar_t_m_1.square()).sqrt()
    # y_t_m_1 posterior mean component coefficients
    gamma_0 = (1 - alpha_t) * sqrt_alpha_bar_t_m_1 / (sqrt_one_minus_alpha_bar_t.square())
    gamma_1 = (sqrt_one_minus_alpha_bar_t_m_1.square()) * (alpha_t.sqrt()) / (sqrt_one_minus_alpha_bar_t.square())
    gamma_2 = 1 + (sqrt_alpha_bar_t - 1) * (alpha_t.sqrt() + sqrt_alpha_bar_t_m_1) / (
        sqrt_one_minus_alpha_bar_t.square())
    eps_theta = model(x, y, t, y_0_hat).to(device).detach()
    # y_0 reparameterization
    y_0_reparam = 1 / sqrt_alpha_bar_t * (
            y - (1 - sqrt_alpha_bar_t) * y_T_mean - eps_theta * sqrt_one_minus_alpha_bar_t)
    # posterior mean
    y_t_m_1_hat = gamma_0 * y_0_reparam + gamma_1 * y + gamma_2 * y_T_mean
    # posterior variance
    beta_t_hat = (sqrt_one_minus_alpha_bar_t_m_1.square()) / (sqrt_one_minus_alpha_bar_t.square()) * (1 - alpha_t)
    y_t_m_1 = y_t_m_1_hat.to(device) + beta_t_hat.sqrt().to(device) * z.to(device)
    return y_t_m_1


# Reverse function -- sample y_0 given y_1
def p_sample_t_1to0(model, x, y, y_0_hat, y_T_mean, one_minus_alphas_bar_sqrt):
    device = next(model.parameters()).device
    t = torch.tensor([0]).to(device)  # corresponding to timestep 1 (i.e., t=1 in diffusion models)
    sqrt_one_minus_alpha_bar_t = extract(one_minus_alphas_bar_sqrt, t, y)
    sqrt_alpha_bar_t = (1 - sqrt_one_minus_alpha_bar_t.square()).sqrt()
    eps_theta = model(x, y, t, y_0_hat).to(device).detach()
    # y_0 reparameterization
    y_0_reparam = 1 / sqrt_alpha_bar_t * (
            y - (1 - sqrt_alpha_bar_t) * y_T_mean - eps_theta * sqrt_one_minus_alpha_bar_t)
    y_t_m_1 = y_0_reparam.to(device)
    return y_t_m_1


def y_0_reparam(model, x, y, y_0_hat, y_T_mean, t, one_minus_alphas_bar_sqrt):
    """
    Obtain y_0 reparameterization from q(y_t | y_0), in which noise term is the eps_theta prediction.
    Algorithm 2 Line 4 in paper.
    """
    device = next(model.parameters()).device
    sqrt_one_minus_alpha_bar_t = extract(one_minus_alphas_bar_sqrt, t, y)
    sqrt_alpha_bar_t = (1 - sqrt_one_minus_alpha_bar_t.square()).sqrt()
    eps_theta = model(x, y, t, y_0_hat).to(device).detach()
    # y_0 reparameterization
    y_0_reparam = 1 / sqrt_alpha_bar_t * (
            y - (1 - sqrt_alpha_bar_t) * y_T_mean - eps_theta * sqrt_one_minus_alpha_bar_t).to(device)
    return y_0_reparam


def p_sample_loop(model, x, y_0_hat, y_T_mean, n_steps, alphas, one_minus_alphas_bar_sqrt,
                  only_last_sample=False):
    num_t, y_p_seq = None, None
    device = next(model.parameters()).device
    z = torch.randn_like(y_T_mean).to(device)
    cur_y = z + y_T_mean  # sampled y_T
    if only_last_sample:
        num_t = 1
    else:
        y_p_seq = [cur_y]
    for t in reversed(range(1, n_steps)):
        y_t = cur_y
        cur_y = p_sample(model, x, y_t, y_0_hat, y_T_mean, t, alphas, one_minus_alphas_bar_sqrt)  # y_{t-1}
        if only_last_sample:
            num_t += 1
        else:
            y_p_seq.append(cur_y)
    if only_last_sample:
        assert num_t == n_steps
        y_0 = p_sample_t_1to0(model, x, cur_y, y_0_hat, y_T_mean, one_minus_alphas_bar_sqrt)
        return y_0
    else:
        assert len(y_p_seq) == n_steps
        y_0 = p_sample_t_1to0(model, x, y_p_seq[-1], y_0_hat, y_T_mean, one_minus_alphas_bar_sqrt)
        y_p_seq.append(y_0)
        return y_p_seq


def compute_kernel(x, y):
    x_size = x.size(0)
    y_size = y.size(0)
    dim = x.size(1)
    x = x.unsqueeze(1) # (x_size, 1, dim)
    y = y.unsqueeze(0) # (1, y_size, dim)
    tiled_x = x.expand(x_size, y_size, dim)
    tiled_y = y.expand(x_size, y_size, dim)
    kernel_input = (tiled_x - tiled_y).pow(2).mean(2)/float(dim)
    return torch.exp(-kernel_input) # (x_size, y_size)

def compute_mmd(x, y):
    x_kernel = compute_kernel(x, x)
    y_kernel = compute_kernel(y, y)
    xy_kernel = compute_kernel(x, y)
    mmd = x_kernel.mean() + y_kernel.mean() - 2*xy_kernel.mean()
    return mmd