FEAT/sampling.py at main · jiajunhe98/FEAT · 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
import torch
from utils import Coef
import numpy as np
from torch.distributions import Normal
from tqdm import tqdm
from networks.egnn import remove_mean


# Controlled Jarzynski with Eular-Maruyama integration
def Jarzynski_integrate(x_init: torch.Tensor,
                        init_logprob: torch.Tensor,
                        eps: Coef,
                        alpha: Coef,
                        beta: Coef,
                        gamma: Coef,
                        vector_field: callable,
                        score_net: callable,
                        collect_interval: int = 10,
                        learn_b: bool = True,
                        times: torch.Tensor = None,
                        forward: bool = True,
                        return_A: bool = False,
                        target_logp: callable = None,
                        center_noise: bool = False,
                        n_particles: int = None,
                        n_dim: int = None,
                        ):

    t_start = times[0]
    t_end = times[-1]

    if forward:
        assert t_start < t_end
    else:
        assert t_start > t_end

    device = x_init.device
    batch_size = x_init.shape[0]
    dim_size = [1 for _ in x_init.shape[1:]]

    # initialize
    A = torch.zeros(batch_size).to(device)
    t = torch.zeros(batch_size).to(device) + t_start

    # sample
    x = x_init
    Xs = [x]
    weighted_Xs = []
    ESS = []

    A -= init_logprob
    for i in range(len(times)-1):

        step_size = np.abs((times[i+1] - times[i]).item())

        if i == 0:
            print (x.shape)
            vf = vector_field(x, t)
            s = score_net.score(x, t)
            if not learn_b:
                vf = vf - gamma.tdt(t).reshape(-1, *dim_size) * s
        std = torch.sqrt(2 * step_size * eps.t(t).reshape(-1, *dim_size))
        if forward:
            noise = torch.randn_like(x)
            if center_noise:
                noise = remove_mean(noise, n_particles, n_dim)
            x_new = x + step_size * vf +  step_size * eps.t(t).reshape(-1, *dim_size) * s + std * noise
            t_new = t*0 + times[i+1]
            A -= Normal(x+step_size*vf+step_size*eps.t(t).reshape(-1, *dim_size)*s, std, validate_args=False).log_prob(x_new).sum(-1)
            vf_new = vector_field(x_new, t_new)
            if not learn_b:
                vf_new = vf_new - gamma.tdt(t_new).reshape(-1, *dim_size) * s
            s_new = score_net.score(x_new, t_new)
            std_new = torch.sqrt(2 * step_size * eps.t(t_new).reshape(-1, *dim_size))
            A += Normal(x_new-step_size*vf_new+step_size*eps.t(t_new).reshape(-1, *dim_size) * s_new, std_new, validate_args=False).log_prob(x).sum(-1)
        else:
            noise = torch.randn_like(x)
            if center_noise:
                noise = remove_mean(noise, n_particles, n_dim)
            x_new = x - step_size * vf +  step_size * eps.t(t).reshape(-1, *dim_size) * s + std * noise
            t_new = t*0 + times[i+1]
            A -= Normal(x-step_size*vf+step_size*eps.t(t).reshape(-1, *dim_size)*s, std, validate_args=False).log_prob(x_new).sum(-1)
            vf_new = vector_field(x_new, t_new)
            if not learn_b:
                vf_new = vf_new - gamma.tdt(t_new).reshape(-1, *dim_size) * s
            s_new = score_net.score(x_new, t_new)
            std_new = torch.sqrt(2 * step_size * eps.t(t_new).reshape(-1, *dim_size))
            A += Normal(x_new+step_size*vf_new+step_size*eps.t(t_new).reshape(-1, *dim_size) * s_new, std_new, validate_args=False).log_prob(x).sum(-1)

        x = x_new
        t = t_new
        vf = vf_new
        s = s_new

        if (i+1) % collect_interval == 0 or i == len(times)-2:
            weighted_Xs.append((x, t.clone()))
        Xs.append(x)

    if return_A:
        A_final = A + target_logp(x).flatten()
        return x, Xs, A_final

    return x, Xs,


# # Controlled Jarzynski
def Jarzynski_integrate_ODE(x1: torch.Tensor, vector_field: callable, n_steps: int = 100, forward: bool = True, use_pyg: bool = False, calculate_div: bool = True):

    if forward:
        t_start = 0
        t_end = 1
    else:
        t_start = 1
        t_end = 0

    step_size = (t_end - t_start) / n_steps
    if use_pyg:
        device = x1.x.device
        batch_size = x1.batch_size
        dim_size = [1 for _ in x1.x.shape[1:]]
    else:
        device = x1.device
        batch_size = x1.shape[0]
        dim_size = [1 for _ in x1.shape[1:]]

    # initialize
    A = torch.zeros(batch_size).to(device)
    t = torch.zeros(batch_size).to(device) + t_start

    # sample
    x = x1
    Xs = [x]
    if use_pyg:
        x.input = x.x
    for i in range(n_steps-1):
        vf = vector_field(x, t).detach()
        if calculate_div:
            A = A + vector_field.div(x, t).detach() * step_size
        if use_pyg:
            x.input = x.input.detach() + step_size * vf.detach()
        else:
            x = x.detach() + step_size * vf.detach()

        t += step_size
        Xs.append(x.clone())


    return x, Xs, A