BG-ODE/model.py at main · Kerr93/BG-ODE · 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
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchdiffeq import odeint_adjoint as odeint

import numpy as np

from data_utils import state_dim
import ipdb


def _get_loss_with_nan(y, y_hat):
    mask = 1 - torch.isnan(y).float()

    y = torch.nan_to_num(y)
    loss = F.mse_loss(y, y_hat, reduction='none')
    loss = torch.sum(loss * mask) / torch.sum(mask)

    return loss


class ODEFunc(nn.Module):
    def __init__(self, hid_dim):
        super().__init__()

        self.f = nn.Sequential(
            nn.Linear(hid_dim, hid_dim),
            nn.GELU(),
            nn.Linear(hid_dim, hid_dim),
            nn.GELU(),
            nn.Linear(hid_dim, hid_dim)
        )

    def forward(self, t, x):
        return self.f(x)


class Model(nn.Module):
    def __init__(self, input_dim, hid_dim):
        super().__init__()

        self.input_dim = input_dim - 1
        self.hid_dim = hid_dim

        self.rnn = nn.GRUCell(self.input_dim, self.hid_dim)

        self.ode_func = ODEFunc(self.hid_dim)

        self.g = nn.Sequential(
            nn.Linear(hid_dim, hid_dim),
            nn.GELU(),
            nn.Linear(hid_dim, hid_dim),
            nn.GELU(),
            nn.Linear(hid_dim, 1)
        )

        self.s = nn.Sequential(
            nn.Linear(hid_dim, hid_dim),
            nn.GELU(),
            nn.Linear(hid_dim, hid_dim),
            nn.GELU(),
            nn.Linear(hid_dim, state_dim)
        )

    def forward(self, inputs, mask, do_interpolation=False):
        if do_interpolation:
            assert(not self.training)
            interpolations = []

        ts = inputs.select(-1, 0).squeeze(dim=0)

        start_time = ts[0]
        ts = (ts - start_time) / 1000

        inputs = inputs.clone()
        inputs = torch.nan_to_num(inputs)

        inputs = inputs[:, :, 1:]

        hid = torch.zeros(inputs.size(0), self.hid_dim).cuda()

        predictions = []
        states = []

        for t in range(inputs.size(1)):
            prediction = self.g(hid.squeeze(dim=1)).squeeze(dim=-1) #(1)
            state = self.s(hid.squeeze(dim=1)) #(1, 13)

            if t > 0:
                # no predictio for t = 0 since we have no prior knowledge
                predictions.append(prediction)
                states.append(state)

            x = inputs.select(1, t)
            m = mask.select(1, t)

            blood, others = x.split([1, self.input_dim - 1], dim=-1)
            # replace blood with prediction for auto-regressive prediction
            blood = m * blood + (1 - m) * prediction

            x = torch.cat([blood, others], dim=-1)

            hid = self.rnn(x, hid)

            if t < inputs.size(1) - 1:
                if do_interpolation:
                    ts_span = ts[[t, t + 1]]
                    steps = np.round((ts[t + 1] - ts[t]).item() * 1e3) + 1
                    ts_span = torch.linspace(
                        ts[t], ts[t + 1], steps=int(steps)
                    )
                else:
                    ts_span = ts[[t, t + 1]]

                hid = odeint(
                    self.ode_func, hid, ts_span,
                    rtol=1e-3, atol=1e-3,
                    adjoint_options=dict(norm='seminorm')
                )

                if do_interpolation:
                    _interpolation = self.g(hid.squeeze(dim=1)).squeeze(dim=-1)
                    if t == 0:
                        _interpolation = _interpolation[[-1]]
                    else:
                        # eliminate the first point since we already know the input (i.e., x)
                        _interpolation = _interpolation[1:]

                    interpolations.append(_interpolation)

                hid = hid[-1]

        predictions = torch.cat(predictions, dim=0)
        states = torch.cat(states, dim=0)

        if do_interpolation:
            interpolations = torch.cat(interpolations, dim=-1)
            return predictions, states, interpolations
        else:
            return predictions, states