STD-Hunter/STD_Hunter.py at main · YangLabHKUST/STD-Hunter · 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
import torch
import torch.nn as nn
from torch.nn import init
import numpy as np
import numbers
import torch.nn.functional as F


class LayerNorm(nn.Module):
    __constants__ = ['normalized_shape', 'weight', 'bias', 'eps', 'elementwise_affine']
    def __init__(self, normalized_shape, eps=1e-5, elementwise_affine=True):
        super(LayerNorm, self).__init__()
        if isinstance(normalized_shape, numbers.Integral):
            normalized_shape = (normalized_shape,)
        self.normalized_shape = tuple(normalized_shape)
        self.eps = eps
        self.elementwise_affine = elementwise_affine
        if self.elementwise_affine:
            self.weight = nn.Parameter(torch.Tensor(*normalized_shape))
            self.bias = nn.Parameter(torch.Tensor(*normalized_shape))
        else:
            self.register_parameter('weight', None)
            self.register_parameter('bias', None)
        self.reset_parameters()


    def reset_parameters(self):
        if self.elementwise_affine:
            init.ones_(self.weight)
            init.zeros_(self.bias)

    def forward(self, input):
        if self.elementwise_affine:
            return F.layer_norm(input, tuple(input.shape[1:]), self.weight, self.bias, self.eps)
        else:
            return F.layer_norm(input, tuple(input.shape[1:]), self.weight, self.bias, self.eps)

    def extra_repr(self):
        return '{normalized_shape}, eps={eps}, ' \
            'elementwise_affine={elementwise_affine}'.format(**self.__dict__)


class STD_Hunter(nn.Module):
    def __init__(self, num_nodes, device, adj, dropout=0.3, channels=32, seq_length=12, in_dim=2, out_dim=1, layers=3, rank=4):
        super(STD_Hunter, self).__init__()
        self.dropout = dropout
        self.channels = channels
        self.adj = adj
        self.device = device
        self.seq_length = seq_length
        self.norm = nn.ModuleList()
        self.skip_convs = nn.ModuleList()
        self.rank = rank
        self.start_conv = nn.Conv2d(in_channels=in_dim, out_channels=channels, kernel_size=(1, 1))
        self.layers = layers

        self.characteristics = nn.Parameter(torch.empty(size=(num_nodes, 12, self.rank)))
        nn.init.xavier_uniform_(self.characteristics.data, gain=1.414)

        self.conv_w = nn.Linear(self.rank, int(channels/4)*channels*18*layers)
        self.conv_b = nn.Linear(self.rank, int(channels/4)*4*layers)
        self.gate_w = nn.Linear(self.rank, int(channels/4)*channels*18*layers)
        self.gate_b = nn.Linear(self.rank, int(channels/4)*4*layers)

        self.res_w = nn.Linear(self.rank, channels*channels*layers)
        self.res_b = nn.Linear(self.rank, channels*layers)

        for j in range(layers):
            self.norm.append(LayerNorm((channels, num_nodes, seq_length),elementwise_affine=True))
            self.skip_convs.append(nn.Conv2d(in_channels=channels, out_channels=channels, kernel_size=(1, seq_length), groups=channels))

        self.end_conv_1 = nn.Conv2d(in_channels=channels, out_channels=channels, kernel_size=(1,1), bias=True)
        self.end_conv_2 = nn.Conv2d(in_channels=channels, out_channels=out_dim, kernel_size=(1,1), bias=True)

        self.idx = torch.arange(num_nodes).to(device)
        self.skip0 = nn.Conv2d(in_channels=in_dim, out_channels=channels, kernel_size=(1, seq_length), bias=True)
        self.skipE = nn.Conv2d(in_channels=channels, out_channels=channels, kernel_size=(1, seq_length), groups=channels)

    def forward(self, input):
        # the prediction month of this batch
        month = input[0, 0, 0, -1]
        input = input[:, :, :, 0:self.seq_length]

        characteristics = self.characteristics
        # graph convolutional layer on spatiotemporal characteristics
        adj_month = np.diag(np.ones(12))/3
        np.fill_diagonal(adj_month[1:], 1/3)
        np.fill_diagonal(adj_month[:, 1:], 1/3)
        adj_month[0, 11] = 1/3
        adj_month[11, 0] = 1/3
        adj_month = torch.Tensor(adj_month).to(self.device)
        characteristics = torch.matmul(adj_month, characteristics)
        characteristics = torch.matmul(self.adj, characteristics[:, torch.tensor(month, dtype=torch.long)])

        # generate parameters for TC layers using characteristics
        conv_w_all = self.conv_w(characteristics)
        conv_b_all = self.conv_b(characteristics)
        gate_w_all = self.gate_w(characteristics)
        gate_b_all = self.gate_b(characteristics)
        res_w_all = self.res_w(characteristics)
        res_b_all = self.res_b(characteristics)

        x = self.start_conv(input)
        skip = self.skip0(F.dropout(input, self.dropout, training=self.training))
        for i in range(self.layers):
            residual = x
            channel_num = int(self.channels/4)
            param_num = channel_num*self.channels
            conv_w = conv_w_all[:, (i*18)*param_num:(i+1)*18*param_num].view(x.shape[2],channel_num,x.shape[1],1,18).flatten(0, 1)
            conv_b = conv_b_all[:, i*4*channel_num:(i+1)*4*channel_num].view(x.shape[2], channel_num*4).flatten(0, 1)
            gate_w = gate_w_all[:, (i*18)*param_num:(i+1)*18*param_num].view(x.shape[2],channel_num,x.shape[1],1,18).flatten(0, 1)
            gate_b = gate_b_all[:, i*4*channel_num:(i+1)*4*channel_num].view(x.shape[2], channel_num*4).flatten(0, 1)
            filter = torch.zeros(x.shape).to(self.device)
            gate = torch.zeros(x.shape).to(self.device)
            shape = x.shape
            group_num = x.shape[2]
            x = x.transpose(1,2).reshape(x.shape[0], x.shape[1]*x.shape[2], 1, x.shape[3])

            filter[:,0:channel_num] = F.conv2d(F.pad(x, (0,1,0,0), mode='circular'),conv_w[...,:2],conv_b[:int(len(conv_b)/4)],groups=group_num).reshape((shape[0], shape[2], channel_num, shape[3])).transpose(1,2)
            filter[:,channel_num:2*channel_num] = F.conv2d(F.pad(x, (0,2,0,0), mode='circular'),conv_w[...,2:5],conv_b[int(len(conv_b)/4):int(len(conv_b)/4)*2],groups=group_num).reshape((shape[0], shape[2], channel_num, shape[3])).transpose(1,2)
            filter[:,2*channel_num:3*channel_num] = F.conv2d(F.pad(x, (0,5,0,0), mode='circular'),conv_w[...,5:11],conv_b[int(len(conv_b)/4)*2:int(len(conv_b)/4)*3],groups=group_num).reshape((shape[0], shape[2], channel_num, shape[3])).transpose(1,2)
            filter[:,3*channel_num:4*channel_num] = F.conv2d(F.pad(x, (0,6,0,0), mode='circular'),conv_w[...,11:],conv_b[int(len(conv_b)/4)*3:],groups=group_num).reshape((shape[0], shape[2], channel_num, shape[3])).transpose(1,2)
            filter = torch.tanh(filter)

            gate[:,0:channel_num] = F.conv2d(F.pad(x, (0,1,0,0), mode='circular'),gate_w[...,:2],gate_b[:int(len(conv_b)/4)],groups=group_num).reshape((shape[0], shape[2], channel_num, shape[3])).transpose(1,2)
            gate[:,channel_num:2*channel_num] = F.conv2d(F.pad(x, (0,2,0,0), mode='circular'),gate_w[...,2:5],gate_b[int(len(conv_b)/4):int(len(conv_b)/4)*2],groups=group_num).reshape((shape[0], shape[2], channel_num, shape[3])).transpose(1,2)
            gate[:,2*channel_num:3*channel_num] = F.conv2d(F.pad(x, (0,5,0,0), mode='circular'),gate_w[...,5:11],gate_b[int(len(conv_b)/4)*2:int(len(conv_b)/4)*3],groups=group_num).reshape((shape[0], shape[2], channel_num, shape[3])).transpose(1,2)
            gate[:,3*channel_num:4*channel_num] = F.conv2d(F.pad(x, (0,6,0,0), mode='circular'),gate_w[...,11:],gate_b[int(len(conv_b)/4)*3:],groups=group_num).reshape((shape[0], shape[2], channel_num, shape[3])).transpose(1,2)
            gate = torch.sigmoid(gate)

            x = filter*gate
            x = F.dropout(x, self.dropout, training=self.training)
            s = self.skip_convs[i](x)
            # skip connection
            skip = s+skip
            param_num = int(self.res_w.weight.shape[0]/self.layers)
            res_w = res_w_all[:, i*param_num:(i+1)*param_num].view(x.shape[2],x.shape[1],x.shape[1],1,1).flatten(0, 1)
            res_b = res_b_all[:, i*x.shape[1]:(i+1)*x.shape[1]].view(x.shape[2], x.shape[1]).flatten(0, 1)
            x = F.conv2d(x.transpose(1,2).reshape(x.shape[0], x.shape[1]*x.shape[2], 1, x.shape[3]),res_w,res_b,groups=x.shape[2]).reshape(x.shape[0], x.shape[2], x.shape[1], x.shape[3]).transpose(1,2)
            # residual connection
            x = x + residual

            x = self.norm[i](x)

        skip = F.relu(self.skipE(x)+skip)
        x = F.relu(self.end_conv_1(skip))
        x = self.end_conv_2(x)
        return x