MotifRGC/models.py at main · RiemannGraph/MotifRGC · 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
167
168
169
170
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from geoopt.manifolds.stereographic.math import project
from geoopt.manifolds.stereographic import StereographicExact
from geoopt import ManifoldTensor
from geoopt import ManifoldParameter
from backbone import GCN, GAT, GraphSAGE


EPS = 1e-5


class RiemannianFeatures(nn.Module):
    def __init__(self, num_nodes, dimensions, init_curvature, num_factors, learnable=True):
        super(RiemannianFeatures, self).__init__()
        self.manifolds = nn.ModuleList()
        self.features = nn.ParameterList()
        for i in range(num_factors):
            if isinstance(dimensions, list):
                d = dimensions[i]
            else:
                d = dimensions
            k = init_curvature * (torch.randn(1) + 1)
            manifold = StereographicExact(k=k, learnable=learnable)
            features = ManifoldParameter(ManifoldTensor(torch.empty(num_nodes, d), manifold=manifold))
            if k != 0:
                self.init_weights(features)
            self.manifolds.append(manifold)
            self.features.append(features)

    @staticmethod
    def init_weights(w, scale=1e-4):
        w.data.uniform_(-scale, scale)
        w_norm = w.data.norm(p=2, dim=-1, keepdim=True) + EPS
        w.data = w.data / w_norm * w.manifold.radius * 0.9 * torch.rand(1)

    @staticmethod
    def normalize(x, manifold):
        x_norm = x.norm(p=2, dim=-1, keepdim=True) + EPS
        if manifold.k != 0:
            x = x / x_norm * 0.9 * torch.rand(1).to(x.device) * manifold.radius
        else:
            x = x / x_norm
        return x

    def forward(self):
        products = []
        for manifold, features in zip(self.manifolds, self.features):
            products.append(project(features, k=manifold.k))
        return products


class Model(nn.Module):
    def __init__(self, backbone, n_layers, in_features, hidden_features, embed_features, n_heads, drop_edge, drop_node,
                 num_factors, dimensions, d_embeds, temperature, device=torch.device('cuda')):
        super(Model, self).__init__()
        d_embeds = dimensions
        if backbone == 'gcn':
            self.encoder = GCN(n_layers, in_features, hidden_features, embed_features, drop_edge, drop_node)
            self.encoder2 = GCN(n_layers, d_embeds, hidden_features, embed_features, drop_edge, drop_node)
        elif backbone == 'gat':
            self.encoder = GAT(n_layers, in_features, hidden_features, embed_features, n_heads, drop_edge, drop_node)
            self.encoder2 = GAT(n_layers, d_embeds, hidden_features, embed_features, n_heads, drop_edge, drop_node)
        elif backbone == 'sage':
            self.encoder = GraphSAGE(n_layers, in_features, hidden_features, embed_features, drop_edge, drop_node)
            self.encoder2 = GraphSAGE(n_layers, d_embeds, hidden_features, embed_features, drop_edge, drop_node)
        else:
            raise NotImplementedError

        self.temperature = temperature
        self.Ws = []
        self.bias = []
        for i in range(num_factors):
            if isinstance(dimensions, list):
                d = dimensions[i]
            else:
                d = dimensions
            pre = torch.randn(d_embeds, d).to(device)
            w = pre / (torch.norm(pre, dim=-1, keepdim=True) + EPS)
            self.Ws.append(w)
            self.bias.append(2 * torch.pi * torch.rand(d_embeds).to(device))

        self.decoder = FermiDiracDecoder(2, 1)
        self.norm = nn.LayerNorm((num_factors+1) * embed_features)

    def forward(self, x, edge_index, motif, neg_motif, rm_features: RiemannianFeatures):
        products = rm_features()
        x = self.encoder(x, edge_index)
        laplacian = self.random_mapping(rm_features.manifolds, products)
        cl_loss = 0
        embeds = []
        for embed in laplacian:
            embed = self.encoder2(embed, edge_index)
            cl_loss = cl_loss + self.cal_cl_loss(x, embed)
            embeds.append(embed)
        loss = cl_loss / len(laplacian) + self.cal_motif_loss(laplacian, motif, neg_motif)
        embeds = torch.concat(embeds, -1)
        return self.norm(torch.concat([x, embeds], -1)), loss

    def random_mapping(self, manifolds, products):
        out = []
        for i in range(len(manifolds)):
            x = products[i]
            w = self.Ws[i]
            b = self.bias[i]
            k = manifolds[i].k
            if k == 0:
                distance = x @ w.t()
            else:
                div = torch.sum((x[:, None] - w[None]) ** 2, dim=-1)
                distance = torch.log((1 + k * torch.sum(x * x, -1, keepdim=True)) / (div + EPS) + EPS)
            n = x.shape[-1]
            z = torch.exp((n - 1) * distance / 2) * torch.cos(distance + b)
            out.append(z)
        return out

    def cal_cl_loss(self, x1, x2):
        norm1 = x1.norm(dim=-1)
        norm2 = x2.norm(dim=-1)
        sim_matrix = torch.einsum('ik,jk->ij', x1, x2) / (torch.einsum('i,j->ij', norm1, norm2) + EPS)
        sim_matrix = torch.exp(sim_matrix / self.temperature)
        pos_sim = sim_matrix.diag()
        loss_1 = pos_sim / (sim_matrix.sum(dim=-2) - pos_sim + EPS)
        loss_2 = pos_sim / (sim_matrix.sum(dim=-1) - pos_sim + EPS)

        loss_1 = -torch.log(loss_1).mean()
        loss_2 = -torch.log(loss_2).mean()
        loss = (loss_1 + loss_2) / 2.
        return loss

    def cal_motif_loss(self, products, pos_motifs, neg_motifs):
        embeddings = torch.concat(products, dim=-1)
        pos_scores1 = self.decoder(torch.sum((embeddings[pos_motifs[0]] - embeddings[pos_motifs[1]])**2, -1))
        pos_scores2 = self.decoder(torch.sum((embeddings[pos_motifs[2]] - embeddings[pos_motifs[1]])**2, -1))
        pos_scores3 = self.decoder(torch.sum((embeddings[pos_motifs[2]] - embeddings[pos_motifs[0]])**2, -1))

        neg_scores1 = self.decoder(torch.sum((embeddings[neg_motifs[0]] - embeddings[neg_motifs[1]])**2, -1))
        neg_scores2 = self.decoder(torch.sum((embeddings[neg_motifs[2]] - embeddings[neg_motifs[1]])**2, -1))
        neg_scores3 = self.decoder(torch.sum((embeddings[neg_motifs[2]] - embeddings[neg_motifs[0]])**2, -1))

        pos1 = pos_scores1 * pos_scores2 * (1 - pos_scores3)
        pos2 = pos_scores1 * pos_scores2 * pos_scores3
        pos0 = 1 - pos1 - pos2
        pos = torch.stack([pos0, pos1, pos2], dim=1)
        p_y = pos_motifs[-1].detach() + 1

        neg1 = neg_scores1 * neg_scores2 * (1 - neg_scores3)
        neg2 = neg_scores1 * neg_scores2 * neg_scores3
        neg0 = 1 - neg1 - neg2
        neg = torch.stack([neg0, neg1, neg2], dim=1)
        n_y = torch.zeros_like(p_y)

        probs = torch.concat([pos, neg], dim=0)
        label = torch.concat([p_y, n_y])

        loss = F.nll_loss(torch.log(probs + 1e-5), label)
        return loss


class FermiDiracDecoder(nn.Module):
    def __init__(self, r, t):
        super(FermiDiracDecoder, self).__init__()
        self.r = r
        self.t = t

    def forward(self, dist):
        probs = torch.sigmoid((self.r - dist) / self.t)
        return probs