-
Notifications
You must be signed in to change notification settings - Fork 5
/
model.py
180 lines (136 loc) · 6.85 KB
/
model.py
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
171
172
173
174
175
176
177
178
179
180
from torch.nn import Sequential, Linear, ReLU
from torch_geometric.nn import GINConv, global_add_pool, GCNConv, global_mean_pool
import torch
import torch.nn.functional as F
import torch.nn as nn
class HCL(nn.Module):
def __init__(self, hidden_dim, num_gc_layers, feat_dim, str_dim):
super(HCL, self).__init__()
self.embedding_dim = hidden_dim * num_gc_layers
self.encoder_feat = Encoder_GIN(feat_dim, hidden_dim, num_gc_layers)
self.encoder_str = Encoder_GIN(str_dim, hidden_dim, num_gc_layers)
self.proj_head_feat_g = nn.Sequential(nn.Linear(self.embedding_dim, self.embedding_dim), nn.ReLU(inplace=True),
nn.Linear(self.embedding_dim, self.embedding_dim))
self.proj_head_str_g = nn.Sequential(nn.Linear(self.embedding_dim, self.embedding_dim), nn.ReLU(inplace=True),
nn.Linear(self.embedding_dim, self.embedding_dim))
self.proj_head_feat_n = nn.Sequential(nn.Linear(self.embedding_dim, self.embedding_dim), nn.ReLU(inplace=True),
nn.Linear(self.embedding_dim, self.embedding_dim))
self.proj_head_str_n = nn.Sequential(nn.Linear(self.embedding_dim, self.embedding_dim), nn.ReLU(inplace=True),
nn.Linear(self.embedding_dim, self.embedding_dim))
self.proj_head_b = nn.Sequential(nn.Linear(self.embedding_dim * 2, self.embedding_dim), nn.ReLU(inplace=True),
nn.Linear(self.embedding_dim, self.embedding_dim))
self.init_emb()
def init_emb(self):
initrange = -1.5 / self.embedding_dim
for m in self.modules():
if isinstance(m, nn.Linear):
torch.nn.init.xavier_uniform_(m.weight.data)
if m.bias is not None:
m.bias.data.fill_(0.0)
def get_b(self, x_f, x_s, edge_index, batch, num_graphs):
g_f, _ = self.encoder_feat(x_f, edge_index, batch)
g_s, _ = self.encoder_str(x_s, edge_index, batch)
b = self.proj_head_b(torch.cat((g_f, g_s), 1))
return b
def forward(self, x_f, x_s, edge_index, batch, num_graphs):
g_f, n_f = self.encoder_feat(x_f, edge_index, batch)
g_s, n_s = self.encoder_str(x_s, edge_index, batch)
b = self.proj_head_b(torch.cat((g_f, g_s), 1))
g_f = self.proj_head_feat_g(g_f)
g_s = self.proj_head_str_g(g_s)
n_f = self.proj_head_feat_n(n_f)
n_s = self.proj_head_str_n(n_s)
return b, g_f, g_s, n_f, n_s
@staticmethod
def scoring_b(b, cluster_result, temperature = 0.2):
im2cluster, prototypes, density = cluster_result['im2cluster'], cluster_result['centroids'], cluster_result['density']
batch_size, _ = b.size()
b_abs = b.norm(dim=1)
prototypes_abs = prototypes.norm(dim=1)
sim_matrix = torch.einsum('ik,jk->ij', b, prototypes) / torch.einsum('i,j->ij', b_abs, prototypes_abs)
sim_matrix = torch.exp(sim_matrix / (temperature * density))
v, id = torch.min(sim_matrix, 1)
return v
@staticmethod
def calc_loss_b(b, index, cluster_result, temperature=0.2):
im2cluster, prototypes, density = cluster_result['im2cluster'], cluster_result['centroids'], cluster_result['density']
pos_proto_id = im2cluster[index].cpu().tolist()
batch_size, _ = b.size()
b_abs = b.norm(dim=1)
prototypes_abs = prototypes.norm(dim=1)
sim_matrix = torch.einsum('ik,jk->ij', b, prototypes) / torch.einsum('i,j->ij', b_abs, prototypes_abs)
sim_matrix = torch.exp(sim_matrix / (temperature * density))
pos_sim = sim_matrix[range(batch_size), pos_proto_id]
loss = pos_sim / (sim_matrix.sum(dim=1) - pos_sim)
loss = - torch.log(loss + 1e-12)
return loss
@staticmethod
def calc_loss_n(x, x_aug, batch, temperature=0.2):
batch_size, _ = x.size()
x_abs = x.norm(dim=1)
x_aug_abs = x_aug.norm(dim=1)
node_belonging_mask = batch.repeat(batch_size,1)
node_belonging_mask = node_belonging_mask == node_belonging_mask.t()
sim_matrix = torch.einsum('ik,jk->ij', x, x_aug) / torch.einsum('i,j->ij', x_abs, x_aug_abs)
sim_matrix = torch.exp(sim_matrix / temperature) * node_belonging_mask
pos_sim = sim_matrix[range(batch_size), range(batch_size)]
loss_0 = pos_sim / (sim_matrix.sum(dim=0) - pos_sim + 1e-12)
loss_1 = pos_sim / (sim_matrix.sum(dim=1) - pos_sim + 1e-12)
loss_0 = - torch.log(loss_0)
loss_1 = - torch.log(loss_1)
loss = (loss_0 + loss_1) / 2.0
loss = global_mean_pool(loss, batch)
return loss
@staticmethod
def calc_loss_g(x, x_aug, temperature=0.2):
batch_size, _ = x.size()
x_abs = x.norm(dim=1)
x_aug_abs = x_aug.norm(dim=1)
sim_matrix = torch.einsum('ik,jk->ij', x, x_aug) / torch.einsum('i,j->ij', x_abs, x_aug_abs)
sim_matrix = torch.exp(sim_matrix / temperature)
pos_sim = sim_matrix[range(batch_size), range(batch_size)]
loss_0 = pos_sim / (sim_matrix.sum(dim=0) - pos_sim)
loss_1 = pos_sim / (sim_matrix.sum(dim=1) - pos_sim)
loss_0 = - torch.log(loss_0)
loss_1 = - torch.log(loss_1)
loss = (loss_0 + loss_1) / 2.0
return loss
class Encoder_GIN(torch.nn.Module):
def __init__(self, num_features, dim, num_gc_layers):
super(Encoder_GIN, self).__init__()
self.num_gc_layers = num_gc_layers
self.convs = torch.nn.ModuleList()
for i in range(num_gc_layers):
if i:
nn = Sequential(Linear(dim, dim), ReLU(), Linear(dim, dim))
else:
nn = Sequential(Linear(num_features, dim), ReLU(), Linear(dim, dim))
conv = GINConv(nn)
self.convs.append(conv)
def forward(self, x, edge_index, batch):
xs = []
for i in range(self.num_gc_layers):
x = F.relu(self.convs[i](x, edge_index))
xs.append(x)
xpool = [global_add_pool(x, batch) for x in xs]
x = torch.cat(xpool, 1)
return x, torch.cat(xs, 1)
class Encoder_GCN(torch.nn.Module):
def __init__(self, num_features, dim, num_gc_layers):
super(Encoder_GCN, self).__init__()
self.num_gc_layers = num_gc_layers
self.convs = torch.nn.ModuleList()
for i in range(num_gc_layers):
if i:
conv = GCNConv(dim, dim)
else:
conv = GCNConv(num_features, dim)
self.convs.append(conv)
def forward(self, x, edge_index, batch):
xs = []
for i in range(self.num_gc_layers):
x = F.relu(self.convs[i](x, edge_index))
xs.append(x)
xpool = [global_mean_pool(x, batch) for x in xs]
x = torch.cat(xpool, 1)
return x, torch.cat(xs, 1)