-
Notifications
You must be signed in to change notification settings - Fork 8
/
models.py
270 lines (231 loc) · 11.4 KB
/
models.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
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
'''
**************************************************************************
The models here are designed to work with datasets having maximum arity 6.
**************************************************************************
'''
import numpy as np
import torch
from torch.nn import functional as F, Parameter
from torch.nn.init import xavier_normal_, xavier_uniform_
import time
import math
class BaseClass(torch.nn.Module):
def __init__(self):
super(BaseClass, self).__init__()
self.cur_itr = torch.nn.Parameter(torch.tensor(0, dtype=torch.int32), requires_grad=False)
self.best_mrr = torch.nn.Parameter(torch.tensor(0, dtype=torch.float64), requires_grad=False)
self.best_itr = torch.nn.Parameter(torch.tensor(0, dtype=torch.int32), requires_grad=False)
class MDistMult(BaseClass):
def __init__(self, dataset, emb_dim, **kwargs):
super(MDistMult, self).__init__()
self.emb_dim = emb_dim
self.E = torch.nn.Embedding(dataset.num_ent(), emb_dim, padding_idx=0)
self.R = torch.nn.Embedding(dataset.num_rel(), emb_dim, padding_idx=0)
self.hidden_drop_rate = kwargs["hidden_drop"]
self.hidden_drop = torch.nn.Dropout(self.hidden_drop_rate)
def init(self):
self.E.weight.data[0] = torch.ones(self.emb_dim)
self.R.weight.data[0] = torch.ones(self.emb_dim)
xavier_normal_(self.E.weight.data[1:])
xavier_normal_(self.R.weight.data[1:])
def forward(self, r_idx, e1_idx, e2_idx, e3_idx, e4_idx, e5_idx, e6_idx):
r = self.R(r_idx)
e1 = self.E(e1_idx)
e2 = self.E(e2_idx)
e3 = self.E(e3_idx)
e4 = self.E(e4_idx)
e5 = self.E(e5_idx)
e6 = self.E(e6_idx)
x = r * e1 * e2 * e3 * e4 * e5 * e6
x = self.hidden_drop(x)
x = torch.sum(x, dim=1)
return x
class MCP(BaseClass):
def __init__(self, dataset, emb_dim, **kwargs):
super(MCP, self).__init__()
self.emb_dim = emb_dim
self.E1 = torch.nn.Embedding(dataset.num_ent(), emb_dim, padding_idx=0)
self.E2 = torch.nn.Embedding(dataset.num_ent(), emb_dim, padding_idx=0)
self.E3 = torch.nn.Embedding(dataset.num_ent(), emb_dim, padding_idx=0)
self.E4 = torch.nn.Embedding(dataset.num_ent(), emb_dim, padding_idx=0)
self.E5 = torch.nn.Embedding(dataset.num_ent(), emb_dim, padding_idx=0)
self.E6 = torch.nn.Embedding(dataset.num_ent(), emb_dim, padding_idx=0)
self.R = torch.nn.Embedding(dataset.num_rel(), emb_dim, padding_idx=0)
self.hidden_drop_rate = kwargs["hidden_drop"]
self.hidden_drop = torch.nn.Dropout(self.hidden_drop_rate)
def init(self):
self.E1.weight.data[0] = torch.ones(self.emb_dim)
self.E2.weight.data[0] = torch.ones(self.emb_dim)
self.E3.weight.data[0] = torch.ones(self.emb_dim)
self.E4.weight.data[0] = torch.ones(self.emb_dim)
self.E5.weight.data[0] = torch.ones(self.emb_dim)
self.E6.weight.data[0] = torch.ones(self.emb_dim)
xavier_normal_(self.E1.weight.data[1:])
xavier_normal_(self.E2.weight.data[1:])
xavier_normal_(self.E3.weight.data[1:])
xavier_normal_(self.E4.weight.data[1:])
xavier_normal_(self.E5.weight.data[1:])
xavier_normal_(self.E6.weight.data[1:])
xavier_normal_(self.R.weight.data)
def forward(self, r_idx, e1_idx, e2_idx, e3_idx, e4_idx, e5_idx, e6_idx):
r = self.R(r_idx)
e1 = self.E1(e1_idx)
e2 = self.E2(e2_idx)
e3 = self.E3(e3_idx)
e4 = self.E4(e4_idx)
e5 = self.E5(e5_idx)
e6 = self.E6(e6_idx)
x = r * e1 * e2 * e3 * e4 * e5 * e6
x = self.hidden_drop(x)
x = torch.sum(x, dim=1)
return x
class HSimplE(BaseClass):
def __init__(self, dataset, emb_dim, **kwargs):
super(HSimplE, self).__init__()
self.emb_dim = emb_dim
self.max_arity = 6
self.E = torch.nn.Embedding(dataset.num_ent(), emb_dim, padding_idx=0)
self.R = torch.nn.Embedding(dataset.num_rel(), emb_dim, padding_idx=0)
self.hidden_drop_rate = kwargs["hidden_drop"]
self.hidden_drop = torch.nn.Dropout(self.hidden_drop_rate)
def init(self):
self.E.weight.data[0] = torch.ones(self.emb_dim)
xavier_normal_(self.E.weight.data[1:])
xavier_normal_(self.R.weight.data)
def shift(self, v, sh):
y = torch.cat((v[:, sh:], v[:, :sh]), dim=1)
return y
def forward(self, r_idx, e1_idx, e2_idx, e3_idx, e4_idx, e5_idx, e6_idx):
r = self.R(r_idx)
e1 = self.E(e1_idx)
e2 = self.shift(self.E(e2_idx), int(1 * self.emb_dim/self.max_arity))
e3 = self.shift(self.E(e3_idx), int(2 * self.emb_dim/self.max_arity))
e4 = self.shift(self.E(e4_idx), int(3 * self.emb_dim/self.max_arity))
e5 = self.shift(self.E(e5_idx), int(4 * self.emb_dim/self.max_arity))
e6 = self.shift(self.E(e6_idx), int(5 * self.emb_dim/self.max_arity))
x = r * e1 * e2 * e3 * e4 * e5 * e6
x = self.hidden_drop(x)
x = torch.sum(x, dim=1)
return x
class HypE(BaseClass):
def __init__(self, d, emb_dim, **kwargs):
super(HypE, self).__init__()
self.cur_itr = torch.nn.Parameter(torch.tensor(0, dtype=torch.int32), requires_grad=False)
self.in_channels = kwargs["in_channels"]
self.out_channels = kwargs["out_channels"]
self.filt_h = kwargs["filt_h"]
self.filt_w = kwargs["filt_w"]
self.stride = kwargs["stride"]
self.hidden_drop_rate = kwargs["hidden_drop"]
self.emb_dim = emb_dim
self.max_arity = 6
rel_emb_dim = emb_dim
self.E = torch.nn.Embedding(d.num_ent(), emb_dim, padding_idx=0)
self.R = torch.nn.Embedding(d.num_rel(), rel_emb_dim, padding_idx=0)
self.bn0 = torch.nn.BatchNorm2d(self.in_channels)
self.inp_drop = torch.nn.Dropout(0.2)
fc_length = (1-self.filt_h+1)*math.floor((emb_dim-self.filt_w)/self.stride + 1)*self.out_channels
self.bn2 = torch.nn.BatchNorm1d(fc_length)
self.hidden_drop = torch.nn.Dropout(self.hidden_drop_rate)
# Projection network
self.fc = torch.nn.Linear(fc_length, emb_dim)
self.device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
# size of the convolution filters outputted by the hypernetwork
fc1_length = self.in_channels*self.out_channels*self.filt_h*self.filt_w
# Hypernetwork
self.fc1 = torch.nn.Linear(rel_emb_dim + self.max_arity + 1, fc1_length)
self.fc2 = torch.nn.Linear(self.max_arity + 1, fc1_length)
def init(self):
self.E.weight.data[0] = torch.ones(self.emb_dim)
self.R.weight.data[0] = torch.ones(self.emb_dim)
xavier_uniform_(self.E.weight.data[1:])
xavier_uniform_(self.R.weight.data[1:])
def convolve(self, r_idx, e_idx, pos):
e = self.E(e_idx).view(-1, 1, 1, self.E.weight.size(1))
r = self.R(r_idx)
x = e
x = self.inp_drop(x)
one_hot_target = (pos == torch.arange(self.max_arity + 1).reshape(self.max_arity + 1)).float().to(self.device)
poses = one_hot_target.repeat(r.shape[0]).view(-1, self.max_arity + 1)
one_hot_target.requires_grad = False
poses.requires_grad = False
k = self.fc2(poses)
k = k.view(-1, self.in_channels, self.out_channels, self.filt_h, self.filt_w)
k = k.view(e.size(0)*self.in_channels*self.out_channels, 1, self.filt_h, self.filt_w)
x = x.permute(1, 0, 2, 3)
x = F.conv2d(x, k, stride=self.stride, groups=e.size(0))
x = x.view(e.size(0), 1, self.out_channels, 1-self.filt_h+1, -1)
x = x.permute(0, 3, 4, 1, 2)
x = torch.sum(x, dim=3)
x = x.permute(0, 3, 1, 2).contiguous()
x = x.view(e.size(0), -1)
x = self.fc(x)
return x
def forward(self, r_idx, e1_idx, e2_idx, e3_idx, e4_idx, e5_idx, e6_idx, ms, bs):
r = self.R(r_idx)
e1 = self.convolve(r_idx, e1_idx, 0) * ms[:,0].view(-1, 1) + bs[:,0].view(-1, 1)
e2 = self.convolve(r_idx, e2_idx, 1) * ms[:,1].view(-1, 1) + bs[:,1].view(-1, 1)
e3 = self.convolve(r_idx, e3_idx, 2) * ms[:,2].view(-1, 1) + bs[:,2].view(-1, 1)
e4 = self.convolve(r_idx, e4_idx, 3) * ms[:,3].view(-1, 1) + bs[:,3].view(-1, 1)
e5 = self.convolve(r_idx, e5_idx, 4) * ms[:,4].view(-1, 1) + bs[:,4].view(-1, 1)
e6 = self.convolve(r_idx, e6_idx, 5) * ms[:,5].view(-1, 1) + bs[:,5].view(-1, 1)
x = e1 * e2 * e3 * e4 * e5 * e6 * r
x = self.hidden_drop(x)
x = torch.sum(x, dim=1)
return x
class MTransH(BaseClass):
def __init__(self, dataset, emb_dim, **kwargs):
super(MTransH, self).__init__()
self.emb_dim = emb_dim
self.E = torch.nn.Embedding(dataset.num_ent(), emb_dim, padding_idx=0)
self.R1 = torch.nn.Embedding(dataset.num_rel(), emb_dim, padding_idx=0)
self.R2 = torch.nn.Embedding(dataset.num_rel(), emb_dim, padding_idx=0)
self.b0 = torch.nn.Embedding(dataset.num_rel(), 1)
self.b1 = torch.nn.Embedding(dataset.num_rel(), 1)
self.b2 = torch.nn.Embedding(dataset.num_rel(), 1)
self.b3 = torch.nn.Embedding(dataset.num_rel(), 1)
self.b4 = torch.nn.Embedding(dataset.num_rel(), 1)
self.b5 = torch.nn.Embedding(dataset.num_rel(), 1)
self.hidden_drop_rate = kwargs["hidden_drop"]
self.hidden_drop = torch.nn.Dropout(self.hidden_drop_rate)
def init(self):
self.E.weight.data[0] = torch.ones(self.emb_dim)
self.R1.weight.data[0] = torch.ones(self.emb_dim)
self.R2.weight.data[0] = torch.ones(self.emb_dim)
xavier_normal_(self.E.weight.data[1:])
xavier_normal_(self.R1.weight.data[1:])
xavier_normal_(self.R2.weight.data[1:])
normalize_entity_emb = F.normalize(self.E.weight.data[1:], p=2, dim=1)
normalize_relation_emb = F.normalize(self.R1.weight.data[1:], p=2, dim=1)
normalize_norm_emb = F.normalize(self.R2.weight.data[1:], p=2, dim=1)
self.E.weight.data[1:] = normalize_entity_emb
self.R1.weight.data[1:] = normalize_relation_emb
self.R2.weight.data[1:] = normalize_norm_emb
xavier_normal_(self.b0.weight.data)
xavier_normal_(self.b1.weight.data)
xavier_normal_(self.b2.weight.data)
xavier_normal_(self.b3.weight.data)
xavier_normal_(self.b4.weight.data)
xavier_normal_(self.b5.weight.data)
def pnr(self, e_idx, r_idx):
original = self.E(e_idx)
norm = self.R2(r_idx)
return original - torch.sum(original * norm, dim=1, keepdim=True) * norm
def forward(self, r_idx, e1_idx, e2_idx, e3_idx, e4_idx, e5_idx, e6_idx, ms):
r = self.R1(r_idx)
e1 = self.pnr(e1_idx, r_idx) * self.b0(r_idx)
e1 = e1 * ms[:,0].unsqueeze(-1).expand_as(e1)
e2 = self.pnr(e2_idx, r_idx) * self.b1(r_idx)
e2 = e2 * ms[:,1].unsqueeze(-1).expand_as(e2)
e3 = self.pnr(e3_idx, r_idx) * self.b2(r_idx)
e3 = e3 * ms[:,2].unsqueeze(-1).expand_as(e3)
e4 = self.pnr(e4_idx, r_idx) * self.b3(r_idx)
e4 = e4 * ms[:,3].unsqueeze(-1).expand_as(e4)
e5 = self.pnr(e5_idx, r_idx) * self.b4(r_idx)
e5 = e5 * ms[:,4].unsqueeze(-1).expand_as(e5)
e6 = self.pnr(e6_idx, r_idx) * self.b5(r_idx)
e6 = e6 * ms[:,5].unsqueeze(-1).expand_as(e6)
x = r + e1 + e2 + e3 + e4 + e5 + e6
x = self.hidden_drop(x)
x = -1 * torch.norm(x, p=2, dim=1)
return x