-
Notifications
You must be signed in to change notification settings - Fork 142
/
Ingram.py
253 lines (195 loc) · 11.4 KB
/
Ingram.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
from . import BaseModel, register_model
import torch
import torch.nn as nn
@register_model('Ingram')
class Ingram(BaseModel):
@classmethod
def build_model_from_args(cls, config):
return cls(config)
def __init__(self, config):
super().__init__()
self.model = Model(config)
def forward(self, *args):
return self.model(*args)
def extra_loss(self):
pass
class Model(nn.Module):
def __init__(self, args):
super().__init__()
self.args = args
layers_ent = []
layers_rel = []
layer_dim_ent = self.args.hdr_e * self.args.d_e
layer_dim_rel = self.args.hdr_r * self.args.d_r
num_layer_ent = self.args.nle
num_head = self.args.num_head
num_layer_rel = self.args.nlr
num_bin = self.args.num_bin
dim_ent = self.args.d_e
dim_rel = self.args.d_r
bias = True
for _ in range(num_layer_ent):
layers_ent.append(InGramEntityLayer(layer_dim_ent, layer_dim_ent, layer_dim_rel, \
bias=bias, num_head=num_head))
for _ in range(num_layer_rel):
layers_rel.append(InGramRelationLayer(layer_dim_rel, layer_dim_rel, num_bin, \
bias=bias, num_head=num_head))
res_proj_ent = []
for _ in range(num_layer_ent):
res_proj_ent.append(nn.Linear(layer_dim_ent, layer_dim_ent, bias=bias))
res_proj_rel = []
for _ in range(num_layer_rel):
res_proj_rel.append(nn.Linear(layer_dim_rel, layer_dim_rel, bias=bias))
self.res_proj_ent = nn.ModuleList(res_proj_ent)
self.res_proj_rel = nn.ModuleList(res_proj_rel)
self.bias = bias
self.ent_proj1 = nn.Linear(dim_ent, layer_dim_ent, bias=bias)
self.ent_proj2 = nn.Linear(layer_dim_ent, dim_ent, bias=bias)
self.layers_ent = nn.ModuleList(layers_ent)
self.layers_rel = nn.ModuleList(layers_rel)
self.rel_proj1 = nn.Linear(dim_rel, layer_dim_rel, bias=bias)
self.rel_proj2 = nn.Linear(layer_dim_rel, dim_rel, bias=bias)
self.rel_proj = nn.Linear(dim_rel, dim_ent, bias=bias)
self.num_layer_ent = num_layer_ent
self.num_layer_rel = num_layer_rel
self.act = nn.ReLU()
self.param_init()
def param_init(self):
nn.init.xavier_normal_(self.ent_proj1.weight, gain=nn.init.calculate_gain('relu'))
nn.init.xavier_normal_(self.ent_proj2.weight, gain=nn.init.calculate_gain('relu'))
nn.init.xavier_normal_(self.rel_proj1.weight, gain=nn.init.calculate_gain('relu'))
nn.init.xavier_normal_(self.rel_proj2.weight, gain=nn.init.calculate_gain('relu'))
nn.init.xavier_normal_(self.rel_proj.weight, gain=nn.init.calculate_gain('relu'))
for layer_idx in range(self.num_layer_ent):
nn.init.xavier_normal_(self.res_proj_ent[layer_idx].weight, gain=nn.init.calculate_gain('relu'))
for layer_idx in range(self.num_layer_rel):
nn.init.xavier_normal_(self.res_proj_rel[layer_idx].weight, gain=nn.init.calculate_gain('relu'))
if self.bias:
nn.init.zeros_(self.ent_proj1.bias)
nn.init.zeros_(self.ent_proj2.bias)
nn.init.zeros_(self.rel_proj1.bias)
nn.init.zeros_(self.rel_proj2.bias)
nn.init.zeros_(self.rel_proj.bias)
for layer_idx in range(self.num_layer_ent):
nn.init.zeros_(self.res_proj_ent[layer_idx].bias)
for layer_idx in range(self.num_layer_rel):
nn.init.zeros_(self.res_proj_rel[layer_idx].bias)
def forward(self, emb_ent, emb_rel, triplets, relation_triplets):
layer_emb_ent = self.ent_proj1(emb_ent)
layer_emb_rel = self.rel_proj1(emb_rel)
for layer_idx, layer in enumerate(self.layers_rel):
layer_emb_rel = layer(layer_emb_rel, relation_triplets) + \
self.res_proj_rel[layer_idx](layer_emb_rel)
layer_emb_rel = self.act(layer_emb_rel)
for layer_idx, layer in enumerate(self.layers_ent):
layer_emb_ent = layer(layer_emb_ent, layer_emb_rel, triplets) + \
self.res_proj_ent[layer_idx](layer_emb_ent)
layer_emb_ent = self.act(layer_emb_ent)
return self.ent_proj2(layer_emb_ent), self.rel_proj2(layer_emb_rel)
def score(self, emb_ent, emb_rel, triplets):
head_idxs = triplets[..., 0]
rel_idxs = triplets[..., 1]
tail_idxs = triplets[..., 2]
head_embs = emb_ent[head_idxs]
tail_embs = emb_ent[tail_idxs]
rel_embs = self.rel_proj(emb_rel[rel_idxs])
output = (head_embs * rel_embs * tail_embs).sum(dim=-1)
return output
class InGramEntityLayer(nn.Module):
def __init__(self, dim_in_ent, dim_out_ent, dim_rel, bias=True, num_head=8):
super(InGramEntityLayer, self).__init__()
self.dim_out_ent = dim_out_ent
self.dim_hid_ent = dim_out_ent // num_head
assert dim_out_ent == self.dim_hid_ent * num_head
self.num_head = num_head
self.attn_proj = nn.Linear(2 * dim_in_ent + dim_rel, dim_out_ent, bias=bias)
self.attn_vec = nn.Parameter(torch.zeros((1, num_head, self.dim_hid_ent)))
self.aggr_proj = nn.Linear(dim_in_ent + dim_rel, dim_out_ent, bias=bias)
self.dim_rel = dim_rel
self.act = nn.LeakyReLU(negative_slope=0.2)
self.bias = bias
self.param_init()
def param_init(self):
nn.init.xavier_normal_(self.attn_proj.weight, gain=nn.init.calculate_gain('relu'))
nn.init.xavier_normal_(self.attn_vec, gain=nn.init.calculate_gain('relu'))
nn.init.xavier_normal_(self.aggr_proj.weight, gain=nn.init.calculate_gain('relu'))
if self.bias:
nn.init.zeros_(self.attn_proj.bias)
nn.init.zeros_(self.aggr_proj.bias)
def forward(self, emb_ent, emb_rel, triplets):
num_ent = len(emb_ent)
num_rel = len(emb_rel)
head_idxs = triplets[..., 0]
rel_idxs = triplets[..., 1]
tail_idxs = triplets[..., 2]
ent_freq = torch.zeros((num_ent,)).cuda().index_add(dim=0, index=tail_idxs, \
source=torch.ones_like(tail_idxs,
dtype=torch.float).cuda()).unsqueeze(
dim=1)
self_rel = torch.zeros((num_ent, self.dim_rel)).cuda().index_add(dim=0, index=tail_idxs,
source=emb_rel[rel_idxs]) / ent_freq
# add self-loops
emb_rels = torch.cat([emb_rel[rel_idxs], self_rel], dim=0)
head_idxs = torch.cat([head_idxs, torch.arange(num_ent).cuda()], dim=0)
tail_idxs = torch.cat([tail_idxs, torch.arange(num_ent).cuda()], dim=0)
concat_mat_att = torch.cat([emb_ent[tail_idxs], emb_ent[head_idxs], \
emb_rels], dim=-1)
attn_val_raw = (self.act(self.attn_proj(concat_mat_att).view(-1, self.num_head, self.dim_hid_ent)) *
self.attn_vec).sum(dim=-1, keepdim=True)
scatter_idx = tail_idxs.unsqueeze(dim=-1).repeat(1, self.num_head).unsqueeze(dim=-1)
attn_val_max = torch.zeros((num_ent, self.num_head, 1)).cuda().scatter_reduce(dim=0, \
index=scatter_idx, \
src=attn_val_raw, reduce='amax', \
include_self=False)
attn_val = torch.exp(attn_val_raw - attn_val_max[tail_idxs])
attn_sums = torch.zeros((num_ent, self.num_head, 1)).cuda().index_add(dim=0, index=tail_idxs, source=attn_val)
beta = attn_val / (attn_sums[tail_idxs] + 1e-16)
concat_mat = torch.cat([emb_ent[head_idxs], emb_rels], dim=-1)
aggr_val = beta * self.aggr_proj(concat_mat).view(-1, self.num_head, self.dim_hid_ent)
output = torch.zeros((num_ent, self.num_head, self.dim_hid_ent)).cuda().index_add(dim=0, index=tail_idxs,
source=aggr_val)
return output.flatten(1, -1)
class InGramRelationLayer(nn.Module):
def __init__(self, dim_in_rel, dim_out_rel, num_bin, bias=True, num_head=8):
super(InGramRelationLayer, self).__init__()
self.dim_out_rel = dim_out_rel
self.dim_hid_rel = dim_out_rel // num_head
assert dim_out_rel == self.dim_hid_rel * num_head
self.attn_proj = nn.Linear(2 * dim_in_rel, dim_out_rel, bias=bias)
self.attn_bin = nn.Parameter(torch.zeros(num_bin, num_head, 1))
self.attn_vec = nn.Parameter(torch.zeros(1, num_head, self.dim_hid_rel))
self.aggr_proj = nn.Linear(dim_in_rel, dim_out_rel, bias=bias)
self.num_head = num_head
self.act = nn.LeakyReLU(negative_slope=0.2)
self.num_bin = num_bin
self.bias = bias
self.param_init()
def param_init(self):
nn.init.xavier_normal_(self.attn_proj.weight, gain=nn.init.calculate_gain('relu'))
nn.init.xavier_normal_(self.attn_vec, gain=nn.init.calculate_gain('relu'))
nn.init.xavier_normal_(self.aggr_proj.weight, gain=nn.init.calculate_gain('relu'))
if self.bias:
nn.init.zeros_(self.attn_proj.bias)
nn.init.zeros_(self.aggr_proj.bias)
def forward(self, emb_rel, relation_triplets):
num_rel = len(emb_rel)
head_idxs = relation_triplets[..., 0]
tail_idxs = relation_triplets[..., 1]
concat_mat = torch.cat([emb_rel[head_idxs], emb_rel[tail_idxs]], dim=-1)
attn_val_raw = (self.act(self.attn_proj(concat_mat).view(-1, self.num_head, self.dim_hid_rel)) * \
self.attn_vec).sum(dim=-1, keepdim=True) + self.attn_bin[relation_triplets[..., 2]]
scatter_idx = head_idxs.unsqueeze(dim=-1).repeat(1, self.num_head).unsqueeze(dim=-1)
attn_val_max = torch.zeros((num_rel, self.num_head, 1)).cuda().scatter_reduce(dim=0, \
index=scatter_idx, \
src=attn_val_raw, reduce='amax', \
include_self=False)
attn_val = torch.exp(attn_val_raw - attn_val_max[head_idxs])
attn_sums = torch.zeros((num_rel, self.num_head, 1)).cuda().index_add(dim=0, index=head_idxs, source=attn_val)
beta = attn_val / (attn_sums[head_idxs] + 1e-16)
output = torch.zeros((num_rel, self.num_head, self.dim_hid_rel)).cuda().index_add(dim=0, \
index=head_idxs,
source=beta * self.aggr_proj(
emb_rel[tail_idxs]).view(
-1, self.num_head,
self.dim_hid_rel))
return output.flatten(1, -1)