-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathmodels.py
217 lines (178 loc) · 9.49 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
from abc import ABC, abstractmethod
from typing import Tuple
import torch
from torch import nn
import numpy as np
class TKBCModel(nn.Module, ABC):
@abstractmethod
def get_rhs(self, chunk_begin: int, chunk_size: int):
pass
@abstractmethod
def get_queries(self, queries: torch.Tensor):
pass
@abstractmethod
def score(self, x: torch.Tensor):
pass
def get_ranking(
self, queries, filters, year2id = {},
batch_size: int = 1000, chunk_size: int = -1
):
"""
Returns filtered ranking for each queries.
:param queries: a torch.LongTensor of quadruples (lhs, rel, rhs, timestam)
:param filters: filters[(lhs, rel, ts)] gives the elements to filter from ranking
:param batch_size: maximum number of queries processed at once
:param chunk_size: maximum number of candidates processed at once
:return:
"""
if chunk_size < 0:
chunk_size = self.sizes[2]
ranks = torch.ones(len(queries))
with torch.no_grad():
c_begin = 0
while c_begin < self.sizes[2]:
b_begin = 0
rhs = self.get_rhs(c_begin, chunk_size) # 将输入进来的训练集分为几个batch
while b_begin < len(queries):
if queries.shape[1] > 4: # time intervals exist 对五元组中时间戳的处理
these_queries = queries[b_begin:b_begin + batch_size]
start_queries = []
end_queries = []
for triple in these_queries:
if triple[3].split('-')[0] == '####':
start_idx = -1
start = -5000
elif triple[3][0] == '-':
start = -int(triple[3].split('-')[1].replace('#', '0'))
elif triple[3][0] != '-':
start = int(triple[3].split('-')[0].replace('#','0'))
if triple[4].split('-')[0] == '####':
end_idx = -1
end = 5000
elif triple[4][0] == '-':
end =-int(triple[4].split('-')[1].replace('#', '0'))
elif triple[4][0] != '-':
end = int(triple[4].split('-')[0].replace('#','0'))
for key, time_idx in sorted(year2id.items(), key=lambda x:x[1]): # 时间戳转换成id
if start>=key[0] and start<=key[1]:
start_idx = time_idx
if end>=key[0] and end<=key[1]:
end_idx = time_idx
if start_idx < 0:
start_queries.append([int(triple[0]), int(triple[1])+self.sizes[1]//4, int(triple[2]), end_idx])
else:
start_queries.append([int(triple[0]), int(triple[1]), int(triple[2]), start_idx])
if end_idx < 0:
end_queries.append([int(triple[0]), int(triple[1]), int(triple[2]), start_idx])
else:
end_queries.append([int(triple[0]), int(triple[1])+self.sizes[1]//4, int(triple[2]), end_idx])
start_queries = torch.from_numpy(np.array(start_queries).astype('int64')).cuda()
end_queries = torch.from_numpy(np.array(end_queries).astype('int64')).cuda()
q_s = self.get_queries(start_queries)
q_e = self.get_queries(end_queries)
scores = q_s @ rhs + q_e @ rhs
targets = self.score(start_queries)+self.score(end_queries)
else:
these_queries = queries[b_begin:b_begin + batch_size] # 500, 4
q = self.get_queries(these_queries) # 500, 400
"""
if use_left_queries:
lhs_queries = torch.ones(these_queries.size()).long().cuda()
lhs_queries[:,1] = (these_queries[:,1]+self.sizes[1]//2)%self.sizes[1]
lhs_queries[:,0] = these_queries[:,2]
lhs_queries[:,2] = these_queries[:,0]
lhs_queries[:,3] = these_queries[:,3]
q_lhs = self.get_lhs_queries(lhs_queries)
scores = q @ rhs + q_lhs @ rhs
targets = self.score(these_queries) + self.score(lhs_queries)
"""
scores = q @ rhs
targets = self.score(these_queries)
assert not torch.any(torch.isinf(scores)), "inf scores"
assert not torch.any(torch.isnan(scores)), "nan scores"
assert not torch.any(torch.isinf(targets)), "inf targets"
assert not torch.any(torch.isnan(targets)), "nan targets"
# set filtered and true scores to -1e6 to be ignored
# take care that scores are chunked
for i, query in enumerate(these_queries):
if queries.shape[1]>4:
filter_out = filters[int(query[0]), int(query[1]), query[3], query[4]]
filter_out += [int(queries[b_begin + i, 2])]
else:
filter_out = filters[(query[0].item(), query[1].item(), query[3].item())]
filter_out += [queries[b_begin + i, 2].item()]
if chunk_size < self.sizes[2]:
filter_in_chunk = [
int(x - c_begin) for x in filter_out
if c_begin <= x < c_begin + chunk_size
]
scores[i, torch.LongTensor(filter_in_chunk)] = -1e6
else:
scores[i, torch.LongTensor(filter_out)] = -1e6
ranks[b_begin:b_begin + batch_size] += torch.sum(
(scores >= targets).float(), dim=1
).cpu()
b_begin += batch_size
c_begin += chunk_size
return ranks
class TCompoundE(TKBCModel):
def __init__(self, sizes: Tuple[int, int, int, int], rank: int,no_time_emb=False, init_size: float = 1e-2):
super(TCompoundE, self).__init__()
self.sizes = sizes
self.rank = rank
self.W = nn.Embedding(2*rank, 1, sparse=True)
self.W.weight.data *= 0
self.embeddings = nn.ModuleList([
nn.Embedding(s, 2 * rank, sparse=True)
for s in [sizes[0], sizes[1], sizes[3]] # without no_time_emb
])
self.embeddings[0].weight.data *= init_size
self.embeddings[1].weight.data *= init_size
self.embeddings[2].weight.data *= init_size
self.no_time_emb = no_time_emb
self.pi = 3.14159265358979323846
@staticmethod
def has_time():
return True
def score(self, x):
lhs = self.embeddings[0](x[:, 0])
rel = self.embeddings[1](x[:, 1])
rhs = self.embeddings[0](x[:, 2])
time = self.embeddings[2](x[:, 3])
lhs = lhs[:, :self.rank], lhs[:, self.rank:]
rel = rel[:, :self.rank] / ( 1 / self.pi), rel[:, self.rank:] / ( 1 / self.pi)
rhs = rhs[:, :self.rank], rhs[:, self.rank:]
time = time[:, :self.rank], time[:, self.rank:]
rt = (rel[0] + time[0]) * time[1], rel[1]
return torch.sum(
( (lhs[0] + rt[1]) * rt[0] ) * rhs[0], 1, keepdim=True)
def forward(self, x):
lhs = self.embeddings[0](x[:, 0])
rel = self.embeddings[1](x[:, 1])
rhs = self.embeddings[0](x[:, 2])
time = self.embeddings[2](x[:, 3])
lhs = lhs[:, :self.rank], lhs[:, self.rank:]
rhs = rhs[:, :self.rank], rhs[:, self.rank:]
rel = rel[:, :self.rank] / ( 1 / self.pi), rel[:, self.rank:] / ( 1 / self.pi)
time = time[:, :self.rank], time[:, self.rank:]
right = self.embeddings[0].weight
right = right[:, :self.rank], right[:, self.rank:]
rt = (rel[0] + time[0]) * time[1], rel[1]
return (
((lhs[0] + rt[1]) * rt[0] ) @ right[0].t()
), (
torch.sqrt(lhs[0] ** 2),
torch.sqrt(rt[0] ** 2 + rt[1] ** 2),
torch.sqrt(rhs[0] ** 2)
), self.embeddings[2].weight[:-1] if self.no_time_emb else self.embeddings[2].weight
def get_rhs(self, chunk_begin: int, chunk_size: int):
return self.embeddings[0].weight.data[chunk_begin:chunk_begin + chunk_size][:, :self.rank].transpose(0, 1)
def get_queries(self, queries: torch.Tensor):
lhs = self.embeddings[0](queries[:, 0])
rel = self.embeddings[1](queries[:, 1])
time = self.embeddings[2](queries[:, 3])
lhs = lhs[:, :self.rank], lhs[:, self.rank:]
rel = rel[:, :self.rank] / ( 1 / self.pi), rel[:, self.rank:] / ( 1 / self.pi)
time = time[:, :self.rank], time[:, self.rank:]
rt = (rel[0] + time[0]) * time[1], rel[1]
return (lhs[0] + rt[1]) * rt[0]