-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
236 lines (193 loc) · 8.43 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
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
"""Metamodel implementation."""
import pickle
import random
import re
from typing import Sequence, Tuple
import numpy as np
import torch
import torch.nn as nn
from torch import Tensor
from data.graph import _reverse_relation
from gnn.gcn import GCNModel
from loader import QueryBatch, QueryTargetInfo, VectorizedQueryBatch
class MetaModel(torch.nn.Module):
def __init__(
self,
data_dir,
embed_dim,
num_bands=4,
num_hyperplanes=4,
gcn_layers=2,
gcn_readout='sum',
gcn_use_bias=True,
gcn_opn='corr',
gcn_dropout=0,
device=None
):
super().__init__()
self.data_dir = data_dir
self.embed_dim = embed_dim
self.num_bands = num_bands
self.num_hyperplanes = num_hyperplanes
self.gcn_layers = gcn_layers
self.gcn_readout = gcn_readout
self.gcn_use_bias = gcn_use_bias
self.gcn_opn = gcn_opn
self.gcn_dropout = gcn_dropout
self.device = device
# instantiate GCN models
self.submodels = nn.ModuleList([
GCNModel(
self.embed_dim,
self.gcn_layers,
self.gcn_readout,
self.gcn_use_bias,
self.gcn_opn,
self.gcn_dropout,
self.device
) for _ in range(self.num_bands * self.num_hyperplanes)
])
# build embeddings
self._build_embeddings()
def _build_embeddings(self):
"""
Builds embeddings for entities, variables and relations.
Embeddings for entities and variables are stored in a dict of embeddings:
self.ent_features = {
'type_1': nn.Embedding(num_ents, embed_dim),
...
'type_n': nn.Embedding(num_ents, embed_dim)
}
self.var_features = {
'type_1': nn.Embedding(1, embed_dim),
...
'type_n': nn.Embedding(1, embed_dim)
}
We define self.node_maps as a mapping from global entity ID to typed entity ID.
Embeddings for relations are stored directly as embedding:
self.rel_features = nn.Embedding(num_rels, embed_dim)
We define self.rel_maps as a mapping from tuple (fr, r, to) to rel_id.
"""
# load data and statistics
rels, _, node_maps = pickle.load(open(self.data_dir+"/graph_data.pkl", "rb"))
self.nodes_per_mode = node_maps
node_mode_counts = {mode: len(node_maps[mode]) for mode in node_maps}
num_nodes = sum(node_mode_counts.values())
# create mapping from global id to type-specific id
new_node_maps = torch.ones(num_nodes, dtype=torch.long).fill_(-1)
for mode, id_list in node_maps.items():
for i, n in enumerate(id_list):
assert new_node_maps[n] == -1
new_node_maps[n] = i
self.node_maps = new_node_maps
# create and initialize entity embeddings. For each type: (num_nodes + 1, embed_dim)
self.ent_features = nn.ModuleDict()
self.var_features = nn.ModuleDict()
for mode in rels:
self.ent_features[mode] = torch.nn.Embedding(node_mode_counts[mode], self.embed_dim)
self.ent_features[mode].weight.data.normal_(0, 1./self.embed_dim)
self.var_features[mode] = torch.nn.Embedding(1, self.embed_dim)
self.var_features[mode].weight.data.normal_(0, 1./self.embed_dim)
print("\nCreated entity embeddings:")
for m, e in self.ent_features.items():
print(f" {m}: {e}")
print("\nCreated variable embeddings:")
for m, e in self.var_features.items():
print(f" {m}: {e}")
# create mapping from rel str to rel ID
rel_maps = {}
rel_counter = 0
for fr in list(rels.keys()):
for to_r in rels[fr]:
to, r = to_r
rel_id = (fr, r, to)
if rel_id not in rel_maps:
rel_maps[rel_id] = rel_counter
rel_counter += 1
self.rel_maps = rel_maps
# create relation and initialize relation embeddings: (num_types, embed_dim)
self.rel_features = torch.nn.Embedding(len(rel_maps), self.embed_dim)
self.rel_features.weight.data.normal_(0, 1./self.embed_dim)
print("\nCreated relation embeddings:")
print(f" {self.rel_features}\n")
def embed_ents(self, mode: str, nodes: Tensor) -> Tensor:
return self.ent_features[mode](self.node_maps[nodes.long()])
def embed_vars(self, mode: str) -> Tensor:
return self.var_features[mode](torch.tensor([0]))
def rel_str_to_id(self, rel: str) -> int:
return self.rel_maps[rel]
def embed_rels(self, rel_types: Tensor) -> Tensor:
return self.rel_features(rel_types)
def vectorize_batch(self, batch: QueryBatch) -> VectorizedQueryBatch:
# embed entities and variables
ent_embed = torch.empty((batch.ent_ids.size(0), self.embed_dim))
for i, (id, mode) in enumerate(zip(batch.ent_ids, batch.ent_modes)):
if "var_" in mode:
emb = self.embed_vars(re.sub("var_", "", mode))
else:
emb = self.embed_ents(mode, id)
ent_embed[i] = emb
# embed relations
num_unique_edges = len(set(batch.edge_ids))
rel_embed = torch.empty((num_unique_edges, self.embed_dim))
inv_rel_embed = torch.empty((num_unique_edges, self.embed_dim))
edge_type = torch.empty((len(batch.edge_ids),), dtype=torch.int64)
all_ids = []
for i, id in enumerate(batch.edge_ids):
if id not in all_ids:
# add to regular edges
rel_id = torch.tensor(self.rel_maps[id])
emb = self.embed_rels(rel_id)
rel_embed[len(all_ids)] = emb
# add to inverse edges
inv_rel_id = torch.tensor(self.rel_maps[_reverse_relation(id)])
inv_emb = self.embed_rels(inv_rel_id)
inv_rel_embed[len(all_ids)] = inv_emb
# keep track of known edges
all_ids.append(id)
new_idx = all_ids.index(id)
edge_type[i] = new_idx
else: # if we've already seen the edge, just refer to index
old_idx = all_ids.index(id)
edge_type[i] = old_idx
# combine regular + inverse edges
rel_embed = torch.cat((rel_embed, inv_rel_embed), dim=0)
assert ent_embed.size(0) == batch.target_idx.size(0)
assert rel_embed.size(0)//2 == max(edge_type)+1
assert num_unique_edges == len(all_ids)
data = VectorizedQueryBatch(
batch_size=batch.batch_size,
batch_idx=batch.batch_idx,
target_idx=batch.target_idx,
ent_embed=ent_embed,
rel_embed=rel_embed,
edge_index=batch.edge_index,
edge_type=edge_type,
q_diameters=batch.q_diameters
)
return data
def forward(self, x_batch: QueryBatch) -> Tensor:
"""
Forwards the query graph batch through the GCN submodels.
Args:
data (QueryBatch):
Contains all information needed for message passing and readout.
Returns:
Tensor: Shape (batch_size, num_bands, num_hyperplanes, embed_dim)
Collection of hyperplanes that demarcate the answer space.
"""
data = self.vectorize_batch(x_batch)
return torch.cat([gcn(data) for gcn in self.submodels], dim=1).reshape(
data.batch_size, self.num_bands, self.num_hyperplanes, self.embed_dim)
def embed_targets(self, y_batch: QueryTargetInfo) -> Tuple[Tensor, Tensor]:
pos_embs = torch.empty((len(y_batch.pos_ids), self.embed_dim))
neg_embs = torch.empty((len(y_batch.pos_ids), self.embed_dim))
for i, (p_id, p_m, n_id) in enumerate(zip(y_batch.pos_ids, y_batch.pos_modes, y_batch.neg_ids)):
pos_embs[i] = self.embed_ents(p_m, p_id)
if not n_id > 0: # no sample found, pick random embedding
neg_embs[i] = self.embed_ents(p_m, torch.tensor([random.choice(self.nodes_per_mode[p_m])], dtype=torch.long))
else:
neg_embs[i] = self.embed_ents(p_m, n_id)
return pos_embs, neg_embs
def predict(self, hyp: Tensor) -> Sequence[Sequence[int]]:
raise NotImplementedError