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
269
270
271
272
273
274
275
276
277
278
279 | class LightEAFrameEncoder(RelationFrameEncoder):
"""Use LightEA algorithm to encode frame.
Args:
ent_dim: int: Entity dimensions
depth: int: Number of hops
mini_dim:int: Mini batching size
rel_dim:int: relation embedding dimensions (same as ent_dim if None)
attribute_encoder: HintOrType[TokenizedFrameEncoder]: Attribute encoder class
attribute_encoder_kwargs: OptionalKwargs: Keyword arguments for initializing attribute encoder class
Quote: Reference
Mao et. al.,"LightEA: A Scalable, Robust, and Interpretable Entity Alignment Framework via Three-view Label Propagation", EMNLP 2022 <https://aclanthology.org/2022.emnlp-main.52.pdf>
"""
def __init__(
self,
ent_dim: int = 256,
depth: int = 2,
mini_dim: int = 16,
rel_dim: Optional[int] = None,
attribute_encoder: HintOrType[TokenizedFrameEncoder] = None,
attribute_encoder_kwargs: OptionalKwargs = None,
):
self.ent_dim = ent_dim
self.depth = depth
self.device = resolve_device()
self.mini_dim = mini_dim
self.rel_dim = ent_dim if rel_dim is None else rel_dim
self.attribute_encoder = tokenized_frame_encoder_resolver.make(
attribute_encoder, attribute_encoder_kwargs
)
def _encode_rel(
self,
rel_triples_left: np.ndarray,
rel_triples_right: np.ndarray,
ent_features: NamedVector,
) -> GeneralVector:
(
node_size,
rel_size,
ent_tuple,
triples_idx,
ent_ent,
ent_ent_val,
rel_ent,
ent_rel,
) = self._transform_graph(rel_triples_left, rel_triples_right)
return self._get_features(
node_size,
rel_size,
ent_tuple,
triples_idx,
ent_ent,
ent_ent_val,
rel_ent,
ent_rel,
ent_features.vectors,
)
def _transform_graph(
self, rel_triples_left: np.ndarray, rel_triples_right: np.ndarray
):
triples = []
rel_size = 0
for line in rel_triples_left:
h, r, t = line
triples.append([h, t, 2 * r])
triples.append([t, h, 2 * r + 1])
rel_size = max(rel_size, 2 * r + 1)
for line in rel_triples_right:
h, r, t = line
triples.append([h, t, 2 * r])
triples.append([t, h, 2 * r + 1])
rel_size = max(rel_size, 2 * r + 1)
triples = np.unique(triples, axis=0)
node_size, rel_size = np.max(triples) + 1, np.max(triples[:, 2]) + 1 # type: ignore
ent_tuple, triples_idx = [], []
ent_ent_s, rel_ent_s, ent_rel_s = {}, set(), set()
last, index = (-1, -1), -1
for i in range(node_size):
ent_ent_s[(i, i)] = 0
for h, t, r in triples:
ent_ent_s[(h, h)] += 1
ent_ent_s[(t, t)] += 1
if (h, t) != last:
last = (h, t)
index += 1
ent_tuple.append([h, t])
ent_ent_s[(h, t)] = 0
triples_idx.append([index, r])
ent_ent_s[(h, t)] += 1
rel_ent_s.add((r, h))
ent_rel_s.add((t, r))
ent_tuple = np.array(ent_tuple) # type: ignore
triples_idx = np.unique(np.array(triples_idx), axis=0) # type: ignore
ent_ent = np.unique(np.array(list(ent_ent_s.keys())), axis=0)
ent_ent_val = np.array([ent_ent_s[(x, y)] for x, y in ent_ent]).astype(
"float32"
)
rel_ent = np.unique(np.array(list(rel_ent_s)), axis=0)
ent_rel = np.unique(np.array(list(ent_rel_s)), axis=0)
return (
node_size,
rel_size,
ent_tuple,
triples_idx,
ent_ent,
ent_ent_val,
rel_ent,
ent_rel,
)
@torch.no_grad()
def _get_features(
self,
node_size,
rel_size,
ent_tuple,
triples_idx,
ent_ent,
ent_ent_val,
rel_ent,
ent_rel,
ent_feature,
):
ent_feature = ent_feature.to(self.device)
rel_feature = torch.zeros((rel_size, ent_feature.shape[-1])).to(self.device)
ent_ent, ent_rel, rel_ent, ent_ent_val, triples_idx, ent_tuple = map(
torch.tensor,
[ent_ent, ent_rel, rel_ent, ent_ent_val, triples_idx, ent_tuple],
)
ent_ent = ent_ent.t()
ent_rel = ent_rel.t()
rel_ent = rel_ent.t()
triples_idx = triples_idx.t()
ent_tuple = ent_tuple.t()
ent_ent_graph = torch.sparse_coo_tensor(
indices=ent_ent, values=ent_ent_val, size=(node_size, node_size)
).to(self.device)
rel_ent_graph = torch.sparse_coo_tensor(
indices=rel_ent,
values=torch.ones(rel_ent.shape[1]),
size=(rel_size, node_size),
).to(self.device)
ent_rel_graph = torch.sparse_coo_tensor(
indices=ent_rel,
values=torch.ones(ent_rel.shape[1]),
size=(node_size, rel_size),
).to(self.device)
ent_list, rel_list = [ent_feature], [rel_feature]
for _ in trange(self.depth):
new_rel_feature = torch.from_numpy(
_batch_sparse_matmul(rel_ent_graph, ent_feature, self.device)
).to(self.device)
new_rel_feature = _my_norm(new_rel_feature)
new_ent_feature = torch.from_numpy(
_batch_sparse_matmul(ent_ent_graph, ent_feature, self.device)
).to(self.device)
new_ent_feature += torch.from_numpy(
_batch_sparse_matmul(ent_rel_graph, rel_feature, self.device)
).to(self.device)
new_ent_feature = _my_norm(new_ent_feature)
ent_feature = new_ent_feature
rel_feature = new_rel_feature
ent_list.append(ent_feature)
rel_list.append(rel_feature)
ent_feature = torch.cat(ent_list, dim=1)
rel_feature = torch.cat(rel_list, dim=1)
ent_feature = _my_norm(ent_feature)
rel_feature = _my_norm(rel_feature)
rel_feature = _random_projection(rel_feature, self.rel_dim, self.device)
batch_size = ent_feature.shape[-1] // self.mini_dim
sparse_graph = torch.sparse_coo_tensor(
indices=triples_idx,
values=torch.ones(triples_idx.shape[1]),
size=(torch.max(triples_idx).item() + 1, rel_size),
).to(self.device)
adj_value = _batch_sparse_matmul(sparse_graph, rel_feature, self.device)
del rel_feature
features_list = []
for batch in trange(self.rel_dim // batch_size + 1):
temp_list = []
for head in trange(batch_size):
if batch * batch_size + head >= self.rel_dim:
break
sparse_graph = torch.sparse_coo_tensor(
indices=ent_tuple,
values=adj_value[:, batch * batch_size + head],
size=(node_size, node_size),
).to(self.device)
feature = _batch_sparse_matmul(
sparse_graph,
_random_projection(ent_feature, self.mini_dim, self.device).to(
self.device
),
self.device,
batch_size=128,
save_mem=True,
)
temp_list.append(feature)
if len(temp_list):
features_list.append(np.concatenate(temp_list, axis=-1))
features = np.concatenate(features_list, axis=-1)
features = normalize(features)
features = np.concatenate([ent_feature.cpu().numpy(), features], axis=-1)
return features
|