File size: 18,398 Bytes
4523bda | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 | import math
import multiprocessing as mp
import random
import time
import gc
import tensorflow as tf
import numpy as np
from sklearn import preprocessing
import openea.modules.train.batch as bat
from openea.modules.finding.evaluation import early_stop
from openea.modules.utils.util import generate_out_folder
from openea.modules.utils.util import load_session
from openea.modules.utils.util import task_divide
from openea.modules.base.initializers import init_embeddings, orthogonal_init
from openea.modules.base.losses import margin_loss
from openea.modules.base.optimizers import generate_optimizer
from utils import MyKGs, find_alignment
from eval import valid, test
from nn_search import generate_neighbours
def search_kg1_to_kg2_1nn_neighbor(embeds1, embeds2, ents2, mapping_mat, return_sim=False, soft_nn=10):
if mapping_mat is not None:
embeds1 = np.matmul(embeds1, mapping_mat)
embeds1 = preprocessing.normalize(embeds1)
sim_mat = np.matmul(embeds1, embeds2.T)
nearest_pairs = find_alignment(sim_mat, soft_nn)
nns = [ents2[x[0][1]] for x in nearest_pairs]
if return_sim:
sim_list = []
for pair in nearest_pairs:
sim_list.append(sim_mat[pair[0][0], pair[0][1]])
return nns, sim_list
return nns
class MTransEV2:
def set_kgs(self, kgs: MyKGs):
self.kgs = kgs
def set_args(self, args):
self.args = args
self.out_folder = generate_out_folder(self.args.output, self.args.training_data, self.args.dataset_division,
self.__class__.__name__)
def init(self):
self._define_variables()
self._define_embed_graph()
self._define_align_graph()
self.session = load_session()
tf.global_variables_initializer().run(session=self.session)
def __init__(self):
self.out_folder = None
self.args = None
self.kgs = None
self.session = None
self.seed_entities1 = None
self.seed_entities2 = None
self.neg_ts = None
self.neg_rs = None
self.neg_hs = None
self.pos_ts = None
self.pos_rs = None
self.pos_hs = None
self.rel_embeds = None
self.ent_embeds = None
self.mapping_mat = None
self.eye_mat = None
self.triple_optimizer = None
self.triple_loss = None
self.mapping_optimizer = None
self.mapping_loss = None
self.mapping_mat = None
self.flag1 = -1
self.flag2 = -1
self.early_stop = False
def _define_variables(self):
with tf.variable_scope('KG' + 'embeddings'):
self.ent_embeds = init_embeddings([self.kgs.entities_num, self.args.dim], 'ent_embeds',
self.args.init, self.args.ent_l2_norm)
self.rel_embeds = init_embeddings([self.kgs.relations_num, self.args.dim], 'rel_embeds',
self.args.init, self.args.rel_l2_norm)
with tf.variable_scope('alignment' + 'mapping'):
self.mapping_mat = orthogonal_init([self.args.dim, self.args.dim], 'mapping_matrix')
self.eye_mat = tf.constant(np.eye(self.args.dim), dtype=tf.float32, name='eye')
def _define_embed_graph(self):
print("build embedding learning graph...")
with tf.name_scope('triple_placeholder'):
self.pos_hs = tf.placeholder(tf.int32, shape=[None])
self.pos_rs = tf.placeholder(tf.int32, shape=[None])
self.pos_ts = tf.placeholder(tf.int32, shape=[None])
self.neg_hs = tf.placeholder(tf.int32, shape=[None])
self.neg_rs = tf.placeholder(tf.int32, shape=[None])
self.neg_ts = tf.placeholder(tf.int32, shape=[None])
with tf.name_scope('triple_lookup'):
phs = tf.nn.embedding_lookup(self.ent_embeds, self.pos_hs)
prs = tf.nn.embedding_lookup(self.rel_embeds, self.pos_rs)
pts = tf.nn.embedding_lookup(self.ent_embeds, self.pos_ts)
nhs = tf.nn.embedding_lookup(self.ent_embeds, self.neg_hs)
nrs = tf.nn.embedding_lookup(self.rel_embeds, self.neg_rs)
nts = tf.nn.embedding_lookup(self.ent_embeds, self.neg_ts)
with tf.name_scope('triple_loss'):
self.triple_loss = margin_loss(phs, prs, pts, nhs, nrs, nts, self.args.embed_margin, self.args.loss_norm)
self.triple_optimizer = generate_optimizer(self.triple_loss, self.args.learning_rate,
opt=self.args.optimizer)
def _mapping_align_loss(self, seed_embed1, seed_embed2):
mapped_seed_embed1 = tf.matmul(seed_embed1, self.mapping_mat)
mapped_seed_embed1 = tf.nn.l2_normalize(mapped_seed_embed1, 1)
with tf.name_scope('mapping_distance'):
distance = mapped_seed_embed1 - seed_embed2
with tf.name_scope('mapping_loss'):
align_loss = tf.reduce_sum(tf.reduce_sum(tf.square(distance), axis=1))
orthogonal_loss = tf.reduce_mean(
tf.reduce_sum(tf.pow(tf.matmul(self.mapping_mat, self.mapping_mat, transpose_b=True) - self.eye_mat, 2), 1))
return align_loss + orthogonal_loss
def _mapping_align_marginal_loss(self, seed_embed1, seed_embed2, pos_embed1, neg_embed2):
mapped_seed_embed1 = tf.matmul(seed_embed1, self.mapping_mat)
mapped_seed_embed1 = tf.nn.l2_normalize(mapped_seed_embed1, 1)
with tf.name_scope('mapping_distance'):
distance = mapped_seed_embed1 - seed_embed2
with tf.name_scope('mapping_loss'):
pos_score = tf.reduce_sum(tf.square(distance), axis=1)
align_loss = tf.reduce_sum(pos_score)
if self.args.mapping_margin > 0.0:
mapped_pos_embed1 = tf.matmul(pos_embed1, self.mapping_mat)
mapped_pos_embed1 = tf.nn.l2_normalize(mapped_pos_embed1, 1)
neg_distance = mapped_pos_embed1 - neg_embed2
neg_score = tf.reduce_sum(tf.square(neg_distance), axis=1)
align_loss += self.args.mapping_neg_weight * \
tf.reduce_sum(tf.nn.relu(tf.constant(self.args.mapping_margin) - neg_score))
orthogonal_loss = tf.reduce_mean(tf.reduce_sum(
tf.square(tf.matmul(self.mapping_mat, self.mapping_mat, transpose_b=True) - self.eye_mat), 1))
return align_loss + orthogonal_loss
def _define_align_graph(self):
print("build alignment learning graph...")
with tf.name_scope('seed_links_placeholder'):
self.seed_entities1 = tf.placeholder(tf.int32, shape=[None])
self.seed_entities2 = tf.placeholder(tf.int32, shape=[None])
self.pos_entities1 = tf.placeholder(tf.int32, shape=[None])
self.neg_entities2 = tf.placeholder(tf.int32, shape=[None])
with tf.name_scope('seed_links_lookup'):
seed_embed1 = tf.nn.embedding_lookup(self.ent_embeds, self.seed_entities1)
seed_embed2 = tf.nn.embedding_lookup(self.ent_embeds, self.seed_entities2)
pos_embed1 = tf.nn.embedding_lookup(self.ent_embeds, self.pos_entities1)
neg_embed2 = tf.nn.embedding_lookup(self.ent_embeds, self.neg_entities2)
with tf.name_scope('mapping_loss'):
self.mapping_loss = self._mapping_align_marginal_loss(seed_embed1, seed_embed2, pos_embed1, neg_embed2)
self.mapping_optimizer = generate_optimizer(self.mapping_loss, self.args.learning_rate,
opt=self.args.optimizer)
def _eval_valid_embeddings(self, remove_dangling=False):
if remove_dangling:
candidate_list = self.kgs.valid_entities2 + list(self.kgs.kg2.entities_set
- set(self.kgs.train_entities2)
- set(self.kgs.valid_entities2)
- set([x for x, _ in self.kgs.train_unlinked_entities2])
- set([x for x, _ in self.kgs.valid_unlinked_entities2]))
else:
candidate_list = self.kgs.valid_entities2 + list(self.kgs.kg2.entities_set
- set(self.kgs.train_entities2)
- set(self.kgs.valid_entities2))
embeds1 = tf.nn.embedding_lookup(self.ent_embeds, self.kgs.valid_entities1).eval(session=self.session)
embeds2 = tf.nn.embedding_lookup(self.ent_embeds, candidate_list).eval(session=self.session)
mapping = self.mapping_mat.eval(session=self.session) if self.mapping_mat is not None else None
return embeds1, embeds2, mapping
def _eval_test_embeddings(self, remove_dangling=False):
if remove_dangling:
candidate_list = self.kgs.test_entities2 + list(self.kgs.kg2.entities_set
- set(self.kgs.train_entities2)
- set(self.kgs.valid_entities2)
- set(self.kgs.test_entities2)
- set([x for x, _ in self.kgs.train_unlinked_entities2])
- set([x for x, _ in self.kgs.valid_unlinked_entities2]))
else:
candidate_list = self.kgs.test_entities2 + list(self.kgs.kg2.entities_set
- set(self.kgs.train_entities2)
- set(self.kgs.valid_entities2)
- set(self.kgs.test_entities2))
embeds1 = tf.nn.embedding_lookup(self.ent_embeds, self.kgs.test_entities1).eval(session=self.session)
embeds2 = tf.nn.embedding_lookup(self.ent_embeds, candidate_list).eval(session=self.session)
mapping = self.mapping_mat.eval(session=self.session) if self.mapping_mat is not None else None
return embeds1, embeds2, mapping
def valid_alignment(self, stop_metric, remove_dangling):
print("\nvalidating entity alignment...")
embeds1, embeds2, mapping = self._eval_valid_embeddings(remove_dangling=remove_dangling)
hits, mrr_12, sim_list = valid(embeds1, embeds2, mapping, self.args.top_k,
self.args.test_threads_num, metric=self.args.eval_metric,
normalize=self.args.eval_norm, csls_k=0, accurate=False)
print()
return hits[0] if stop_metric == 'hits1' else mrr_12
def test(self):
embeds1, embeds2, mapping = self._eval_test_embeddings(remove_dangling=False)
_, _, _, sim_list = test(embeds1, embeds2, mapping, self.args.top_k, self.args.test_threads_num,
metric=self.args.eval_metric, normalize=self.args.eval_norm, csls_k=0, accurate=True)
print()
def eval_embeddings(self, entity_list):
embeds = tf.nn.embedding_lookup(self.ent_embeds, entity_list)
return embeds.eval(session=self.session)
def launch_training_1epo(self, epoch, triple_steps, steps_tasks, training_batch_queue, neighbors1, neighbors2):
self.launch_embed_training_1epo(epoch, triple_steps, steps_tasks, training_batch_queue, neighbors1, neighbors2)
self.launch_align_training_1epo(epoch, triple_steps)
def launch_embed_training_1epo(self, epoch, triple_steps, steps_tasks, batch_queue, neighbors1, neighbors2):
start = time.time()
for steps_task in steps_tasks:
mp.Process(target=bat.generate_relation_triple_batch_queue,
args=(self.kgs.kg1.relation_triples_list, self.kgs.kg2.relation_triples_list,
self.kgs.kg1.relation_triples_set, self.kgs.kg2.relation_triples_set,
self.kgs.kg1.entities_list, self.kgs.kg2.entities_list,
self.args.batch_size, steps_task,
batch_queue, neighbors1, neighbors2, self.args.neg_triple_num)).start()
epoch_loss = 0
trained_samples_num = 0
for i in range(triple_steps):
batch_pos, batch_neg = batch_queue.get()
batch_loss, _ = self.session.run(fetches=[self.triple_loss, self.triple_optimizer],
feed_dict={self.pos_hs: [x[0] for x in batch_pos],
self.pos_rs: [x[1] for x in batch_pos],
self.pos_ts: [x[2] for x in batch_pos],
self.neg_hs: [x[0] for x in batch_neg],
self.neg_rs: [x[1] for x in batch_neg],
self.neg_ts: [x[2] for x in batch_neg]})
trained_samples_num += len(batch_pos)
epoch_loss += batch_loss
epoch_loss /= trained_samples_num
random.shuffle(self.kgs.kg1.relation_triples_list)
random.shuffle(self.kgs.kg2.relation_triples_list)
print('epoch {}, avg. triple loss: {:.4f}, cost time: {:.1f}s'.format(epoch, epoch_loss, time.time() - start))
def launch_align_training_1epo(self, epoch, triple_steps):
start = time.time()
epoch_loss = 0
trained_samples_num = 0
batch_size = 2 * len(self.kgs.train_links) // triple_steps
kg_training_ents2 = list(self.kgs.kg2.entities_set - set(self.kgs.train_entities2))
neg_batch_size = batch_size * self.args.mapping_neg_num
for i in range(triple_steps):
links_batch = random.sample(self.kgs.train_links, batch_size)
pos_entities1 = random.sample(self.kgs.kg1.entities_list, neg_batch_size)
neg_entities2 = random.sample(kg_training_ents2, neg_batch_size)
batch_loss, _ = self.session.run(fetches=[self.mapping_loss, self.mapping_optimizer],
feed_dict={self.seed_entities1: [x[0] for x in links_batch],
self.seed_entities2: [x[1] for x in links_batch],
self.pos_entities1: pos_entities1,
self.neg_entities2: neg_entities2})
epoch_loss += batch_loss
trained_samples_num += len(links_batch)
epoch_loss /= trained_samples_num
print('epoch {}, avg. mapping loss: {:.4f}, cost time: {:.1}s'.format(epoch, epoch_loss, time.time() - start))
def generate_neighbors(self):
t1 = time.time()
assert 0.0 < self.args.truncated_epsilon < 1.0
num1 = len(self.kgs.kg1.entities_list) // 2
if len(self.kgs.kg1.entities_list) > 200000:
num1 = len(self.kgs.kg1.entities_list) // 3
if num1 > len(self.kgs.useful_entities_list1):
kg1_random_ents = self.kgs.useful_entities_list1 + \
random.sample(list(set(self.kgs.kg1.entities_list) - set(self.kgs.useful_entities_list1)),
num1 - len(self.kgs.useful_entities_list1))
else:
kg1_random_ents = self.kgs.useful_entities_list1
embeds1 = self.eval_embeddings(kg1_random_ents)
num2 = len(self.kgs.kg2.entities_list) // 2
if len(self.kgs.kg2.entities_list) > 200000:
num2 = len(self.kgs.kg2.entities_list) // 3
if num2 > len(self.kgs.useful_entities_list2):
kg2_random_ents = self.kgs.useful_entities_list2 + \
random.sample(list(set(self.kgs.kg2.entities_list) - set(self.kgs.useful_entities_list2)),
num2 - len(self.kgs.useful_entities_list2))
else:
kg2_random_ents = self.kgs.useful_entities_list2
embeds2 = self.eval_embeddings(kg2_random_ents)
neighbors_num1 = int((1 - self.args.truncated_epsilon) * num1)
neighbors_num2 = int((1 - self.args.truncated_epsilon) * num2)
print("generating neighbors...")
neighbors1 = generate_neighbours(embeds1, kg1_random_ents, neighbors_num1,
frags_num=self.args.batch_threads_num)
neighbors2 = generate_neighbours(embeds2, kg2_random_ents, neighbors_num2,
frags_num=self.args.batch_threads_num)
print("generating neighbors ({}, {}) costs {:.3f} s.".format(num1, num2, time.time() - t1))
gc.collect()
return neighbors1, neighbors2
def run(self):
t = time.time()
triples_num = self.kgs.kg1.relation_triples_num + self.kgs.kg2.relation_triples_num
triple_steps = int(math.ceil(triples_num / self.args.batch_size))
steps_tasks = task_divide(list(range(triple_steps)), self.args.batch_threads_num)
training_batch_manager = mp.Manager()
training_batch_queue = training_batch_manager.Queue()
neighbors1, neighbors2 = dict(), dict()
# training
for i in range(1, self.args.max_epoch + 1):
self.launch_training_1epo(i, triple_steps, steps_tasks, training_batch_queue, neighbors1, neighbors2)
# validation
if i >= self.args.start_valid and i % self.args.eval_freq == 0:
flag = self.valid_alignment(self.args.stop_metric, False)
self.flag1, self.flag2, self.early_stop = early_stop(self.flag1, self.flag2, flag)
if self.early_stop or i == self.args.max_epoch:
break
# truncated sampling cache
if self.args.neg_sampling == 'truncated' and i % self.args.truncated_freq == 0:
if neighbors1 is not None:
del neighbors1, neighbors2
gc.collect()
neighbors1, neighbors2 = self.generate_neighbors()
print("Training ends. Total time = {:.1f} s.".format(time.time() - t))
|