-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathtrain_align.py
155 lines (133 loc) · 5.99 KB
/
train_align.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
'''
code for entity alignment task
'''
from __future__ import division
from __future__ import print_function
import time
import tensorflow as tf
from utils import *
from metrics import *
from models import AutoRGCN_Align
import logging
import os
# Settings
flags = tf.app.flags
FLAGS = flags.FLAGS
flags.DEFINE_string('dataset', 'zh_en', 'Dataset name: zh_en, ja_en, fr_en')
flags.DEFINE_string('mode', 'None', 'KE method for GCN: TransE, TransH, TransD, DistMult, RotatE, QuatE')
flags.DEFINE_string('optim', 'Adam', 'Optimizer: GD, Adam')
flags.DEFINE_float('learning_rate', 0.001, 'Initial learning rate.')
flags.DEFINE_integer('epochs', 2000, 'Number of epochs for training.')
flags.DEFINE_float('dropout', 0., 'Dropout rate (1 - keep probability).')
flags.DEFINE_float('gamma', 3.0, 'Hyper-parameter for margin based loss.')
flags.DEFINE_integer('num_negs', 5, 'Number of negative samples for each positive seed.')
flags.DEFINE_float('alpha', 0.5, 'Weight of entity conv update.')
flags.DEFINE_float('beta', 0.5, 'Weight of relation conv update.')
flags.DEFINE_integer('layer', 0, 'number of hidden layers')
flags.DEFINE_integer('dim', 200, 'hidden Dimension')
flags.DEFINE_integer('seed', 3, 'Proportion of seeds, 3 means 30%')
flags.DEFINE_boolean('rel_align', True, 'If true, use relation alignment information.')
flags.DEFINE_boolean('rel_update', False, 'If true, use graph conv for rel update.')
flags.DEFINE_integer('randomseed', 12306, 'seed for randomness')
flags.DEFINE_boolean('valid', False, 'If true, split validation data.')
flags.DEFINE_boolean('save', False, 'If true, save the print')
flags.DEFINE_string('metric', "cityblock", 'metric for testing')
flags.DEFINE_string('loss_mode', "L1", 'mode for loss calculation')
flags.DEFINE_string('embed', "random", 'init embedding for entities') # random, text
np.random.seed(FLAGS.randomseed)
random.seed(FLAGS.randomseed)
tf.set_random_seed(FLAGS.randomseed)
if FLAGS.save:
nsave = "log/{}/{}".format(FLAGS.dataset, FLAGS.mode)
else:
print("not saving file")
nsave = "log/trash"
create_exp_dir(nsave)
log_format = '%(asctime)s %(message)s'
logging.basicConfig(stream=sys.stdout, level=logging.INFO,
format=log_format, datefmt='%m/%d %I:%M:%S', filemode="w")
save_fname = 'alpha{}-beta{}-layer{}-sdim{}-lr{}-seed{}'.format(
FLAGS.alpha, FLAGS.beta, FLAGS.layer, FLAGS.dim,
FLAGS.learning_rate, FLAGS.randomseed)
save_fname = save_fname + "-rel"
save_fname = "auto-" + save_fname
if not FLAGS.valid:
save_fname = "test-" + save_fname
fh = logging.FileHandler(os.path.join(nsave, save_fname + ".txt"), "w")
# model_file = os.path.join(nsave, save_fname + ".pt")
fh.setFormatter(logging.Formatter(log_format))
logging.getLogger().addHandler(fh)
logging.getLogger().setLevel(logging.INFO)
# Load data
adj, num_ent, train, test, valid = load_data_align(FLAGS)
rel_num = np.max(adj[2][:, 1]) + 1
print("Relation num: ", rel_num)
# process graph to fit into later computation
support = [preprocess_adj(adj)]
num_supports = 1
model_func = AutoRGCN_Align
num_negs = FLAGS.num_negs
print("Entity num: ", num_ent)
# Define placeholders
placeholders = {
'features': tf.placeholder(tf.float32),
'dropout': tf.placeholder_with_default(0., shape=()),
'num_features_nonzero': tf.placeholder_with_default(0, shape=())
}
# graph structure data
placeholders['support'] = [[tf.placeholder(tf.float32, shape=[None, 1]),
tf.placeholder(tf.float32, shape=[None, 1]), \
tf.placeholder(tf.int32)] for _ in range(num_supports)]
# Create model
input_dim = [num_ent, rel_num]
hidden_dim = [FLAGS.dim, FLAGS.dim]
output_dim = [FLAGS.dim, FLAGS.dim]
if FLAGS.mode == "TransH":
hidden_dim[1] *= 2
output_dim[1] *= 2
elif FLAGS.mode == "TransD":
hidden_dim[0] *= 2
hidden_dim[1] *= 2
output_dim[0] *= 2
output_dim[1] *= 2
names_neg = [["left", "neg_right", "neg_left", "right"]]
model = model_func(placeholders, input_dim, hidden_dim, output_dim, dataset=FLAGS.dataset,
train_labels=train, REL=None, mode=FLAGS.mode, embed=FLAGS.embed, alpha=FLAGS.alpha,
beta=FLAGS.beta, layer_num=FLAGS.layer, sparse_inputs=False, featureless=True,
logging=True, rel_update=FLAGS.rel_update, task="align", names_neg=names_neg)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
# generate positive examples
num_labels = len(train)
left_labels = np.ones((num_labels, num_negs)) * (train[:, 0].reshape((num_labels, 1)))
left_labels = left_labels.reshape((num_labels * num_negs,))
right_labels = np.ones((num_labels, num_negs)) * (train[:, 1].reshape((num_labels, 1)))
right_labels = right_labels.reshape((num_labels * num_negs,))
# Train model
for epoch in range(FLAGS.epochs):
# generate negtive examples
if epoch % 10 == 0:
left_neg = np.random.choice(num_ent, num_labels * num_negs)
right_neg = np.random.choice(num_ent, num_labels * num_negs)
feed_dict = construct_feed_dict(1.0, support, placeholders)
feed_dict.update({placeholders['dropout']: FLAGS.dropout})
for i,labels in enumerate([left_labels,right_neg,left_neg,right_labels]):
feed_dict.update({names_neg[0][i]+":0": labels})
# Training step
outputs = sess.run([model.opt_op, model.loss], feed_dict=feed_dict)
# Print results
if epoch % 10 == 0:
logging.info("Epoch: {} train_loss= {:.5f}".format(epoch+1, outputs[1]))
if epoch % 100 == 0 and valid is not None:
output_embeddings = sess.run(model.outputs, feed_dict=feed_dict)
get_align(output_embeddings[0], valid, logging, FLAGS.metric)
if epoch % 2000 == 0 and epoch > 0 and valid is None:
output_embeddings = sess.run(model.outputs, feed_dict=feed_dict)
get_align(output_embeddings[0], test, logging, FLAGS.metric)
print("Optimization Finished!")
if valid is not None:
exit()
# test
feed_dict = construct_feed_dict(1.0, support, placeholders)
output_embeddings = sess.run(model.outputs, feed_dict=feed_dict)
get_align(output_embeddings[0], test, logging, FLAGS.metric)