-
Notifications
You must be signed in to change notification settings - Fork 5
/
model4.py
144 lines (109 loc) · 5.84 KB
/
model4.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
# -*- coding: utf-8 -*-
from data import *
from glove import *
import tensorflow as tf
#import tensorflow.nn.rnn_cell as rnn_cell
from tensorflow.contrib import rnn
from sklearn.cross_validation import train_test_split
import os
import glob
import numpy as np
import h5py
from random import random
from numpy import array
from numpy import cumsum
from matplotlib import pyplot
from pandas import DataFrame
from keras.layers import Dense, Input, LSTM, Embedding, Dropout, Activation
from keras.layers.merge import concatenate
from keras.models import Model
from keras.layers.normalization import BatchNormalization
from keras.callbacks import EarlyStopping, ModelCheckpoint
from keras import optimizers
import keras.backend as K
#from keras.losses import categorical_crossentropy
# rm old log files
for file in glob.glob('/home/salomons/tmp/tf.log/*'):
os.remove(file)
# config
train_path = '/data/senseval2/eng-lex-sample.training.xml'
test_path = '/data/senseval2/eng-lex-samp.evaluation.xml'
# load data
#train_data = load_senteval2_data(train_path, True)
#test_data = load_senteval2_data(test_path, False)
train_data_ = load_train_data(23)
test_data = load_test_data(2)
print('Dataset size (train/test): %d / %d' % (len(train_data_), len(test_data)))
EMBEDDING_DIM = 100
print('Embedding vector: %d' % EMBEDDING_DIM)
# build vocab utils
word_to_id = build_vocab(train_data_)
target_word_to_id, target_sense_to_id, n_words, n_senses_from_target_id = build_sense_ids(train_data_)
print('Vocabulary size: %d' % len(word_to_id))
#build context vocab of the target sense
train_target_sense_to_context = build_context(train_data_, word_to_id)
#build context embeddings of the target sense
embedding_matrix = fill_with_gloves(word_to_id, 100)
target_sense_to_context_embedding = build_embedding(train_target_sense_to_context, embedding_matrix, len(word_to_id), EMBEDDING_DIM)
# make numeric
train_ndata = convert_to_numeric(train_data_, word_to_id, target_word_to_id, target_sense_to_id, n_senses_from_target_id, target_sense_to_context_embedding, is_training = True)
#test_ndata = convert_to_numeric(test_data, word_to_id, target_word_to_id, target_sense_to_id, n_senses_from_target_id, target_sense_to_context_embedding, is_training = False)
n_step_f = 40
n_step_b = 40
print('n_step forward/backward: %d / %d' % (n_step_f, n_step_b))
MAX_SEQUENCE_LENGTH = 40
act = 'relu'
#STAMP = 'lstm_%d_%d_%.2f_%.2f'%(100, 2, 0.2, 0.5)
#bst_model_path = h5py.File("weights.best.hdf5", "w")
def cos_distance(y_true, y_pred):
y_true = K.l2_normalize(y_true, axis=-1)
y_pred = K.l2_normalize(y_pred, axis=-1)
return K.mean(1 - K.sum((y_true * y_pred), axis=-1))
def own_model(train_forward_data, train_backward_data, train_sense_embedding,
val_forward_data=None, val_backward_data=None, val_sense_embedding=None,
n_units=100, dense_unints=256, is_training=True, EMBEDDING_DIM=100, epochs=100, batch_size=2048, init_word_vecs=None):
embedding_layer = Embedding(len(word_to_id),
EMBEDDING_DIM,
weights=[init_word_vecs],
input_length=MAX_SEQUENCE_LENGTH,
trainable=False)
lstm_layer = LSTM(n_units, dropout=0.2, recurrent_dropout=0.2)
forward_input = Input(shape=(MAX_SEQUENCE_LENGTH, ), dtype='int32', name='forward_input')
embedded_forward = embedding_layer(forward_input)
forward_lstm = lstm_layer(embedded_forward)
backward_input = Input(shape=(MAX_SEQUENCE_LENGTH, ), dtype='int32', name='backward_input')
embedded_backward = embedding_layer(backward_input)
backward_lstm = lstm_layer(embedded_backward)
merged = concatenate([forward_lstm, backward_lstm])
#merged = lstm_layer(merged)
merged = Dropout(0.2)(merged) if is_training else merged
merged = BatchNormalization()(merged)
merged = Dense(units=dense_unints, activation=act)(merged)
merged = Dropout(0.2)(merged) if is_training else merged
merged = BatchNormalization()(merged)
preds = Dense(EMBEDDING_DIM)(merged)
## train the model
model = Model(inputs=[forward_input, backward_input], outputs=preds)
nadam = optimizers.Nadam(clipnorm=1.) #, clipvalue=0.5
model.compile(loss='mse', optimizer=nadam, metrics=[cos_distance])
early_stopping =EarlyStopping(monitor='val_loss', patience=10)
#bst_model_path = STAMP + '.h5'
bst_model_path = "weights.best.hdf5"
model_checkpoint = ModelCheckpoint(bst_model_path, save_best_only=True, save_weights_only=True, verbose=1)
hist = model.fit([train_forward_data, train_backward_data], train_sense_embedding,
validation_data=([val_forward_data, val_backward_data], val_sense_embedding),
epochs=epochs, batch_size=batch_size, shuffle=True,
callbacks=[early_stopping, model_checkpoint])
model.load_weights(bst_model_path)
bst_val_score = min(hist.history['val_loss'])
print(hist.history.keys())
print('min val loss is: %f' % (bst_val_score))
if __name__ == '__main__':
grouped_by_target = group_by_target(train_ndata)
train_data, val_data = split_grouped(grouped_by_target, 0)
init_emb = fill_with_gloves(word_to_id, EMBEDDING_DIM)
train_forward_data, train_backward_data, train_target_sense_ids, train_sense_embedding = get_data(train_data, n_step_f, n_step_b)
val_forward_data, val_backward_data, val_target_sense_ids, val_sense_embedding = get_data(val_data, n_step_f, n_step_b)
own_model(train_forward_data, train_backward_data, train_sense_embedding,
val_forward_data, val_backward_data, val_sense_embedding,
init_word_vecs=init_emb)