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
|
import argparse
import csv
import numpy as np
import keras
import keras.backend as K
from Helpers import Data, Get_Embedding
from keras.layers import LSTM, Embedding, Input, Lambda, concatenate
from keras.models import Model
def get_learning_rate(epoch=None, model=None):
return np.round(float(K.get_value(model.optimizer.lr)), 5)
def make_cosine_func(hidden_size=50):
def exponent_neg_cosine_similarity(x):
""" Helper function for the similarity estimate of the LSTMs outputs """
leftNorm = K.l2_normalize(x[:, :hidden_size], axis=-1)
rightNorm = K.l2_normalize(x[:, hidden_size:], axis=-1)
return K.sum(K.prod([leftNorm, rightNorm], axis=0), axis=1, keepdims=True)
return exponent_neg_cosine_similarity
def main(args):
source_lang = args.source_lang
target_lang = args.target_lang
hidden_size = args.hidden_size
max_len = args.max_len
num_iters = args.num_iters
data_file = args.data_file
learning_rate = args.learning_rate
batch = args.batch
data = Data(source_lang, target_lang, data_file, max_len)
x_train = data.x_train
y_train = data.y_train
x_predict = data.x_val
y_predict = data.y_val
vocab_size = data.vocab_size
max_len = data.max_len
# https://stackoverflow.com/a/10741692/3005749
x = data.y_val
y = np.bincount(x.astype(np.int32))
ii = np.nonzero(y)[0]
assert ii == 1
assert y[ii] == 1000 # hardcoded for now
if not batch:
print(f"Source Lang: {source_lang}")
print(f"Target Lang: {target_lang}")
print(f"Using {len(x_train[0])} pairs to learn")
print(f"Predicting {len(y_predict)} pairs")
print(f"Vocabulary size: {vocab_size}")
print(f"Maximum sequence length: {max_len}")
source_emb_file = args.source_emb_file
target_emb_file = args.target_emb_file
embedding = Get_Embedding(
source_lang, target_lang, source_emb_file, target_emb_file, data.word_to_id
)
embedding_size = embedding.embedding_matrix.shape[1]
seq_1 = Input(shape=(max_len,), dtype="int32", name="sequence1")
seq_2 = Input(shape=(max_len,), dtype="int32", name="sequence2")
embed_layer = Embedding(
output_dim=embedding_size,
input_dim=vocab_size + 1,
input_length=max_len,
trainable=False,
)
embed_layer.build((None,))
embed_layer.set_weights([embedding.embedding_matrix])
input_1 = embed_layer(seq_1)
input_2 = embed_layer(seq_2)
l1 = LSTM(units=hidden_size)
l1_out = l1(input_1)
l2_out = l1(input_2)
concats = concatenate([l1_out, l2_out], axis=-1)
out_func = make_cosine_func(hidden_size)
main_output = Lambda(out_func, output_shape=(1,))(concats)
model = Model(inputs=[seq_1, seq_2], outputs=[main_output])
opt = keras.optimizers.Adadelta(lr=learning_rate, clipnorm=1.25)
model.compile(optimizer=opt, loss="mean_squared_error", metrics=["accuracy"])
model.summary()
adjuster = keras.callbacks.ReduceLROnPlateau(
monitor="val_acc", patience=5, verbose=1, factor=0.5, min_lr=0.0001
)
history = model.fit(
x_train,
y_train,
validation_data=(x_predict, y_predict),
epochs=num_iters,
batch_size=32,
verbose=1,
callbacks=[adjuster],
)
target_sents = x_predict[1]
precision_at_one = 0
precision_at_ten = 0
for index, sent in enumerate(x_predict[0]):
source_sents = np.array([sent] * 1000)
to_predict = [source_sents, target_sents]
preds = model.predict(to_predict)
ind = np.argpartition(preds.ravel(), -10)[-10:]
if index in ind:
precision_at_ten += 1
if np.argmax(preds.ravel()) == index:
precision_at_one += 1
training_samples = len(x_train[0])
validation_samples = len(y_predict)
fields = [
source_lang,
target_lang,
training_samples,
validation_samples,
precision_at_one,
precision_at_ten,
]
if not batch:
print(f"P@1: {precision_at_one/1000}, {precision_at_one} defs")
else:
with open("supervised.csv", "a") as f:
writer = csv.writer(f)
writer.writerow(fields)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(
"-sl", "--source_lang", type=str, help="Source language.", default="english"
)
parser.add_argument(
"-tl", "--target_lang", type=str, help="Target language.", default="italian"
)
parser.add_argument("-df", "--data_file", type=str, help="Path to dataset.")
parser.add_argument(
"-es",
"--source_emb_file",
type=str,
help="Path to Source (English) Embedding File.",
)
parser.add_argument(
"-et", "--target_emb_file", type=str, help="Path to Target Embedding File."
)
parser.add_argument(
"-l",
"--max_len",
type=int,
help="Maximum number of words in a sentence.",
default=20,
)
parser.add_argument(
"-z",
"--hidden_size",
type=int,
help="Number of Units in LSTM layer.",
default=50,
)
parser.add_argument(
"-b",
"--batch",
action="store_true",
help="running in batch (store results to csv) or"
+ "running in a single instance (output the results)",
)
parser.add_argument(
"-n", "--num_iters", type=int, help="Number of iterations/epochs.", default=7
)
parser.add_argument(
"-lr",
"--learning_rate",
type=float,
help="Learning rate for optimizer.",
default=1.0,
)
args = parser.parse_args()
main(args)
|
