-
Notifications
You must be signed in to change notification settings - Fork 0
/
dataloaders_processed.py
160 lines (123 loc) · 6.89 KB
/
dataloaders_processed.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
156
157
158
159
160
import io
import pickle
import re
import tensorflow as tf
import numpy as np
import unicodedata
import config as cfg
from tensorflow.keras.preprocessing.sequence import pad_sequences
from sklearn.model_selection import train_test_split
def unicode_to_ascii(s):
return ''.join(c for c in unicodedata.normalize('NFD', s) if unicodedata.category(c) != 'Mn')
def preprocess_sentence(w, lang, aligned=True, add_special_tag=True):
w = w.strip()
if lang == "en" and not aligned:
# This part is required only for english unaligned samples in word2vec
w = unicode_to_ascii(w.lower())
# Removing everything except(letters)
w = re.sub(r"[^a-z]+", " ", w)
if lang == "fr" and not aligned:
# Adding space with punctuation for easy split.
w = re.sub(r"([?.!,¿])", r" \1 ", w)
w = re.sub(r'[" "]+', " ", w)
w = w.strip()
# currently keeping these tags for both type of data
if add_special_tag:
w = '<start> ' + w + ' <end>'
return w
def load_lang(lang_path):
return io.open(cfg.root_path + lang_path, encoding='UTF-8').read().strip().split('\n')
def load_test_generator(lang_path, input_tokenizer, batch_size):
test_data = io.open(lang_path, encoding='UTF-8').read().strip().split('\n')
test_data = [preprocess_sentence(x, "en", aligned=True) for x in test_data]
test_tensor = input_tokenizer.texts_to_sequences(test_data)
test_tensor = pad_sequences(test_tensor, padding='post')
test_dataset = tf.data.Dataset.from_tensor_slices(test_tensor).batch(batch_size)
return test_dataset
def dataloader_unaligned(preprocess=True):
unaligned_en = load_lang(cfg.unaligned_en_path)
unaligned_fr = load_lang(cfg.unaligned_fr_path)
if preprocess:
unaligned_en = [preprocess_sentence(x, "en", aligned=False) for x in unaligned_en]
unaligned_fr = [preprocess_sentence(x, "fr", aligned=False) for x in unaligned_fr]
return unaligned_en, unaligned_fr
def dataloader_aligned(preprocess=True, add_special_tag=True):
aligned_en = load_lang(cfg.aligned_en_path)
aligned_fr = load_lang(cfg.aligned_fr_path)
if preprocess:
aligned_en = [preprocess_sentence(x, "en", aligned=True, add_special_tag=add_special_tag) for x in aligned_en]
aligned_fr = [preprocess_sentence(x, "fr", aligned=True, add_special_tag=add_special_tag) for x in aligned_fr]
return aligned_en, aligned_fr
def dataloader_aligned_synthetic():
try:
aligned_en = load_lang(cfg.aligned_en_synth_path)
aligned_fr = load_lang(cfg.aligned_fr_synth_path)
aligned_en = [preprocess_sentence(x, "en", aligned=True) for x in aligned_en]
aligned_fr = [preprocess_sentence(x, "fr", aligned=True) for x in aligned_fr]
return aligned_en, aligned_fr
except Exception as e:
print(e)
print("No synthetic data present.")
def tokenize(aligned_lang, unaligned_lang, num_words):
lang_tokenizer = tf.keras.preprocessing.text.Tokenizer(filters=' ', lower=False, num_words=num_words)
lang_tokenizer.fit_on_texts(aligned_lang + unaligned_lang)
tensor = lang_tokenizer.texts_to_sequences(aligned_lang)
tensor = tf.keras.preprocessing.sequence.pad_sequences(tensor, padding='post')
return tensor, lang_tokenizer
def train_val_split(input_lang, output_lang):
return train_test_split(input_lang, output_lang, test_size=cfg.train_val_split_ratio,
random_state=cfg.random_seed_for_split)
def load_embeddings(emb_path):
try:
with open(cfg.root_path + emb_path, "rb") as f:
emb = pickle.load(f)
return emb
except:
print("Embeddings does not exist: ", emb_path)
def load_data(reverse_translate=False, add_synthetic_data=False, load_emb=False, inp_vocab_size=20000,
tar_vocab_size=20000, emb_size=128):
print("-- Loading Datafiles --")
unaligned_en, unaligned_fr = dataloader_unaligned()
aligned_en, aligned_fr = dataloader_aligned()
input_aligned_train, input_aligned_val, output_aligned_train, output_aligned_val = train_val_split(aligned_en,
aligned_fr)
print("Training Data: ", len(input_aligned_train))
print("Validation Data: ", len(input_aligned_val))
# To use for Bleu score in the end
aligned_en_org, aligned_fr_org = dataloader_aligned(preprocess=False)
_, input_aligned_val_org, _, output_aligned_val_org = train_val_split(aligned_en_org, aligned_fr_org)
print("-- Creating Vocabulary and Tokenizing --")
input_train, input_tokenizer = tokenize(input_aligned_train, unaligned_en, inp_vocab_size)
output_train, output_tokenizer = tokenize(output_aligned_train, unaligned_fr, tar_vocab_size)
if add_synthetic_data:
print("-- Adding Synthetic Data --")
synth_en, synth_fr = dataloader_aligned_synthetic()
synth_en = input_tokenizer.texts_to_sequences(synth_en)
synth_fr = output_tokenizer.texts_to_sequences(synth_fr)
synth_en = pad_sequences(synth_en, padding='post', maxlen=input_train.shape[1])
synth_fr = pad_sequences(synth_fr, padding='post', maxlen=output_train.shape[1])
input_train = np.concatenate((input_train, synth_en, input_train), axis=0)
output_train = np.concatenate((output_train, synth_fr, output_train), axis=0)
print("Updated Training Data: ", input_train.shape[0])
if load_emb:
print("-- Loading embedding --")
e_emb = load_embeddings(cfg.emb_path_en)
d_emb = load_embeddings(cfg.emb_path_fr)
else:
print("-- Skipping embeddings --")
e_emb, d_emb = None, None
if reverse_translate:
input_train, input_tokenizer, input_aligned_val, input_aligned_val_org, e_emb, output_train, \
output_tokenizer, output_aligned_val, output_aligned_val_org, d_emb = output_train, output_tokenizer, \
output_aligned_val, \
output_aligned_val_org, d_emb, \
input_train, input_tokenizer, \
input_aligned_val, \
input_aligned_val_org, e_emb
print("-- Tokenizing Validation set --")
input_val = input_tokenizer.texts_to_sequences(input_aligned_val)
input_val = pad_sequences(input_val, padding='post', maxlen=input_train.shape[1])
output_val = output_tokenizer.texts_to_sequences(output_aligned_val)
output_val = pad_sequences(output_val, padding='post', maxlen=output_train.shape[1])
return input_train, input_val, input_aligned_val_org, input_tokenizer, e_emb, output_train, output_val, \
output_aligned_val_org, output_tokenizer, d_emb