NeuralMachineTranslation/utils.py at master · gugundi/NeuralMachineTranslation · GitHub

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
from device import with_cpu
import itertools
import os
import pandas as pd
import random
from sklearn.model_selection import train_test_split
import torchtext


def get_or_create_dir(base_path, dir_name):
    out_directory = os.path.join(base_path, dir_name)
    if not os.path.exists(out_directory):
        os.makedirs(out_directory)
    return out_directory


def create_debug_csv():
    n_lines = 1000
    with open('.data/iwslt/de-en/train.de-en.de') as f:
        source = itertools.islice(f, n_lines)
        source = map(lambda sentence: sentence.replace('\n', ''), source)
        source = list(source)
    with open('.data/iwslt/de-en/train.de-en.en') as f:
        target = itertools.islice(f, n_lines)
        target = map(lambda sentence: sentence.replace('\n', ''), target)
        target = list(target)
    data = {"src": source, "trg": target}
    dataframe = pd.DataFrame(data, columns=['src', 'trg'])
    train, val = train_test_split(dataframe, test_size=0.1)
    train.to_csv('.data/debug/train.csv', index=False)
    val.to_csv('.data/debug/val.csv', index=False)


def create_dummy_fixed_length_csv():
    n_observations = 10000
    source_length = 10
    target_length = 5
    condition = 5
    max_int = 10
    src = []
    trg = []
    for i in range(n_observations):
        positions = range(source_length)
        source = [random.randint(condition, max_int) for _ in range(source_length)]
        for j in range(target_length):
            position = random.choice(positions)
            source[position] = random.randint(1, condition - 1)
            positions = list(set(positions) - set([position]))
        target = filter(lambda x: x < condition, source)
        source = map(str, source)
        target = map(str, target)
        src.append(" ".join(source))
        trg.append(" ".join(target))
    data = {"src": src, "trg": trg}
    dataframe = pd.DataFrame(data, columns=['src', 'trg'])
    train, val = train_test_split(dataframe, test_size=0.1)
    train.to_csv('.data/dummy_fixed_length/train.csv', index=False)
    val.to_csv('.data/dummy_fixed_length/val.csv', index=False)


def create_iwslt():
    with open('.data/iwslt/de-en/train.de-en.de') as f:
        source = f.readlines()
        source = map(lambda sentence: sentence.replace('\n', ''), source)
        source = filter(lambda sentence: sentence != '', source)
        source = list(source)
    with open('.data/iwslt/de-en/train.de-en.en') as f:
        target = f.readlines()
        target = map(lambda sentence: sentence.replace('\n', ''), target)
        target = filter(lambda sentence: sentence != '', target)
        target = list(target)
    data = {"src": source, "trg": target}
    dataframe = pd.DataFrame(data, columns=['src', 'trg'])
    train, val = train_test_split(dataframe, test_size=0.1)
    train.to_csv('.data/iwslt/train.csv', index=False)
    val.to_csv('.data/iwslt/val.csv', index=False)


def create_multi30k():
    with open('.data/multi30k/train.en') as f:
        train_src = f.readlines()
        train_src = map(lambda sentence: sentence.replace('\n', ''), train_src)
        train_src = filter(lambda sentence: sentence != '', train_src)
        train_src = list(train_src)
    with open('.data/multi30k/train.de') as f:
        train_trg = f.readlines()
        train_trg = map(lambda sentence: sentence.replace('\n', ''), train_trg)
        train_trg = filter(lambda sentence: sentence != '', train_trg)
        train_trg = list(train_trg)
    train = {"src": train_src, "trg": train_trg}
    train = pd.DataFrame(train, columns=['src', 'trg'])

    with open('.data/multi30k/val.en') as f:
        val_src = f.readlines()
        val_src = map(lambda sentence: sentence.replace('\n', ''), val_src)
        val_src = filter(lambda sentence: sentence != '', val_src)
        val_src = list(val_src)
    with open('.data/multi30k/val.de') as f:
        val_trg = f.readlines()
        val_trg = map(lambda sentence: sentence.replace('\n', ''), val_trg)
        val_trg = filter(lambda sentence: sentence != '', val_trg)
        val_trg = list(val_trg)
    val = {"src": val_src, "trg": val_trg}
    val = pd.DataFrame(val, columns=['src', 'trg'])

    with open('.data/multi30k/test2016.en') as f:
        test_src = f.readlines()
        test_src = map(lambda sentence: sentence.replace('\n', ''), test_src)
        test_src = filter(lambda sentence: sentence != '', test_src)
        test_src = list(test_src)
    with open('.data/multi30k/test2016.de') as f:
        test_trg = f.readlines()
        test_trg = map(lambda sentence: sentence.replace('\n', ''), test_trg)
        test_trg = filter(lambda sentence: sentence != '', test_trg)
        test_trg = list(test_trg)
    test = {"src": test_src, "trg": test_trg}
    test = pd.DataFrame(test, columns=['src', 'trg'])

    train.to_csv('.data/multi30k/train.csv', index=False)
    val.to_csv('.data/multi30k/val.csv', index=False)
    test.to_csv('.data/multi30k/test.csv', index=False)


def create_dummy_variable_length_csv():
    n_observations = 10000
    min_source_length = 3
    max_source_length = 15
    condition = 5
    max_int = 10
    src = []
    trg = []
    for i in range(n_observations):
        source_length = random.randint(min_source_length, max_source_length)
        source = [random.randint(1, max_int) for _ in range(source_length)]
        target = filter(lambda x: x < condition, source)
        source = map(str, source)
        target = map(str, target)
        src.append(" ".join(source))
        trg.append(" ".join(target))
    data = {"src": src, "trg": trg}
    dataframe = pd.DataFrame(data, columns=['src', 'trg'])
    train, val = train_test_split(dataframe, test_size=0.1)
    train.to_csv('.data/dummy_variable_length/train.csv', index=False)
    val.to_csv('.data/dummy_variable_length/val.csv', index=False)


def load_from_csv(config, csv_dir_path, source_tokenizer, target_tokenizer, device):
    print(f'Data loader: Started ({csv_dir_path}).')

    EOS_token = config.get('EOS_token')
    PAD_token = config.get('PAD_token')
    SOS_token = config.get('SOS_token')

    source_field = torchtext.data.Field(
        tokenize=source_tokenizer,
        init_token=SOS_token,
        eos_token=EOS_token,
        pad_token=PAD_token,
        include_lengths=True
    )
    target_field = torchtext.data.Field(
        tokenize=target_tokenizer,
        init_token=SOS_token,
        eos_token=EOS_token,
        pad_token=PAD_token,
        include_lengths=True
    )
    data_fields = [('src', source_field), ('trg', target_field)]

    print('Data loader: Making splits.')
    train, val = torchtext.data.TabularDataset.splits(
        path=csv_dir_path,
        train='train.csv',
        validation='val.csv',
        format='csv',
        fields=data_fields,
        skip_header=True
    )
    source_vocabulary_size = config.get('source_vocabulary_size')
    target_vocabulary_size = config.get('target_vocabulary_size')

    print('Data loader: Building vocabulary.')
    source_field.build_vocab(train, val, max_size=source_vocabulary_size)
    target_field.build_vocab(train, val, max_size=target_vocabulary_size)

    print('Data loader: Iterator splits splits.')
    train_iter, val_iter = torchtext.data.BucketIterator.splits(
        (train, val),
        batch_size=config.get('batch_size'),
        device=device,
        shuffle=True,
        sort_key=lambda x: len(x.src)
    )

    print('Data loader: Finished.')

    return train_iter, val_iter, source_field.vocab, target_field.vocab, val


def list2words(language, sentence):
    sentence = map(lambda idx: language.itos[idx], sentence)
    return list(sentence)


def torch2words(language, sentence):
    sentence = sentence.squeeze()
    sentence = with_cpu(sentence)
    sentence = map(lambda idx: language.itos[idx], sentence)
    return list(sentence)


def filter_words(words, SOS_token, EOS_token, PAD_token):
    return filter(lambda word: word != SOS_token and word != EOS_token and word != PAD_token, words)


def words2text(words, SOS_token, EOS_token, PAD_token):
    sentence = filter_words(words, SOS_token, EOS_token, PAD_token)
    sentence = " ".join(sentence)
    return sentence


def get_text(source_words, target_words, translation_words, SOS_token, EOS_token, PAD_token):
    source = words2text(source_words, SOS_token, EOS_token, PAD_token)
    target = words2text(target_words, SOS_token, EOS_token, PAD_token)
    translation = words2text(translation_words, SOS_token, EOS_token, PAD_token)
    return f"""
    Source: \"{source}\"
    Target: \"{target}\"
    Translation: \"{translation}\"
    """