-
Notifications
You must be signed in to change notification settings - Fork 0
/
embed_subwords.py
229 lines (202 loc) · 7.85 KB
/
embed_subwords.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
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
import argparse
import json
import fasttext
import numpy as np
import os
from subprocess import call
import sys
sys.path.append('subword-embedding')
from subword_corpus import MLFDataset
from visualise import visualise_embedding, label_maps_from_file
sys.path.append('utils')
from utils import write_to_file
RES_DIR = 'results'
def parse_arguments(args_to_parse):
""" Parse the command line arguments.
Arguments:
args_to_parse: CLI arguments to parse
"""
description = "Generate a subword-based corpus from which to generate subword embeddings"
parser = argparse.ArgumentParser(description=description)
# General options
general = parser.add_argument_group('General options')
general.add_argument(
'-i', '--mlf-file', type=str, required=True,
help="Input MLF file from which to generate the corpus."
)
general.add_argument(
'--subword-corpus', type=str,
default='{}/subword-corpus.dat'.format(RES_DIR),
help='The path name of the subword corpus produced.'
)
general.add_argument(
'-u', '--unique-subwords', type=str,
default='{}/unique-subword-list.json'.format(RES_DIR),
help='Path name of the file where the the unique subword unit list will be saved.'
)
general.add_argument(
'-c', '--subword-context', type=int,
default=1, help='The subword context: monophone (1), biphone (2), triphone (3), etc'
)
general.add_argument(
'--loc-info', dest='subword_loc_info', action='store_true'
)
general.add_argument(
'--no-loc-info', dest='subword_loc_info', action='store_false'
)
general.set_defaults(subword_loc_info=False)
# Visualisation options
visuals = parser.add_argument_group('Visualisation options')
visuals.add_argument(
'--only-viz', dest='only_viz', action='store_true'
)
visuals.set_defaults(only_viz=False)
# Map the phone / grapheme names
visuals.add_argument(
'--no-map', dest='map_label', action='store_const', const=0
)
visuals.add_argument(
'--map-to-en', dest='map_label', action='store_const', const=1
)
visuals.add_argument(
'--map-to-native', dest='map_label', action='store_const', const=2
)
visuals.set_defaults(
map_label=0
)
visuals.add_argument(
'-v', '--emb-visual', type=str,
default='{}/embedding/visualisation.png'.format(RES_DIR),
help='The path name to the visualisation image of the subword embeddings'
)
visuals.add_argument(
'-s', '--summary-file', type=str,
default='data/summary.txt',
help='The path name to the summary file or saved map'
)
# Embedding options
embedding = parser.add_argument_group('Embedding options')
model_opts = ['word2vec', 'fastText']
embedding.add_argument(
'-m', '--model', type=str, default='word2vec', choices=model_opts,
help='The model to use: word2vec or fastText'
)
embedding.add_argument(
'-e', '--embedding', type=str,
default='{}/embedding/'.format(RES_DIR),
help='Path name of the embedding directory'
)
embedding.add_argument(
'-w2v', '--word2vec-dir', type=str,
default='/home/dawna/ar527/word2vec/word2vec',
help='Path to local word2vec directory'
)
embedding.add_argument(
'-l', '--vec-length', type=int, default=4,
help='The length of the vector representing each embedded subword unit.'
)
embedding.add_argument(
'--save-to-npy', dest='embed_to_npy', action='store_true'
)
embedding.set_defaults(embed_to_npy=False)
embedding.add_argument(
'--apostrophe-embedding', dest='apostrophe_embedding', action='store_true'
)
embedding.set_defaults(apostrophe_embedding=False)
# t-SNE representation
embedding.add_argument(
'-p', '--perplexity', type=float,
default=5,
help='The perplexity for t-SNE'
)
embedding.add_argument(
'-lr', '--learning-rate', type=int,
default=200,
help='The t-SNE learning rate.'
)
args = parser.parse_args(args_to_parse)
return args
def word2vec_embed(corpus_path, vector_length, target_dir, word2vec_dir):
call("subword-embedding/word2vec_embed.sh {} {} {} {}".format(
vector_length, target_dir, corpus_path, word2vec_dir), shell=True
)
def fasttext_embed(corpus_path, vector_length, unique_subwords_path, target_dir):
model = fasttext.train_unsupervised(corpus_path, dim=vector_length)
print('In-vocab subwords: {}'.format(model.words))
with open(unique_subwords_path) as unique_subwords_file:
unique_subwords = json.load(unique_subwords_file)
output_list = []
# Add header
output_list.append('{} {}'.format(len(model.words), vector_length))
unique_subwords = ['</s>'] + unique_subwords
for subword in unique_subwords:
embedding = ' '.join([str(element) for element in model[subword].tolist()])
line = '{} {}'.format(subword, embedding)
output_list.append(line)
# Check that nothing has been left out
for subword in model.words:
if not subword in unique_subwords:
# Need to add subword
print('Adding {}'.format(subword))
embedding = ' '.join([str(element) for element in model[subword].tolist()])
line = '{} {}'.format(subword, embedding)
output_list.append(line)
with open(os.path.join(target_dir, 'embedding.txt'), 'w') as embedding_file:
embedding_file.write('\n'.join(output_list))
def save_embedding_to_npy(embedding_dict,
npy_embedding_file='{}/embedding/embedding.npy'.format(RES_DIR)):
for key, val in embedding_dict.items():
embedding_dict[key] = np.array(val)
np.save(npy_embedding_file, embedding_dict)
def main(args):
""" Primary point of entry for generating sub-word (phone or grapheme) level embeddings
"""
if not args.only_viz:
if not os.path.exists(RES_DIR):
os.makedirs(RES_DIR)
print('Generating corpus...')
subword_dataset = MLFDataset(
path_to_mlf=args.mlf_file,
subword_context_width=args.subword_context,
incl_posn_info=args.subword_loc_info,
separate_apostrophe_embedding=args.apostrophe_embedding
)
subword_dataset.save_unique_subwords(target_file=args.unique_subwords)
write_to_file(content=subword_dataset.corpus(), target=args.subword_corpus)
print('Using {} to generate embeddings...'.format(args.model))
if args.model == 'word2vec':
word2vec_embed(
corpus_path=args.subword_corpus,
vector_length=args.vec_length,
target_dir=args.embedding,
word2vec_dir=args.word2vec_dir
)
elif args.model == 'fastText':
if not os.path.exists(args.embedding):
os.makedirs(args.embedding)
fasttext_embed(
corpus_path=args.subword_corpus,
vector_length=args.vec_length,
unique_subwords_path=args.unique_subwords,
target_dir=args.embedding,
)
print('Creating visualisation using t-SNE...')
label_map = label_maps_from_file(
path_to_summary=args.summary_file,
label_mapping_code=args.map_label,
separate_apostrophe_embedding=args.apostrophe_embedding,
saved_dict=False
)
embedding_dict = visualise_embedding(
embedding_dir=os.path.join(args.embedding, 'embedding.txt'),
perplexity=args.perplexity,
learning_rate=args.learning_rate,
image_path_name=args.emb_visual,
label_mapping=label_map
)
if args.embed_to_npy:
save_embedding_to_npy(embedding_dict)
print('Complete')
if __name__=='__main__':
args = parse_arguments(sys.argv[1:])
main(args)