-
Notifications
You must be signed in to change notification settings - Fork 1
/
piecelearn.py
177 lines (148 loc) · 5.59 KB
/
piecelearn.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
"""Learn wordpiece embeddings."""
import sentencepiece as spm
import logging
import numpy as np
from gensim.models import Word2Vec
NO_OPT = {"input", "model_prefix", "vocab_size", "model_type"}
logger = logging.getLogger(__name__)
logger.setLevel(logging.INFO)
class Sentences(object):
"""
An iterator over an arbitrary generator expression.
adapted from here:
https://jacopofarina.eu/posts/gensim-generator-is-not-iterator/
"""
def __init__(self, filename, sentencepieceprocessor):
"""
parameters
----------
generator_expression : generator
A generator.
"""
self.filename = filename
self.sentencepieceprocessor = sentencepieceprocessor
def __iter__(self):
"""Start from a new iterator over the generator."""
self.generator_expression = open(self.filename)
return self
def __next__(self):
return self.sentencepieceprocessor.encode_as_pieces(
next(self.generator_expression)
)
def reorder_embeddings(sp, wv):
"""reorder embeddings."""
vecs = wv.vectors
# Create new random vectors with same mean and std
mean, std = vecs.mean(0), vecs.std(0)
new_vecs = np.random.normal(mean, std, size=(len(sp), vecs.shape[1]))
new_indices = np.asarray([sp.piece_to_id(x) for x in wv.index_to_key])
mask = new_indices != 0
masked = new_indices[mask]
assert np.all(np.unique(masked, return_counts=True)[1] == 1)
new_vecs[masked] = vecs[mask]
return [sp.id_to_piece(x) for x in range(len(sp))], new_vecs
def train_spm(path_to_corpus, modelname, vocab_size=30000, model_type="bpe", **kwargs):
"""
Train an SPM model.
This simply calls the `spm_train` command, with all flags passed in as a
string. This function thus simply assembles such a string from the
path_to_corpus, modelname, vocab_size arguments, and any other arguments
the user passes in through kwargs.
Make sure all the kwargs that are passed in match the names of the
arguments of spm_train exactly, no checking is done.
Parameters
----------
path_to_corpus : str
The path to the input corpus.
modelname : str
The name of the spm model.
vocab_size : int, default 30000
The number of wordpieces to make
"""
intersection = set(kwargs.keys()) & NO_OPT
if intersection:
raise ValueError(f"{intersection} can not be assigned via kwargs")
opts = [
f"--input={path_to_corpus}",
f"--model_prefix={modelname}",
f"--vocab_size={vocab_size}",
f"--model_type={model_type}",
]
for k, v in kwargs.items():
opts.append(f"--{k}={v}")
logger.info(f"spm options: {opts}")
spm.SentencePieceTrainer.Train(" ".join(opts))
logger.info("Finished training sentencepiece model")
def train_word2vec(lines, spm_path, **kwargs):
"""
Train a Word2Vec model on lines with an encoded model.
All parameters are exposed through kwargs. For a full list of parameters,
see the gensim Word2Vec documentation:
https://radimrehurek.com/gensim/models/word2vec.html
Parameters
----------
lines : generator
A generator expression over lines. Usually passing open(filename) is
sufficient.
spm_path : str
The path to the model file created by the spm model.
"""
sp = spm.SentencePieceProcessor()
# return value is boolean.
sp.load(spm_path)
s = Sentences(lines, sp)
model = Word2Vec(min_count=0, **kwargs)
logger.info("building vocab")
model.build_vocab(s)
logger.info("starting training")
model.train(s, total_examples=model.corpus_count, epochs=model.epochs)
logger.info("done training")
return reorder_embeddings(sp, model.wv)
def train(
path_to_corpus,
spm_model_name,
vocab_size,
word2vec_path,
spm_kwargs=None,
w2v_kwargs=None,
):
"""
Train an spm and a word2vec model.
Parameters
----------
path_to_corpus : str
Path to a single text file, with one sentence/document per line.
Any preprocessing to this file must be done beforehand, e.g, neither
SPM nor word2vec lowercase sentences.
spm_model_name : str
The model name under which to save the sentencepiece model.
vocab_size : int
The vocab size of the spm model.
spm_kwargs : dict
Extra command line options to pass to the spm_train command. This takes
the form of a dict. Command line options should be spelled exactly as
in the official documentation. For a list of options, see:
https://github.com/google/sentencepiece#train-sentencepiece-model
or run `spm_train --help`
Note that the model name, vocab size, input file and model type
parameters are set by other flags or hardcoded.
word2vec_path : str
The path to which to save the word2vec model as a .vec file.
w2v_kwargs : dict
The kwargs to pass to the word2vec initialization function.
The min_count parameter is always set to 0, because we want to have
vectors for all our wordpieces.
"""
if spm_kwargs is None:
spm_kwargs = {}
if w2v_kwargs is None:
w2v_kwargs = {}
train_spm(path_to_corpus, spm_model_name, vocab_size=vocab_size, **spm_kwargs)
words, vectors = train_word2vec(
path_to_corpus, f"{spm_model_name}.model", **w2v_kwargs
)
with open(word2vec_path, "w") as f:
shape = " ".join([str(x) for x in vectors.shape])
f.write(f"{shape}\n")
for word, vec in zip(words, vectors):
f.write(f"{word} {' '.join(str(x) for x in vec)}\n")