/
models.py
180 lines (160 loc) · 9.01 KB
/
models.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
from ...imports import *
from ... import utils as U
from . import preprocessor as pp
from .anago.models import BiLSTMCRF
BILSTM_CRF = 'bilstm-crf'
BILSTM = 'bilstm'
BILSTM_ELMO = 'bilstm-elmo'
BILSTM_CRF_ELMO = 'bilstm-crf-elmo'
BILSTM_TRANSFORMER = 'bilstm-bert'
SEQUENCE_TAGGERS = {
BILSTM: 'Bidirectional LSTM (https://arxiv.org/abs/1603.01360)',
BILSTM_TRANSFORMER: 'Bidirectional LSTM w/ BERT embeddings',
BILSTM_CRF: 'Bidirectional LSTM-CRF (https://arxiv.org/abs/1603.01360)',
BILSTM_ELMO: 'Bidirectional LSTM w/ Elmo embeddings [English only]',
BILSTM_CRF_ELMO: 'Bidirectional LSTM-CRF w/ Elmo embeddings [English only]',
}
V1_ONLY_MODELS = [BILSTM_CRF, BILSTM_CRF_ELMO]
TRANSFORMER_MODELS = [BILSTM_TRANSFORMER]
ELMO_MODELS = [BILSTM_ELMO, BILSTM_CRF_ELMO]
def print_sequence_taggers():
for k,v in SEQUENCE_TAGGERS.items():
print("%s: %s" % (k,v))
def sequence_tagger(name, preproc,
wv_path_or_url=None,
bert_model = 'bert-base-multilingual-cased',
bert_layers_to_use = U.DEFAULT_TRANSFORMER_LAYERS,
word_embedding_dim=100,
char_embedding_dim=25,
word_lstm_size=100,
char_lstm_size=25,
fc_dim=100,
dropout=0.5,
verbose=1):
"""
Build and return a sequence tagger (i.e., named entity recognizer).
Args:
name (string): one of:
- 'bilstm-crf' for Bidirectional LSTM-CRF model
- 'bilstm' for Bidirectional LSTM (no CRF layer)
preproc(NERPreprocessor): an instance of NERPreprocessor
wv_path_or_url(str): either a URL or file path toa fasttext word vector file (.vec or .vec.zip or .vec.gz)
Example valid values for wv_path_or_url:
Randomly-initialized word embeeddings:
set wv_path_or_url=None
English pretrained word vectors:
https://dl.fbaipublicfiles.com/fasttext/vectors-english/crawl-300d-2M.vec.zip
Chinese pretrained word vectors:
https://dl.fbaipublicfiles.com/fasttext/vectors-crawl/cc.zh.300.vec.gz
Russian pretrained word vectors:
https://dl.fbaipublicfiles.com/fasttext/vectors-crawl/cc.ru.300.vec.gz
Dutch pretrained word vectors:
https://dl.fbaipublicfiles.com/fasttext/vectors-crawl/cc.nl.300.vec.gz
See these two Web pages for a full list of URLs to word vector files for
different languages:
1. https://fasttext.cc/docs/en/english-vectors.html (for English)
2. https://fasttext.cc/docs/en/crawl-vectors.html (for non-English langages)
Default:None (randomly-initialized word embeddings are used)
bert_model_name(str): the name of the BERT model. default: 'bert-base-multilingual-cased'
This parameter is only used if bilstm-bert is selected for name parameter.
The value of this parameter is a name of BERT model from here:
https://huggingface.co/transformers/pretrained_models.html
or a community-uploaded BERT model from here:
https://huggingface.co/models
Example values:
bert-base-multilingual-cased: Multilingual BERT (157 languages) - this is the default
bert-base-cased: English BERT
bert-base-chinese: Chinese BERT
distilbert-base-german-cased: German DistilBert
albert-base-v2: English ALBERT model
monologg/biobert_v1.1_pubmed: community uploaded BioBERT (pretrained on PubMed)
bert_layers_to_use(list): indices of hidden layers to use. default:[-2] # second-to-last layer
To use the concatenation of last 4 layers: use [-1, -2, -3, -4]
word_embedding_dim (int): word embedding dimensions.
char_embedding_dim (int): character embedding dimensions.
word_lstm_size (int): character LSTM feature extractor output dimensions.
char_lstm_size (int): word tagger LSTM output dimensions.
fc_dim (int): output fully-connected layer size.
dropout (float): dropout rate.
verbose (boolean): verbosity of output
Return:
model (Model): A Keras Model instance
"""
if name not in SEQUENCE_TAGGERS:
raise ValueError('invalid name: %s' % (name))
# check BERT
if name in TRANSFORMER_MODELS and not bert_model:
raise ValueError('bert_model is required for bilstm-bert models')
if name in TRANSFORMER_MODELS and DISABLE_V2_BEHAVIOR:
raise ValueError('BERT and other transformer models cannot be used with DISABLE_v2_BEHAVIOR')
# check CRF
if not DISABLE_V2_BEHAVIOR and name in V1_ONLY_MODELS:
warnings.warn('Falling back to BiLSTM (no CRF) because DISABLE_V2_BEHAVIOR=False')
msg = "\nIMPORTANT NOTE: ktrain uses the CRF module from keras_contrib, which is not yet\n" +\
"fully compatible with TensorFlow 2. You can still use the BiLSTM-CRF model\n" +\
"in ktrain for sequence tagging with TensorFlow 2, but you must add the\n" +\
"following to the top of your script or notebook BEFORE you import ktrain:\n\n" +\
"import os\n" +\
"os.environ['DISABLE_V2_BEHAVIOR'] = '1'\n\n" +\
"For this run, a vanilla BiLSTM model (with no CRF layer) will be used.\n"
print(msg)
name = BILSTM if name == BILSTM_CRF else BILSTM_ELMO
# check for use_char=True
if not DISABLE_V2_BEHAVIOR and preproc.p._use_char:
# turn off masking due to open TF2 issue ##33148: https://github.com/tensorflow/tensorflow/issues/33148
warnings.warn('Setting use_char=False: character embeddings cannot be used in TF2 due to open TensorFlow 2 bug (#33148).\n' +\
'Add os.environ["DISABLE_V2_BEHAVIOR"] = "1" to the top of script if you really want to use it.')
preproc.p._use_char=False
if verbose:
emb_names = []
if wv_path_or_url is not None:
emb_names.append('word embeddings initialized with fasttext word vectors (%s)' % (os.path.basename(wv_path_or_url)))
else:
emb_names.append('word embeddings initialized randomly')
if name in TRANSFORMER_MODELS: emb_names.append('BERT embeddings with ' + bert_model)
if name in ELMO_MODELS: emb_names.append('Elmo embeddings for English')
if preproc.p._use_char: emb_names.append('character embeddings')
if len(emb_names) > 1:
print('Embedding schemes employed (combined with concatenation):')
else:
print('embedding schemes employed:')
for emb_name in emb_names:
print('\t%s' % (emb_name))
print()
# setup embedding
if wv_path_or_url is not None:
wv_model, word_embedding_dim = preproc.get_wv_model(wv_path_or_url, verbose=verbose)
else:
wv_model = None
if name == BILSTM_CRF:
use_crf = False if not DISABLE_V2_BEHAVIOR else True # fallback to bilstm
elif name == BILSTM_CRF_ELMO:
use_crf = False if not DISABLE_V2_BEHAVIOR else True # fallback to bilstm
preproc.p.activate_elmo()
elif name == BILSTM:
use_crf = False
elif name == BILSTM_ELMO:
use_crf = False
preproc.p.activate_elmo()
elif name == BILSTM_TRANSFORMER:
use_crf = False
preproc.p.activate_transformer(bert_model, layers=bert_layers_to_use, force=True)
else:
raise ValueError('Unsupported model name')
model = BiLSTMCRF(char_embedding_dim=char_embedding_dim,
word_embedding_dim=word_embedding_dim,
char_lstm_size=char_lstm_size,
word_lstm_size=word_lstm_size,
fc_dim=fc_dim,
char_vocab_size=preproc.p.char_vocab_size,
word_vocab_size=preproc.p.word_vocab_size,
num_labels=preproc.p.label_size,
dropout=dropout,
use_crf=use_crf,
use_char=preproc.p._use_char,
embeddings=wv_model,
use_elmo=preproc.p.elmo_is_activated(),
use_transformer_with_dim=preproc.p.get_transformer_dim())
model, loss = model.build()
model.compile(loss=loss, optimizer=U.DEFAULT_OPT)
return model