def build_and_train_conll03en_flair_sequence_tagger(corpus,tag_type,tag_dictionary):
'''
do not change!
same configuration as described in
file: "flair/resources/docs/EXPERIMENTS.md"
section: "CoNLL-03 Named Entity Recognition (English)"
'''
embeddings: StackedEmbeddings = StackedEmbeddings(
embeddings=[
WordEmbeddings("glove"),
PooledFlairEmbeddings("news-forward", pooling="min"),
PooledFlairEmbeddings("news-backward", pooling="min"),
]
)
from flair.models import SequenceTagger
tagger: SequenceTagger = SequenceTagger(
hidden_size=256,
embeddings=embeddings,
tag_dictionary=tag_dictionary,
tag_type=tag_type,
)
from flair.trainers import ModelTrainer
corpus = Corpus(train=corpus.train, dev=corpus.dev,test=[])
trainer: ModelTrainer = ModelTrainer(tagger, corpus)
# trainer.train("resources/taggers/example-ner", train_with_dev=True, max_epochs=150) # original
trainer.train("flair_checkpoints", train_with_dev=False, max_epochs=40,save_final_model=False) # original
return tagger
def main(base_path, output_dir, nb_epochs):
# parser = argparse.ArgumentParser()
# parser.add_argument("--data_dir", default='./', type=str, required=True, help="The parent dir of inpu data, must contain folder name `conll_03`")
# parser.add_argument("--output_dir", default=None, required=True, help="The output directory where is going to store the trained model")
# parser.add_argument("--train_epochs", default=3, type=int, required=True, help="Number of epochs to train")
# args = parser.parse_args()
# base_path = args.data_dir
corpus: Corpus = CONLL_03(base_path=base_path)
tag_type = 'ner'
tag_dictionary = corpus.make_tag_dictionary(tag_type=tag_type)
embedding_types: List[TokenEmbeddings] = [
WordEmbeddings('glove'),
PooledFlairEmbeddings('news-forward', pooling='min'),
PooledFlairEmbeddings('news-backward', pooling='min'),
]
embeddings: StackedEmbeddings = StackedEmbeddings(
embeddings=embedding_types)
tagger: SequenceTagger = SequenceTagger(hidden_size=256,
embeddings=embeddings,
tag_dictionary=tag_dictionary,
tag_type=tag_type)
trainer: ModelTrainer = ModelTrainer(tagger, corpus)
# output_dir = args.output_dir
# nb_epochs = args.train_epochs
# output_dir =
# nb_epochs = 10
trainer.train(output_dir, train_with_dev=False,
max_epochs=nb_epochs) # 150
def load_flair(mode = 'flair'):
if mode == 'flair':
stacked_embeddings = StackedEmbeddings([
WordEmbeddings('glove'),
PooledFlairEmbeddings('news-forward', pooling='min'),
PooledFlairEmbeddings('news-backward', pooling='min')
])
else:##bert
stacked_embeddings = BertEmbeddings('bert-base-uncased') ##concat last 4 layers give the best
return stacked_embeddings
def __init__(self, hidden_dim: int, rnn_type: str, vocab_size: int,
tagset_size: int, task_type: str):
super(TaskLearner, self).__init__()
self.task_type = task_type
self.rnn_type = rnn_type
self.bidirectional = True
self.num_layers = 2
self.num_directions = 2 if self.bidirectional else 1
# Word Embeddings (TODO: Implement pre-trained word embeddings)
# self.word_embeddings = nn.Embedding(num_embeddings=vocab_size, embedding_dim=embedding_dim) # TODO: Implement padding_idx=self.pad_idx
embedding_types: List[TokenEmbeddings] = [
WordEmbeddings('glove'),
PooledFlairEmbeddings('news-forward', pooling='min'),
PooledFlairEmbeddings('news-backward', pooling='min')
]
embeddings: StackedEmbeddings = StackedEmbeddings(
embeddings=embedding_types)
self.embeddings = embeddings
self.embedding_dim: int = self.embeddings.embedding_length
if self.rnn_type == 'gru':
rnn = nn.GRU
elif self.rnn_type == 'lstm':
rnn = nn.LSTM
elif self.rnn_type == 'rnn':
rnn = nn.RNN
else:
raise ValueError
# Sequence tagger
self.rnn = rnn(input_size=self.embedding_dim,
hidden_size=hidden_dim,
num_layers=self.num_layers,
dropout=0.0 if self.num_layers == 1 else 0.5,
bidirectional=self.bidirectional,
batch_first=True)
if self.task_type == 'SEQ':
# Linear layer that maps hidden state space from rnn to tag space
self.hidden2tag = nn.Linear(in_features=hidden_dim *
self.num_directions,
out_features=tagset_size)
if self.task_type == 'CLF':
# COME BACK LATER...
self.drop = nn.Dropout(p=0.5)
self.hidden2tag = nn.Linear(in_features=hidden_dim *
self.num_directions,
out_features=1)
def train_flair(self):
# Flair Model Initialisation and Training
# # 1. get the corpus
# corpus: Corpus = ColumnCorpus(os.path.join(os.getcwd(), 'results', '10'),
# {0: 'text', 1: 'ner'},
# train_file='train.txt',
# test_file='test.txt',
# dev_file='valid.txt',
# column_delimiter=' ')
corpus: Corpus = CONLL_03(
base_path=os.path.join(os.getcwd(), 'results', '10'))
corpus.dev_file = 'valid.txt' # rather than 'dev.txt'
# 2. what tag do we want to predict?
tag_type = 'ner'
# 3. make the tag dictionary from the corpus
tag_dictionary = corpus.make_tag_dictionary(tag_type=tag_type)
# initialize embeddings
embedding_types: List[TokenEmbeddings] = [
# GloVe embeddings
WordEmbeddings('glove'),
# contextual string embeddings, forward
PooledFlairEmbeddings('news-forward', pooling='min'),
# contextual string embeddings, backward
PooledFlairEmbeddings('news-backward', pooling='min'),
]
embeddings: StackedEmbeddings = StackedEmbeddings(
embeddings=embedding_types)
# initialize sequence tagger
tagger: SequenceTagger = SequenceTagger(hidden_size=256,
embeddings=embeddings,
tag_dictionary=tag_dictionary,
tag_type=tag_type)
# initialize trainer
trainer: ModelTrainer = ModelTrainer(tagger, corpus)
results = trainer.train(os.path.join(os.getcwd(), 'results', '10',
'tagger'),
train_with_dev=False,
max_epochs=50)
print(results)
def post_init(self):
import flair
flair.device = self.device
from flair.embeddings import WordEmbeddings, FlairEmbeddings, BytePairEmbeddings, PooledFlairEmbeddings, \
DocumentPoolEmbeddings
embeddings_list = []
for e in self.embeddings:
model_name, model_id = e.split(':', maxsplit=1)
emb = None
try:
if model_name == 'flair':
emb = FlairEmbeddings(model_id)
elif model_name == 'pooledflair':
emb = PooledFlairEmbeddings(model_id)
elif model_name == 'word':
emb = WordEmbeddings(model_id)
elif model_name == 'byte-pair':
emb = BytePairEmbeddings(model_id)
except ValueError:
self.logger.error(f'embedding not found: {e}')
continue
if emb is not None:
embeddings_list.append(emb)
if embeddings_list:
self.model = DocumentPoolEmbeddings(embeddings_list,
pooling=self.pooling_strategy)
self.logger.info(
f'flair encoder initialized with embeddings: {self.embeddings}'
)
else:
self.logger.error('flair encoder initialization failed.')
def post_init(self):
from flair.embeddings import WordEmbeddings, FlairEmbeddings, BytePairEmbeddings, PooledFlairEmbeddings, \
DocumentPoolEmbeddings
if self.model is not None:
return
embeddings_list = []
for e in self.embeddings:
model_name, model_id = e.split(':', maxsplit=1)
emb = None
try:
if model_name == 'flair':
emb = FlairEmbeddings(model_id)
elif model_name == 'pooledflair':
emb = PooledFlairEmbeddings(model_id)
elif model_name == 'word':
emb = WordEmbeddings(model_id)
elif model_name == 'byte-pair':
emb = BytePairEmbeddings(model_id)
except ValueError:
self.logger.error('embedding not found: {}'.format(e))
continue
if emb is not None:
embeddings_list.append(emb)
if embeddings_list:
self.model = DocumentPoolEmbeddings(embeddings_list,
pooling=self.pooling_strategy)
self.logger.info(
'initialize flair encoder with embeddings: {}'.format(
self.embeddings))
def _train(
self,
output_dir: Union[str, Path],
corpus: Optional[ColumnCorpus] = None,
tagger: Optional[SequenceTagger] = None,
hidden_size: int = 256,
learning_rate: float = 0.1,
mini_batch_size: int = 32,
max_epochs: int = 100,
use_crf: bool = True,
) -> SequenceTagger:
tag_dictionary = corpus.make_tag_dictionary(tag_type="ner")
if not tagger:
tagger = SequenceTagger(
hidden_size=hidden_size,
embeddings=PooledFlairEmbeddings("news-forward"),
tag_dictionary=tag_dictionary,
tag_type="ner",
use_crf=use_crf,
)
trainer = ModelTrainer(tagger, corpus)
trainer.train(
output_dir,
learning_rate=learning_rate,
mini_batch_size=mini_batch_size,
max_epochs=max_epochs,
)
model_path = Path(output_dir, "best-model.pt")
return SequenceTagger.load(model_path)
def use_flair_to_extract_context_embeddings(file, dest_folder, embedding_type, embedding_size, pretrained_model=None):
if embedding_type.lower() == 'elmo':
context_embedding = ELMoEmbeddings(model='pubmed')
elif embedding_type.lower() == 'elmo_transformer':
context_embedding = ELMoTransformerEmbeddings()
elif embedding_type.lower() == 'flair':
context_embedding = PooledFlairEmbeddings()
elif embedding_type.lower() == 'bioflair':
flair_1 = PooledFlairEmbeddings('pubmed-forward')
flair_2 = PooledFlairEmbeddings('pubmed-backward')
elmo = ELMoEmbeddings(model='pubmed')
#bert = BertEmbeddings(bert_model_or_path='bert-base-multilingual-cased', layers='-1')
context_embedding = StackedEmbeddings(embeddings=[flair_1, flair_2, elmo])
elif embedding_type.lower() == 'biobert' or embedding_type.lower() == 'bert':
context_embedding = BertEmbeddings(bert_model_or_path=pretrained_model, layers='-1')
data = {}
dest_name = os.path.basename(file).split('.')
print(dest_folder)
with open(file, 'r') as f, open('{}/{}.pickle'.format(dest_folder, dest_name[0]), 'wb') as d:
sentence = ''
instance = []
j = 0
for i in f.readlines():
if i != '\n':
i = i.split()
sentence += ' '+i[0]
elif i == '\n':
sent = Sentence(sentence.strip())
context_embedding.embed(sent)
v = ''
for i in sent:
instance.append((i.text, i.embedding[:embedding_size]))
sentence = ''
if instance:
data[j] = list(zip(*(instance.copy())))
j += 1
instance.clear()
pickle.dump(data, d)
f.close()
d.close()
def get_embeddings(pooling_op='min'):
return StackedEmbeddings(embeddings=[
# pre-trained embeddings
PooledFlairEmbeddings(
'es-forward',
pooling=pooling_op,
),
PooledFlairEmbeddings(
'es-backward',
pooling=pooling_op,
),
BytePairEmbeddings(
language='es',
dim=300,
),
# self-trained embeddings
SpanishHealthCorpusEmbeddings('wang2vec'),
# SpanishHealthCorpusEmbeddings('fastText'),
])
if args.task == 'ner':
flair_corpus = NLPTask.CONLL_03
tag_type = 'ner'
embedding_types = [WordEmbeddings('glove')]
else:
flair_corpus = NLPTask.CONLL_2000
tag_type = 'np'
embedding_types = [WordEmbeddings('extvec')]
_base_path = 'resources/taggers/{}-{}'.format(args.task, args.model_path)
if args.use_flair_embeddings is True:
embedding_types.extend([
# contextual string embeddings, forward
PooledFlairEmbeddings('news-forward'),
# contextual string embeddings, backward
PooledFlairEmbeddings('news-backward')
])
if args.model_path != 'NA':
embedding_types.append(EyeTrackingFeatureEmbedding(args.model_path))
corpus = NLPTaskDataFetcher.load_corpus(flair_corpus,
base_path=TASK_DATASET_DIR)
tag_dictionary = corpus.make_tag_dictionary(tag_type=tag_type)
embeddings = StackedEmbeddings(embeddings=embedding_types)
tagger = SequenceTagger(hidden_size=256,
embeddings=embeddings,
tag_dictionary=tag_dictionary,
tag_type=tag_type)
search_space = SearchSpace()
#Create or embedding stacks
#Flair recommends adding GLoVe to their character-level embeddings
flair_normal = StackedEmbeddings([
WordEmbeddings('glove'),
FlairEmbeddings('mix-forward'),
FlairEmbeddings('mix-backward')
])
bert = BertEmbeddings()
elmo = ELMoEmbeddings('original')
flair_pooled = StackedEmbeddings([
WordEmbeddings('glove'),
PooledFlairEmbeddings('mix-forward'),
PooledFlairEmbeddings('mix-backward')
])
search_space.add(Parameter.EMBEDDINGS,
hp.choice,
options=[bert, elmo, flair_normal, flair_pooled])
#other hyperparams are kept fixed for this excercise.
#Add to the lists to add to grid
#unfortunately for small grids, Flair picks random search instead of true
#grid search
search_space.add(Parameter.HIDDEN_SIZE, hp.choice, options=[384])
search_space.add(Parameter.RNN_LAYERS, hp.choice, options=[1])
search_space.add(Parameter.DROPOUT, hp.choice, options=[0.0])
# this is the folder in which train, test and dev files reside
data_folder = 'data/ner/bc5dr'
# init a corpus using column format, data folder and the names of the train, dev and test files
corpus: Corpus = ColumnCorpus(data_folder,
columns,
train_file='train.txt',
test_file='test.txt',
dev_file='dev.txt')
tag_type = 'ner'
tag_dictionary = corpus.make_tag_dictionary(tag_type=tag_type)
from flair.embeddings import TokenEmbeddings, WordEmbeddings, StackedEmbeddings, PooledFlairEmbeddings, ELMoEmbeddings
from typing import List
embedding_types: List[TokenEmbeddings] = [
PooledFlairEmbeddings('pubmed-forward'),
PooledFlairEmbeddings('pubmed-backward'),
ELMoEmbeddings('pubmed'),
]
embeddings: StackedEmbeddings = StackedEmbeddings(embeddings=embedding_types)
from flair.models import SequenceTagger
tagger: SequenceTagger = SequenceTagger(hidden_size=256,
embeddings=embeddings,
tag_dictionary=tag_dictionary,
tag_type=tag_type,
use_crf=True)
# initialize trainer
from flair.trainers import ModelTrainer
def train(model, selected_embeddings):
# 1. get the corpus
if model == 'AMT':
corpus = read_in_AMT()
elif model == 'CADEC':
corpus = read_in_CADEC()
elif model == 'TwitterADR':
corpus = read_in_TwitterADR()
elif model == 'Micromed':
corpus = read_in_Micromed()
print(corpus)
# 2. what tag do we want to predict?
tag_type = 'ner'
# 3. make the tag dictionary from the corpus
tag_dictionary = corpus.make_tag_dictionary(tag_type=tag_type)
print(tag_dictionary.idx2item)
embedding_types: List[TokenEmbeddings] = [
]
if selected_embeddings['glove']:
embedding_types.append(WordEmbeddings('glove'))
if selected_embeddings['twitter']:
embedding_types.append(WordEmbeddings('twitter'))
if selected_embeddings['char']:
embedding_types.append(CharacterEmbeddings())
# FlairEmbeddings
if selected_embeddings['flair']:
embedding_types.append(FlairEmbeddings('news-forward'))
# sFlairEmbeddings
if selected_embeddings['flair']:
embedding_types.append(FlairEmbeddings('news-backward'))
# PooledFlairEmbeddings
if selected_embeddings['pooled-flair']:
embedding_types.append(PooledFlairEmbeddings('news-forward', pooling='mean'))
# PooledFlairEmbeddings
if selected_embeddings['pooled-flair']:
embedding_types.append(PooledFlairEmbeddings('news-backward', pooling='mean'))
# init BERT
if selected_embeddings['bert']:
embedding_types.append(BertEmbeddings())
# init roberta
if selected_embeddings['roberta']:
embedding_types.append(RoBERTaEmbeddings())
# init BioBERT
if selected_embeddings['biobert']:
embedding_types.append(BertEmbeddings("data/embeddings/biobert-pubmed-pmc-cased"))
# init clinical BERT
if selected_embeddings['clinicalbiobert']:
embedding_types.append(BertEmbeddings("data/embeddings/pretrained_bert_tf/biobert-base-clinical-cased"))
# init multilingual ELMo
if selected_embeddings['elmo']:
embedding_types.append(ELMoEmbeddings())
embeddings: StackedEmbeddings = StackedEmbeddings(embeddings=embedding_types)
tagger: SequenceTagger = SequenceTagger(hidden_size=256,
embeddings=embeddings,
tag_dictionary=tag_dictionary,
tag_type=tag_type,
use_crf=True
)
trainer: ModelTrainer = ModelTrainer(tagger, corpus)
selected_embeddings_text = [key for key in selected_embeddings if selected_embeddings[key]]
selected_embeddings_text = '_'.join(selected_embeddings_text)
model_dir = 'resources/taggers/FA_' + model + selected_embeddings_text
# 7. start training
trainer.train(model_dir,
train_with_dev=True,
learning_rate=0.1,
mini_batch_size=4,
max_epochs=200,
checkpoint=True)
# 8. plot training curves (optional)
from flair.visual.training_curves import Plotter
plotter = Plotter()
plotter.plot_training_curves(model_dir + '/loss.tsv')
plotter.plot_weights(model_dir + '/weights.txt')
def read_dataset(self, file_dict, dataset_name, *args, **kwargs):
"""
:param file_dict: Will have just one key:value
file_dict['base_path'] = <base_path>
base_path will have the path to the directory that will have the structure :
conll_03 directory
conll_03/eng.testa
conll_03/eng.testb
conll_03/eng.train
onto-ner directory
onto-ner/eng.testa
onto-ner/eng.testb
onto-ner/eng.train
:param dataset_name: Could be one of the constants from NLPTask class(only NLPTask.CONLL_03 and
NLPTask.ONTONER are used)
:param args:
:param kwargs:
:return:
"""
base_path = file_dict['base_path']
self.dataset = dataset_name
# 1. get the corpus
corpus: TaggedCorpus = NLPTaskDataFetcher.load_corpus(dataset_name, base_path)
# 2. what tag do we want to predict?
tag_type = 'ner'
# 3. make the tag dictionary from the corpus
tag_dictionary = corpus.make_tag_dictionary(tag_type=tag_type)
if dataset_name == NLPTask.CONLL_03:
# initialize embeddings
embedding_types: List[TokenEmbeddings] = [
# GloVe embeddings
WordEmbeddings('glove'),
# contextual string embeddings, forward
PooledFlairEmbeddings('news-forward', pooling='min'),
# contextual string embeddings, backward
PooledFlairEmbeddings('news-backward', pooling='min'),
]
elif dataset_name == NLPTask.ONTONER:
# initialize embeddings
embedding_types: List[TokenEmbeddings] = [
WordEmbeddings('crawl'),
FlairEmbeddings('news-forward'),
FlairEmbeddings('news-backward'),
]
embeddings: StackedEmbeddings = StackedEmbeddings(embeddings=embedding_types)
# initialize sequence tagger
from flair.models import SequenceTagger
tagger: SequenceTagger = SequenceTagger(hidden_size=256,
embeddings=embeddings,
tag_dictionary=tag_dictionary,
tag_type=tag_type)
self.corpus = corpus
self.embeddings = embeddings
self.tag_dictionary = tag_dictionary
self.embedding_types = embedding_types
self.tagger = tagger
# -*- coding: utf-8 -*-
from flair.datasets import CONLL_03
from flair.embeddings import PooledFlairEmbeddings, StackedEmbeddings, WordEmbeddings
from flair.models import SequenceTagger
from flair.trainers import ModelTrainer
corpus = CONLL_03(base_path="data/conll-2003")
embedding_types = [
WordEmbeddings("glove"),
PooledFlairEmbeddings("news-forward", pooling="min"),
PooledFlairEmbeddings("news-backward", pooling="min"),
]
embeddings: StackedEmbeddings = StackedEmbeddings(embeddings=embedding_types)
tagger: SequenceTagger = SequenceTagger(
hidden_size=256,
embeddings=embeddings,
tag_dictionary=corpus.make_tag_dictionary(tag_type="ner"),
tag_type="ner",
)
trainer: ModelTrainer = ModelTrainer(tagger, corpus)
trainer.train("models/checkpoints", train_with_dev=True, max_epochs=150)
def use_flair_to_extract_context_embeddings(files,
file_name,
dest_folder,
layer,
embedding_type,
embedding_size,
pretrained_model=None):
if embedding_type.lower() == 'elmo':
context_embedding = ELMoEmbeddings(model='pubmed')
elif embedding_type.lower() == 'elmo_transformer':
context_embedding = ELMoTransformerEmbeddings()
elif embedding_type.lower() == 'flair':
context_embedding = PooledFlairEmbeddings()
elif embedding_type.lower() == 'bioflair':
flair_1 = PooledFlairEmbeddings('pubmed-forward')
flair_2 = PooledFlairEmbeddings('pubmed-backward')
elmo = ELMoEmbeddings(model='pubmed')
context_embedding = StackedEmbeddings(
embeddings=[flair_1, flair_2, elmo])
elif embedding_type.lower() == 'biobert' or embedding_type.lower(
) == 'bert':
context_embedding = TransformerWordEmbeddings(pretrained_model,
layers=layer)
data = []
for i in files:
open_f = open(i, 'r')
data += open_f.readlines()
open_f.close()
with open('{}/{}1.pickle'.format(dest_folder, file_name),
'wb') as store, open(
'{}/ebm_comet_multilabels_p1.txt'.format(dest_folder),
'w') as file:
#sentence = ''
sentence = []
multi_labels = ''
instance = []
label_representations = {}
#fetch outcome phrase vector representations grouped in their respective outcome domain labels
if file_name.lower() == 'ebm-comet':
label_representations = ebm_comet_preprocessing(
data=data,
context_embedding=context_embedding,
sentence=[],
label_representations={},
file=file)
elif file_name.lower() == 'ebm-nlp':
label_representations, domain_label_count = ebm_nlp_processing(
data=data,
context_embedding=context_embedding,
sentence=[],
label_representations={})
label_centroids = {}
print(label_representations.keys())
print([i.shape for i in list(label_representations.values())])
print(domain_label_count)
#find the centroid of each group of outcome phrases vectors to represent each label
for lab in label_representations:
label_centroids[lab] = torch.mean(label_representations[lab], 0)
pickle.dump(label_centroids, store)
store.close
# corpus: Corpus = NLPTaskDataFetcher.load_corpus(NLPTask.CONLL_03, base_path='resources/tasks')
# 2. what tag do we want to predict?
tag_type = 'ner'
# 3. make the tag dictionary from the corpus
tag_dictionary = corpus.make_tag_dictionary(tag_type=tag_type)
# initialize embeddings
embedding_types: List[TokenEmbeddings] = [
# GloVe embeddings
WordEmbeddings('glove'),
# contextual string embeddings, forward
PooledFlairEmbeddings('news-forward', pooling='min'),
# contextual string embeddings, backward
PooledFlairEmbeddings('news-backward', pooling='min'),
]
embeddings: StackedEmbeddings = StackedEmbeddings(embeddings=embedding_types)
# initialize sequence tagger
tagger: SequenceTagger = SequenceTagger(hidden_size=256,
embeddings=embeddings,
tag_dictionary=tag_dictionary,
tag_type=tag_type)
# initialize trainer
def resume(model1, selected_embeddings, model2):
# 1. get the corpus
if model2 == 'AMT':
corpus = read_in_AMT()
elif model2 == 'CADEC':
corpus = read_in_CADEC()
elif model2 == 'TwitterADR':
corpus = read_in_TwitterADR()
elif model2 == 'Micromed':
corpus = read_in_Micromed()
print(corpus)
# 2. what tag do we want to predict?
tag_type = 'ner'
# 3. make the tag dictionary from the corpus
tag_dictionary = corpus.make_tag_dictionary(tag_type=tag_type)
print(tag_dictionary.idx2item)
embedding_types: List[TokenEmbeddings] = [
]
if selected_embeddings['glove']:
embedding_types.append(WordEmbeddings('glove'))
if selected_embeddings['twitter']:
embedding_types.append(WordEmbeddings('twitter'))
if selected_embeddings['char']:
embedding_types.append(CharacterEmbeddings())
if selected_embeddings['flair']:
embedding_types.append(FlairEmbeddings('news-forward'))
if selected_embeddings['flair']:
embedding_types.append(FlairEmbeddings('news-backward'))
if selected_embeddings['pooled-flair']:
embedding_types.append(PooledFlairEmbeddings('news-forward', pooling='mean'))
if selected_embeddings['pooled-flair']:
embedding_types.append(PooledFlairEmbeddings('news-backward', pooling='mean'))
# init multilingual BERT
if selected_embeddings['bert']:
embedding_types.append(BertEmbeddings())
# init multilingual ELMo
if selected_embeddings['elmo']:
embedding_types.append(ELMoEmbeddings())
embeddings: StackedEmbeddings = StackedEmbeddings(embeddings=embedding_types)
# tagger: SequenceTagger = SequenceTagger(hidden_size=256,
# embeddings=embeddings,
# tag_dictionary=tag_dictionary,
# tag_type=tag_type,
# use_crf=True)
selected_embeddings_text = [key for key in selected_embeddings if selected_embeddings[key]]
selected_embeddings_text = '_'.join(selected_embeddings_text)
model_dir1 = 'resources/taggers/to_resume_CoNLL-03_' + model1 + selected_embeddings_text
#checkpoint = tagger.load_checkpoint(Path(model_dir1+ '/checkpoint.pt'))
#trainer = ModelTrainer.load_from_checkpoint(checkpoint, corpus)
best_model = SequenceTagger.load(Path(model_dir1+ '/best-model.pt'))
trainer: ModelTrainer = ModelTrainer(best_model, corpus)
# resources/taggers/to_resume_CADECglove_char_flair/
model_dir2 = 'resources/taggers/train_with_dev_from_' + model1 + '_to_' + model2 + selected_embeddings_text + '_fine-tuned7s'
trainer.train(model_dir2,
EvaluationMetric.MICRO_F1_SCORE,
train_with_dev=True,
learning_rate=0.1,
mini_batch_size=8,
max_epochs=150,
checkpoint=True)
