def test_text_classifier_transformer_finetune(results_base_path,
tasks_base_path):
flair.set_seed(123)
corpus = ClassificationCorpus(
tasks_base_path / "trivial" / "trivial_text_classification_single",
label_type="city",
)
label_dict = corpus.make_label_dictionary(label_type="city")
model: TextClassifier = TextClassifier(
document_embeddings=TransformerDocumentEmbeddings(
"distilbert-base-uncased"),
label_dictionary=label_dict,
label_type="city",
multi_label=False,
)
trainer = ModelTrainer(model, corpus)
trainer.fine_tune(
results_base_path,
mini_batch_size=2,
max_epochs=10,
shuffle=True,
learning_rate=0.5e-5,
num_workers=2,
)
# check if model can predict
sentence = Sentence("this is Berlin")
sentence_empty = Sentence(" ")
model.predict(sentence)
model.predict([sentence, sentence_empty])
model.predict([sentence_empty])
# load model
loaded_model = TextClassifier.load(results_base_path / "final-model.pt")
# chcek if model predicts correct label
sentence = Sentence("this is Berlin")
sentence_empty = Sentence(" ")
loaded_model.predict([sentence, sentence_empty])
values = []
for label in sentence.labels:
assert label.value is not None
assert 0.0 <= label.score <= 1.0
assert type(label.score) is float
values.append(label.value)
assert "Berlin" in values
# check if loaded model successfully fit the training data
result: Result = loaded_model.evaluate(corpus.test, gold_label_type="city")
assert result.classification_report["micro avg"]["f1-score"] == 1.0
del loaded_model
def prepare_resources(self):
# turn off INFO and DEBUG logging
import flair # KEEP THIS IMPORT HERE! (it initialises 'flair' logger)
import logging
logger = logging.getLogger('flair')
logger.setLevel(logging.WARNING)
if self.seed:
flair.set_seed(self.seed)
def test_text_classifier_multi(results_base_path, tasks_base_path):
flair.set_seed(123)
corpus = ClassificationCorpus(
tasks_base_path / "trivial" / "trivial_text_classification_multi",
label_type="city",
)
label_dict = corpus.make_label_dictionary(label_type="city")
model: TextClassifier = TextClassifier(
document_embeddings=DocumentPoolEmbeddings([turian_embeddings],
fine_tune_mode="linear"),
label_dictionary=label_dict,
label_type="city",
multi_label=True,
)
trainer = ModelTrainer(model, corpus)
trainer.train(results_base_path,
mini_batch_size=2,
max_epochs=50,
shuffle=True)
# check if model can predict
sentence = Sentence("this is Berlin")
sentence_empty = Sentence(" ")
model.predict(sentence)
model.predict([sentence, sentence_empty])
model.predict([sentence_empty])
# load model
loaded_model = TextClassifier.load(results_base_path / "final-model.pt")
# chcek if model predicts correct label
sentence = Sentence("this is Berlin")
sentence_double = Sentence("this is Berlin and pizza")
loaded_model.predict([sentence, sentence_double])
values = []
for label in sentence_double.labels:
assert label.value is not None
assert 0.0 <= label.score <= 1.0
assert type(label.score) is float
values.append(label.value)
assert "Berlin" in values
assert "pizza" in values
# check if loaded model successfully fit the training data
result: Result = loaded_model.evaluate(corpus.test, gold_label_type="city")
print(result.classification_report)
assert result.classification_report["micro avg"]["f1-score"] == 1.0
del loaded_model
def run_experiment(seed, batch_size, epoch, learning_rate, hipe_datasets,
json_config):
# Config values
# Replace it with more Pythonic solutions later!
word_embedding = json_config["word_embedding"]
use_crf = json_config["use_crf"] if "use_crf" in json_config else False
# Set seed for reproducibility
set_seed(seed)
corpus_list = []
# Dataset-related
for dataset in hipe_datasets:
dataset_name, language = dataset.split("/")
corpus_list.append(
NER_HIPE_2022(dataset_name=dataset_name,
language=language,
add_document_separator=True))
print("Use CRF:", use_crf)
corpora: MultiCorpus = MultiCorpus(corpora=corpus_list,
sample_missing_splits=False)
label_dictionary = corpora.make_label_dictionary(label_type="ner")
print("Label Dictionary:", label_dictionary.get_items())
# FastText Embeddings
embeddings = FastTextEmbeddings(embeddings=word_embedding)
tagger: SequenceTagger = SequenceTagger(
hidden_size=256,
embeddings=embeddings,
tag_dictionary=label_dictionary,
tag_type="ner",
use_crf=use_crf,
)
# Trainer
trainer: ModelTrainer = ModelTrainer(tagger, corpora)
datasets = "-".join([dataset for dataset in hipe_datasets])
trainer.train(
f"hipe2022-flert-we-trainer-{datasets}-{word_embedding}-bs{batch_size}-wsFalse-e{epoch}-lr{learning_rate}-crf{use_crf}-{seed}",
mini_batch_size=batch_size,
mini_batch_chunk_size=2,
patience=3,
max_epochs=epoch,
shuffle=True,
learning_rate=learning_rate,
)
# Finally, print model card for information
tagger.print_model_card()
def test_sequence_tagger_transformer_finetune(results_base_path,
tasks_base_path):
flair.set_seed(123)
# load dataset
corpus: Corpus = ColumnCorpus(
data_folder=tasks_base_path / "trivial" / "trivial_bioes",
column_format={
0: "text",
1: "ner"
},
)
tag_dictionary = corpus.make_label_dictionary("ner")
# tagger without CRF
tagger: SequenceTagger = SequenceTagger(
hidden_size=64,
embeddings=TransformerWordEmbeddings("distilbert-base-uncased",
fine_tune=True),
tag_dictionary=tag_dictionary,
tag_type="ner",
use_crf=False,
use_rnn=False,
reproject_embeddings=False,
)
# train
trainer = ModelTrainer(tagger, corpus)
trainer.fine_tune(
results_base_path,
mini_batch_size=2,
max_epochs=10,
shuffle=True,
learning_rate=0.5e-4,
)
loaded_model: SequenceTagger = SequenceTagger.load(results_base_path /
"final-model.pt")
sentence = Sentence("this is New York")
sentence_empty = Sentence(" ")
loaded_model.predict(sentence)
loaded_model.predict([sentence, sentence_empty])
loaded_model.predict([sentence_empty])
# check if loaded model can predict
entities = [span.text for span in sentence.get_spans("ner")]
assert "New York" in entities
# check if loaded model successfully fit the training data
result: Result = loaded_model.evaluate(corpus.test, gold_label_type="ner")
assert result.classification_report["micro avg"]["f1-score"] == 1.0
del loaded_model
def run_experiment(data_folder, task_name, model_name, split):
# Adjust logging level
logging.getLogger("flair").setLevel(level="ERROR")
set_seed(1)
if task_name in ["lft", "onb"]:
# Configuration
column_format = {0: "token", 1: "ner"}
# We use official data from Riedl and Padó
train_file = f"enp_DE.{task_name}.mr.tok.train.bio"
dev_file = f"enp_DE.{task_name}.mr.tok.dev.bio"
test_file = f"enp_DE.{task_name}.mr.tok.test.bio"
# Corpus
corpus = ColumnCorpus(
data_folder=data_folder,
column_format=column_format,
train_file=train_file,
dev_file=dev_file,
test_file=test_file,
tag_to_bioes="ner",
)
tagger: SequenceTagger = SequenceTagger.load(model_name)
ds = corpus.test if split == "test" else corpus.dev
for sentence in ds:
tagger.predict(sentence, label_name="predicted")
gold_spans = sentence.get_spans("ner")
pred_spans = sentence.get_spans("predicted")
for token in sentence:
gold_tag = "O"
for span in gold_spans:
if token in span:
gold_tag = "B-" + span.tag if token == span[
0] else "I-" + span.tag
pred_tag = "O"
for span in pred_spans:
if token in span:
pred_tag = "B-" + span.tag if token == span[
0] else "I-" + span.tag
print(f"{token.text} {gold_tag} {pred_tag}")
print("")
def test_sequence_tagger_with_crf(results_base_path, tasks_base_path):
flair.set_seed(123)
# load dataset
corpus: Corpus = ColumnCorpus(data_folder=tasks_base_path / "trivial" /
"trivial_bioes",
column_format={
0: "text",
1: "ner"
})
tag_dictionary = corpus.make_label_dictionary("ner")
# tagger without CRF
tagger: SequenceTagger = SequenceTagger(
hidden_size=64,
embeddings=turian_embeddings,
tag_dictionary=tag_dictionary,
tag_type="ner",
use_crf=True,
)
# train
trainer: ModelTrainer = ModelTrainer(tagger, corpus)
trainer.train(
results_base_path,
learning_rate=0.1,
mini_batch_size=2,
max_epochs=10,
shuffle=False,
)
loaded_model: SequenceTagger = SequenceTagger.load(results_base_path /
"final-model.pt")
sentence = Sentence("this is New York")
sentence_empty = Sentence(" ")
loaded_model.predict(sentence)
loaded_model.predict([sentence, sentence_empty])
loaded_model.predict([sentence_empty])
# check if loaded model can predict
entities = [span.text for span in sentence.get_spans('ner')]
assert "New York" in entities
# check if loaded model successfully fit the training data
result: Result = loaded_model.evaluate(corpus.test, gold_label_type='ner')
assert result.classification_report["micro avg"]["f1-score"] == 1.
# clean up results directory
shutil.rmtree(results_base_path)
del loaded_model
def test_text_classifier_multi(results_base_path, tasks_base_path):
flair.set_seed(123)
flair_embeddings = FlairEmbeddings("news-forward-fast")
corpus = ClassificationCorpus(
tasks_base_path / "trivial" / "trivial_text_classification_single",
label_type="city",
)
label_dict = corpus.make_label_dictionary(label_type="city")
model: TextClassifier = TextClassifier(
document_embeddings=DocumentPoolEmbeddings([flair_embeddings], fine_tune_mode="linear"),
label_dictionary=label_dict,
label_type="city",
)
trainer = ModelTrainer(model, corpus)
trainer.train(results_base_path, mini_batch_size=2, max_epochs=1, shuffle=True)
del model
train_log_file = results_base_path / "training.log"
assert train_log_file.exists()
lines = train_log_file.read_text(encoding="utf-8").split("\n")
expected_substrings = [
"Device: ",
"Corpus: ",
"Parameters:",
"- learning_rate: ",
"- patience: ",
"Embeddings storage mode:",
"epoch 1 - iter",
"EPOCH 1 done: loss",
"Results:",
]
for expected_substring in expected_substrings:
assert any(expected_substring in line for line in lines), expected_substring
import sys
import os
sys.path.insert(0, "/vol/fob-vol7/mi19/harnisph/flair")
import flair
import torch
from flair.models import TARSSequenceTagger2
from flair.data import Sentence
from flair.datasets import CONLL_03
tagger = TARSSequenceTagger2.load(
"resources/v3/moviecomplex-long/final-model.pt")
flair.set_seed(3)
label_name_map = {
"LOC": "Location",
"PER": "Person",
"ORG": "Organization",
"MISC": "Miscellaneous"
}
print(label_name_map)
corpus = CONLL_03(
tag_to_bioes=None,
tag_to_bio2="ner",
label_name_map=label_name_map,
base_path="/vol/fob-vol7/mi19/harnisph/studienprojekt-dokumentation")
corpus = corpus.downsample(0.1)
tag_type = "ner"
tag_dictionary = corpus.make_label_dictionary(tag_type)
tagger.add_and_switch_to_new_task("zeroshot-moviecomplex-long-to-conll3",
def run_experiment(seed, batch_size, epoch, learning_rate, hipe_datasets,
json_config):
# Config values
# Replace it with more Pythonic solutions later!
hf_model = json_config["hf_model"]
context_size = json_config["context_size"]
layers = json_config["layers"] if "layers" in json_config else "-1"
use_crf = json_config["use_crf"] if "use_crf" in json_config else False
# Set seed for reproducibility
set_seed(seed)
corpus_list = []
# Dataset-related
for dataset in hipe_datasets:
dataset_name, language = dataset.split("/")
current_corpus = NER_HIPE_2022(dataset_name=dataset_name,
language=language,
add_document_separator=True)
for split in ["train", "dev"]:
kb_data = []
print(f"Loading KB contexts for {dataset}...")
with open(f"kb_data/ajmc/{language}/{language}_{split}.jsonl",
"rt") as f_p:
for line in f_p:
kb_data.append(line)
corpus_split = current_corpus.train if split == "train" else current_corpus.dev
for index, sent in enumerate(corpus_split):
jsonl = json.loads(kb_data[index])
kb_context = " ".join(jsonl["contexts"]).split(" ")
sent.kb_context = kb_context
corpus_list.append(current_corpus)
if context_size == 0:
context_size = False
print("FLERT Context:", context_size)
print("Layers:", layers)
print("Use CRF:", use_crf)
corpora: MultiCorpus = MultiCorpus(corpora=corpus_list,
sample_missing_splits=False)
label_dictionary = corpora.make_label_dictionary(label_type="ner")
print("Label Dictionary:", label_dictionary.get_items())
# Embeddings
embeddings = KBTransformerEmbeddings(
model=hf_model,
layers=layers,
subtoken_pooling="first",
fine_tune=True,
use_context=context_size,
)
tagger: SequenceTagger = SequenceTagger(
hidden_size=256,
embeddings=embeddings,
tag_dictionary=label_dictionary,
tag_type="ner",
use_crf=use_crf,
use_rnn=False,
reproject_embeddings=False,
)
# Trainer
trainer: ModelTrainer = ModelTrainer(tagger, corpora)
datasets = "-".join([dataset for dataset in hipe_datasets])
trainer.fine_tune(
f"hipe2022-flert-fine-tune-kb-{datasets}-{hf_model}-bs{batch_size}-ws{context_size}-e{epoch}-lr{learning_rate}-layers{layers}-crf{use_crf}-{seed}",
learning_rate=learning_rate,
mini_batch_size=batch_size,
max_epochs=epoch,
shuffle=True,
embeddings_storage_mode='none',
weight_decay=0.,
use_final_model_for_eval=False,
)
# Finally, print model card for information
tagger.print_model_card()
from flair.training_utils import EvaluationMetric
from flair.visual.training_curves import Plotter
#get the corpus
#flair.set_seed(1)
import flair
# flair.set_seed(2)
# flair.set_seed(3)
from flair.datasets import CONLL_03
from mapping import (twitter_ner_mapped, onto_ner_mapped, wikigold_ner_mapped,
webpages_ner_mapped)
dataset_name = "conll3"
for seed in [1, 2, 3]:
flair.set_seed(123)
if dataset_name == "onto_ner":
corpus = onto_ner_mapped()
elif dataset_name == "conll3":
corpus = CONLL_03()
elif dataset_name == "wikipedia":
corpus = wikigold_ner_mapped()
elif dataset_name == "webpages":
corpus = webpages_ner_mapped()
elif dataset_name == "twitter":
corpus = twitter_ner_mapped()
flair.set_seed(seed)
import sys
import os
#sys.path.insert(0, os.path.join('C:/', 'Users', 'pharn', 'flair'))
#sys.path.insert(0, os.path.join('C:/', 'Users', 'pharn', 'AppData', 'Local', 'Packages', 'PythonSoftwareFoundation.Python.3.8_qbz5n2kfra8p0', 'LocalCache', 'local-packages', 'Python38', 'site-packages'))
sys.path.insert(0, "/vol/fob-vol7/mi19/harnisph/flair")
sys.path.insert(0, os.path.join("vol", "fob-vol7", "mi19", "harnisph",
"flair"))
import flair
from flair.models import TARSSequenceTagger2
from flair.data import Sentence
#from flair.datasets import CONLL_3, MIT_MOVIE_NER_COMPLEX
from flair.datasets import WNUT_2020_NER
flair.set_seed(1)
tagger = TARSSequenceTagger2.load(
"resources/testfaelle-studproj/conll_3/final-model.pt")
sentences = [
Sentence(
"The Parlament of the United Kingdom is discussing a variety of topics."
),
Sentence(
"A man fell in love with a woman. This takes place in the last century. The film received the Golden Love Film Award."
),
Sentence("The Company of Coca Cola was invented in 1901."),
Sentence("This is very frustrating! I was smiling since I saw you."),
Sentence("The Green Party received only a small percentage of the vote."),
Sentence(
"Bayern Munich won the german soccer series the sixth time in a row.")
def main():
parser = HfArgumentParser((ModelArguments, TrainingArguments, FlertArguments, DataArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
(
model_args,
training_args,
flert_args,
data_args,
) = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
(
model_args,
training_args,
flert_args,
data_args,
) = parser.parse_args_into_dataclasses()
set_seed(training_args.seed)
flair.device = training_args.device
corpus = get_flair_corpus(data_args)
logger.info(corpus)
tag_type: str = "ner"
tag_dictionary = corpus.make_label_dictionary(tag_type)
logger.info(tag_dictionary)
embeddings = TransformerWordEmbeddings(
model=model_args.model_name_or_path,
layers=model_args.layers,
subtoken_pooling=model_args.subtoken_pooling,
fine_tune=True,
use_context=flert_args.context_size,
respect_document_boundaries=flert_args.respect_document_boundaries,
)
tagger = SequenceTagger(
hidden_size=model_args.hidden_size,
embeddings=embeddings,
tag_dictionary=tag_dictionary,
tag_type=tag_type,
use_crf=model_args.use_crf,
use_rnn=False,
reproject_embeddings=False,
)
trainer = ModelTrainer(tagger, corpus)
trainer.fine_tune(
data_args.output_dir,
learning_rate=training_args.learning_rate,
mini_batch_size=training_args.batch_size,
mini_batch_chunk_size=training_args.mini_batch_chunk_size,
max_epochs=training_args.num_epochs,
embeddings_storage_mode=training_args.embeddings_storage_mode,
weight_decay=training_args.weight_decay,
)
torch.save(model_args, os.path.join(data_args.output_dir, "model_args.bin"))
torch.save(training_args, os.path.join(data_args.output_dir, "training_args.bin"))
# finally, print model card for information
tagger.print_model_card()
def run_experiment(seed, batch_size, epoch, learning_rate, hipe_datasets,
json_config):
# Config values
# Replace it with more Pythonic solutions later!
hf_model = json_config["hf_model"]
context_size = json_config["context_size"]
layers = json_config["layers"] if "layers" in json_config else "-1"
use_crf = json_config["use_crf"] if "use_crf" in json_config else False
additional_hipe_datasets = json_config[
"additional_hipe_datasets"] if "additional_hipe_datasets" in json_config else None
label_name_map = json_config[
"label_name_map"] if "label_name_map" in json_config else None
# Set seed for reproducibility
set_seed(seed)
corpus_list = []
# Dataset-related
for dataset in hipe_datasets:
dataset_name, language = dataset.split("/")
preproc_fn = None
if dataset_name == "ajmc":
preproc_fn = prepare_ajmc_corpus
corpus_list.append(
NER_HIPE_2022(dataset_name=dataset_name,
language=language,
preproc_fn=preproc_fn,
add_document_separator=True))
if additional_hipe_datasets and label_name_map:
# Special case: do not use Dev data from additional datasets
# This makes evaluation and comparison much more easier!
for dataset in additional_hipe_datasets:
dataset_name, language = dataset.split("/")
preproc_fn = None
if dataset_name == "hipe2020":
print("Using own HIPE-2020 Preprocessing function.")
print(
"Please make sure that Flair Datasets folder was cleaned before!"
)
preproc_fn = prepare_clef_2020_corpus
additional_corpus = NER_HIPE_2022(dataset_name=dataset_name,
label_name_map=label_name_map,
language=language,
add_document_separator=True,
preproc_fn=preproc_fn)
additional_corpus._dev = []
corpus_list.append(additional_corpus)
if context_size == 0:
context_size = False
print("FLERT Context:", context_size)
print("Layers:", layers)
print("Use CRF:", use_crf)
corpora: MultiCorpus = MultiCorpus(corpora=corpus_list,
sample_missing_splits=False)
label_dictionary = corpora.make_label_dictionary(label_type="ner")
print("Label Dictionary:", label_dictionary.get_items())
# Embeddings
embeddings = TransformerWordEmbeddings(
model=hf_model,
layers=layers,
subtoken_pooling="first",
fine_tune=True,
use_context=context_size,
)
tagger: SequenceTagger = SequenceTagger(
hidden_size=256,
embeddings=embeddings,
tag_dictionary=label_dictionary,
tag_type="ner",
use_crf=use_crf,
use_rnn=False,
reproject_embeddings=False,
)
# Trainer
trainer: ModelTrainer = ModelTrainer(tagger, corpora)
datasets = "-".join([dataset for dataset in hipe_datasets])
trainer.fine_tune(
f"hipe2022-flert-fine-tune-{datasets}-{hf_model}-bs{batch_size}-ws{context_size}-e{epoch}-lr{learning_rate}-layers{layers}-crf{use_crf}-{seed}",
learning_rate=learning_rate,
mini_batch_size=batch_size,
max_epochs=epoch,
shuffle=True,
embeddings_storage_mode='none',
weight_decay=0.,
use_final_model_for_eval=False,
)
# Finally, print model card for information
tagger.print_model_card()
def run_experiment(data_folder, task_name, model_name, run_id, use_context):
# Set seed for reproducibility
set_seed(int(run_id))
if use_context == 0:
use_context = False
print("FLERT Context:", use_context)
if task_name in ["lft", "onb"]:
# Configuration
column_format = {0: "token", 1: "ner"}
# We use official data from Riedl and Padó
train_file = f"enp_DE.{task_name}.mr.tok.train.bio"
dev_file = f"enp_DE.{task_name}.mr.tok.dev.bio"
test_file = f"enp_DE.{task_name}.mr.tok.test.bio"
# Corpus
corpus = ColumnCorpus(
data_folder=data_folder,
column_format=column_format,
train_file=train_file,
dev_file=dev_file,
test_file=test_file,
tag_to_bioes="ner",
)
# Corpus configuration
tag_type = "ner"
tag_dictionary = corpus.make_tag_dictionary(tag_type=tag_type)
print(tag_dictionary.idx2item)
# Embeddings
embedding_types: List[TokenEmbeddings] = [
TransformerWordEmbeddings(model=model_name,
layers="all",
layer_mean=True,
use_context=use_context)
]
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
trainer: ModelTrainer = ModelTrainer(tagger, corpus)
trainer.train(
f"resources/taggers/ner-{task_name}-{model_name}-context{use_context}-{run_id}",
learning_rate=0.1,
mini_batch_size=16,
max_epochs=200,
shuffle=True,
)
def run_experiment(seed, batch_size, epoch, learning_rate, hipe_datasets,
json_config):
# Config values
# Replace it with more Pythonic solutions later!
best_model = json_config["best_model"]
context_size = json_config["context_size"]
layers = json_config["layers"] if "layers" in json_config else "-1"
use_crf = json_config["use_crf"] if "use_crf" in json_config else False
# Set seed for reproducibility
set_seed(seed)
corpus_list = []
# Dataset-related
for dataset in hipe_datasets:
dataset_name, language = dataset.split("/")
preproc_fn = None
if dataset_name == "ajmc":
preproc_fn = prepare_ajmc_corpus
corpus_list.append(
NER_HIPE_2022(dataset_name=dataset_name,
language=language,
preproc_fn=preproc_fn,
add_document_separator=True))
if context_size == 0:
context_size = False
print("FLERT Context:", context_size)
print("Layers:", layers)
print("Use CRF:", use_crf)
corpora: MultiCorpus = MultiCorpus(corpora=corpus_list,
sample_missing_splits=False)
label_dictionary = corpora.make_label_dictionary(label_type="ner")
print("Label Dictionary:", label_dictionary.get_items())
print("Loading model from stage 1:", best_model)
tagger: SequenceTagger = SequenceTagger.load(best_model)
# Trainer
trainer: ModelTrainer = ModelTrainer(tagger, corpora)
datasets = "-".join([dataset for dataset in hipe_datasets])
best_model_name = best_model.replace("/", "_")
trainer.fine_tune(
f"hipe2022-flert-fine-tune-multistage-{datasets}-{best_model_name}-bs{batch_size}-ws{context_size}-e{epoch}-lr{learning_rate}-layers{layers}-crf{use_crf}-{seed}",
learning_rate=learning_rate,
mini_batch_size=batch_size,
max_epochs=epoch,
shuffle=True,
embeddings_storage_mode='none',
weight_decay=0.,
use_final_model_for_eval=False,
)
# Finally, print model card for information
tagger.print_model_card()
def run_experiment(seed, batch_size, epoch, learning_rate, json_config):
# Config values
# Replace it with more Pythonic solutions later!
hf_model = json_config["hf_model"]
context_size = json_config["context_size"]
layers = json_config["layers"] if "layers" in json_config else "-1"
use_crf = json_config["use_crf"] if "use_crf" in json_config else False
task_name = json_config["task_name"]
# Dataset-related
data_folder = json_config["data_folder"]
train_file = json_config["train_file"]
dev_file = json_config["dev_file"]
test_file = json_config["test_file"]
# Set seed for reproducibility
set_seed(seed)
if context_size == 0:
context_size = False
print("FLERT Context:", context_size)
print("Layers:", layers)
print("Use CRF:", use_crf)
# Configuration
column_format = {0: "text", 1: "ner"}
# Corpus
corpus = ColumnCorpus(data_folder=data_folder,
column_format=column_format,
train_file=train_file,
dev_file=dev_file,
test_file=test_file,
tag_to_bioes="ner",
)
# Corpus configuration
tag_type = "ner"
tag_dictionary = corpus.make_tag_dictionary(tag_type=tag_type)
print(tag_dictionary.idx2item)
# Embeddings
embeddings = TransformerWordEmbeddings(
model=hf_model,
layers=layers,
subtoken_pooling="first",
fine_tune=True,
use_context=context_size,
)
tagger: SequenceTagger = SequenceTagger(
hidden_size=256,
embeddings=embeddings,
tag_dictionary=tag_dictionary,
tag_type=tag_type,
use_crf=use_crf,
use_rnn=False,
reproject_embeddings=False,
)
# Trainer
trainer: ModelTrainer = ModelTrainer(tagger, corpus)
trainer.fine_tune(
f"histo-flert-fine-tuning-{task_name}-{hf_model}-bs{batch_size}-ws{context_size}-e{epoch}-lr{learning_rate}-layers{layers}-crf{use_crf}-{seed}",
learning_rate=learning_rate,
mini_batch_size=batch_size,
max_epochs=epoch,
shuffle=True,
embeddings_storage_mode='none',
weight_decay=0.,
use_final_model_for_eval=False,
)
help="Model name (such as Hugging Face model hub name")
parser.add_argument(
"-d",
"--dataset",
type=str,
help="Defines dataset, choose between imst, boun or xtreme")
# Parse experimental arguments
args = parser.parse_args()
# use cuda device as passed
flair.device = f'cuda:{str(args.cuda)}'
# for each passed seed, do one experimental run
for seed in args.seeds:
flair.set_seed(seed)
# model
hf_model = args.model
# initialize embeddings
embeddings = TransformerWordEmbeddings(
model=hf_model,
layers="-1",
subtoken_pooling="first",
fine_tune=True,
use_context=False,
respect_document_boundaries=False,
)
# select dataset depending on which language variable is passed
