tomaarsen
/

span-marker-bert-base-orgs

Token Classification

named-entity-recognition

generated_from_span_marker_trainer

Carbon Emissions

Model card Files Files and versions Metrics Training metrics Community

span-marker-bert-base-orgs / train.py

tomaarsen's picture

tomaarsen HF staff

Upload train.py with huggingface_hub

0bed352 6 months ago

raw history blame contribute delete

No virus

3.38 kB

	from pathlib import Path
	import random
	import shutil
	from datasets import load_dataset, concatenate_datasets, Features, Sequence, ClassLabel, Value, DatasetDict
	from transformers import TrainingArguments
	from span_marker import SpanMarkerModel, Trainer
	from span_marker.model_card import SpanMarkerModelCardData
	from huggingface_hub import upload_folder, upload_file



	def main() -> None:
	# Load the dataset, ensure "tokens" and "ner_tags" columns, and get a list of labels
	labels = ["O", "B-ORG", "I-ORG"]
	dataset_id = "tomaarsen/ner-orgs"
	dataset = load_dataset(dataset_id)

	train_dataset = dataset["train"]
	eval_dataset = dataset["validation"]
	eval_dataset = eval_dataset.select(random.sample(range(len(eval_dataset)), k=3000))
	test_dataset = dataset["test"]

	# Initialize a SpanMarker model using a pretrained BERT-style encoder
	encoder_id = "bert-base-cased"
	model_id = f"tomaarsen/span-marker-bert-base-orgs"
	model = SpanMarkerModel.from_pretrained(
	encoder_id,
	labels=labels,
	# SpanMarker hyperparameters:
	model_max_length=256,
	marker_max_length=128,
	entity_max_length=8,
	# Model card variables
	model_card_data=SpanMarkerModelCardData(
	model_id=model_id,
	dataset_id=dataset_id,
	encoder_id=encoder_id,
	dataset_name="FewNERD, CoNLL2003, and OntoNotes v5",
	license="cc-by-sa-4.0",
	language=["en"],
	),
	)

	# Prepare the 🤗 transformers training arguments
	output_dir = Path("models") / model_id
	args = TrainingArguments(
	output_dir=output_dir,
	run_name=model_id,
	# Training Hyperparameters:
	learning_rate=5e-5,
	per_device_train_batch_size=32,
	per_device_eval_batch_size=32,
	num_train_epochs=3,
	weight_decay=0.01,
	warmup_ratio=0.1,
	bf16=True, # Replace `bf16` with `fp16` if your hardware can't use bf16.
	# Other Training parameters
	logging_first_step=True,
	logging_steps=100,
	evaluation_strategy="steps",
	save_strategy="steps",
	eval_steps=3000,
	save_total_limit=1,
	dataloader_num_workers=4,
	)

	# Initialize the trainer using our model, training args & dataset, and train
	trainer = Trainer(
	model=model,
	args=args,
	train_dataset=train_dataset,
	eval_dataset=eval_dataset,
	)
	trainer.train()

	# Compute & save the metrics on the test set
	metrics = trainer.evaluate(test_dataset, metric_key_prefix="test")
	trainer.save_metrics("test", metrics)

	# Save the model & training script locally
	trainer.save_model(output_dir / "checkpoint-final")
	shutil.copy2(__file__, output_dir / "checkpoint-final" / "train.py")

	# Upload everything to the Hub
	breakpoint()
	model.push_to_hub(model_id, private=True)
	upload_folder(folder_path=output_dir / "runs", path_in_repo="runs", repo_id=model_id)
	upload_file(path_or_fileobj=__file__, path_in_repo="train.py", repo_id=model_id)
	upload_file(path_or_fileobj=output_dir / "all_results.json", path_in_repo="all_results.json", repo_id=model_id)
	upload_file(path_or_fileobj=output_dir / "emissions.csv", path_in_repo="emissions.csv", repo_id=model_id)


	if __name__ == "__main__":
	main()