def run_training(args: dict):
word_lexicon = load_vocab(args['lex_size'])
train_data, dev_data, test_data, entities_lexicon = instantiate_datasets(
word_lexicon, args
)
print("Instantiating model")
pyramid_ner = Pyramid(
word_lexicon=word_lexicon,
entities_lexicon=entities_lexicon,
classifier_type=args['classifier_type'],
word_embeddings=args['word_embeddings'],
language_model=args['language_model'],
char_embeddings_dim=args['char_embeddings_dim'],
encoder_hidden_size=args['encoder_hidden_size'],
embedding_encoder_type=args['embedding_encoder_type'],
embedding_encoder_output_size=args['embedding_encoder_output_size'],
decoder_hidden_size=args['decoder_hidden_size'],
inverse_pyramid=args['inverse_pyramid'],
custom_tokenizer=args['custom_tokenizer'],
pyramid_max_depth=args['pyramid_max_depth'],
decoder_dropout=args['decoder_dropout'],
encoder_dropout=args['encoder_dropout'],
use_pre=args['use_pre'],
use_post=args['use_post'],
hidden_size=args['hidden_size'],
rnn_layers=args['rnn_layers'],
reproject_words=args['reproject_words'],
reproject_words_dimension=args['reproject_words_dimension'],
bidirectional=args['bidirectional'],
dropout=args['dropout'],
word_dropout=args['word_dropout'],
locked_dropout=args['locked_dropout'],
device=DEVICE
)
print(pyramid_ner.nnet)
trainer = MultiLabelTrainer(pyramid_ner)
optimizer, scheduler = get_default_sgd_optim(pyramid_ner.nnet.parameters())
ner_model, train_report = trainer.train(
train_data,
optimizer=optimizer,
scheduler=scheduler,
dev_data=dev_data,
epochs=args['epochs'],
patience=args['patience'],
grad_clip=args['grad_clip'],
restore_weights_on=args['restore_weights_on']
)
train_report.plot_loss_report()
train_report.plot_custom_report('micro_f1')
formatted_report = trainer.test_model(test_data, out_dict=False)
write_report(args, train_report, formatted_report)
def run_sigmoid_test(args: dict):
word_lexicon = load_vocab(args['lex_size'])
print("Loading data")
train_data = list(
itertools.islice(wrg_reader("data/nne_concat/train.txt"), 256))
dev_data = list(itertools.islice(wrg_reader("data/nne_concat/dev.txt"),
64))
test_data = list(
itertools.islice(wrg_reader("data/nne_concat/test.txt"), 64))
print("Generating entity lexicon")
entities_lexicon = list(
sorted({
entity.name
for data in (train_data, dev_data, test_data)
for data_point in data for entity in data_point.entities
}))
print(entities_lexicon)
print("Instantiating train dataset")
train_dataset = SigmoidMultiLabelNerDataset(
train_data,
pyramid_max_depth=args['pyramid_max_depth'],
token_lexicon=word_lexicon,
entities_lexicon=entities_lexicon,
custom_tokenizer=None,
char_vectorizer=True,
)
print("Instantiating dev dataset")
dev_dataset = SigmoidMultiLabelNerDataset(
dev_data,
pyramid_max_depth=args['pyramid_max_depth'],
token_lexicon=word_lexicon,
entities_lexicon=entities_lexicon,
custom_tokenizer=None,
char_vectorizer=True,
)
print("Instantiating test dataset")
test_dataset = SigmoidMultiLabelNerDataset(
test_data,
pyramid_max_depth=args['pyramid_max_depth'],
token_lexicon=word_lexicon,
entities_lexicon=entities_lexicon,
custom_tokenizer=None,
char_vectorizer=True,
)
dev_dataloader, train_dataloader, test_dataloader = get_data_loader(
train_dataset, dev_dataset, test_dataset, args)
print("Instantiating SigmoidMultiLabelPyramid")
pyramid_ner = SigmoidMultiLabelPyramid(
word_lexicon=word_lexicon,
entities_lexicon=entities_lexicon,
classifier_type=args['classifier_type'],
word_embeddings=args['word_embeddings'],
language_model=args['language_model'],
char_embeddings_dim=args['char_embeddings_dim'],
encoder_hidden_size=args['encoder_hidden_size'],
decoder_hidden_size=args['decoder_hidden_size'],
inverse_pyramid=args['inverse_pyramid'],
custom_tokenizer=args['custom_tokenizer'],
pyramid_max_depth=args['pyramid_max_depth'],
decoder_dropout=args['decoder_dropout'],
encoder_dropout=args['encoder_dropout'],
device=DEVICE)
print(pyramid_ner.nnet)
run_training(pyramid_ner, train_dataloader, dev_dataloader,
test_dataloader, args)
def run_token_window_test(args: dict):
word_lexicon = load_vocab(args['lex_size'])
print("Loading data")
train_data = list(
itertools.islice(
wrg_token_window_reader("data/nne_raw/train/",
args['token_window']), 256))
dev_data = list(
itertools.islice(
wrg_token_window_reader("data/nne_raw/dev/", args['token_window']),
64 * 10))
test_data = list(
itertools.islice(
wrg_token_window_reader("data/nne_raw/test/",
args['token_window']), 64))
print("Generating entity lexicon")
entities_lexicon = list(
sorted({
entity.name
for data in (train_data, dev_data, test_data)
for data_point in data for entity in data_point.entities
}))
print(entities_lexicon)
print("Instantiating train dataset")
train_dataset = TokenWindowMultiLabelNerDataset(
train_data,
pyramid_max_depth=args['pyramid_max_depth'],
token_lexicon=word_lexicon,
entities_lexicon=entities_lexicon,
custom_tokenizer=None,
char_vectorizer=True,
)
print("Instantiating dev dataset")
dev_dataset = TokenWindowMultiLabelNerDataset(
dev_data,
pyramid_max_depth=args['pyramid_max_depth'],
token_lexicon=word_lexicon,
entities_lexicon=entities_lexicon,
custom_tokenizer=None,
char_vectorizer=True,
)
print("Instantiating test dataset")
test_dataset = TokenWindowMultiLabelNerDataset(
test_data,
pyramid_max_depth=args['pyramid_max_depth'],
token_lexicon=word_lexicon,
entities_lexicon=entities_lexicon,
custom_tokenizer=None,
char_vectorizer=True,
)
dev_dataloader, train_dataloader, test_dataloader = get_data_loader(
train_dataset, dev_dataset, test_dataset, args)
test_document_rnn(word_lexicon, train_dataloader, dev_dataloader,
test_dataloader, entities_lexicon, args)
dev_dataloader, train_dataloader, test_dataloader = get_data_loader(
train_dataset, dev_dataset, test_dataset, args)
test_sentence_transformer(word_lexicon, train_dataloader, dev_dataloader,
test_dataloader, entities_lexicon, args)