def test_lm_finetuning(caplog):
caplog.set_level(logging.CRITICAL)
set_all_seeds(seed=42)
device, n_gpu = initialize_device_settings(use_cuda=True)
n_epochs = 1
batch_size = 5
evaluate_every = 2
lang_model = "bert-base-cased"
tokenizer = BertTokenizer.from_pretrained(
pretrained_model_name_or_path=lang_model, do_lower_case=False)
processor = BertStyleLMProcessor(
data_dir="samples/lm_finetuning",
train_filename="train-sample.txt",
test_filename="test-sample.txt",
dev_filename=None,
tokenizer=tokenizer,
max_seq_len=64,
)
data_silo = DataSilo(processor=processor, batch_size=batch_size)
language_model = Bert.load(lang_model)
lm_prediction_head = BertLMHead.load(lang_model)
next_sentence_head = NextSentenceHead.load(lang_model)
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[lm_prediction_head, next_sentence_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_token", "per_sequence"],
device=device,
)
optimizer, warmup_linear = initialize_optimizer(
model=model,
learning_rate=2e-5,
warmup_proportion=0.1,
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
)
trainer = Trainer(
optimizer=optimizer,
data_silo=data_silo,
epochs=n_epochs,
n_gpu=n_gpu,
warmup_linear=warmup_linear,
evaluate_every=evaluate_every,
device=device,
)
model = trainer.train(model)
save_dir = "testsave/lm_finetuning"
model.save(save_dir)
processor.save(save_dir)
# 2. Create a DataProcessor that handles all the conversion from raw text into a pytorch Dataset
processor = BertStyleLMProcessor(data_dir="../data/lm_finetune_nips",
tokenizer=tokenizer,
max_seq_len=128,
max_docs=30)
# 3. Create a DataSilo that loads several datasets (train/dev/test), provides DataLoaders for them and calculates a few descriptive statistics of our datasets
data_silo = DataSilo(processor=processor,
batch_size=batch_size,
max_multiprocessing_chunksize=20)
# 4. Create an AdaptiveModel
# a) which consists of a pretrained language model as a basis
language_model = LanguageModel.load(lang_model)
# b) and *two* prediction heads on top that are suited for our task => Language Model finetuning
lm_prediction_head = BertLMHead.load(lang_model)
next_sentence_head = NextSentenceHead.load(lang_model)
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[lm_prediction_head, next_sentence_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_token", "per_sequence"],
device=device,
)
# 5. Create an optimizer
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=2e-5,
device=device,
n_batches=len(data_silo.loaders["train"]),
def test_lm_finetuning(caplog):
caplog.set_level(logging.CRITICAL)
set_all_seeds(seed=42)
device, n_gpu = initialize_device_settings(use_cuda=False)
n_epochs = 1
batch_size = 1
evaluate_every = 2
lang_model = "bert-base-cased"
tokenizer = BertTokenizer.from_pretrained(
pretrained_model_name_or_path=lang_model,
do_lower_case=False,
never_split_chars=["-", "_"])
processor = BertStyleLMProcessor(
data_dir="samples/lm_finetuning",
train_filename="train-sample.txt",
test_filename="test-sample.txt",
dev_filename=None,
tokenizer=tokenizer,
max_seq_len=12,
)
data_silo = DataSilo(processor=processor, batch_size=batch_size)
language_model = Bert.load(lang_model)
lm_prediction_head = BertLMHead.load(lang_model)
next_sentence_head = NextSentenceHead.load(lang_model)
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[lm_prediction_head, next_sentence_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_token", "per_sequence"],
device=device,
)
optimizer, warmup_linear = initialize_optimizer(
model=model,
learning_rate=2e-5,
warmup_proportion=0.1,
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
)
trainer = Trainer(
optimizer=optimizer,
data_silo=data_silo,
epochs=n_epochs,
n_gpu=n_gpu,
warmup_linear=warmup_linear,
evaluate_every=evaluate_every,
device=device,
)
model = trainer.train(model)
save_dir = "testsave/lm_finetuning"
model.save(save_dir)
processor.save(save_dir)
basic_texts = [
{
"text": "Farmer's life is great."
},
{
"text": "It's nothing for big city kids though."
},
]
model = Inferencer.load(save_dir, embedder_only=True)
result = model.extract_vectors(dicts=basic_texts)
assert result[0]["context"] == [
'Farmer', "'", 's', 'life', 'is', 'great', '.'
]
assert result[0]["vec"].shape == (768, )
# TODO check why results vary accross runs with same seed
assert isinstance(result[0]["vec"][0], np.float32)
def test_lm_finetuning_custom_vocab(caplog):
caplog.set_level(logging.CRITICAL)
set_all_seeds(seed=42)
device, n_gpu = initialize_device_settings(use_cuda=False)
n_epochs = 1
batch_size = 1
evaluate_every = 2
lang_model = "bert-base-cased"
tokenizer = Tokenizer.load(
pretrained_model_name_or_path=lang_model, do_lower_case=False
)
tokenizer.add_tokens(["aaaaaaaaaaaaaaaa", "bbbbbbbbbbbbbbbbbbbbb", "ccccccccccccccccccccccc"])
processor = BertStyleLMProcessor(
data_dir="samples/lm_finetuning",
train_filename="train-sample.txt",
test_filename="test-sample.txt",
dev_filename=None,
tokenizer=tokenizer,
max_seq_len=12,
next_sent_pred=True
)
data_silo = DataSilo(processor=processor, batch_size=batch_size, max_processes=1)
language_model = LanguageModel.load(lang_model, n_added_tokens=len(tokenizer.added_tokens_decoder))
lm_prediction_head = BertLMHead.load(lang_model, n_added_tokens=len(tokenizer.added_tokens_decoder))
next_sentence_head = NextSentenceHead.load(lang_model)
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[lm_prediction_head, next_sentence_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_token", "per_sequence"],
device=device
)
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=2e-5,
#optimizer_opts={'name': 'AdamW', 'lr': 2E-05},
n_batches=len(data_silo.loaders["train"]),
n_epochs=1,
device=device,
schedule_opts={'name': 'CosineWarmup', 'warmup_proportion': 0.1}
)
trainer = Trainer(
optimizer=optimizer,
data_silo=data_silo,
epochs=n_epochs,
n_gpu=n_gpu,
lr_schedule=lr_schedule,
evaluate_every=evaluate_every,
device=device,
)
model = trainer.train(model)
# LM embeddings and weight of decoder in head are shared and should therefore be equal
assert torch.all(
torch.eq(model.language_model.model.embeddings.word_embeddings.weight, model.prediction_heads[0].decoder.weight))
save_dir = "testsave/lm_finetuning"
model.save(save_dir)
processor.save(save_dir)
basic_texts = [
{"text": "Farmer's life is great."},
{"text": "It's nothing for big city kids though."},
]
model = Inferencer.load(save_dir, embedder_only=True)
result = model.extract_vectors(dicts=basic_texts)
assert result[0]["context"] == ['Farmer', "'", 's', 'life', 'is', 'great', '.']
assert result[0]["vec"].shape == (768,)
# TODO check why results vary accross runs with same seed
assert isinstance(result[0]["vec"][0], np.float32)
def lm_finetuning():
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
set_all_seeds(seed=42)
ml_logger = MLFlowLogger(tracking_uri="https://public-mlflow.deepset.ai/")
ml_logger.init_experiment(experiment_name="Public_FARM",
run_name="Run_minimal_example_lm")
##########################
########## Settings
##########################
device, n_gpu = initialize_device_settings(use_cuda=False)
n_epochs = 1
batch_size = 32
evaluate_every = 30
lang_model = "bert-base-cased"
do_lower_case = False
# 1.Create a tokenizer
tokenizer = Tokenizer.load(pretrained_model_name_or_path=lang_model,
do_lower_case=do_lower_case)
# 2. Create a DataProcessor that handles all the conversion from raw text into a pytorch Dataset
processor = BertStyleLMProcessor(data_dir=Path("../data/lm_finetune_nips"),
tokenizer=tokenizer,
max_seq_len=128,
max_docs=20)
# 3. Create a DataSilo that loads several datasets (train/dev/test), provides DataLoaders for them and calculates a few descriptive statistics of our datasets
data_silo = DataSilo(processor=processor,
batch_size=batch_size,
max_multiprocessing_chunksize=20)
# 4. Create an AdaptiveModel
# a) which consists of a pretrained language model as a basis
language_model = LanguageModel.load(lang_model)
# b) and *two* prediction heads on top that are suited for our task => Language Model finetuning
lm_prediction_head = BertLMHead.load(lang_model)
next_sentence_head = NextSentenceHead.load(lang_model)
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[lm_prediction_head, next_sentence_head],
embeds_dropout_prob=0.1,
lm_output_types=["per_token", "per_sequence"],
device=device,
)
# 5. Create an optimizer
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=2e-5,
device=device,
n_batches=len(data_silo.loaders["train"]),
n_epochs=n_epochs,
)
# 6. Feed everything to the Trainer, which keeps care of growing our model into powerful plant and evaluates it from time to time
trainer = Trainer(
model=model,
optimizer=optimizer,
data_silo=data_silo,
epochs=n_epochs,
n_gpu=n_gpu,
lr_schedule=lr_schedule,
evaluate_every=evaluate_every,
device=device,
)
# 7. Let it grow! Watch the tracked metrics live on the public mlflow server: https://public-mlflow.deepset.ai
trainer.train()
# 8. Hooray! You have a model. Store it:
save_dir = Path("saved_models/bert-english-lm-tutorial")
model.save(save_dir)
processor.save(save_dir)
def outcome_pretraining(task_config,
model_name,
cache_dir,
run_name="0",
lr=1e-05,
warmup_steps=5000,
embeds_dropout=0.1,
epochs=200, # large because we use early stopping by default
batch_size=20,
grad_acc_steps=1,
early_stopping_metric="loss",
early_stopping_mode="min",
early_stopping_patience=10,
model_class="Bert",
tokenizer_class="BertTokenizer",
do_lower_case=True,
do_train=True,
do_eval=True,
do_hpo=False,
max_seq_len=512,
seed=11,
eval_every=500,
use_amp=False,
use_cuda=True,
):
# Load task config
task_config = yaml.safe_load(open(task_config))
data_dir = Path(task_config["data"]["data_dir"])
# General Settings
set_all_seeds(seed=seed)
device, n_gpu = initialize_device_settings(use_cuda=use_cuda, use_amp=use_amp)
# 1.Create a tokenizer
tokenizer = Tokenizer.load(pretrained_model_name_or_path=model_name, tokenizer_class=tokenizer_class,
do_lower_case=do_lower_case)
# 2. Create a DataProcessor that handles all the conversion from raw text into a pytorch Dataset
processor = OutcomePretrainingProcessor(tokenizer=tokenizer,
max_seq_len=max_seq_len,
data_dir=data_dir,
train_filename=task_config["data"]["train_filename"],
dev_filename=task_config["data"]["dev_filename"],
seed=seed,
max_size_admission=50,
max_size_discharge=50,
cache_dir=cache_dir)
# 3. Create a DataSilo that loads several datasets (train/dev/test), provides DataLoaders for them and calculates a
# few descriptive statistics of our datasets
data_silo = OutcomePretrainingDataSilo(
processor=processor,
caching=True,
cache_dir=cache_dir,
batch_size=batch_size,
max_multiprocessing_chunksize=200)
if do_train:
# Set save dir for experiment output
save_dir = Path(task_config["output_dir"]) / f'{task_config["experiment_name"]}_{run_name}'
# Use HPO config args if config is passed
if do_hpo:
save_dir = save_dir / tune.session.get_trial_name()
else:
exp_name = f"exp_{random.randint(100000, 999999)}"
save_dir = save_dir / exp_name
# Create save dir
if not os.path.exists(save_dir):
os.makedirs(save_dir)
# Setup MLFlow logger
ml_logger = MLFlowLogger(tracking_uri=task_config["log_dir"])
ml_logger.init_experiment(experiment_name=task_config["experiment_name"],
run_name=f'{task_config["experiment_name"]}_{run_name}')
# 4. Create an AdaptiveModel
# a) which consists of a pretrained language model as a basis
language_model = LanguageModel.load(model_name, language_model_class=model_class)
# b) and NextSentenceHead prediction head or TextClassificationHead if it's not a Bert Model
if model_class == "Bert":
next_sentence_head = NextSentenceHead.load(model_class)
else:
next_sentence_head = TextClassificationHead(num_labels=2)
model = AdaptiveModel(
language_model=language_model,
prediction_heads=[next_sentence_head],
embeds_dropout_prob=embeds_dropout,
lm_output_types=["per_sequence"],
device=device,
)
# 5. Create an optimizer
schedule_opts = {"name": "LinearWarmup",
"num_warmup_steps": warmup_steps}
model, optimizer, lr_schedule = initialize_optimizer(
model=model,
learning_rate=lr,
device=device,
n_batches=len(data_silo.loaders["train"]),
n_epochs=epochs,
use_amp=use_amp,
grad_acc_steps=grad_acc_steps,
schedule_opts=schedule_opts)
# 6. Create an early stopping instance
early_stopping = None
if early_stopping_mode != "none":
early_stopping = EarlyStopping(
mode=early_stopping_mode,
min_delta=0.0001,
save_dir=save_dir,
metric=early_stopping_metric,
patience=early_stopping_patience
)
# 7. Feed everything to the Trainer, which keeps care of growing our model into powerful plant and evaluates it
# from time to time
trainer = ExtendedTrainer(
model=model,
optimizer=optimizer,
data_silo=data_silo,
epochs=epochs,
n_gpu=n_gpu,
lr_schedule=lr_schedule,
evaluate_every=eval_every,
early_stopping=early_stopping,
device=device,
grad_acc_steps=grad_acc_steps,
evaluator_test=do_eval
)
def score_callback(eval_score, train_loss):
tune.report(roc_auc_dev=eval_score, train_loss=train_loss)
# 8. Train the model
trainer.train(score_callback=score_callback if do_hpo else None)
# 9. Save model if not saved in early stopping
model.save(save_dir / "final_model")
processor.save(save_dir / "final_model")
if do_eval:
# Load newly trained model or existing model
if do_train:
model_dir = save_dir
else:
model_dir = Path(model_name)
logger.info("###### Eval on TEST SET #####")
evaluator_test = Evaluator(
data_loader=data_silo.get_data_loader("test"),
tasks=data_silo.processor.tasks,
device=device
)
# Load trained model for evaluation
model = AdaptiveModel.load(model_dir, device)
model.connect_heads_with_processor(data_silo.processor.tasks, require_labels=True)
# Evaluate
results = evaluator_test.eval(model, return_preds_and_labels=True)
# Log results
utils.log_results(results, dataset_name="test", steps=len(evaluator_test.data_loader),
save_path=model_dir / "eval_results.txt")