def pretrain_and_evaluate(args, model, eval_only):
val_dataset = tf.data.TFRecordDataset(file_path=args.val_datapath)
if eval_only:
train_dataset = val_dataset
else:
logger.info(
f'Loading and tokenizing training data is usually slow: {args.train_datapath}'
)
train_dataset = tf.data.Dataset(file_path=args.train_datapath)
trainer = TFTrainer(model=model,
args=args,
train_dataset=train_dataset,
eval_dataset=val_dataset)
eval_loss = trainer.evaluate()
eval_loss = eval_loss['eval_loss']
logger.info(f'Initial eval bpc: {eval_loss / math.log(2)}')
if not eval_only:
trainer.train()
trainer.save_model()
eval_loss = trainer.evaluate()
eval_loss = eval_loss['eval_loss']
logger.info(f'Eval bpc after pretraining: {eval_loss / math.log(2)}')
async def accuracy(self, sources: Sources):
if not os.path.isfile(
os.path.join(self.parent.config.output_dir, "tf_model.h5")):
raise ModelNotTrained("Train model before assessing for accuracy.")
config = self.parent.config._asdict()
self.tokenizer = AutoTokenizer.from_pretrained(
self.parent.config.output_dir)
eval_features = await self._preprocess_data(sources)
eval_dataset = await self.example_features_to_dataset(eval_features)
def compute_metrics(p: EvalPrediction) -> Dict:
preds = self.np.argmax(p.predictions, axis=1)
return classification_compute_metrics(preds, p.label_ids)
with self.parent.config.strategy.scope():
self.model = TFAutoModelForSequenceClassification.from_pretrained(
config["directory"])
trainer = TFTrainer(
model=self.model,
args=self.parent.config,
eval_dataset=eval_dataset,
compute_metrics=compute_metrics,
)
result = trainer.evaluate()
return Accuracy(result["eval_acc"])
async def accuracy(self, sources: Sources):
if not os.path.isfile(
os.path.join(self.parent.config.output_dir, "tf_model.h5")
):
raise ModelNotTrained("Train model before assessing for accuracy.")
data_df = await self._preprocess_data(sources)
eval_dataset = self.get_dataset(data_df, self.tokenizer, mode="eval",)
with self.parent.config.strategy.scope():
self.model = TFAutoModelForTokenClassification.from_pretrained(
self.parent.config.output_dir,
config=self.config,
cache_dir=self.parent.config.cache_dir,
)
trainer = TFTrainer(
model=self.model,
args=self.parent.config,
train_dataset=None,
eval_dataset=eval_dataset.get_dataset(),
compute_metrics=self.compute_metrics,
)
result = trainer.evaluate()
return Accuracy(result["eval_f1"])
def finetune_with_tftrainer():
""" Fine tune with TFTrainer but it's not working with some package error
Found an issue from github of the same problem I have, follow up there:
https://github.com/huggingface/transformers/issues/5151
"""
tokenizer = AutoTokenizer.from_pretrained(os.getenv('MODEL_NAME'))
# Get data for fine-tuning
dataset, dataset_size, num_labels = build_dataset(tokenizer)
config = AutoConfig.from_pretrained(os.getenv('MODEL_NAME'),
num_labels=num_labels)
training_args = TFTrainingArguments(output_dir=os.getenv('OUTPUT_DIR'),
logging_dir=os.getenv('OUTPUT_DIR'),
logging_first_step=True,
logging_steps=1,
overwrite_output_dir=True,
do_train=True,
do_eval=True,
learning_rate=2e-5,
debug=True)
with training_args.strategy.scope():
# model = TFAutoModel.from_pretrained(os.getenv('MODEL_NAME'), config=config, cache_dir=os.getenv('OUTPUT_DIR'))
model = TCPMDistilBertClassification.from_pretrained(
os.getenv('MODEL_NAME'), config=config)
# shuffle and split train/test tasks manuanly
dataset = dataset.shuffle(dataset_size)
train_size = int(dataset_size * (4 / 5))
# test_size = dataset_size - train_size # didn't use so commented out.
train_data = dataset.take(train_size) #.batch(TRAIN_BATCH_SIZE)
test_data = dataset.skip(train_size) #.batch(TEST_BATCH_SIZE)
trainer = TFTrainer(model=model,
args=training_args,
train_dataset=train_data,
eval_dataset=test_data,
compute_metrics=compute_metrics)
print('Sample training data:')
for i in train_data.take(2):
print(f'Input #{i}:')
pprint(i[0])
print(f'Target #{i}:')
pprint(i[1])
print('Sample test data:')
for i in train_data.take(2):
print(f'Input #{i}:')
pprint(i[0])
print(f'Output #{i}:')
pprint(i[1])
print('TF Trainer args:')
pprint(training_args)
print('TCPM Model config:')
pprint(config)
print('Trainning start..')
# Train the model
trainer.train()
trainer.save_model()
tokenizer.save_pretrained(os.getenv('OUTPUT_DIR'))
# Evaluate the model
result = trainer.evaluate()
pprint(result)
with open(os.path.join(os.getenv('OUTPUT_DIR'), 'eval_results.json'),
'w') as fwrite:
json.dump(result, fwrite, indent=4)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TFTrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir) and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(
"n_replicas: %s, distributed training: %s, 16-bits training: %s",
training_args.n_replicas,
bool(training_args.n_replicas > 1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
train_dataset, eval_dataset, test_ds, label2id = get_tfds(
train_file=data_args.train_file,
eval_file=data_args.dev_file,
test_file=data_args.test_file,
tokenizer=tokenizer,
label_column_id=data_args.label_column_id,
max_seq_length=data_args.max_seq_length,
)
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
num_labels=len(label2id),
label2id=label2id,
id2label={id: label
for label, id in label2id.items()},
finetuning_task="text-classification",
cache_dir=model_args.cache_dir,
)
with training_args.strategy.scope():
model = TFAutoModelForSequenceClassification.from_pretrained(
model_args.model_name_or_path,
from_pt=bool(".bin" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
def compute_metrics(p: EvalPrediction) -> Dict:
preds = np.argmax(p.predictions, axis=1)
return {"acc": (preds == p.label_ids).mean()}
# Initialize our Trainer
trainer = TFTrainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
trainer.train()
trainer.save_model()
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
result = trainer.evaluate()
trainer.log_metrics("eval", result)
trainer.save_metrics("eval", result)
results.update(result)
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, GlueDataTrainingArguments, TFTrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(
"n_replicas: %s, distributed training: %s, 16-bits training: %s",
training_args.n_replicas,
bool(training_args.n_replicas > 1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
try:
num_labels = glue_tasks_num_labels["mnli" if data_args.task_name == "mnli-mm" else data_args.task_name]
output_mode = glue_output_modes[data_args.task_name]
except KeyError:
raise ValueError("Task not found: %s" % (data_args.task_name))
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
finetuning_task=data_args.task_name,
cache_dir=model_args.cache_dir,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
with training_args.strategy.scope():
model = TFAutoModelForSequenceClassification.from_pretrained(
model_args.model_name_or_path,
from_pt=bool(".bin" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
# Get datasets
train_dataset = (
get_tfds(
task_name=data_args.task_name,
tokenizer=tokenizer,
max_seq_length=data_args.max_seq_length,
data_dir=data_args.data_dir,
)
if training_args.do_train
else None
)
eval_dataset = (
get_tfds(
task_name=data_args.task_name,
tokenizer=tokenizer,
max_seq_length=data_args.max_seq_length,
mode=Split.dev,
data_dir=data_args.data_dir,
)
if training_args.do_eval
else None
)
def compute_metrics(p: EvalPrediction) -> Dict:
if output_mode == "classification":
preds = np.argmax(p.predictions, axis=1)
elif output_mode == "regression":
preds = np.squeeze(p.predictions)
return glue_compute_metrics(data_args.task_name, preds, p.label_ids)
# Initialize our Trainer
trainer = TFTrainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
trainer.train()
trainer.save_model()
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
result = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in result.items():
logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
results.update(result)
return results
def run_bert_regression_trainer():
""" Run bert single class classification(a.k.a regression) model."""
print('START TRAINNING FOR REGRESSION')
log_dir = os.path.join(os.getenv('OUTPUT_DIR'), 'hf_trainer')
tokenizer = AutoTokenizer.from_pretrained(os.getenv('MODEL_NAME'))
config = AutoConfig.from_pretrained(os.getenv('MODEL_NAME'), num_labels=1)
training_args = TFTrainingArguments(
output_dir=log_dir,
logging_dir=log_dir,
logging_first_step=True,
logging_steps=1,
overwrite_output_dir=True,
do_train=True,
do_eval=True,
learning_rate=2e-5,
debug=True,
)
with training_args.strategy.scope():
# model = TFDistilBertForSequenceClassification.from_pretrained(os.getenv('MODEL_NAME'), config=config)
model = TCPMDistilBertRegression.from_pretrained(
os.getenv('MODEL_NAME'), config=config)
print('\nModel Config:')
print(config)
print('Tokenizer: ', tokenizer)
print('Model: ', model)
print('\nTFTraingArguments:')
print(training_args)
tc = TopCoder()
encoded_text = tc.get_bert_encoded_txt_features(tokenizer)
metadata = tc.get_meta_data_features(encoded_tech=True, softmax_tech=True)
target = tc.get_target()
split = int((4 / 5) * len(target))
dataset = tf.data.Dataset.from_tensor_slices(
(dict(**encoded_text, meta_input=metadata), target))
dataset = dataset.shuffle(len(target))
train_ds, test_ds = dataset.take(split), dataset.skip(split)
print('\nTrain dataset samples:')
for el in train_ds.take(3):
pprint(el)
print('\nTest dataset samples:')
for el in test_ds.take(3):
pprint(el)
trainer = TFTrainer(
model=model,
args=training_args,
train_dataset=train_ds,
eval_dataset=test_ds,
compute_metrics=compute_metrics,
)
trainer.train()
trainer.save_model()
tokenizer.save_pretrained(log_dir)
result = trainer.evaluate()
print('\nTrainning eval:')
pprint(result)
with open(os.path.join(log_dir, 'eval_results.json'), 'w') as fwrite:
json.dump(result, fwrite, indent=4)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TFTrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(
"n_replicas: %s, distributed training: %s, 16-bits training: %s",
training_args.n_replicas,
bool(training_args.n_replicas > 1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Prepare Token Classification task
labels = get_labels(data_args.labels)
label_map: Dict[int, str] = {i: label for i, label in enumerate(labels)}
num_labels = len(labels)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
id2label=label_map,
label2id={label: i for i, label in enumerate(labels)},
cache_dir=model_args.cache_dir,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast,
)
with training_args.strategy.scope():
model = TFAutoModelForTokenClassification.from_pretrained(
model_args.model_name_or_path,
from_pt=bool(".bin" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
# Get datasets
train_dataset = (
TFNerDataset(
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.train,
)
if training_args.do_train
else None
)
eval_dataset = (
TFNerDataset(
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.dev,
)
if training_args.do_eval
else None
)
def align_predictions(predictions: np.ndarray, label_ids: np.ndarray) -> Tuple[List[int], List[int]]:
preds = np.argmax(predictions, axis=2)
batch_size, seq_len = preds.shape
out_label_list = [[] for _ in range(batch_size)]
preds_list = [[] for _ in range(batch_size)]
for i in range(batch_size):
for j in range(seq_len):
if label_ids[i, j] != -1:
out_label_list[i].append(label_map[label_ids[i][j]])
preds_list[i].append(label_map[preds[i][j]])
return preds_list, out_label_list
def compute_metrics(p: EvalPrediction) -> Dict:
preds_list, out_label_list = align_predictions(p.predictions, p.label_ids)
return {
"precision": precision_score(out_label_list, preds_list),
"recall": recall_score(out_label_list, preds_list),
"f1": f1_score(out_label_list, preds_list),
}
# Initialize our Trainer
trainer = TFTrainer(
model=model,
args=training_args,
train_dataset=train_dataset.get_dataset() if train_dataset else None,
eval_dataset=eval_dataset.get_dataset() if eval_dataset else None,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
trainer.train()
trainer.save_model()
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
result = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in result.items():
logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
results.update(result)
# Predict
if training_args.do_predict:
test_dataset = TFNerDataset(
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.test,
)
predictions, label_ids, metrics = trainer.predict(test_dataset.get_dataset())
preds_list, labels_list = align_predictions(predictions, label_ids)
report = classification_report(labels_list, preds_list)
logger.info("\n%s", report)
output_test_results_file = os.path.join(training_args.output_dir, "test_results.txt")
with open(output_test_results_file, "w") as writer:
writer.write("%s\n" % report)
# Save predictions
output_test_predictions_file = os.path.join(training_args.output_dir, "test_predictions.txt")
with open(output_test_predictions_file, "w") as writer:
with open(os.path.join(data_args.data_dir, "test.txt"), "r") as f:
example_id = 0
for line in f:
if line.startswith("-DOCSTART-") or line == "" or line == "\n":
writer.write(line)
if not preds_list[example_id]:
example_id += 1
elif preds_list[example_id]:
output_line = line.split()[0] + " " + preds_list[example_id].pop(0) + "\n"
writer.write(output_line)
else:
logger.warning("Maximum sequence length exceeded: No prediction for '%s'.", line.split()[0])
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TFTrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.warning(
"device: %s, n_replicas: %s, 16-bits training: %s",
training_args.device,
training_args.n_replicas,
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Set seed
set_seed(training_args.seed)
try:
processor = processors[data_args.task_name]()
label_list = processor.get_labels()
num_labels = len(label_list)
except KeyError:
raise ValueError("Task not found: %s" % (data_args.task_name))
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
finetuning_task=data_args.task_name,
cache_dir=model_args.cache_dir,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
with training_args.strategy.scope():
model = TFAutoModelForMultipleChoice.from_pretrained(
model_args.model_name_or_path,
from_pt=bool(".bin" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
# Get datasets
train_dataset = (
TFMultipleChoiceDataset(
data_dir=data_args.data_dir,
tokenizer=tokenizer,
task=data_args.task_name,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.train,
)
if training_args.do_train
else None
)
eval_dataset = (
TFMultipleChoiceDataset(
data_dir=data_args.data_dir,
tokenizer=tokenizer,
task=data_args.task_name,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.dev,
)
if training_args.do_eval
else None
)
def compute_metrics(p: EvalPrediction) -> Dict:
preds = np.argmax(p.predictions, axis=1)
return {"acc": simple_accuracy(preds, p.label_ids)}
# Initialize our Trainer
trainer = TFTrainer(
model=model,
args=training_args,
train_dataset=train_dataset.get_dataset() if train_dataset else None,
eval_dataset=eval_dataset.get_dataset() if eval_dataset else None,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
trainer.train()
trainer.save_model()
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
result = trainer.evaluate()
trainer.log_metrics("eval", results)
trainer.save_metrics("eval", results)
results.update(result)
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TFTrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
module = import_module("tasks")
try:
token_classification_task_clazz = getattr(module, model_args.task_type)
token_classification_task: TokenClassificationTask = token_classification_task_clazz()
except AttributeError:
raise ValueError(
f"Task {model_args.task_type} needs to be defined as a TokenClassificationTask subclass in {module}. "
f"Available tasks classes are: {TokenClassificationTask.__subclasses__()}"
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.DEBUG,
)
"""
logging.basicConfig(
filename="/scratch/project_2001426/harttu/july-2020/transformers-ner/test.log",
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.DEBUG,
filemode='w'
)
"""
logger.info("FROM run_tf_ner.py")
logger.info(
"n_replicas: %s, distributed training: %s, 16-bits training: %s",
training_args.n_replicas,
bool(training_args.n_replicas > 1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Prepare Token Classification task
labels = token_classification_task.get_labels(data_args.labels)
print("LABELS:")
print(labels)
label_map: Dict[int, str] = {i: label for i, label in enumerate(labels)}
print("LABEL_MAP:")
print(label_map)
num_labels = len(labels)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
id2label=label_map,
label2id={label: i for i, label in enumerate(labels)},
cache_dir=model_args.cache_dir,
)
print("CONFIG")
print(config)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast,
)
with training_args.strategy.scope():
model = TFAutoModelForTokenClassification.from_pretrained(
model_args.model_name_or_path,
from_pt=True,#bool(".bin" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
print("SETTING DATASETS")
# Get datasets
train_dataset = (
TFTokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.train,
)
if training_args.do_train
else None
)
eval_dataset = (
TFTokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.dev,
)
if training_args.do_eval
else None
)
def align_predictions(predictions: np.ndarray, label_ids: np.ndarray) -> Tuple[List[int], List[int]]:
preds = np.argmax(predictions, axis=2)
batch_size, seq_len = preds.shape
out_label_list = [[] for _ in range(batch_size)]
preds_list = [[] for _ in range(batch_size)]
for i in range(batch_size):
for j in range(seq_len):
if label_ids[i, j] != -100:
out_label_list[i].append(label_map[label_ids[i][j]])
preds_list[i].append(label_map[preds[i][j]])
return preds_list, out_label_list
def compute_metrics(p: EvalPrediction) -> Dict:
preds_list, out_label_list = align_predictions(p.predictions, p.label_ids)
return {
"precision": precision_score(out_label_list, preds_list),
"recall": recall_score(out_label_list, preds_list),
"f1": f1_score(out_label_list, preds_list),
}
"""
# Prepare optimizer and schedule (linear warmup and decay)
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
"weight_decay": args.weight_decay,
},
{"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
)
"""
import tensorflow as tf
initial_learning_rate=0.001
decay_steps=10000
print("INITIALIZING TRAINER")
# Initialize our Trainer
trainer = TFTrainer(
model=model,
#optimizers=(tf.keras.optimizers.Adam(
#learning_rate=initial_learning_rate, beta_1=0.9,
#beta_2=0.999, epsilon=1e-07, amsgrad=False,
#name='Adam'),tf.keras.optimizers.schedules.PolynomialDecay(
#initial_learning_rate, decay_steps,
#end_learning_rate=0.0001, power=1.0,
#cycle=False, name=None
#)),
args=training_args,
train_dataset=train_dataset.get_dataset() if train_dataset else None,
eval_dataset=eval_dataset.get_dataset() if eval_dataset else None,
compute_metrics=compute_metrics,
)
print("TRAINING")
# Training
if training_args.do_train:
trainer.train()
trainer.save_model()
tokenizer.save_pretrained(training_args.output_dir)
print("EVALUATING")
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
result = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in result.items():
logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
results.update(result)
# Predict
print("PREDICTING")
if training_args.do_predict:
test_dataset = TFTokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.test,
)
predictions, label_ids, metrics = trainer.predict(test_dataset.get_dataset())
preds_list, labels_list = align_predictions(predictions, label_ids)
report = classification_report(labels_list, preds_list)
logger.info("\n%s", report)
output_test_results_file = os.path.join(training_args.output_dir, "test_results.txt")
with open(output_test_results_file, "w") as writer:
writer.write("%s\n" % report)
# Save predictions
output_test_predictions_file = os.path.join(training_args.output_dir, "test_predictions.txt")
with open(output_test_predictions_file, "w") as writer:
with open(os.path.join(data_args.data_dir, "test.txt"), "r") as f:
example_id = 0
for line in f:
if line.startswith("-DOCSTART-") or line == "" or line == "\n":
writer.write(line)
if not preds_list[example_id]:
example_id += 1
elif preds_list[example_id]:
output_line = line.split()[0] + " " + preds_list[example_id].pop(0) + "\n"
writer.write(output_line)
else:
logger.warning("Maximum sequence length exceeded: No prediction for '%s'.", line.split()[0])
return results
def main(self):
model_args = ModelArguments(model_name_or_path=self.model_name_or_path)
data_args = DataTrainingArguments(data_dir=self.data_dir,
labels='./labels.txt',
max_seq_length=self.max_seq_length)
training_args = TFTrainingArguments(
output_dir=self.output_dir,
do_eval=self.do_eval,
do_predict=self.do_predict,
do_train=self.do_train,
per_device_train_batch_size=self.per_device_train_batch_size,
save_steps=self.save_steps,
seed=self.seed)
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
module = import_module("tasks")
try:
token_classification_task_clazz = getattr(module,
model_args.task_type)
token_classification_task: TokenClassificationTask = token_classification_task_clazz(
)
except AttributeError:
raise ValueError(
f"Task {model_args.task_type} needs to be defined as a TokenClassificationTask subclass in {module}. "
f"Available tasks classes are: {TokenClassificationTask.__subclasses__()}"
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(
"n_replicas: %s, distributed training: %s, 16-bits training: %s",
training_args.n_replicas,
bool(training_args.n_replicas > 1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Prepare Token Classification task
labels = token_classification_task.get_labels(data_args.labels)
# JQ
labels.extend('PAD')
label_map: Dict[int,
str] = {i: label
for i, label in enumerate(labels)}
# JQ
label_map[9] = 'PAD'
num_labels = len(labels)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
# START HERE
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
id2label=label_map,
label2id={label: i
for i, label in enumerate(labels)},
cache_dir=model_args.cache_dir,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast,
)
with training_args.strategy.scope():
model = TFAutoModelForTokenClassification.from_pretrained(
model_args.model_name_or_path,
from_pt=bool(".bin" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
# Get datasets
train_dataset = (TFTokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.train,
) if training_args.do_train else None)
eval_dataset = (TFTokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.dev,
) if training_args.do_eval else None)
def align_predictions(
predictions: np.ndarray,
label_ids: np.ndarray) -> Tuple[List[int], List[int]]:
preds = np.argmax(predictions, axis=2)
batch_size, seq_len = preds.shape
out_label_list = [[] for _ in range(batch_size)]
preds_list = [[] for _ in range(batch_size)]
for i in range(batch_size):
for j in range(seq_len):
if label_ids[i, j] != -100:
out_label_list[i].append(label_map[label_ids[i][j]])
preds_list[i].append(label_map[preds[i][j]])
return preds_list, out_label_list
def compute_metrics(p: EvalPrediction) -> Dict:
preds_list, out_label_list = align_predictions(
p.predictions, p.label_ids)
return {
"precision": precision_score(out_label_list, preds_list),
"recall": recall_score(out_label_list, preds_list),
"f1": f1_score(out_label_list, preds_list),
}
# Initialize our Trainer
trainer = TFTrainer(
model=model,
args=training_args,
train_dataset=train_dataset.get_dataset()
if train_dataset else None,
eval_dataset=eval_dataset.get_dataset() if eval_dataset else None,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
trainer.train()
trainer.save_model()
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
result = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir,
"eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in result.items():
logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
results.update(result)
# Predict
if training_args.do_predict:
test_dataset = TFTokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.test,
)
predictions, label_ids, metrics = trainer.predict(
test_dataset.get_dataset())
preds_list, labels_list = align_predictions(predictions, label_ids)
report = classification_report(labels_list, preds_list)
logger.info("\n%s", report)
output_test_results_file = os.path.join(training_args.output_dir,
"test_results.txt")
with open(output_test_results_file, "w") as writer:
writer.write("%s\n" % report)
# Save predictions
output_test_predictions_file = os.path.join(
training_args.output_dir, "test_predictions.txt")
with open(output_test_predictions_file, "w") as writer:
with open(os.path.join(data_args.data_dir, "test.txt"),
"r") as f:
example_id = 0
for line in f:
if line.startswith(
"-DOCSTART-") or line == "" or line == "\n":
writer.write(line)
if not preds_list[example_id]:
example_id += 1
elif preds_list[example_id]:
output_line = line.split(
)[0] + " " + preds_list[example_id].pop(0) + "\n"
writer.write(output_line)
else:
logger.warning(
"Maximum sequence length exceeded: No prediction for '%s'.",
line.split()[0])
return results
