def main():
parser = PdArgumentParser(
(ModelArguments, DataArguments, TrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# Log model and data config
training_args.print_config(model_args, "Model")
training_args.print_config(data_args, "Data")
paddle.set_device(training_args.device)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, world_size: {training_args.world_size}, "
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# set_seed(args)
data_args.dataset = data_args.dataset.strip()
if data_args.dataset not in ALL_DATASETS:
raise ValueError("Not found dataset {}".format(data_args.dataset))
if data_args.dataset in ALL_DATASETS:
# if you custom you hyper-parameters in yaml config, it will overwrite all args.
config = ALL_DATASETS[data_args.dataset]
for args in (model_args, data_args, training_args):
for arg in vars(args):
if arg in config.keys():
setattr(args, arg, config[arg])
training_args.per_device_train_batch_size = config["batch_size"]
training_args.per_device_eval_batch_size = config["batch_size"]
dataset_config = data_args.dataset.split(" ")
raw_datasets = load_dataset(
dataset_config[0],
None if len(dataset_config) <= 1 else dataset_config[1],
)
label_list = getattr(raw_datasets['train'], "label_list", None)
data_args.label_list = label_list
data_args.ignore_label = -100
data_args.no_entity_id = len(data_args.label_list) - 1
num_classes = 1 if raw_datasets["train"].label_list == None else len(
raw_datasets['train'].label_list)
# Define tokenizer, model, loss function.
tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path)
model = AutoModelForTokenClassification.from_pretrained(
model_args.model_name_or_path, num_classes=num_classes)
class criterion(nn.Layer):
def __init__(self):
super(criterion, self).__init__()
self.loss_fn = paddle.nn.loss.CrossEntropyLoss(
ignore_index=data_args.ignore_label)
def forward(self, *args, **kwargs):
return paddle.mean(self.loss_fn(*args, **kwargs))
loss_fct = criterion()
# Define dataset pre-process function
trans_fn = partial(ner_trans_fn, tokenizer=tokenizer, args=data_args)
# Define data collector
data_collator = DataCollatorForTokenClassification(tokenizer)
# Dataset pre-process
if training_args.do_train:
train_dataset = raw_datasets["train"].map(trans_fn)
if training_args.do_eval:
eval_dataset = raw_datasets["dev"].map(trans_fn)
if training_args.do_predict:
test_dataset = raw_datasets["test"].map(trans_fn)
# Define the metrics of tasks.
# Metrics
metric = load_metric("seqeval")
def compute_metrics(p):
predictions, labels = p
predictions = np.argmax(predictions, axis=2)
# Remove ignored index (special tokens)
true_predictions = [[
label_list[p] for (p, l) in zip(prediction, label) if l != -100
] for prediction, label in zip(predictions, labels)]
true_labels = [[
label_list[l] for (p, l) in zip(prediction, label) if l != -100
] for prediction, label in zip(predictions, labels)]
results = metric.compute(predictions=true_predictions,
references=true_labels)
return {
"precision": results["overall_precision"],
"recall": results["overall_recall"],
"f1": results["overall_f1"],
"accuracy": results["overall_accuracy"],
}
trainer = Trainer(
model=model,
criterion=loss_fct,
args=training_args,
data_collator=data_collator,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
compute_metrics=compute_metrics,
)
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
elif last_checkpoint is not None:
checkpoint = last_checkpoint
# Training
if training_args.do_train:
train_result = trainer.train(resume_from_checkpoint=checkpoint)
metrics = train_result.metrics
trainer.save_model()
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluate and tests model
if training_args.do_eval:
eval_metrics = trainer.evaluate()
trainer.log_metrics("eval", eval_metrics)
if training_args.do_predict:
test_ret = trainer.predict(test_dataset)
trainer.log_metrics("test", test_ret.metrics)
if test_ret.label_ids is None:
paddle.save(
test_ret.predictions,
os.path.join(training_args.output_dir,
"test_results.pdtensor"),
)
# export inference model
if training_args.do_export:
# You can also load from certain checkpoint
# trainer.load_state_dict_from_checkpoint("/path/to/checkpoint/")
input_spec = [
paddle.static.InputSpec(shape=[None, None],
dtype="int64"), # input_ids
paddle.static.InputSpec(shape=[None, None],
dtype="int64") # segment_ids
]
if model_args.export_model_dir is None:
model_args.export_model_dir = os.path.join(
training_args.output_dir, "export")
paddlenlp.transformers.export_model(model=trainer.model,
input_spec=input_spec,
path=model_args.export_model_dir)
def main():
parser = PdArgumentParser(
(ModelArguments, DataArguments, TrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# Log model and data config
training_args.print_config(model_args, "Model")
training_args.print_config(data_args, "Data")
paddle.set_device(training_args.device)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, world_size: {training_args.world_size}, "
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
data_args.dataset = data_args.dataset.strip()
dataset_config = data_args.dataset.split(" ")
print(dataset_config)
raw_datasets = load_dataset(
dataset_config[0],
name=None if len(dataset_config) <= 1 else dataset_config[1],
splits=('train', 'dev'))
data_args.label_list = getattr(raw_datasets['train'], "label_list", None)
num_classes = 1 if raw_datasets["train"].label_list == None else len(
raw_datasets['train'].label_list)
# Define tokenizer, model, loss function.
tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path)
model = AutoModelForSequenceClassification.from_pretrained(
model_args.model_name_or_path, num_classes=num_classes)
criterion = nn.loss.CrossEntropyLoss(
) if data_args.label_list else nn.loss.MSELoss()
# Define dataset pre-process function
trans_fn = partial(clue_trans_fn, tokenizer=tokenizer, args=data_args)
# Define data collector
data_collator = DataCollatorWithPadding(tokenizer)
# Dataset pre-process
if training_args.do_train:
train_dataset = raw_datasets["train"].map(trans_fn)
if training_args.do_eval:
eval_dataset = raw_datasets["dev"].map(trans_fn)
if training_args.do_predict:
test_dataset = raw_datasets["test"].map(trans_fn)
# Define the metrics of tasks.
def compute_metrics(p):
preds = p.predictions[0] if isinstance(p.predictions,
tuple) else p.predictions
preds = paddle.to_tensor(preds)
label = paddle.to_tensor(p.label_ids)
probs = F.softmax(preds, axis=1)
metric = Accuracy()
metric.reset()
result = metric.compute(preds, label)
metric.update(result)
accu = metric.accumulate()
metric.reset()
return {"accuracy": accu}
trainer = Trainer(
model=model,
criterion=criterion,
args=training_args,
data_collator=data_collator,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
compute_metrics=compute_metrics,
)
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
elif last_checkpoint is not None:
checkpoint = last_checkpoint
# Training
if training_args.do_train:
train_result = trainer.train(resume_from_checkpoint=checkpoint)
metrics = train_result.metrics
trainer.save_model()
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluate and tests model
if training_args.do_eval:
eval_metrics = trainer.evaluate()
trainer.log_metrics("eval", eval_metrics)
if training_args.do_predict:
test_ret = trainer.predict(test_dataset)
trainer.log_metrics("test", test_ret.metrics)
if test_ret.label_ids is None:
paddle.save(
test_ret.predictions,
os.path.join(training_args.output_dir,
"test_results.pdtensor"),
)
# export inference model
if training_args.do_export:
# You can also load from certain checkpoint
# trainer.load_state_dict_from_checkpoint("/path/to/checkpoint/")
input_spec = [
paddle.static.InputSpec(shape=[None, None],
dtype="int64"), # input_ids
paddle.static.InputSpec(shape=[None, None],
dtype="int64") # segment_ids
]
if model_args.export_model_dir is None:
model_args.export_model_dir = os.path.join(
training_args.output_dir, "export")
paddlenlp.transformers.export_model(model=trainer.model,
input_spec=input_spec,
path=model_args.export_model_dir)
def main():
parser = PdArgumentParser(
(ModelArguments, DataArguments, PreTrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
set_seed(training_args)
paddle.set_device(training_args.device)
if paddle.distributed.get_world_size() > 1:
paddle.distributed.init_parallel_env()
training_args.eval_iters = 10
training_args.test_iters = training_args.eval_iters * 10
# Log model and data config
training_args.print_config(model_args, "Model")
training_args.print_config(data_args, "Data")
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, world_size: {training_args.world_size}, "
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
# if last_checkpoint is None and len(
# os.listdir(training_args.output_dir)) > 1:
# raise ValueError(
# f"Output directory ({training_args.output_dir}) already exists and is not empty. "
# "Use --overwrite_output_dir to overcome.")
if last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
base_class, model_class, criterion_class, tokenizer_class = MODEL_CLASSES[
model_args.model_type]
pretrained_models_list = list(
model_class.pretrained_init_configuration.keys())
if model_args.model_name_or_path in pretrained_models_list:
model_config = model_class.pretrained_init_configuration[
model_args.model_name_or_path]
model_config["hidden_dropout_prob"] = model_args.hidden_dropout_prob
model_config[
"attention_probs_dropout_prob"] = model_args.attention_probs_dropout_prob
model = model_class(base_class(**model_config))
else:
model = model_class.from_pretrained(
model_args.model_name_or_path,
hidden_dropout_prob=model_args.hidden_dropout_prob,
attention_probs_dropout_prob=model_args.
attention_probs_dropout_prob)
class CriterionWrapper(paddle.nn.Layer):
"""
"""
def __init__(self):
"""CriterionWrapper
"""
super(CriterionWrapper, self).__init__()
self.criterion = criterion_class()
def forward(self, output, labels):
"""forward function
Args:
output (tuple): prediction_scores, seq_relationship_score
labels (tuple): masked_lm_labels, next_sentence_labels
Returns:
Tensor: final loss.
"""
prediction_scores, seq_relationship_score = output
masked_lm_labels, next_sentence_labels = labels
lm_loss, sop_loss = self.criterion(prediction_scores,
seq_relationship_score,
masked_lm_labels,
next_sentence_labels)
loss = lm_loss + sop_loss
return loss
# Create the learning_rate sheduler and optimizer
if training_args.decay_steps is None:
training_args.decay_steps = training_args.max_steps
warmup_steps = training_args.warmup_ratio * training_args.max_steps
lr_scheduler = LinearAnnealingWithWarmupDecay(
training_args.learning_rate,
training_args.min_learning_rate,
warmup_step=warmup_steps,
decay_step=training_args.decay_steps)
data_file = get_train_data_file(data_args)
tokenizer = tokenizer_class.from_pretrained(model_args.model_name_or_path)
train_dataset, eval_dataset, test_dataset, data_collator = create_pretrained_dataset(
data_args, training_args, data_file, tokenizer)
trainer = PretrainingTrainer(
model=model,
criterion=CriterionWrapper(),
args=training_args,
data_collator=data_collator,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
optimizers=(None, lr_scheduler),
tokenizer=tokenizer,
)
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
elif last_checkpoint is not None:
checkpoint = last_checkpoint
# Training
if training_args.do_train:
train_result = trainer.train(resume_from_checkpoint=checkpoint)
metrics = train_result.metrics
trainer.save_model()
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
if training_args.do_predict:
test_ret = trainer.predict(test_dataset)
trainer.log_metrics("test", test_ret.metrics)
def main():
parser = PdArgumentParser(
(ModelArguments, DataArguments, TrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
training_args.eval_iters = 10
training_args.test_iters = training_args.eval_iters * 10
# Log model and data config
training_args.print_config(model_args, "Model")
training_args.print_config(data_args, "Data")
paddle.set_device(training_args.device)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, world_size: {training_args.world_size}, "
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 1:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
model_class, tokenizer_class = MODEL_CLASSES['ernie-health']
# Loads or initialize a model.
pretrained_models = list(
tokenizer_class.pretrained_init_configuration.keys())
if model_args.model_name_or_path in pretrained_models:
tokenizer = tokenizer_class.from_pretrained(
model_args.model_name_or_path)
generator = ElectraGenerator(
ElectraModel(**model_class.pretrained_init_configuration[
model_args.model_name_or_path + "-generator"]))
discriminator = ErnieHealthDiscriminator(
ElectraModel(**model_class.pretrained_init_configuration[
model_args.model_name_or_path + "-discriminator"]))
model = model_class(generator, discriminator)
else:
raise ValueError("Only support %s" % (", ".join(pretrained_models)))
# Loads dataset.
tic_load_data = time.time()
logger.info("start load data : %s" %
(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())))
train_dataset = MedicalCorpus(data_path=data_args.input_dir,
tokenizer=tokenizer)
logger.info("load data done, total : %s s" % (time.time() - tic_load_data))
# Reads data and generates mini-batches.
data_collator = DataCollatorForErnieHealth(
tokenizer=tokenizer,
max_seq_length=data_args.max_seq_length,
mlm_prob=data_args.masked_lm_prob,
return_dict=True)
class CriterionWrapper(paddle.nn.Layer):
"""
"""
def __init__(self):
"""CriterionWrapper
"""
super(CriterionWrapper, self).__init__()
self.criterion = ErnieHealthPretrainingCriterion(
getattr(
model.generator,
ElectraGenerator.base_model_prefix).config["vocab_size"],
model.gen_weight)
def forward(self, output, labels):
"""forward function
Args:
output (tuple): generator_logits, logits_rtd, logits_mts, logits_csp, disc_labels, mask
labels (tuple): generator_labels
Returns:
Tensor: final loss.
"""
generator_logits, logits_rtd, logits_mts, logits_csp, disc_labels, masks = output
generator_labels = labels
loss, gen_loss, rtd_loss, mts_loss, csp_loss = self.criterion(
generator_logits, generator_labels, logits_rtd, logits_mts,
logits_csp, disc_labels, masks)
return loss
trainer = Trainer(
model=model,
criterion=CriterionWrapper(),
args=training_args,
data_collator=data_collator,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=None,
tokenizer=tokenizer,
)
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
elif last_checkpoint is not None:
checkpoint = last_checkpoint
# Training
if training_args.do_train:
train_result = trainer.train(resume_from_checkpoint=checkpoint)
metrics = train_result.metrics
trainer.save_model()
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
def main():
parser = PdArgumentParser(
(ModelArguments, DataArguments, TrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# Log model and data config
training_args.print_config(model_args, "Model")
training_args.print_config(data_args, "Data")
paddle.set_device(training_args.device)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, world_size: {training_args.world_size}, "
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# set_seed(args)
data_args.dataset = data_args.dataset.strip()
if data_args.dataset in ALL_DATASETS:
# if you custom you hyper-parameters in yaml config, it will overwrite all args.
config = ALL_DATASETS[data_args.dataset]
for args in (model_args, data_args, training_args):
for arg in vars(args):
if arg in config.keys():
setattr(args, arg, config[arg])
training_args.per_device_train_batch_size = config["batch_size"]
training_args.per_device_eval_batch_size = config["batch_size"]
dataset_config = data_args.dataset.split(" ")
raw_datasets = load_dataset(
dataset_config[0],
None if len(dataset_config) <= 1 else dataset_config[1],
cache_dir=model_args.cache_dir)
label_list = getattr(raw_datasets['train'], "label_list", None)
data_args.label_list = label_list
# Define tokenizer, model, loss function.
tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path)
model = AutoModelForQuestionAnswering.from_pretrained(
model_args.model_name_or_path)
loss_fct = CrossEntropyLossForSQuAD()
# Preprocessing the datasets.
# Preprocessing is slighlty different for training and evaluation.
if training_args.do_train:
column_names = raw_datasets["train"].column_names
elif training_args.do_eval:
column_names = raw_datasets["validation"].column_names
else:
column_names = raw_datasets["test"].column_names
if training_args.do_train:
train_dataset = raw_datasets["train"]
# Create train feature from dataset
with training_args.main_process_first(
desc="train dataset map pre-processing"):
# Dataset pre-process
train_dataset = train_dataset.map(
partial(prepare_train_features,
tokenizer=tokenizer,
args=data_args),
batched=True,
num_proc=4,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on train dataset",
)
if training_args.do_eval:
eval_examples = raw_datasets["validation"]
with training_args.main_process_first(
desc="evaluate dataset map pre-processing"):
eval_dataset = eval_examples.map(
partial(prepare_validation_features,
tokenizer=tokenizer,
args=data_args),
batched=True,
num_proc=4,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on validation dataset",
)
if training_args.do_predict:
predict_examples = raw_datasets["test"]
with training_args.main_process_first(
desc="test dataset map pre-processing"):
predict_dataset = predict_examples.map(
partial(prepare_validation_features,
tokenizer=tokenizer,
args=data_args),
batched=True,
num_proc=4,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on prediction dataset",
)
# Define data collector
data_collator = DataCollatorWithPadding(tokenizer)
# Post-processing:
def post_processing_function(examples,
features,
predictions,
stage="eval"):
# Post-processing: we match the start logits and end logits to answers in the original context.
predictions, all_nbest_json, scores_diff_json = compute_prediction(
examples=examples,
features=features,
predictions=predictions,
n_best_size=data_args.n_best_size,
max_answer_length=data_args.max_answer_length,
null_score_diff_threshold=data_args.null_score_diff_threshold,
)
# # Format the result to the format the metric expects.
# formatted_predictions = [{
# "id": k,
# "prediction_text": v
# } for k, v in predictions.items()]
references = [{
"id": ex["id"],
"answers": ex["answers"]
} for ex in examples]
return EvalPrediction(predictions=predictions, label_ids=references)
def compute_metrics(p: EvalPrediction):
ret = squad_evaluate(examples=p.label_ids,
preds=p.predictions,
is_whitespace_splited=False)
return dict(ret)
# return metric.compute(predictions=p.predictions, references=p.label_ids)
trainer = QuestionAnsweringTrainer(
model=model,
criterion=loss_fct,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
eval_examples=eval_examples if training_args.do_eval else None,
data_collator=data_collator,
post_process_function=post_processing_function,
tokenizer=tokenizer,
compute_metrics=compute_metrics,
)
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
elif last_checkpoint is not None:
checkpoint = last_checkpoint
if training_args.do_train:
# Training
train_result = trainer.train(resume_from_checkpoint=checkpoint)
metrics = train_result.metrics
trainer.save_model()
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# model.set_state_dict(paddle.load("tmp/model_state.pdparams"))
# Evaluate and tests model
if training_args.do_eval:
eval_metrics = trainer.evaluate()
trainer.log_metrics("eval", eval_metrics)
if training_args.do_predict:
test_ret = trainer.predict(predict_dataset, predict_examples)
trainer.log_metrics("predict", test_ret.metrics)
if test_ret.label_ids is None:
paddle.save(
test_ret.predictions,
os.path.join(training_args.output_dir,
"test_results.pdtensor"),
)
# export inference model
if training_args.do_export:
# You can also load from certain checkpoint
# trainer.load_state_dict_from_checkpoint("/path/to/checkpoint/")
input_spec = [
paddle.static.InputSpec(shape=[None, None],
dtype="int64"), # input_ids
paddle.static.InputSpec(shape=[None, None],
dtype="int64") # segment_ids
]
if model_args.export_model_dir is None:
model_args.export_model_dir = os.path.join(
training_args.output_dir, "export")
paddlenlp.transformers.export_model(model=trainer.model,
input_spec=input_spec,
path=model_args.export_model_dir)