def test_data_collator_with_padding(self):
tokenizer = BertTokenizer(self.vocab_file)
features = [{
"input_ids": [0, 1, 2]
}, {
"input_ids": [0, 1, 2, 3, 4, 5]
}]
data_collator = DataCollatorWithPadding(tokenizer, return_tensors="np")
batch = data_collator(features)
self.assertEqual(batch["input_ids"].shape, (2, 6))
self.assertEqual(batch["input_ids"][0].tolist(),
[0, 1, 2] + [tokenizer.pad_token_id] * 3)
data_collator = DataCollatorWithPadding(tokenizer,
padding="max_length",
max_length=10,
return_tensors="np")
batch = data_collator(features)
self.assertEqual(batch["input_ids"].shape, (2, 10))
data_collator = DataCollatorWithPadding(tokenizer,
pad_to_multiple_of=8,
return_tensors="np")
batch = data_collator(features)
self.assertEqual(batch["input_ids"].shape, (2, 8))
def main(train_args: TrainingArguments, args: Args):
log.info(f"Parsed args: {args}")
log.info(f"Parsed training args: {train_args}")
# https://huggingface.co/docs/datasets/loading#json
dataset = load_dataset("json",
data_files=join(args.input_dir, args.input_file))
coalesced_dataset = dataset.map(coalesce)
tokenizer = AutoTokenizer.from_pretrained(args.model)
def preprocess_function(examples):
return tokenizer(examples["text"], truncation=True)
tokenized_dataset = coalesced_dataset.map(preprocess_function,
batched=True)
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
model = AutoModelForSequenceClassification.from_pretrained(args.model,
num_labels=2)
# TODO: separate train and eval inputs.
trainer = Trainer(
model=model,
args=train_args,
train_dataset=tokenized_dataset["train"],
eval_dataset=tokenized_dataset["train"],
tokenizer=tokenizer,
data_collator=data_collator,
)
trainer.train()
def init_dataset_for_squad(model_args,
data_args,
training_args,
last_checkpoint=None):
datasets = init_datasets_squad(data_args, model_args)
# Place holder for now
extra_config_kwargs = {}
config = init_config(model_args, extra_config_kwargs=extra_config_kwargs)
tokenizer = init_tokenizer(model_args)
model = init_model(model_args,
config,
tokenizer,
finetuning=True,
squad=True)
check_sparsity_callback(model, model_args)
logging.info(f"Tokenizing datasets for squad ...")
(train_dataset,
eval_dataset,
eval_examples,
answer_column_name) = \
preprocess_datasets_squad(datasets, tokenizer, training_args, data_args)
if data_args.pad_to_max_length:
data_collator = default_data_collator
else:
pad_to_multiple_of = 8 if training_args.fp16 else None
data_collator = \
DataCollatorWithPadding(tokenizer,
pad_to_multiple_of=pad_to_multiple_of)
return (tokenizer, data_collator, train_dataset, eval_dataset,
eval_examples, model, answer_column_name)
def collate(self, samples: Any) -> Tensor:
"""Override to convert a set of samples to a batch."""
if self.padding != "max_length":
data_collator = DataCollatorWithPadding(
AutoTokenizer.from_pretrained(self.backbone, use_fast=True))
return data_collator(samples)
return default_data_collator(samples)
def katas(self):
if not (os.path.exists(self.folder)):
logger.info(f"Download GermEval18 dataset")
_download_extract_downstream_data(self.folder + "/train.tsv",
proxies=None)
def preprocess_function(examples):
labels = GermEval18.labels(True)
examples["coarse_label"] = [
labels.index(x) for x in examples["coarse_label"]
]
labels = GermEval18.labels(False)
examples["fine_label"] = [
labels.index(x) for x in examples["fine_label"]
]
assert callable(self.tokenizer), "tokenizer is not callable"
return self.tokenizer(
examples["text"],
truncation=True,
return_token_type_ids=True, # FIXME
max_length=self.spec.max_seq_length)
spec = KataSpec(remove_columns=["text"],
max_seq_length=self.spec.max_seq_length)
collator = DataCollatorWithPadding(self.tokenizer,
padding="max_length",
max_length=self.spec.max_seq_length)
return (TsvKata(spec, preprocess_function, collator,
"../data/germeval18/train.tsv"),
TsvKata(spec, preprocess_function, collator,
"../data/germeval18/test.tsv"))
def get_dataloader_and_optimizer(args, tokenizer, model, train_dataset, eval_dataset):
data_collator = DataCollatorWithPadding(tokenizer)
train_dataloader = DataLoader(train_dataset, shuffle=True, collate_fn=data_collator,
batch_size=args.batch_size)
eval_dataloader = DataLoader(eval_dataset, collate_fn=data_collator,
batch_size=args.batch_size)
optimizer = AdamW(model.parameters(), lr=args.learning_rate)
return optimizer, train_dataloader, eval_dataloader, data_collator
def init_dataset_for_finetuning(model_args,
data_args,
training_args,
last_checkpoint=None):
datasets = init_datasets_task(data_args, training_args)
is_regression, label_list, num_labels = get_labels(datasets, data_args)
logging.info(f"Training {data_args.task_name} with {num_labels} labels")
# For finetuning required to add labels and task name to config kwargs
extra_config_kwargs = dict(
num_labels=num_labels,
finetuning_task=data_args.task_name,
)
config = init_config(model_args, extra_config_kwargs=extra_config_kwargs)
tokenizer = init_tokenizer(model_args)
model = init_model(model_args, config, tokenizer, finetuning=True)
check_sparsity_callback(model, model_args)
# Tokenizing and preprocessing the datasets for downstream tasks
# TODO: load from cached tokenized datasets for finetuning as well
logging.info(f"Tokenizing datasets for finetuning ...")
tokenized_datasets = preprocess_datasets_task(datasets, tokenizer,
data_args, model, num_labels,
label_list, is_regression)
# Separate into train, eval and test
train_dataset = tokenized_datasets["train"]
eval_dataset = tokenized_datasets["validation_matched" if data_args.
task_name == "mnli" else "validation"]
test_dataset = None
if (data_args.task_name is not None or data_args.test_file is not None):
if training_args.do_predict:
test_dataset = tokenized_datasets["test_matched" if data_args.
task_name == "mnli" else "test"]
# Log fingerprint used in HF smart caching
logging.info(f"Dataset fingerprint: {train_dataset._fingerprint}")
# Data collator will default to DataCollatorWithPadding,
# so we change it if we already did the padding.
if data_args.pad_to_max_length:
data_collator = default_data_collator
elif training_args.fp16:
data_collator = DataCollatorWithPadding(tokenizer,
pad_to_multiple_of=8)
else:
data_collator = None
return (tokenizer, data_collator, train_dataset, eval_dataset,
test_dataset, model, is_regression, tokenized_datasets, label_list,
config)
def __init__(self, args, model, tokenizer, eval_answers):
self.args = args
self.metric = load_metric("squad")
self.model, self.tokenizer = model, tokenizer
self.data_collator = DataCollatorWithPadding(
self.tokenizer,
pad_to_multiple_of=8 if self.args.train.fp16 else None)
self.train_dataset = None # 필요한거: train pp된거, eval 원본, eval retrieve 된거, eval retrieve 되고 pp된거
self.eval_dataset = None
self.eval_examples = None
self.eval_answers = eval_answers
def explain(model, test_str):
data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=None)
# test_str = "The new design is awful!"
ex_tokens = tokenizer([test_str])
str_list = tokenizer.convert_ids_to_tokens(ex_tokens['input_ids'][0])
# print(str_list)
ex_tokens = data_collator(ex_tokens)
result = model(output_hidden_states=True, **ex_tokens)
logits = result.logits
pred = torch.argmax(logits)
layer_ids = [-2, -3, -4, -5, -6, -7]
expln_ = []
for layer_id in layer_ids:
# layer_id = -2
if layer_id == -2:
result.hidden_states[layer_id].retain_grad()
model.zero_grad()
logits[0][pred.item()].backward(retain_graph=True)
hs_grad = result.hidden_states[layer_id].grad
expln = (hs_grad * result.hidden_states[layer_id]).sum(dim=-1)
cls_grad = hs_grad[:, 0, :]
cls_hs = result.hidden_states[layer_id][:, 0, :]
sep_grad = hs_grad[:, -1, :]
sep_hs = result.hidden_states[layer_id][:, -1, :]
expln_.append(expln)
else:
result.hidden_states[layer_id].retain_grad()
model.zero_grad()
cls_hs.backward(cls_grad, retain_graph=True)
hs_grad = result.hidden_states[layer_id].grad
expln = (hs_grad * result.hidden_states[layer_id]).sum(dim=-1)
cls_grad = hs_grad[:, 0, :]
cls_hs = result.hidden_states[layer_id][:, 0, :]
model.zero_grad()
sep_hs.backward(sep_grad, retain_graph=True)
hs_grad = result.hidden_states[layer_id].grad
expln += (hs_grad * result.hidden_states[layer_id]).sum(dim=-1)
sep_grad = hs_grad[:, -1, :]
sep_hs = result.hidden_states[layer_id][:, -1, :]
expln_.append(expln)
expln = sum(expln_)
return pred, expln, str_list
def __call__(self):
dataset = load_dataset(path="glue", name="cola", cache_dir=self.spec.cache_dir)
encoded = dataset.map(function=self.preprocess_function,
batched=self.spec.batch_preprocess,
batch_size=self.spec.batch_preprocess_size,
remove_columns=["sentence"])
data_collator = DataCollatorWithPadding(self.tokenizer, "max_length", self.spec.max_seq_length)
return (
DatasetKata(self.spec, encoded["train"], data_collator),
DatasetKata(self.spec, encoded["validation"], data_collator),
DatasetKata(self.spec, encoded["test"], data_collator)
)
def run_glue(task_key,
cfg,
model,
model_args,
training_args,
tokenizer,
mode="train",
extract=False,
**kwargs):
r"""
cfg: YACS cfg node
ckpt_path: Unsupported
"""
task_name = TASK_KEY_TO_NAME[task_key]
data_args = DataTrainingArguments(task_name=task_name,
data_dir=cfg.DATA.DATAPATH)
glue_dataset = load_features_dict(tokenizer, cfg)
# print(glue_dataset.keys())
train_dataset = glue_dataset[task_key]['train']
split_key = cfg.EVAL.SPLIT
if task_key == "mnli":
split_key = f"{split_key}_matched"
eval_dataset = glue_dataset[task_key][split_key]
# eval_dataset = glue_dataset[task_key]['validation_mismached']
# train_dataset = GlueDataset(data_args, tokenizer=tokenizer, limit_length=cfg.TRAIN.TASK_LIMIT)
# eval_dataset = GlueDataset(data_args, tokenizer=tokenizer, mode='dev')
collator = DataCollatorWithPadding(tokenizer)
trainer = FixedTrainer(model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
compute_metrics=get_eval_metrics_func(task_key),
data_collator=collator,
config=cfg)
if mode == "train":
trainer.train()
if mode != "train" or cfg.EVAL_ON_COMPLETION:
extract_path = None
if extract:
extract_path = get_extract_path(cfg, model_args)
metrics = trainer.evaluate(extract_path=extract_path,
cache_path=osp.join(
cfg.TASK.EXTRACT_TOKENS_MASK_CACHE,
task_key))
metrics_file = get_metrics_path(cfg, model_args)
torch.save(metrics, metrics_file)
def of(self, *args):
def preprocess_function(examples):
assert callable(self.tokenizer), "tokenizer is not callable"
return self.tokenizer(examples["text"],
truncation=True,
max_length=self.max_length)
values = args[0]
spec = KataSpec(remove_columns=["text"],
max_seq_length=self.max_length)
collator = DataCollatorWithPadding(self.tokenizer,
padding="max_length",
max_length=self.max_length)
return DictKata(spec, preprocess_function, collator, {"text": values})
def loaders(self):
if self._loaders is None:
ps, mgr = self.params, self.mgr
if ps.pad_to_max_length:
c = default_data_collator
else:
c = DataCollatorWithPadding(
self.tokenizer, pad_to_multiple_of=(8 if mgr.use_fp16 else None)
)
t = DataLoader(
self.train_ds, shuffle=True, collate_fn=c, batch_size=ps.train_batch_size
)
e = DataLoader(self.eval_ds, collate_fn=c, batch_size=ps.eval_batch_size)
self._loaders = {TRAIN: t, EVAL: e}
return self._loaders
def __init__(
self,
# The abbreviation/name of your Hugging Face transformer architecture (e.b., bert, bart, etc..)
hf_arch: str,
# A specific configuration instance you want to use
hf_config: PretrainedConfig,
# A Hugging Face tokenizer
hf_tokenizer: PreTrainedTokenizerBase,
# A Hugging Face model
hf_model: PreTrainedModel,
# Defaults to use Hugging Face's DataCollatorWithPadding(tokenizer=hf_tokenizer)
data_collator: Type = None,
):
store_attr()
self.data_collator = data_collator if (
data_collator) else DataCollatorWithPadding(tokenizer=hf_tokenizer)
def loaders(self):
if self._loaders is None:
ps = self.params
if ps.pad_to_max_length:
c = default_data_collator
elif ps.fp16:
c = DataCollatorWithPadding(self.tokenizer, pad_to_multiple_of=8)
else:
c = None
t = DataLoader(
self.train_ds, shuffle=True, collate_fn=c, batch_size=ps.train_batch_size
)
e = DataLoader(self.eval_ds, collate_fn=c, batch_size=ps.eval_batch_size)
self._loaders = {TRAIN: t, EVAL: e}
if ps.do_test:
p = DataLoader(self.test_ds, collate_fn=c, batch_size=ps.eval_batch_size)
self._loaders[TEST] = p
return self._loaders
def get_data(
model_args,
training_args,
tokenizer,
text_data_path="../data/test_dataset"): # 경로 변경 ../data/test_dataset
"""
get data
Args:
model_args: model arguments
training_args: training arguments
tokenizer: tokenizer
text_data_path: Defaults to "../data/test_dataset"
Returns:
text_data, val_iter, val_dataset, scores
"""
text_data = load_from_disk(text_data_path)
# run_ lasticsearch
if "elastic" in model_args.retrieval_type:
is_sentence_trainformer = False
if "sentence_trainformer" in model_args.retrieval_type:
is_sentence_trainformer = True
# number of text to concat
concat_num = model_args.retrieval_elastic_num
text_data, scores = run_elasticsearch(text_data, concat_num,
model_args,
is_sentence_trainformer)
elif model_args.retrieval_type == "dense":
concat_num = model_args.retrieval_elastic_num
text_data, scores = run_concat_dense_retrival(text_data, concat_num)
column_names = text_data["validation"].column_names
data_collator = (DataCollatorWithPadding(
tokenizer, pad_to_multiple_of=8 if training_args.fp16 else None))
# 데이터 tokenize(mrc 모델안에 들어 갈 수 있도록)
data_processor = DataProcessor(tokenizer)
val_text = text_data["validation"]
val_dataset = data_processor.val_tokenzier(val_text, column_names)
val_iter = DataLoader(val_dataset, collate_fn=data_collator, batch_size=1)
return text_data, val_iter, val_dataset, scores
def test_strategies_with_dataset(self, args=args):
"""
(Constraint)
- num_train_epoch 1
- random seed 1
- dataset fragment (rows : 100)
(Caution)
ERROR가 표시된다면, 상위 단위 테스트 결과를 확인하세요.
"""
for seed, strategy in [(SEED, strategy) for strategy in strategies]:
wandb.init(project="p-stage-3-test", reinit=True)
args = update_args(args, strategy)
args.strategy, args.seed = strategy, seed
set_seed(seed)
datasets = prepare_dataset(args, is_train=True)
model, tokenizer = get_reader_model(args)
train_dataset, post_processing_function = preprocess_dataset(
args, datasets, tokenizer, is_train=True)
train_dataset = train_dataset.select(range(100)) # select 100
data_collator = DataCollatorWithPadding(
tokenizer, pad_to_multiple_of=8 if args.train.fp16 else None)
args.train.do_train = True
args.train.run_name = "_".join(
[strategy, args.alias, str(seed), "test"])
wandb.run.name = args.train.run_name
# TRAIN MRC
args.train.num_train_epochs = 1.0 # fix epoch 1
trainer = QuestionAnsweringTrainer(
model=model,
args=args.train, # training_args
custom_args=args,
train_dataset=train_dataset,
tokenizer=tokenizer,
data_collator=data_collator,
post_process_function=post_processing_function,
compute_metrics=compute_metrics,
)
trainer.train()
def train(self, input_filepath, dataclass_args: TCTrainArgs):
if not dataclass_args.load_preprocessed_data:
self.logger.info("Preprocessing dataset...")
contexts, labels = self._get_data(input_filepath)
train_encodings = self.tokenizer(contexts, truncation=True, padding=True)
else:
self.logger.info("Loading dataset from %s...", dataclass_args.load_preprocessed_data_path)
train_encodings, labels = self._get_preprocessed_data(dataclass_args.load_preprocessed_data_path)
if dataclass_args.save_preprocessed_data:
self.logger.info("Saving training dataset to %s...", dataclass_args.save_preprocessed_data_path)
input_ids = train_encodings["input_ids"]
attention_mask = train_encodings["attention_mask"]
self._generate_json(dataclass_args.save_preprocessed_data_path, input_ids, attention_mask, labels, "train")
train_dataset = TextClassificationDataset(train_encodings, labels)
data_collator = DataCollatorWithPadding(self.tokenizer)
self._run_train(train_dataset, dataclass_args, data_collator)
def __call__(self):
if not (os.path.exists(self.folder)):
logger.info(f"Download ToxicComments dataset")
_download_extract_downstream_data(self.folder + "/train.tsv",
proxies=None)
def preprocess_function(examples):
# TODO: use to utility class
def hot_encoding(labels: Optional[str]):
label_ids = [0] * len(ToxicComments.labels())
if labels is None:
return label_ids
for l in labels.split(","):
if l != "":
label_ids[ToxicComments.labels().index(l)] = 1
return label_ids
examples["label"] = [hot_encoding(x) for x in examples["label"]]
assert callable(self.tokenizer), "tokenizer is not callable"
return self.tokenizer(
examples["text"],
truncation=True,
return_token_type_ids=True, # FIXME
max_length=self.spec.max_seq_length)
spec = KataSpec(remove_columns=["text"],
max_seq_length=self.spec.max_seq_length)
collator = DataCollatorWithPadding(self.tokenizer,
padding="max_length",
max_length=self.spec.max_seq_length)
return (TsvKata(spec,
preprocess_function,
collator,
"../data/toxic-comments/train.tsv",
quote_char='"'),
TsvKata(spec,
preprocess_function,
collator,
"../data/toxic-comments/val.tsv",
quote_char='"'))
def loaders(self):
if self._loaders is None:
ps, mgr = self.params, self.mgr
if ps.pad_to_max_length:
c = default_data_collator
else:
c = DataCollatorWithPadding(
self.tokenizer, pad_to_multiple_of=(8 if mgr.use_fp16 else None)
)
t = DataLoader(
self.train_ds, shuffle=True, collate_fn=c, batch_size=ps.train_batch_size
)
x = self.eval_ds.remove_columns(["example_id", "offset_mapping"])
e = DataLoader(x, collate_fn=c, batch_size=ps.eval_batch_size)
self._loaders = {TRAIN: t, EVAL: e}
if ps.do_test:
x = self.test_ds.remove_columns(["example_id", "offset_mapping"])
p = DataLoader(x, collate_fn=c, batch_size=ps.eval_batch_size)
self._loaders[TEST] = p
return self._loaders
def eval(self, input_filepath, dataclass_args: TCEvalArgs):
if not dataclass_args.load_preprocessed_data:
self.logger.info("Preprocessing dataset...")
contexts, labels = self._get_data(input_filepath)
eval_encodings = self.tokenizer(contexts, truncation=True, padding=True)
else:
self.logger.info("Loading dataset from %s...", dataclass_args.load_preprocessed_data_path)
eval_encodings, labels = self._get_preprocessed_data(dataclass_args.load_preprocessed_data_path)
if dataclass_args.save_preprocessed_data:
self.logger.info("Saving training dataset to %s...", dataclass_args.save_preprocessed_data_path)
input_ids = eval_encodings["input_ids"]
attention_mask = eval_encodings["attention_mask"]
self._generate_json(dataclass_args.save_preprocessed_data_path, input_ids, attention_mask, labels, "train")
eval_dataset = TextClassificationDataset(eval_encodings, labels)
data_collator = DataCollatorWithPadding(self.tokenizer)
result = self._run_eval(eval_dataset, data_collator, dataclass_args)
return EvalResult(loss=result["eval_loss"])
def train(self, input_filepath, dataclass_args: QATrainArgs):
"""
See docstring in HappyQuestionAnswering.train()
"""
if dataclass_args.save_preprocessed_data:
self.logger.info("Saving preprocessed data is currently "
"not available for question answering models. "
"It will be added soon. ")
if dataclass_args.load_preprocessed_data:
self.logger.info("Loading preprocessed data is currently "
"not available for question answering models. "
"It will be added soon. ")
self.logger.info("Preprocessing dataset...")
contexts, questions, answers = self._get_data(input_filepath)
self.__add_end_idx(contexts, answers)
encodings = self.tokenizer(contexts, questions, truncation=True, padding=True)
self.__add_token_positions(encodings, answers)
dataset = QuestionAnsweringDataset(encodings)
data_collator = DataCollatorWithPadding(self.tokenizer)
self._run_train(dataset, dataclass_args, data_collator)
def create_trainer(self):
# Data collator
# We have already padded to max length if the corresponding flag is True, otherwise we need to pad in the data
# collator.
training_args = self.training_args
data_args = self.data_args
data_collator = (default_data_collator if data_args.pad_to_max_length
else DataCollatorWithPadding(self.tokenizer))
# TODO: Once the fix lands in a Datasets release, remove the _local here and the squad_v2_local folder.
current_dir = os.path.sep.join(
os.path.join(__file__).split(os.path.sep)[:-1])
metric = load_metric(
os.path.join(current_dir, "squad_v2_local") if data_args.
version_2_with_negative else "squad")
def compute_metrics(p: EvalPrediction):
return metric.compute(predictions=p.predictions,
references=p.label_ids)
all_args = self.get_all_args(exclude_base=True)
# Initialize our Trainer
self.trainer = self.QA_TRAINER_CLASS(
model=None,
args=training_args,
train_dataset=self.train_dataset
if training_args.do_train else None,
eval_dataset=self.validation_dataset
if training_args.do_eval else None,
eval_examples=self.datasets["validation"]
if training_args.do_eval else None,
tokenizer=self.tokenizer,
data_collator=data_collator,
post_process_function=self._post_processing_function,
compute_metrics=compute_metrics,
model_init=self.model_init,
**all_args,
)
def load_data():
"""Load IMDB data (training and eval)"""
raw_datasets = load_dataset("imdb")
raw_datasets = raw_datasets.shuffle(seed=42)
# remove unnecessary data split
del raw_datasets["unsupervised"]
tokenizer = AutoTokenizer.from_pretrained(CHECKPOINT)
def tokenize_function(examples):
return tokenizer(examples["text"], truncation=True)
# random 100 samples
population = random.sample(range(len(raw_datasets["train"])), 100)
tokenized_datasets = raw_datasets.map(tokenize_function, batched=True)
tokenized_datasets["train"] = tokenized_datasets["train"].select(
population)
tokenized_datasets["test"] = tokenized_datasets["test"].select(population)
tokenized_datasets = tokenized_datasets.remove_columns("text")
tokenized_datasets = tokenized_datasets.rename_column("label", "labels")
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
trainloader = DataLoader(
tokenized_datasets["train"],
shuffle=True,
batch_size=32,
collate_fn=data_collator,
)
testloader = DataLoader(tokenized_datasets["test"],
batch_size=32,
collate_fn=data_collator)
return trainloader, testloader
def create_trainer(self):
# Data collator
# We have already padded to max length if the corresponding flag is True, otherwise we need to pad in the data
# collator.
training_args = self.training_args
data_args = self.data_args
data_collator = (default_data_collator if data_args.pad_to_max_length
else DataCollatorWithPadding(self.tokenizer))
metric = load_metric(
"squad_v2" if data_args.version_2_with_negative else "squad")
def compute_metrics(p: EvalPrediction):
return metric.compute(predictions=p.predictions,
references=p.label_ids)
all_args = self.get_all_args(exclude_base=True)
# Initialize our Trainer
self.trainer = self.QA_TRAINER_CLASS(
model=None,
args=training_args,
train_dataset=self.train_dataset
if training_args.do_train else None,
eval_dataset=self.validation_dataset
if training_args.do_eval else None,
eval_examples=self.datasets["validation"]
if training_args.do_eval else None,
tokenizer=self.tokenizer,
data_collator=data_collator,
post_process_function=self._post_processing_function,
compute_metrics=compute_metrics,
model_init=self.model_init,
**all_args,
)
def eval(self, input_filepath, dataclass_args: QAEvalArgs):
"""
See docstring in HappyQuestionAnswering.eval()
"""
if dataclass_args.save_preprocessed_data:
self.logger.info("Saving preprocessed data is currently "
"not available for question answering models. "
"It will be added soon. ")
if dataclass_args.load_preprocessed_data:
self.logger.info("Loading preprocessed data is currently "
"not available for question answering models. "
"It will be added soon. ")
contexts, questions, answers = self._get_data(input_filepath)
self.__add_end_idx(contexts, answers)
encodings = self.tokenizer(contexts, questions, truncation=True, padding=True)
self.__add_token_positions(encodings, answers)
eval_dataset = QuestionAnsweringDataset(encodings)
data_collator = DataCollatorWithPadding(self.tokenizer)
result = self._run_eval(eval_dataset, data_collator, dataclass_args)
return EvalResult(loss=result["eval_loss"])
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, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
# 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:
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."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info(f"Training/evaluation parameters {training_args}")
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
# or specify a GLUE benchmark task (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use as labels the column called 'label' and as pair of sentences the
# sentences in columns called 'sentence1' and 'sentence2' if such column exists or the first two columns not named
# label if at least two columns are provided.
#
# If the CSVs/JSONs contain only one non-label column, the script does single sentence classification on this
# single column. You can easily tweak this behavior (see below)
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.task_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset("glue", data_args.task_name)
else:
# Loading a dataset from your local files.
# CSV/JSON training and evaluation files are needed.
data_files = {
"train": data_args.train_file,
"validation": data_args.validation_file
}
# Get the test dataset: you can provide your own CSV/JSON test file (see below)
# when you use `do_predict` without specifying a GLUE benchmark task.
if training_args.do_predict:
if data_args.test_file is not None:
train_extension = data_args.train_file.split(".")[-1]
test_extension = data_args.test_file.split(".")[-1]
assert (
test_extension == train_extension
), "`test_file` should have the same extension (csv or json) as `train_file`."
data_files["test"] = data_args.test_file
else:
raise ValueError(
"Need either a GLUE task or a test file for `do_predict`.")
for key in data_files.keys():
logger.info(f"load a local file for {key}: {data_files[key]}")
if data_args.train_file.endswith(".csv"):
# Loading a dataset from local csv files
datasets = load_dataset("csv", data_files=data_files)
else:
# Loading a dataset from local json files
datasets = load_dataset("json", data_files=data_files)
# See more about loading any type of standard or custom dataset at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Labels
if data_args.task_name is not None:
is_regression = data_args.task_name == "stsb"
if not is_regression:
label_list = datasets["train"].features["label"].names
num_labels = len(label_list)
else:
num_labels = 1
else:
# Trying to have good defaults here, don't hesitate to tweak to your needs.
is_regression = datasets["train"].features["label"].dtype in [
"float32", "float64"
]
if is_regression:
num_labels = 1
else:
# A useful fast method:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.unique
label_list = datasets["train"].unique("label")
label_list.sort() # Let's sort it for determinism
num_labels = len(label_list)
# Load pretrained model and tokenizer
#
# In 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,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
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_tokenizer,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForSequenceClassification.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# Preprocessing the datasets
if data_args.task_name is not None:
sentence1_key, sentence2_key = task_to_keys[data_args.task_name]
else:
# Again, we try to have some nice defaults but don't hesitate to tweak to your use case.
non_label_column_names = [
name for name in datasets["train"].column_names if name != "label"
]
if "sentence1" in non_label_column_names and "sentence2" in non_label_column_names:
sentence1_key, sentence2_key = "sentence1", "sentence2"
else:
if len(non_label_column_names) >= 2:
sentence1_key, sentence2_key = non_label_column_names[:2]
else:
sentence1_key, sentence2_key = non_label_column_names[0], None
# Padding strategy
if data_args.pad_to_max_length:
padding = "max_length"
else:
# We will pad later, dynamically at batch creation, to the max sequence length in each batch
padding = False
# Some models have set the order of the labels to use, so let's make sure we do use it.
label_to_id = None
if (model.config.label2id !=
PretrainedConfig(num_labels=num_labels).label2id
and data_args.task_name is not None and not is_regression):
# Some have all caps in their config, some don't.
label_name_to_id = {
k.lower(): v
for k, v in model.config.label2id.items()
}
if list(sorted(label_name_to_id.keys())) == list(sorted(label_list)):
label_to_id = {
i: int(label_name_to_id[label_list[i]])
for i in range(num_labels)
}
else:
logger.warn(
"Your model seems to have been trained with labels, but they don't match the dataset: ",
f"model labels: {list(sorted(label_name_to_id.keys()))}, dataset labels: {list(sorted(label_list))}."
"\nIgnoring the model labels as a result.",
)
elif data_args.task_name is None and not is_regression:
label_to_id = {v: i for i, v in enumerate(label_list)}
if data_args.max_seq_length > tokenizer.model_max_length:
logger.warn(
f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the"
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
def preprocess_function(examples):
# Tokenize the texts
args = ((examples[sentence1_key], ) if sentence2_key is None else
(examples[sentence1_key], examples[sentence2_key]))
result = tokenizer(*args,
padding=padding,
max_length=max_seq_length,
truncation=True)
# Map labels to IDs (not necessary for GLUE tasks)
if label_to_id is not None and "label" in examples:
result["label"] = [(label_to_id[l] if l != -1 else -1)
for l in examples["label"]]
return result
datasets = datasets.map(preprocess_function,
batched=True,
load_from_cache_file=not data_args.overwrite_cache)
if training_args.do_train:
if "train" not in datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = datasets["train"]
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
if training_args.do_eval:
if "validation" not in datasets and "validation_matched" not in datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = datasets["validation_matched" if data_args.task_name ==
"mnli" else "validation"]
if data_args.max_val_samples is not None:
eval_dataset = eval_dataset.select(range(
data_args.max_val_samples))
if (training_args.do_predict or data_args.task_name is not None
or data_args.test_file is not None):
if "test" not in datasets and "test_matched" not in datasets:
raise ValueError("--do_predict requires a test dataset")
test_dataset = datasets["test_matched" if data_args.task_name ==
"mnli" else "test"]
if data_args.max_test_samples is not None:
test_dataset = test_dataset.select(
range(data_args.max_test_samples))
# Log a few random samples from the training set:
if training_args.do_train:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(
f"Sample {index} of the training set: {train_dataset[index]}.")
# Get the metric function
if data_args.task_name is not None:
metric = load_metric("glue", data_args.task_name)
# TODO: When datasets metrics include regular accuracy, make an else here and remove special branch from
# compute_metrics
# You can define your custom compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with a
# predictions and label_ids field) and has to return a dictionary string to float.
def compute_metrics(p: EvalPrediction):
preds = p.predictions[0] if isinstance(p.predictions,
tuple) else p.predictions
preds = np.squeeze(preds) if is_regression else np.argmax(preds,
axis=1)
if data_args.task_name is not None:
result = metric.compute(predictions=preds, references=p.label_ids)
if len(result) > 1:
result["combined_score"] = np.mean(list(
result.values())).item()
return result
elif is_regression:
return {"mse": ((preds - p.label_ids)**2).mean().item()}
else:
return {
"accuracy":
(preds == p.label_ids).astype(np.float32).mean().item()
}
# Data collator will default to DataCollatorWithPadding, so we change it if we already did the padding.
if data_args.pad_to_max_length:
data_collator = default_data_collator
elif training_args.fp16:
data_collator = DataCollatorWithPadding(tokenizer,
pad_to_multiple_of=8)
else:
data_collator = None
# Initialize our Trainer
trainer = Trainer(
model=model,
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,
compute_metrics=compute_metrics,
tokenizer=tokenizer,
data_collator=data_collator,
)
# Training
if training_args.do_train:
checkpoint = None
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
# Check the config from that potential checkpoint has the right number of labels before using it as a
# checkpoint.
if AutoConfig.from_pretrained(
model_args.model_name_or_path).num_labels == num_labels:
checkpoint = model_args.model_name_or_path
train_result = trainer.train(resume_from_checkpoint=checkpoint)
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(train_dataset))
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.save_model() # Saves the tokenizer too for easy upload
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
if training_args.do_eval:
logger.info("*** Evaluate ***")
# Loop to handle MNLI double evaluation (matched, mis-matched)
tasks = [data_args.task_name]
eval_datasets = [eval_dataset]
if data_args.task_name == "mnli":
tasks.append("mnli-mm")
eval_datasets.append(datasets["validation_mismatched"])
for eval_dataset, task in zip(eval_datasets, tasks):
metrics = trainer.evaluate(eval_dataset=eval_dataset)
max_val_samples = (data_args.max_val_samples
if data_args.max_val_samples is not None else
len(eval_dataset))
metrics["eval_samples"] = min(max_val_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
if training_args.do_predict:
logger.info("*** Test ***")
# Loop to handle MNLI double evaluation (matched, mis-matched)
tasks = [data_args.task_name]
test_datasets = [test_dataset]
if data_args.task_name == "mnli":
tasks.append("mnli-mm")
test_datasets.append(datasets["test_mismatched"])
for test_dataset, task in zip(test_datasets, tasks):
# Removing the `label` columns because it contains -1 and Trainer won't like that.
test_dataset.remove_columns_("label")
predictions = trainer.predict(
test_dataset=test_dataset).predictions
predictions = (np.squeeze(predictions) if is_regression else
np.argmax(predictions, axis=1))
output_test_file = os.path.join(training_args.output_dir,
f"test_results_{task}.txt")
if trainer.is_world_process_zero():
with open(output_test_file, "w") as writer:
logger.info(f"***** Test results {task} *****")
writer.write("index\tprediction\n")
for index, item in enumerate(predictions):
if is_regression:
writer.write(f"{index}\t{item:3.3f}\n")
else:
item = label_list[item]
writer.write(f"{index}\t{item}\n")
def main():
args = parse_args()
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
accelerator = Accelerator()
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state)
# Setup logging, we only want one process per machine to log things on the screen.
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(
logging.INFO if accelerator.is_local_main_process else logging.ERROR)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(args.dataset_name,
args.dataset_config_name)
else:
data_files = {}
if args.train_file is not None:
data_files["train"] = args.train_file
if args.validation_file is not None:
data_files["validation"] = args.validation_file
if args.test_file is not None:
data_files["test"] = args.test_file
extension = args.train_file.split(".")[-1]
raw_datasets = load_dataset(extension,
data_files=data_files,
field="data")
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if args.config_name:
config = AutoConfig.from_pretrained(args.config_name)
elif args.model_name_or_path:
config = AutoConfig.from_pretrained(args.model_name_or_path)
else:
config = CONFIG_MAPPING[args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
if args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name,
use_fast=True)
elif args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path,
use_fast=True)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported by this script."
"You can do it from another script, save it, and load it from here, using --tokenizer_name."
)
if args.model_name_or_path:
model = AutoModelForQuestionAnswering.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForQuestionAnswering.from_config(config)
# Preprocessing the datasets.
# Preprocessing is slighlty different for training and evaluation.
column_names = raw_datasets["train"].column_names
question_column_name = "question" if "question" in column_names else column_names[
0]
context_column_name = "context" if "context" in column_names else column_names[
1]
answer_column_name = "answers" if "answers" in column_names else column_names[
2]
# Padding side determines if we do (question|context) or (context|question).
pad_on_right = tokenizer.padding_side == "right"
if args.max_seq_length > tokenizer.model_max_length:
logger.warning(
f"The max_seq_length passed ({args.max_seq_length}) is larger than the maximum length for the"
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(args.max_seq_length, tokenizer.model_max_length)
# Training preprocessing
def prepare_train_features(examples):
# Some of the questions have lots of whitespace on the left, which is not useful and will make the
# truncation of the context fail (the tokenized question will take a lots of space). So we remove that
# left whitespace
examples[question_column_name] = [
q.lstrip() for q in examples[question_column_name]
]
# Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
# in one example possible giving several features when a context is long, each of those features having a
# context that overlaps a bit the context of the previous feature.
tokenized_examples = tokenizer(
examples[
question_column_name if pad_on_right else context_column_name],
examples[
context_column_name if pad_on_right else question_column_name],
truncation="only_second" if pad_on_right else "only_first",
max_length=max_seq_length,
stride=args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
padding="max_length" if args.pad_to_max_length else False,
)
# Since one example might give us several features if it has a long context, we need a map from a feature to
# its corresponding example. This key gives us just that.
sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
# The offset mappings will give us a map from token to character position in the original context. This will
# help us compute the start_positions and end_positions.
offset_mapping = tokenized_examples.pop("offset_mapping")
# Let's label those examples!
tokenized_examples["start_positions"] = []
tokenized_examples["end_positions"] = []
for i, offsets in enumerate(offset_mapping):
# We will label impossible answers with the index of the CLS token.
input_ids = tokenized_examples["input_ids"][i]
cls_index = input_ids.index(tokenizer.cls_token_id)
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples.sequence_ids(i)
# One example can give several spans, this is the index of the example containing this span of text.
sample_index = sample_mapping[i]
answers = examples[answer_column_name][sample_index]
# If no answers are given, set the cls_index as answer.
if len(answers["answer_start"]) == 0:
tokenized_examples["start_positions"].append(cls_index)
tokenized_examples["end_positions"].append(cls_index)
else:
# Start/end character index of the answer in the text.
start_char = answers["answer_start"][0]
end_char = start_char + len(answers["text"][0])
# Start token index of the current span in the text.
token_start_index = 0
while sequence_ids[token_start_index] != (1 if pad_on_right
else 0):
token_start_index += 1
# End token index of the current span in the text.
token_end_index = len(input_ids) - 1
while sequence_ids[token_end_index] != (1 if pad_on_right else
0):
token_end_index -= 1
# Detect if the answer is out of the span (in which case this feature is labeled with the CLS index).
if not (offsets[token_start_index][0] <= start_char
and offsets[token_end_index][1] >= end_char):
tokenized_examples["start_positions"].append(cls_index)
tokenized_examples["end_positions"].append(cls_index)
else:
# Otherwise move the token_start_index and token_end_index to the two ends of the answer.
# Note: we could go after the last offset if the answer is the last word (edge case).
while token_start_index < len(offsets) and offsets[
token_start_index][0] <= start_char:
token_start_index += 1
tokenized_examples["start_positions"].append(
token_start_index - 1)
while offsets[token_end_index][1] >= end_char:
token_end_index -= 1
tokenized_examples["end_positions"].append(
token_end_index + 1)
return tokenized_examples
if "train" not in raw_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = raw_datasets["train"]
if args.max_train_samples is not None:
# We will select sample from whole data if agument is specified
train_dataset = train_dataset.select(range(args.max_train_samples))
# Create train feature from dataset
train_dataset = train_dataset.map(
prepare_train_features,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on train dataset",
)
if args.max_train_samples is not None:
# Number of samples might increase during Feature Creation, We select only specified max samples
train_dataset = train_dataset.select(range(args.max_train_samples))
# Validation preprocessing
def prepare_validation_features(examples):
# Some of the questions have lots of whitespace on the left, which is not useful and will make the
# truncation of the context fail (the tokenized question will take a lots of space). So we remove that
# left whitespace
examples[question_column_name] = [
q.lstrip() for q in examples[question_column_name]
]
# Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
# in one example possible giving several features when a context is long, each of those features having a
# context that overlaps a bit the context of the previous feature.
tokenized_examples = tokenizer(
examples[
question_column_name if pad_on_right else context_column_name],
examples[
context_column_name if pad_on_right else question_column_name],
truncation="only_second" if pad_on_right else "only_first",
max_length=max_seq_length,
stride=args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
padding="max_length" if args.pad_to_max_length else False,
)
# Since one example might give us several features if it has a long context, we need a map from a feature to
# its corresponding example. This key gives us just that.
sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
# For evaluation, we will need to convert our predictions to substrings of the context, so we keep the
# corresponding example_id and we will store the offset mappings.
tokenized_examples["example_id"] = []
for i in range(len(tokenized_examples["input_ids"])):
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples.sequence_ids(i)
context_index = 1 if pad_on_right else 0
# One example can give several spans, this is the index of the example containing this span of text.
sample_index = sample_mapping[i]
tokenized_examples["example_id"].append(
examples["id"][sample_index])
# Set to None the offset_mapping that are not part of the context so it's easy to determine if a token
# position is part of the context or not.
tokenized_examples["offset_mapping"][i] = [
(o if sequence_ids[k] == context_index else None)
for k, o in enumerate(tokenized_examples["offset_mapping"][i])
]
return tokenized_examples
if "validation" not in raw_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_examples = raw_datasets["validation"]
if args.max_eval_samples is not None:
# We will select sample from whole data
eval_examples = eval_examples.select(range(args.max_eval_samples))
# Validation Feature Creation
eval_dataset = eval_examples.map(
prepare_validation_features,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on validation dataset",
)
if args.max_eval_samples is not None:
# During Feature creation dataset samples might increase, we will select required samples again
eval_dataset = eval_dataset.select(range(args.max_eval_samples))
if args.do_predict:
if "test" not in raw_datasets:
raise ValueError("--do_predict requires a test dataset")
predict_examples = raw_datasets["test"]
if args.max_predict_samples is not None:
# We will select sample from whole data
predict_examples = predict_examples.select(
range(args.max_predict_samples))
# Predict Feature Creation
predict_dataset = predict_examples.map(
prepare_validation_features,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on prediction dataset",
)
if args.max_predict_samples is not None:
# During Feature creation dataset samples might increase, we will select required samples again
predict_dataset = predict_dataset.select(
range(args.max_predict_samples))
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(
f"Sample {index} of the training set: {train_dataset[index]}.")
# DataLoaders creation:
if args.pad_to_max_length:
# If padding was already done ot max length, we use the default data collator that will just convert everything
# to tensors.
data_collator = default_data_collator
else:
# Otherwise, `DataCollatorWithPadding` will apply dynamic padding for us (by padding to the maximum length of
# the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple
# of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).
data_collator = DataCollatorWithPadding(
tokenizer,
pad_to_multiple_of=(8 if accelerator.use_fp16 else None))
train_dataloader = DataLoader(train_dataset,
shuffle=True,
collate_fn=data_collator,
batch_size=args.per_device_train_batch_size)
eval_dataset_for_model = eval_dataset.remove_columns(
["example_id", "offset_mapping"])
eval_dataloader = DataLoader(eval_dataset_for_model,
collate_fn=data_collator,
batch_size=args.per_device_eval_batch_size)
if args.do_predict:
predict_dataset_for_model = predict_dataset.remove_columns(
["example_id", "offset_mapping"])
predict_dataloader = DataLoader(
predict_dataset_for_model,
collate_fn=data_collator,
batch_size=args.per_device_eval_batch_size)
# 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 = postprocess_qa_predictions(
examples=examples,
features=features,
predictions=predictions,
version_2_with_negative=args.version_2_with_negative,
n_best_size=args.n_best_size,
max_answer_length=args.max_answer_length,
null_score_diff_threshold=args.null_score_diff_threshold,
output_dir=args.output_dir,
prefix=stage,
)
# Format the result to the format the metric expects.
if args.version_2_with_negative:
formatted_predictions = [{
"id": k,
"prediction_text": v,
"no_answer_probability": 0.0
} for k, v in predictions.items()]
else:
formatted_predictions = [{
"id": k,
"prediction_text": v
} for k, v in predictions.items()]
references = [{
"id": ex["id"],
"answers": ex[answer_column_name]
} for ex in examples]
return EvalPrediction(predictions=formatted_predictions,
label_ids=references)
metric = load_metric(
"squad_v2" if args.version_2_with_negative else "squad")
# Create and fill numpy array of size len_of_validation_data * max_length_of_output_tensor
def create_and_fill_np_array(start_or_end_logits, dataset, max_len):
"""
Create and fill numpy array of size len_of_validation_data * max_length_of_output_tensor
Args:
start_or_end_logits(:obj:`tensor`):
This is the output predictions of the model. We can only enter either start or end logits.
eval_dataset: Evaluation dataset
max_len(:obj:`int`):
The maximum length of the output tensor. ( See the model.eval() part for more details )
"""
step = 0
# create a numpy array and fill it with -100.
logits_concat = np.full((len(dataset), max_len),
-100,
dtype=np.float64)
# Now since we have create an array now we will populate it with the outputs gathered using accelerator.gather
for i, output_logit in enumerate(
start_or_end_logits): # populate columns
# We have to fill it such that we have to take the whole tensor and replace it on the newly created array
# And after every iteration we have to change the step
batch_size = output_logit.shape[0]
cols = output_logit.shape[1]
if step + batch_size < len(dataset):
logits_concat[step:step + batch_size, :cols] = output_logit
else:
logits_concat[step:, :cols] = output_logit[:len(dataset) -
step]
step += batch_size
return logits_concat
# Optimizer
# Split weights in two groups, one with weight decay and the other not.
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)
# Prepare everything with our `accelerator`.
model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader)
# Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be
# shorter in multiprocess)
# Scheduler and math around the number of training steps.
num_update_steps_per_epoch = math.ceil(
len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
else:
args.num_train_epochs = math.ceil(args.max_train_steps /
num_update_steps_per_epoch)
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=args.num_warmup_steps,
num_training_steps=args.max_train_steps,
)
# Train!
total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(
f" Instantaneous batch size per device = {args.per_device_train_batch_size}"
)
logger.info(
f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}"
)
logger.info(
f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
# Only show the progress bar once on each machine.
progress_bar = tqdm(range(args.max_train_steps),
disable=not accelerator.is_local_main_process)
completed_steps = 0
for epoch in range(args.num_train_epochs):
model.train()
for step, batch in enumerate(train_dataloader):
outputs = model(**batch)
loss = outputs.loss
loss = loss / args.gradient_accumulation_steps
accelerator.backward(loss)
if step % args.gradient_accumulation_steps == 0 or step == len(
train_dataloader) - 1:
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
progress_bar.update(1)
completed_steps += 1
if completed_steps >= args.max_train_steps:
break
# Evaluation
logger.info("***** Running Evaluation *****")
logger.info(f" Num examples = {len(eval_dataset)}")
logger.info(f" Batch size = {args.per_device_eval_batch_size}")
all_start_logits = []
all_end_logits = []
for step, batch in enumerate(eval_dataloader):
with torch.no_grad():
outputs = model(**batch)
start_logits = outputs.start_logits
end_logits = outputs.end_logits
if not args.pad_to_max_length: # necessary to pad predictions and labels for being gathered
start_logits = accelerator.pad_across_processes(start_logits,
dim=1,
pad_index=-100)
end_logits = accelerator.pad_across_processes(end_logits,
dim=1,
pad_index=-100)
all_start_logits.append(
accelerator.gather(start_logits).cpu().numpy())
all_end_logits.append(accelerator.gather(end_logits).cpu().numpy())
max_len = max([x.shape[1] for x in all_start_logits
]) # Get the max_length of the tensor
# concatenate the numpy array
start_logits_concat = create_and_fill_np_array(all_start_logits,
eval_dataset, max_len)
end_logits_concat = create_and_fill_np_array(all_end_logits, eval_dataset,
max_len)
# delete the list of numpy arrays
del all_start_logits
del all_end_logits
outputs_numpy = (start_logits_concat, end_logits_concat)
prediction = post_processing_function(eval_examples, eval_dataset,
outputs_numpy)
eval_metric = metric.compute(predictions=prediction.predictions,
references=prediction.label_ids)
logger.info(f"Evaluation metrics: {eval_metric}")
# Prediction
if args.do_predict:
logger.info("***** Running Prediction *****")
logger.info(f" Num examples = {len(predict_dataset)}")
logger.info(f" Batch size = {args.per_device_eval_batch_size}")
all_start_logits = []
all_end_logits = []
for step, batch in enumerate(predict_dataloader):
with torch.no_grad():
outputs = model(**batch)
start_logits = outputs.start_logits
end_logits = outputs.end_logits
if not args.pad_to_max_length: # necessary to pad predictions and labels for being gathered
start_logits = accelerator.pad_across_processes(
start_logits, dim=1, pad_index=-100)
end_logits = accelerator.pad_across_processes(
start_logits, dim=1, pad_index=-100)
all_start_logits.append(
accelerator.gather(start_logits).cpu().numpy())
all_end_logits.append(
accelerator.gather(end_logits).cpu().numpy())
max_len = max([x.shape[1] for x in all_start_logits
]) # Get the max_length of the tensor
# concatenate the numpy array
start_logits_concat = create_and_fill_np_array(all_start_logits,
predict_dataset,
max_len)
end_logits_concat = create_and_fill_np_array(all_end_logits,
predict_dataset, max_len)
# delete the list of numpy arrays
del all_start_logits
del all_end_logits
outputs_numpy = (start_logits_concat, end_logits_concat)
prediction = post_processing_function(predict_examples,
predict_dataset, outputs_numpy)
predict_metric = metric.compute(predictions=prediction.predictions,
references=prediction.label_ids)
logger.info(f"Predict metrics: {predict_metric}")
if args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir,
save_function=accelerator.save)
def finetune(accelerator, model_name_or_path, train_file, output_dir,
**kwargs):
"""Fine-tuning a pre-trained model on a downstream task.
Args:
accelerator: An instance of an accelerator for distributed training (on
multi-GPU, TPU) or mixed precision training.
model_name_or_path: Path to pretrained model or model identifier from
huggingface.co/models.
train_file: A csv or a json file containing the training data.
output_dir: The output directory where the model predictions and checkpoints
will be written.
**kwargs: Dictionary of key/value pairs with which to update the
configuration object after loading. The values in kwargs of any keys which
are configuration attributes will be used to override the loaded values.
"""
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO,
)
logger.info(accelerator.state)
# Setup logging, we only want one process per machine to log things on the
# screen. accelerator.is_local_main_process is only True for one process per
# machine.
logger.setLevel(
logging.INFO if accelerator.is_local_main_process else logging.ERROR)
model_args = FTModelArguments(model_name_or_path=model_name_or_path)
data_args = FTDataArguments(train_file=train_file)
training_args = FTTrainingArguments(output_dir=output_dir)
args = argparse.Namespace()
for arg_class in (model_args, data_args, training_args):
for key, value in vars(arg_class).items():
setattr(args, key, value)
for key, value in kwargs.items():
if hasattr(args, key):
setattr(args, key, value)
# Sanity checks
data_files = {}
args.data_file_extension = None
# You need to provide the training data as we always run training
args.do_train = True
assert args.train_file is not None
data_files[Split.TRAIN.value] = args.train_file
if args.do_eval or args.evaluation_strategy != IntervalStrategy.NO.value:
assert args.eval_file is not None
data_files[Split.EVAL.value] = args.eval_file
if args.do_eval and args.test_file is not None:
data_files[Split.TEST.value] = args.test_file
if args.do_predict:
assert args.infer_file is not None
data_files[Split.INFER.value] = args.infer_file
for key in data_files:
extension = data_files[key].split('.')[-1]
assert extension in ['csv', 'json'
], f'`{key}_file` should be a csv or a json file.'
if args.data_file_extension is None:
args.data_file_extension = extension
else:
assert (
extension == args.data_file_extension
), f'`{key}_file` should be a {args.data_file_extension} file`.'
assert (
args.eval_metric in datasets.list_metrics()
), f'{args.eval_metric} not in the list of supported metrics {datasets.list_metrics()}.'
# Handle the output directory creation
if accelerator.is_main_process:
if args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
accelerator.wait_for_everyone()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# You need to provide your CSV/JSON data files.
#
# For CSV/JSON files, this script will use as labels the column called 'label'
# and as pair of sentences the sentences in columns called 'sentence1' and
# 'sentence2' if these columns exist or the first two columns not named
# 'label' if at least two columns are provided.
#
# If the CSVs/JSONs contain only one non-label column, the script does single
# sentence classification on this single column.
#
# In distributed training, the load_dataset function guarantees that only one
# local process can download the dataset.
# Loading the dataset from local csv or json files.
raw_datasets = load_dataset(args.data_file_extension,
data_files=data_files)
# Labels
is_regression = raw_datasets[
Split.TRAIN.value].features['label'].dtype in ['float32', 'float64']
args.is_regression = is_regression
if args.is_regression:
label_list = None
num_labels = 1
else:
label_list = args.label_list
assert label_list is not None
label_list.sort() # Let's sort it for determinism
num_labels = len(label_list)
args.num_labels = num_labels
# Load pre-trained model
config, tokenizer, model = load_from_pretrained(args,
args.model_name_or_path)
# Preprocessing the datasets
non_label_column_names = [
name for name in raw_datasets[Split.TRAIN.value].column_names
if name != 'label'
]
if 'sentence1' in non_label_column_names and 'sentence2' in non_label_column_names:
sentence1_key, sentence2_key = 'sentence1', 'sentence2'
else:
if len(non_label_column_names) >= 2:
sentence1_key, sentence2_key = non_label_column_names[:2]
else:
sentence1_key, sentence2_key = non_label_column_names[0], None
label_to_id = {v: i for i, v in enumerate(label_list)}
config.label2id = label_to_id
config.id2label = {id: label for label, id in config.label2id.items()}
padding = 'max_length' if args.pad_to_max_length else False
def preprocess_function(examples):
# Tokenize the texts
texts = ((examples[sentence1_key], ) if sentence2_key is None else
(examples[sentence1_key], examples[sentence2_key]))
result = tokenizer(*texts,
padding=padding,
max_length=args.max_length,
truncation=True)
if 'label' in examples:
if label_to_id is not None:
# Map labels to IDs (not necessary for GLUE tasks)
result['labels'] = [label_to_id[l] for l in examples['label']]
else:
# In all cases, rename the column to labels because the model will
# expect that.
result['labels'] = examples['label']
return result
with accelerator.main_process_first():
processed_datasets = raw_datasets.map(
preprocess_function,
batched=True,
remove_columns=raw_datasets[Split.TRAIN.value].column_names,
desc='Running tokenizer on dataset',
)
num_examples = {}
splits = [s.value for s in Split]
for split in splits:
if split in processed_datasets:
num_examples[split] = len(processed_datasets[split])
args.num_examples = num_examples
train_dataset = processed_datasets[Split.TRAIN.value]
eval_dataset = processed_datasets[
Split.EVAL.value] if Split.EVAL.value in processed_datasets else None
test_dataset = processed_datasets[
Split.TEST.value] if Split.TEST.value in processed_datasets else None
infer_dataset = processed_datasets[
Split.INFER.value] if Split.INFER.value in processed_datasets else None
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info('Sample %d of the training set: %s.', index,
train_dataset[index])
# DataLoaders creation:
if args.pad_to_max_length:
# If padding was already done ot max length, we use the default data
# collator that will just convert everything to tensors.
data_collator = default_data_collator
else:
# Otherwise, `DataCollatorWithPadding` will apply dynamic padding for us (by
# padding to the maximum length of the samples passed). When using mixed
# precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple of
# 8s, which will enable the use of Tensor Cores on NVIDIA hardware with
# compute capability >= 7.5 (Volta).
data_collator = DataCollatorWithPadding(
tokenizer,
pad_to_multiple_of=(8 if accelerator.use_fp16 else None))
train_dataloader = DataLoader(
train_dataset,
batch_size=args.per_device_train_batch_size,
shuffle=True,
collate_fn=data_collator,
)
eval_dataloader, test_dataloader, infer_dataloader = None, None, None
if eval_dataset is not None:
eval_dataloader = DataLoader(
eval_dataset,
batch_size=args.per_device_eval_batch_size,
collate_fn=data_collator)
if test_dataset is not None:
test_dataloader = DataLoader(
test_dataset,
batch_size=args.per_device_eval_batch_size,
collate_fn=data_collator)
if infer_dataset is not None:
infer_dataloader = DataLoader(
infer_dataset,
batch_size=args.per_device_eval_batch_size,
collate_fn=data_collator)
# Optimizer
# Split weights in two groups, one with weight decay and the other not.
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)
# Prepare everything with our `accelerator`.
model, optimizer, train_dataloader, eval_dataloader, test_dataloader, infer_dataloader = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader, test_dataloader,
infer_dataloader)
# Note -> the training dataloader needs to be prepared before we grab its
# length below (cause its length will be shorter in multiprocess)
# Scheduler and math around the number of training steps.
num_update_steps_per_epoch = math.ceil(
len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_steps == -1:
args.max_steps = args.num_train_epochs * num_update_steps_per_epoch
else:
args.num_train_epochs = math.ceil(args.max_steps /
num_update_steps_per_epoch)
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=args.warmup_steps,
num_training_steps=args.max_steps,
)
# Train
completed_steps, avg_train_loss = train(args, accelerator, model,
tokenizer, train_dataloader,
optimizer, lr_scheduler,
eval_dataloader)
accelerator.wait_for_everyone()
logger.info(
'Training job completed: completed_steps = %d, avg_train_loss = %f',
completed_steps, avg_train_loss)
args.model_name_or_path = os.path.join(args.output_dir, 'best-checkpoint')
logger.info('Loading the best checkpoint: %s', args.model_name_or_path)
config, tokenizer, model = load_from_pretrained(args,
args.model_name_or_path)
model = accelerator.prepare(model)
if args.do_eval:
# Evaluate
if eval_dataloader is not None:
logger.info(
'***** Running evaluation on the eval data using the best checkpoint *****'
)
eval_results = evaluate(args, accelerator, eval_dataloader,
Split.EVAL.value, model, 'best-checkpoint')
avg_eval_loss = eval_results['avg_eval_loss']
eval_metric = eval_results[args.eval_metric]
logger.info('Evaluation job completed: avg_eval_loss = %f',
avg_eval_loss)
logger.info('Evaluation result for the best checkpoint: %s = %f',
args.eval_metric, eval_metric)
if test_dataloader is not None:
logger.info(
'***** Running evaluation on the test data using the best checkpoint *****'
)
eval_results = evaluate(args, accelerator, test_dataloader,
Split.TEST.value, model, 'best-checkpoint')
avg_eval_loss = eval_results['avg_eval_loss']
eval_metric = eval_results[args.eval_metric]
logger.info('Test job completed: avg_test_loss = %f',
avg_eval_loss)
logger.info('Test result for the best checkpoint: %s = %f',
args.eval_metric, eval_metric)
if args.do_predict:
# Predict
if infer_dataloader is not None:
logger.info(
'***** Running inference using the best checkpoint *****')
evaluate(args,
accelerator,
infer_dataloader,
Split.INFER.value,
model,
'best-checkpoint',
has_labels=False)
logger.info('Inference job completed.')
# Release all references to the internal objects stored and call the garbage
# collector. You should call this method between two trainings with different
# models/optimizers.
accelerator.free_memory()
def main():
args = parse_args()
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
accelerator = Accelerator()
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state)
# Setup logging, we only want one process per machine to log things on the screen.
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(
logging.INFO if accelerator.is_local_main_process else logging.ERROR)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
# or specify a GLUE benchmark task (the dataset will be downloaded automatically from the datasets Hub).
# For CSV/JSON files, this script will use as labels the column called 'label' and as pair of sentences the
# sentences in columns called 'sentence1' and 'sentence2' if such column exists or the first two columns not named
# label if at least two columns are provided.
# If the CSVs/JSONs contain only one non-label column, the script does single sentence classification on this
# single column. You can easily tweak this behavior (see below)
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if args.task_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset("glue", args.task_name)
else:
# Loading the dataset from local csv or json file.
data_files = {}
if args.train_file is not None:
data_files["train"] = args.train_file
if args.validation_file is not None:
data_files["validation"] = args.validation_file
extension = (args.train_file if args.train_file is not None else
args.valid_file).split(".")[-1]
raw_datasets = load_dataset(extension, data_files=data_files)
# See more about loading any type of standard or custom dataset at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Labels
if args.task_name is not None:
is_regression = args.task_name == "stsb"
if not is_regression:
label_list = raw_datasets["train"].features["label"].names
num_labels = len(label_list)
else:
num_labels = 1
else:
# Trying to have good defaults here, don't hesitate to tweak to your needs.
is_regression = raw_datasets["train"].features["label"].dtype in [
"float32", "float64"
]
if is_regression:
num_labels = 1
else:
# A useful fast method:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.unique
label_list = raw_datasets["train"].unique("label")
label_list.sort() # Let's sort it for determinism
num_labels = len(label_list)
# Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(args.model_name_or_path,
num_labels=num_labels,
finetuning_task=args.task_name)
tokenizer = AutoTokenizer.from_pretrained(
args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
model = AutoModelForSequenceClassification.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
)
# Preprocessing the datasets
if args.task_name is not None:
sentence1_key, sentence2_key = task_to_keys[args.task_name]
else:
# Again, we try to have some nice defaults but don't hesitate to tweak to your use case.
non_label_column_names = [
name for name in raw_datasets["train"].column_names
if name != "label"
]
if "sentence1" in non_label_column_names and "sentence2" in non_label_column_names:
sentence1_key, sentence2_key = "sentence1", "sentence2"
else:
if len(non_label_column_names) >= 2:
sentence1_key, sentence2_key = non_label_column_names[:2]
else:
sentence1_key, sentence2_key = non_label_column_names[0], None
# Some models have set the order of the labels to use, so let's make sure we do use it.
label_to_id = None
if (model.config.label2id !=
PretrainedConfig(num_labels=num_labels).label2id
and args.task_name is not None and not is_regression):
# Some have all caps in their config, some don't.
label_name_to_id = {
k.lower(): v
for k, v in model.config.label2id.items()
}
if list(sorted(label_name_to_id.keys())) == list(sorted(label_list)):
logger.info(
f"The configuration of the model provided the following label correspondence: {label_name_to_id}. "
"Using it!")
label_to_id = {
i: label_name_to_id[label_list[i]]
for i in range(num_labels)
}
else:
logger.warn(
"Your model seems to have been trained with labels, but they don't match the dataset: ",
f"model labels: {list(sorted(label_name_to_id.keys()))}, dataset labels: {list(sorted(label_list))}."
"\nIgnoring the model labels as a result.",
)
elif args.task_name is None:
label_to_id = {v: i for i, v in enumerate(label_list)}
padding = "max_length" if args.pad_to_max_length else False
def preprocess_function(examples):
# Tokenize the texts
texts = ((examples[sentence1_key], ) if sentence2_key is None else
(examples[sentence1_key], examples[sentence2_key]))
result = tokenizer(*texts,
padding=padding,
max_length=args.max_length,
truncation=True)
if "label" in examples:
if label_to_id is not None:
# Map labels to IDs (not necessary for GLUE tasks)
result["labels"] = [label_to_id[l] for l in examples["label"]]
else:
# In all cases, rename the column to labels because the model will expect that.
result["labels"] = examples["label"]
return result
processed_datasets = raw_datasets.map(
preprocess_function,
batched=True,
remove_columns=raw_datasets["train"].column_names)
train_dataset = processed_datasets["train"]
eval_dataset = processed_datasets["validation_matched" if args.task_name ==
"mnli" else "validation"]
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(
f"Sample {index} of the training set: {train_dataset[index]}.")
# DataLoaders creation:
if args.pad_to_max_length:
# If padding was already done ot max length, we use the default data collator that will just convert everything
# to tensors.
data_collator = default_data_collator
else:
# Otherwise, `DataCollatorWithPadding` will apply dynamic padding for us (by padding to the maximum length of
# the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple
# of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).
data_collator = DataCollatorWithPadding(
tokenizer,
pad_to_multiple_of=(8 if accelerator.use_fp16 else None))
train_dataloader = DataLoader(train_dataset,
shuffle=True,
collate_fn=data_collator,
batch_size=args.per_device_train_batch_size)
eval_dataloader = DataLoader(eval_dataset,
collate_fn=data_collator,
batch_size=args.per_device_eval_batch_size)
# Optimizer
# Split weights in two groups, one with weight decay and the other not.
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)
# Prepare everything with our `accelerator`.
model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader)
# Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be
# shorter in multiprocess)
# Scheduler and math around the number of training steps.
num_update_steps_per_epoch = math.ceil(
len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
else:
args.num_train_epochs = math.ceil(args.max_train_steps /
num_update_steps_per_epoch)
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=args.num_warmup_steps,
num_training_steps=args.max_train_steps,
)
# Get the metric function
if args.task_name is not None:
metric = load_metric("glue", args.task_name)
# Train!
total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(
f" Instantaneous batch size per device = {args.per_device_train_batch_size}"
)
logger.info(
f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}"
)
logger.info(
f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
# Only show the progress bar once on each machine.
progress_bar = tqdm(range(args.max_train_steps),
disable=not accelerator.is_local_main_process)
completed_steps = 0
for epoch in range(args.num_train_epochs):
model.train()
for step, batch in enumerate(train_dataloader):
outputs = model(**batch)
loss = outputs.loss
loss = loss / args.gradient_accumulation_steps
accelerator.backward(loss)
if step % args.gradient_accumulation_steps == 0 or step == len(
train_dataloader) - 1:
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
progress_bar.update(1)
completed_steps += 1
if completed_steps >= args.max_train_steps:
break
model.eval()
for step, batch in enumerate(eval_dataloader):
outputs = model(**batch)
predictions = outputs.logits.argmax(dim=-1)
metric.add_batch(
predictions=accelerator.gather(predictions),
references=accelerator.gather(batch["labels"]),
)
eval_metric = metric.compute()
logger.info(f"epoch {epoch}: {eval_metric}")
if args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir,
save_function=accelerator.save)
if args.task_name == "mnli":
# Final evaluation on mismatched validation set
eval_dataset = processed_datasets["validation_mismatched"]
eval_dataloader = DataLoader(
eval_dataset,
collate_fn=data_collator,
batch_size=args.per_device_eval_batch_size)
eval_dataloader = accelerator.prepare(eval_dataloader)
model.eval()
for step, batch in enumerate(eval_dataloader):
outputs = model(**batch)
predictions = outputs.logits.argmax(dim=-1)
metric.add_batch(
predictions=accelerator.gather(predictions),
references=accelerator.gather(batch["labels"]),
)
eval_metric = metric.compute()
logger.info(f"mnli-mm: {eval_metric}")
