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(
)
if (
os.path.exists(training_args.output_dir) and
os.listdir(training_args.output_dir) and training_args.do_train and
not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO
if training_args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.local_rank != -1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Set seed
set_seed(training_args.seed)
# Prepare the task
label_set = LabelCollection(
data_args.task_type, transform=data_args.label_transform
)
# data_args.num_labels = len(label_set.label_list)
token_classification_task = processors.DeidProcessor(
data_args.data_dir, label_set
)
labels = token_classification_task.label_set.labels
label_map: Dict[int, str] = {i: label for i, label in enumerate(labels)}
num_labels = len(labels)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
id2label=label_map,
label2id={label: i
for i, label in enumerate(labels)},
cache_dir=model_args.cache_dir,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
# we always use fast tokenizers as we need offsets from tokenization
use_fast=True,
)
model = AutoModelForTokenClassification.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,
)
# Get datasets
train_dataset = (
TokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.train,
) if training_args.do_train else None
)
eval_dataset = (
TokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.dev,
) if training_args.do_eval else None
)
def compute_metrics(p: EvalPrediction) -> Dict:
align = partial(align_predictions, label_map=label_map)
preds_list, out_label_list = align(p.predictions, p.label_ids)
return {
"accuracy_score": accuracy_score(out_label_list, preds_list),
"precision": precision_score(out_label_list, preds_list),
"recall": recall_score(out_label_list, preds_list),
"f1": f1_score(out_label_list, preds_list),
}
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
# validate the labels
for feature in train_dataset.features:
assert all(
[
t == -100 or (t >= 0 and t <= model.num_labels)
for t in feature.label_ids
]
), 'label_ids outside valid range for num_labels, check cache'
trainer.train(
model_path=model_args.model_name_or_path if os.path.
isdir(model_args.model_name_or_path) else None
)
trainer.save_model()
# TODO: save label set as well
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
if trainer.is_world_master():
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
result = trainer.evaluate()
output_eval_file = os.path.join(
training_args.output_dir, "eval_results.txt"
)
if trainer.is_world_master():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in result.items():
logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
results.update(result)
# Predict
if training_args.do_predict:
test_dataset = TokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.test,
)
predictions, label_ids, metrics = trainer.predict(test_dataset)
preds_list, out_label_list = align_predictions(
predictions, label_ids, label_map
)
output_test_results_file = os.path.join(
training_args.output_dir, "test_results.txt"
)
if trainer.is_world_master():
with open(output_test_results_file, "w") as writer:
for key, value in metrics.items():
logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
# Save predictions
output_test_predictions_file = os.path.join(
training_args.output_dir, "test_predictions.txt"
)
if trainer.is_world_master():
with open(output_test_predictions_file, "w") as writer:
for i in range(len(preds_list)):
"""
labels_to_write = [
'PAD' if label_ids[i][j] == -100 else
token_classification_task.label_set.id_to_label[
label_ids[i][j]] for j in range(len(preds_list[i]))
]
"""
writer.write(
'\n'.join(
[
f'{i},{preds_list[i][j]},{out_label_list[i][j]}'
for j in range(len(preds_list[i]))
]
)
)
writer.write('\n')
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, 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()
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets:
eval_dataset = datasets.load_dataset(CORPORA_PATH,
split="test",
cache_dir=model_args.cache_dir)
processor = CustomWav2Vec2Processor.from_pretrained(
model_args.model_name_or_path)
model = Wav2VecClassifierModel.from_pretrained(
model_args.model_name_or_path,
attention_dropout=0.01,
hidden_dropout=0.01,
feat_proj_dropout=0.0,
mask_time_prob=0.05,
layerdrop=0.01,
gradient_checkpointing=True,
)
if data_args.max_val_samples is not None:
max_val_samples = min(data_args.max_val_samples, len(eval_dataset))
eval_dataset = eval_dataset.select(range(max_val_samples))
# Preprocessing the datasets.
# We need to read the aduio files as arrays and tokenize the targets.
def speech_file_to_array_fn(batch, start_param, stop_param):
speech_array, sampling_rate = torchaudio.load(batch["file"])
speech_array = speech_array[0].numpy()[(start_param *
sampling_rate):(stop_param *
sampling_rate)]
batch["speech"] = librosa.resample(np.asarray(speech_array),
sampling_rate, S_RATE)
batch["sampling_rate"] = S_RATE
batch["parent"] = batch["label"]
return batch
def filter_null(batch):
return not (batch['speech'] == np.array([0])).all()
eval_dataset_array = []
stop = 0
for i in range(WINDOW_COUNT):
start = 0 if i == 0 else stop
stop = start + WINDOW_LENGTH
arguments = {'start_param': start, 'stop_param': stop}
eval_dataset_array.append(
eval_dataset.map(speech_file_to_array_fn,
remove_columns=eval_dataset.column_names,
num_proc=data_args.preprocessing_num_workers,
fn_kwargs=arguments).filter(filter_null))
eval_dataset = datasets.concatenate_datasets(eval_dataset_array)
def prepare_dataset(batch):
# check that all files have the correct sampling rate
assert (
len(set(batch["sampling_rate"])) == 1
), f"Make sure all inputs have the same sampling rate of {processor.feature_extractor.sampling_rate}."
batch["input_values"] = processor(
batch["speech"],
sampling_rate=batch["sampling_rate"][0]).input_values
batch["labels"] = batch["parent"]
return batch
eval_dataset = eval_dataset.map(
prepare_dataset,
remove_columns=eval_dataset.column_names,
batch_size=training_args.per_device_train_batch_size,
batched=True,
num_proc=data_args.preprocessing_num_workers,
)
from sklearn.metrics import classification_report, confusion_matrix
def compute_metrics(pred):
label_idx = LABEL_IDX
label_names = LABEL_NAMES
labels = pred.label_ids.argmax(-1)
preds = pred.predictions.argmax(-1)
acc = accuracy_score(labels, preds)
f1 = f1_score(labels, preds, average='macro')
report = classification_report(y_true=labels,
y_pred=preds,
labels=label_idx,
target_names=label_names)
matrix = confusion_matrix(y_true=labels, y_pred=preds)
print(report)
print(matrix)
wandb.log({
"conf_mat":
wandb.plot.confusion_matrix(probs=None,
y_true=labels,
preds=preds,
class_names=label_names)
})
wandb.log({
"precision_recall":
wandb.plot.pr_curve(y_true=labels,
y_probas=pred.predictions,
labels=label_names)
})
return {"accuracy": acc, "f1_score": f1}
wandb.init(name=training_args.output_dir, config=training_args)
# Data collator
data_collator = DataCollatorCTCWithPadding(processor=processor,
padding=True)
# Initialize our Trainer
trainer = CTCTrainer(
model=model,
data_collator=data_collator,
args=training_args,
compute_metrics=compute_metrics,
train_dataset=None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=processor.feature_extractor,
)
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
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)
return results
def main():
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If passed only one argument to the script and it's the path to a json file
# then parse it to get 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()
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
logging.basicConfig(
format = "%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt = "%m%d%Y %H:%M:%S",
level = logging.INFO if training_args.local_rank in [-1,0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpus: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.local_rank != -1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Set random seed
set_seed(training_args.seed)
# Prepare data task
labels = get_labels(data_args.labels)
label_map: Dict[int, str] = {i: label for i, label in enumerate(labels)}
num_labels = len(labels)
# Load pretrained model and tokenizer
# Distributed training: the .from_pretrained methods guarantee that only one local process can concurrently download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
id2label=label_map,
label2id={label: i for i, label in enumerate(labels)},
cache_dir=model_args.cache_dir,
hidden_dropout_prob=data_args.embedding_dropout,
hidden_size=data_args.hidden_size,
)
model = AutoModelForTokenClassification.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
config=config,
cache_dir=model_args.cache_dir,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast,
)
# print("config type: {}; tokenizer type: {}; model type: {}".format(type(config), type(tokenizer), type(model)))
print('--MODEL METADATA-- ', data_args.model_metadata)
# Get datasets
train_dataset = (
SRLDataset(
data_path=data_args.train_data_path,
tokenizer=tokenizer,
model_type=config.model_type,
labels_file=data_args.labels.rsplit('/', 1)[-1],
labels=labels,
predict_input=False,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
metadata=data_args.model_metadata,
)
if training_args.do_train
else None
)
eval_dataset = (
SRLDataset(
data_path=data_args.dev_data_path,
tokenizer=tokenizer,
model_type=config.model_type,
labels_file=data_args.labels.rsplit('/', 1)[-1],
labels=labels,
predict_input=False,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
metadata=data_args.model_metadata,
)
if training_args.do_eval
else None
)
def align_predictions(predictions: np.ndarray, label_ids: np.ndarray) -> Tuple[List[int], List[int]]:
"""
Align predictions and true tags to mask out tags to be ignored.
Inputs:
predictions: `np.ndarray`
Input of size (batch_size, seq_len, num labels) representing output of model.
label_ids: `np.ndarray`
Input of size (batch_size, seq_len) representing true tags.
Outputs:
preds_list: `List[int]`
List of predicted label tags.
out_label_list: `List[int]`
List of true label tags.
"""
preds = np.argmax(predictions, axis=2)
batch_size, seq_len = preds.shape
out_label_list = [[] for _ in range(batch_size)]
preds_list = [[] for _ in range(batch_size)]
for i in range(batch_size):
for j in range(seq_len):
# If this label is not masked over, lookup the corresponding tag and append it to outputs.
if label_ids[i, j] != nn.CrossEntropyLoss().ignore_index:
out_label_list[i].append(label_map[label_ids[i][j]])
preds_list[i].append(label_map[preds[i][j]])
return preds_list, out_label_list
def compute_metrics(p: EvalPrediction) -> Dict:
preds_list, out_label_list = align_predictions(p.predictions, p.label_ids)
report = classification_report(out_label_list, preds_list)
output_report_file = os.path.join(training_args.output_dir, "classification_report.txt")
with open(output_report_file, "w") as writer:
writer.write(report)
return {
"precision": precision_score(out_label_list, preds_list),
"recall": recall_score(out_label_list, preds_list),
"f1": f1_score(out_label_list, preds_list),
}
# Initialize trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
compute_metrics=compute_metrics,
)
# Perform training
if training_args.do_train:
trainer.train()
trainer.save_model()
# Can re-save tokenizer to same directory so taht can share model on huggingface.co/models
if trainer.is_world_master():
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
result = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir, "eval_results.txt")
if trainer.is_world_master():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in result.items():
logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
results.update(result)
# Predict
if training_args.do_predict:
test_dataset = SRLDataset(
data_path=data_args.test_data_path,
tokenizer=tokenizer,
model_type=config.model_type,
labels_file=data_args.labels.rsplit('/', 1)[-1],
labels=labels,
predict_input=False,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
metadata=data_args.model_metadata,
)
predictions, label_ids, metrics = trainer.predict(test_dataset)
preds_list, label_list = align_predictions(predictions, label_ids)
output_test_results_file = os.path.join(training_args.output_dir, "test_results.txt")
if trainer.is_world_master():
with open(output_test_results_file, "w") as writer:
for key, value in metrics.items():
logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
# Save predictions
output_test_predictions_file = os.path.join(training_args.output_dir, "test_predictions.txt")
if not os.path.isdir(data_args.test_data_path):
adc_output_file = os.path.join(training_args.output_dir, "adc_output.txt")
if trainer.is_world_master():
with open(output_test_predictions_file, "w") as output_writer, open(adc_output_file, "w") as adc_writer:
with open(data_args.test_data_path, "r") as f:
example_id = 0
for line in f:
str_list = line.strip().split()
separate_index = str_list.index("|||")
sentence = str_list[1:separate_index]
tags = str_list[separate_index+1:]
predicate_idx = str_list[0]
gold_srl = label_list[example_id]
pred_srl = preds_list[example_id]
gold_srl = gold_srl[:-1]
pred_srl = pred_srl[:-1]
output_line = "{0} {1} ||| {2}\n".format(str(predicate_idx), " ".join(sentence), " ".join(pred_srl))
adc_writer.write(output_line)
output_writer.write("sentence: " + " ".join(sentence) + "\n")
output_writer.write("input tags: " + " ".join(tags) + "\n")
output_writer.write("gold srl: " + " ".join(gold_srl) + "\n")
output_writer.write("pred srl: " + " ".join(pred_srl) + "\n")
example_id += 1
else:
if trainer.is_world_master():
with open(output_test_predictions_file, "w") as output_writer:
for index in range(len(preds_list)):
gold_srl = label_list[index]
pred_srl = preds_list[index]
output_writer.write("new sentence:\n") # TODO replace with actual sentence if can find it
output_writer.write(" ".join(gold_srl) + "\n")
output_writer.write(" ".join(pred_srl) + "\n")
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, 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(
)
# 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,
)
# Setup logging, we only want one process per machine to log things on the screen.
logger.setLevel(logging.INFO if jax.process_index() ==
0 else logging.ERROR)
if jax.process_index() == 0:
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()
# Handle the repository creation
if training_args.push_to_hub:
if training_args.hub_model_id is None:
repo_name = get_full_repo_name(Path(
training_args.output_dir).absolute().name,
token=training_args.hub_token)
else:
repo_name = training_args.hub_model_id
repo = Repository(training_args.output_dir, clone_from=repo_name)
# 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 for token classification task 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 'tokens' or the first column if no column called
# 'tokens' 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 data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
else:
# Loading the dataset from local csv or json file.
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = (data_args.train_file if data_args.train_file is not None
else data_args.valid_file).split(".")[-1]
raw_datasets = load_dataset(
extension,
data_files=data_files,
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
# See more about loading any type of standard or custom dataset at
# https://huggingface.co/docs/datasets/loading_datasets.html.
if raw_datasets["train"] is not None:
column_names = raw_datasets["train"].column_names
features = raw_datasets["train"].features
else:
column_names = raw_datasets["validation"].column_names
features = raw_datasets["validation"].features
if data_args.text_column_name is not None:
text_column_name = data_args.text_column_name
elif "tokens" in column_names:
text_column_name = "tokens"
else:
text_column_name = column_names[0]
if data_args.label_column_name is not None:
label_column_name = data_args.label_column_name
elif f"{data_args.task_name}_tags" in column_names:
label_column_name = f"{data_args.task_name}_tags"
else:
label_column_name = column_names[1]
# In the event the labels are not a `Sequence[ClassLabel]`, we will need to go through the dataset to get the
# unique labels.
def get_label_list(labels):
unique_labels = set()
for label in labels:
unique_labels = unique_labels | set(label)
label_list = list(unique_labels)
label_list.sort()
return label_list
if isinstance(features[label_column_name].feature, ClassLabel):
label_list = features[label_column_name].feature.names
# No need to convert the labels since they are already ints.
label_to_id = {i: i for i in range(len(label_list))}
else:
label_list = get_label_list(raw_datasets["train"][label_column_name])
label_to_id = {l: i for i, l in enumerate(label_list)}
num_labels = len(label_list)
# Load pretrained model and tokenizer
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
label2id=label_to_id,
id2label={i: l
for l, i in label_to_id.items()},
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_name_or_path = model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path
if config.model_type in {"gpt2", "roberta"}:
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
add_prefix_space=True,
)
else:
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = FlaxAutoModelForTokenClassification.from_pretrained(
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
# Tokenize all texts and align the labels with them.
def tokenize_and_align_labels(examples):
tokenized_inputs = tokenizer(
examples[text_column_name],
max_length=data_args.max_seq_length,
padding="max_length",
truncation=True,
# We use this argument because the texts in our dataset are lists of words (with a label for each word).
is_split_into_words=True,
)
labels = []
for i, label in enumerate(examples[label_column_name]):
word_ids = tokenized_inputs.word_ids(batch_index=i)
previous_word_idx = None
label_ids = []
for word_idx in word_ids:
# Special tokens have a word id that is None. We set the label to -100 so they are automatically
# ignored in the loss function.
if word_idx is None:
label_ids.append(-100)
# We set the label for the first token of each word.
elif word_idx != previous_word_idx:
label_ids.append(label_to_id[label[word_idx]])
# For the other tokens in a word, we set the label to either the current label or -100, depending on
# the label_all_tokens flag.
else:
label_ids.append(label_to_id[label[word_idx]] if data_args.
label_all_tokens else -100)
previous_word_idx = word_idx
labels.append(label_ids)
tokenized_inputs["labels"] = labels
return tokenized_inputs
processed_raw_datasets = raw_datasets.map(
tokenize_and_align_labels,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
remove_columns=raw_datasets["train"].column_names,
desc="Running tokenizer on dataset",
)
train_dataset = processed_raw_datasets["train"]
eval_dataset = processed_raw_datasets["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]}.")
# Define a summary writer
has_tensorboard = is_tensorboard_available()
if has_tensorboard and jax.process_index() == 0:
try:
from flax.metrics.tensorboard import SummaryWriter
summary_writer = SummaryWriter(training_args.output_dir)
summary_writer.hparams({
**training_args.to_dict(),
**vars(model_args),
**vars(data_args)
})
except ImportError as ie:
has_tensorboard = False
logger.warning(
f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
)
else:
logger.warning(
"Unable to display metrics through TensorBoard because the package is not installed: "
"Please run pip install tensorboard to enable.")
def write_train_metric(summary_writer, train_metrics, train_time, step):
summary_writer.scalar("train_time", train_time, step)
train_metrics = get_metrics(train_metrics)
for key, vals in train_metrics.items():
tag = f"train_{key}"
for i, val in enumerate(vals):
summary_writer.scalar(tag, val, step - len(vals) + i + 1)
def write_eval_metric(summary_writer, eval_metrics, step):
for metric_name, value in eval_metrics.items():
summary_writer.scalar(f"eval_{metric_name}", value, step)
num_epochs = int(training_args.num_train_epochs)
rng = jax.random.PRNGKey(training_args.seed)
dropout_rngs = jax.random.split(rng, jax.local_device_count())
train_batch_size = training_args.per_device_train_batch_size * jax.local_device_count(
)
eval_batch_size = training_args.per_device_eval_batch_size * jax.local_device_count(
)
learning_rate_fn = create_learning_rate_fn(
len(train_dataset),
train_batch_size,
training_args.num_train_epochs,
training_args.warmup_steps,
training_args.learning_rate,
)
state = create_train_state(model,
learning_rate_fn,
num_labels=num_labels,
training_args=training_args)
# define step functions
def train_step(
state: train_state.TrainState, batch: Dict[str, Array],
dropout_rng: PRNGKey) -> Tuple[train_state.TrainState, float]:
"""Trains model with an optimizer (both in `state`) on `batch`, returning a pair `(new_state, loss)`."""
dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
targets = batch.pop("labels")
def loss_fn(params):
logits = state.apply_fn(**batch,
params=params,
dropout_rng=dropout_rng,
train=True)[0]
loss = state.loss_fn(logits, targets)
return loss
grad_fn = jax.value_and_grad(loss_fn)
loss, grad = grad_fn(state.params)
grad = jax.lax.pmean(grad, "batch")
new_state = state.apply_gradients(grads=grad)
metrics = jax.lax.pmean(
{
"loss": loss,
"learning_rate": learning_rate_fn(state.step)
},
axis_name="batch")
return new_state, metrics, new_dropout_rng
p_train_step = jax.pmap(train_step,
axis_name="batch",
donate_argnums=(0, ))
def eval_step(state, batch):
logits = state.apply_fn(**batch, params=state.params, train=False)[0]
return state.logits_fn(logits)
p_eval_step = jax.pmap(eval_step, axis_name="batch")
metric = load_metric("seqeval")
def get_labels(y_pred, y_true):
# Transform predictions and references tensos to numpy arrays
# Remove ignored index (special tokens)
true_predictions = [[
label_list[p] for (p, l) in zip(pred, gold_label) if l != -100
] for pred, gold_label in zip(y_pred, y_true)]
true_labels = [[
label_list[l] for (p, l) in zip(pred, gold_label) if l != -100
] for pred, gold_label in zip(y_pred, y_true)]
return true_predictions, true_labels
def compute_metrics():
results = metric.compute()
if data_args.return_entity_level_metrics:
# Unpack nested dictionaries
final_results = {}
for key, value in results.items():
if isinstance(value, dict):
for n, v in value.items():
final_results[f"{key}_{n}"] = v
else:
final_results[key] = value
return final_results
else:
return {
"precision": results["overall_precision"],
"recall": results["overall_recall"],
"f1": results["overall_f1"],
"accuracy": results["overall_accuracy"],
}
logger.info(f"===== Starting training ({num_epochs} epochs) =====")
train_time = 0
# make sure weights are replicated on each device
state = replicate(state)
train_time = 0
step_per_epoch = len(train_dataset) // train_batch_size
total_steps = step_per_epoch * num_epochs
epochs = tqdm(range(num_epochs),
desc=f"Epoch ... (1/{num_epochs})",
position=0)
for epoch in epochs:
train_start = time.time()
train_metrics = []
# Create sampling rng
rng, input_rng = jax.random.split(rng)
# train
for step, batch in enumerate(
tqdm(
train_data_collator(input_rng, train_dataset,
train_batch_size),
total=step_per_epoch,
desc="Training...",
position=1,
)):
state, train_metric, dropout_rngs = p_train_step(
state, batch, dropout_rngs)
train_metrics.append(train_metric)
cur_step = (epoch * step_per_epoch) + (step + 1)
if cur_step % training_args.logging_steps == 0 and cur_step > 0:
# Save metrics
train_metric = unreplicate(train_metric)
train_time += time.time() - train_start
if has_tensorboard and jax.process_index() == 0:
write_train_metric(summary_writer, train_metrics,
train_time, cur_step)
epochs.write(
f"Step... ({cur_step}/{total_steps} | Training Loss: {train_metric['loss']}, Learning Rate:"
f" {train_metric['learning_rate']})")
train_metrics = []
if cur_step % training_args.eval_steps == 0 and cur_step > 0:
eval_metrics = {}
# evaluate
for batch in tqdm(
eval_data_collator(eval_dataset, eval_batch_size),
total=len(eval_dataset) // eval_batch_size,
desc="Evaluating ...",
position=2,
):
labels = batch.pop("labels")
predictions = p_eval_step(state, batch)
predictions = np.array(
[pred for pred in chain(*predictions)])
labels = np.array([label for label in chain(*labels)])
labels[np.array(chain(
*batch["attention_mask"])) == 0] = -100
preds, refs = get_labels(predictions, labels)
metric.add_batch(
predictions=preds,
references=refs,
)
# evaluate also on leftover examples (not divisible by batch_size)
num_leftover_samples = len(eval_dataset) % eval_batch_size
# make sure leftover batch is evaluated on one device
if num_leftover_samples > 0 and jax.process_index() == 0:
# take leftover samples
batch = eval_dataset[-num_leftover_samples:]
batch = {k: np.array(v) for k, v in batch.items()}
labels = batch.pop("labels")
predictions = eval_step(unreplicate(state), batch)
labels = np.array(labels)
labels[np.array(batch["attention_mask"]) == 0] = -100
preds, refs = get_labels(predictions, labels)
metric.add_batch(
predictions=preds,
references=refs,
)
eval_metrics = compute_metrics()
if data_args.return_entity_level_metrics:
logger.info(
f"Step... ({cur_step}/{total_steps} | Validation metrics: {eval_metrics}"
)
else:
logger.info(
f"Step... ({cur_step}/{total_steps} | Validation f1: {eval_metrics['f1']}, Validation Acc:"
f" {eval_metrics['accuracy']})")
if has_tensorboard and jax.process_index() == 0:
write_eval_metric(summary_writer, eval_metrics, cur_step)
if (cur_step % training_args.save_steps == 0
and cur_step > 0) or (cur_step == total_steps):
# save checkpoint after each epoch and push checkpoint to the hub
if jax.process_index() == 0:
params = jax.device_get(unreplicate(state.params))
model.save_pretrained(training_args.output_dir,
params=params)
tokenizer.save_pretrained(training_args.output_dir)
if training_args.push_to_hub:
repo.push_to_hub(
commit_message=
f"Saving weights and logs of step {cur_step}",
blocking=False)
epochs.desc = f"Epoch ... {epoch + 1}/{num_epochs}"
# Eval after training
if training_args.do_eval:
eval_metrics = {}
eval_loader = eval_data_collator(eval_dataset, eval_batch_size)
for batch in tqdm(eval_loader,
total=len(eval_dataset) // eval_batch_size,
desc="Evaluating ...",
position=2):
labels = batch.pop("labels")
predictions = p_eval_step(state, batch)
predictions = np.array([pred for pred in chain(*predictions)])
labels = np.array([label for label in chain(*labels)])
labels[np.array(chain(*batch["attention_mask"])) == 0] = -100
preds, refs = get_labels(predictions, labels)
metric.add_batch(predictions=preds, references=refs)
# evaluate also on leftover examples (not divisible by batch_size)
num_leftover_samples = len(eval_dataset) % eval_batch_size
# make sure leftover batch is evaluated on one device
if num_leftover_samples > 0 and jax.process_index() == 0:
# take leftover samples
batch = eval_dataset[-num_leftover_samples:]
batch = {k: np.array(v) for k, v in batch.items()}
labels = np.array(batch.pop("labels"))
predictions = eval_step(unreplicate(state), batch)
labels[np.array(batch["attention_mask"]) == 0] = -100
preds, refs = get_labels(predictions, labels)
metric.add_batch(predictions=preds, references=refs)
eval_metrics = compute_metrics()
if jax.process_index() == 0:
eval_metrics = {
f"eval_{metric_name}": value
for metric_name, value in eval_metrics.items()
}
path = os.path.join(training_args.output_dir, "eval_results.json")
with open(path, "w") as f:
json.dump(eval_metrics, f, indent=4, sort_keys=True)
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 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."
)
# 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/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 data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir)
if "validation" not in datasets.keys():
datasets["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[:{data_args.validation_split_percentage}%]",
cache_dir=model_args.cache_dir,
)
datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[{data_args.validation_split_percentage}%:]",
cache_dir=model_args.cache_dir,
)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = (
data_args.train_file.split(".")[-1]
if data_args.train_file is not None
else data_args.validation_file.split(".")[-1]
)
if extension == "txt":
extension = "text"
datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
# 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
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config_kwargs = {
"cache_dir": model_args.cache_dir,
"revision": model_args.model_revision,
"use_auth_token": True if model_args.use_auth_token else None,
}
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name, **config_kwargs)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning("You are instantiating a new config instance from scratch.")
tokenizer_kwargs = {
"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,
}
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, **tokenizer_kwargs)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path, **tokenizer_kwargs)
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 model_args.model_name_or_path:
model = AutoModelForCausalLM.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,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForCausalLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
# since this will be pickled to avoid _LazyModule error in Hasher force logger loading before tokenize_function
tok_logger = transformers.utils.logging.get_logger("transformers.tokenization_utils_base")
def tokenize_function(examples):
with CaptureLogger(tok_logger) as cl:
tokenizer.pad_token = tokenizer.eos_token
output = tokenizer(examples[text_column_name],
padding=True,
max_length=data_args.block_size,
truncation=True)
# clm input could be much much longer than block_size
if "Token indices sequence length is longer than the" in cl.out:
tok_logger.warning(
"^^^^^^^^^^^^^^^^ Please ignore the warning above - this long input will be chunked into smaller bits before being passed to the model."
)
return output
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if data_args.block_size is None:
block_size = tokenizer.model_max_length
if block_size > 1024:
logger.warning(
f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
"Picking 1024 instead. You can change that default value by passing --block_size xxx."
)
block_size = 1024
else:
if data_args.block_size > tokenizer.model_max_length:
logger.warning(
f"The block_size passed ({data_args.block_size}) is larger than the maximum length for the model"
f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}."
)
block_size = min(data_args.block_size, tokenizer.model_max_length)
# Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
def group_texts(examples):
# # Concatenate all texts.
# concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
result = examples
# # print(concatenated_examples)
# total_length = len(concatenated_examples[list(examples.keys())[0]])
# # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
# # customize this part to your needs.
# total_length = (total_length // block_size) * block_size
# # Split by chunks of max_len.
# result = {
# k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
# for k, t in concatenated_examples.items()
# }
result["labels"] = result["input_ids"].copy()
return result
# Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder
# for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value might be slower
# to preprocess.
#
# To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
lm_datasets = tokenized_datasets.map(
group_texts,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_train:
if "train" not in tokenized_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = lm_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 tokenized_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = lm_datasets["validation"]
if data_args.max_eval_samples is not None:
eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))
# 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,
tokenizer=tokenizer,
# Data collator will default to DataCollatorWithPadding, so we change it.
data_collator=default_data_collator,
)
# Training
if training_args.do_train:
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
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
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.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(eval_dataset)
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
perplexity = math.exp(metrics["eval_loss"])
metrics["perplexity"] = perplexity
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
if training_args.push_to_hub:
kwargs = {"finetuned_from": model_args.model_name_or_path, "tags": "text-generation"}
if data_args.dataset_name is not None:
kwargs["dataset_tags"] = data_args.dataset_name
if data_args.dataset_config_name is not None:
kwargs["dataset_args"] = data_args.dataset_config_name
kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
else:
kwargs["dataset"] = data_args.dataset_name
trainer.push_to_hub(**kwargs)
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()
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty."
"Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO 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("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_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.
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.train_file.split(".")[-1]
if extension == "txt":
extension = "text"
datasets = load_dataset(extension, data_files=data_files)
# 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
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning("You are instantiating a new config instance from scratch.")
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
)
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 model_args.model_name_or_path:
model = AutoModelForMaskedLM.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,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForMaskedLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
padding = "max_length" if data_args.pad_to_max_length else False
def tokenize_function(examples):
# Remove empty lines
examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
return tokenizer(examples["text"], padding=padding, truncation=True, max_length=data_args.max_seq_length)
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=[text_column_name],
load_from_cache_file=not data_args.overwrite_cache,
)
# Add the chinese references if provided
if data_args.train_ref_file is not None:
tokenized_datasets["train"] = add_chinese_references(tokenized_datasets["train"], data_args.train_ref_file)
if data_args.validation_ref_file is not None:
tokenized_datasets["validation"] = add_chinese_references(
tokenized_datasets["validation"], data_args.validation_ref_file
)
# Data collator
# This one will take care of randomly masking the tokens.
data_collator = DataCollatorForWholeWordMask(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_datasets["train"] if training_args.do_train else None,
eval_dataset=tokenized_datasets["validation"] if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
)
# Training
if training_args.do_train:
model_path = (
model_args.model_name_or_path
if (model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path))
else None
)
trainer.train(model_path=model_path)
trainer.save_model() # Saves the tokenizer too for easy upload
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["eval_loss"])
results["perplexity"] = perplexity
output_eval_file = os.path.join(training_args.output_dir, "eval_results_mlm_wwm.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in results.items():
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, 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()
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.local_rank != -1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Prepare Question-Answering task
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
model = AutoModelForQuestionAnswering.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,
)
# Get datasets
is_language_sensitive = hasattr(model.config, "lang2id")
train_dataset = (
SquadDataset(
data_args, tokenizer=tokenizer, is_language_sensitive=is_language_sensitive, cache_dir=model_args.cache_dir
)
if training_args.do_train
else None
)
eval_dataset = (
SquadDataset(
data_args,
tokenizer=tokenizer,
mode="dev",
is_language_sensitive=is_language_sensitive,
cache_dir=model_args.cache_dir,
)
if training_args.do_eval
else None
)
# Initialize our Trainer
trainer = Trainer(model=model, args=training_args, train_dataset=train_dataset, eval_dataset=eval_dataset,)
# Training
if training_args.do_train:
trainer.train(
model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
)
trainer.save_model()
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
if trainer.is_world_master():
tokenizer.save_pretrained(training_args.output_dir)
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, Seq2SeqTrainingArguments))
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(
)
# 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)],
)
log_level = training_args.get_process_log_level()
logger.setLevel(log_level)
datasets.utils.logging.set_verbosity(log_level)
transformers.utils.logging.set_verbosity(log_level)
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# 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}"
)
logger.info(f"Training/evaluation parameters {training_args}")
if data_args.source_prefix is None and model_args.model_name_or_path in [
"t5-small",
"t5-base",
"t5-large",
"t5-3b",
"t5-11b",
]:
logger.warning(
"You're running a t5 model but didn't provide a source prefix, which is the expected, e.g. with "
"`--source_prefix 'summarize: ' `")
# 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 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 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 first column for the full texts and the second column for the
# summaries (unless you specify column names for this with the `text_column` and `summary_column` arguments).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name,
cache_dir=model_args.cache_dir)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
raw_datasets = load_dataset(extension,
data_files=data_files,
cache_dir=model_args.cache_dir)
# 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
#
# 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,
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 = AutoModelForSeq2SeqLM.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,
)
model.resize_token_embeddings(len(tokenizer))
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
if (hasattr(model.config, "max_position_embeddings")
and model.config.max_position_embeddings <
data_args.max_source_length):
if model_args.resize_position_embeddings is None:
logger.warning(
f"Increasing the model's number of position embedding vectors from {model.config.max_position_embeddings} "
f"to {data_args.max_source_length}.")
model.resize_position_embeddings(data_args.max_source_length)
elif model_args.resize_position_embeddings:
model.resize_position_embeddings(data_args.max_source_length)
else:
raise ValueError(
f"`--max_source_length` is set to {data_args.max_source_length}, but the model only has {model.config.max_position_embeddings}"
f" position encodings. Consider either reducing `--max_source_length` to {model.config.max_position_embeddings} or to automatically "
"resize the model's position encodings by passing `--resize_position_embeddings`."
)
prefix = data_args.source_prefix if data_args.source_prefix is not None else ""
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
if training_args.do_train:
column_names = raw_datasets["train"].column_names
elif training_args.do_eval:
column_names = raw_datasets["validation"].column_names
elif training_args.do_predict:
column_names = raw_datasets["test"].column_names
else:
logger.info(
"There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`."
)
return
# Get the column names for input/target.
dataset_columns = summarization_name_mapping.get(data_args.dataset_name,
None)
if data_args.text_column is None:
text_column = dataset_columns[
0] if dataset_columns is not None else column_names[0]
else:
text_column = data_args.text_column
if text_column not in column_names:
raise ValueError(
f"--text_column' value '{data_args.text_column}' needs to be one of: {', '.join(column_names)}"
)
if data_args.summary_column is None:
summary_column = dataset_columns[
1] if dataset_columns is not None else column_names[1]
else:
summary_column = data_args.summary_column
if summary_column not in column_names:
raise ValueError(
f"--summary_column' value '{data_args.summary_column}' needs to be one of: {', '.join(column_names)}"
)
# Temporarily set max_target_length for training.
max_target_length = data_args.max_target_length
padding = "max_length" if data_args.pad_to_max_length else False
if training_args.label_smoothing_factor > 0 and not hasattr(
model, "prepare_decoder_input_ids_from_labels"):
logger.warning(
"label_smoothing is enabled but the `prepare_decoder_input_ids_from_labels` method is not defined for"
f"`{model.__class__.__name__}`. This will lead to loss being calculated twice and will take up more memory"
)
def preprocess_function(examples):
inputs = examples[text_column]
targets = examples[summary_column]
inputs = [prefix + inp for inp in inputs]
model_inputs = tokenizer(inputs,
max_length=data_args.max_source_length,
padding=padding,
truncation=True)
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets,
max_length=max_target_length,
padding=padding,
truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [[
(l if l != tokenizer.pad_token_id else -100) for l in label
] for label in labels["input_ids"]]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
if training_args.do_train:
if "train" not in raw_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = raw_datasets["train"]
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
with training_args.main_process_first(
desc="train dataset map pre-processing"):
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on train dataset",
)
if training_args.do_eval:
max_target_length = data_args.val_max_target_length
if "validation" not in raw_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = raw_datasets["validation"]
if data_args.max_eval_samples is not None:
eval_dataset = eval_dataset.select(
range(data_args.max_eval_samples))
with training_args.main_process_first(
desc="validation dataset map pre-processing"):
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on validation dataset",
)
if training_args.do_predict:
max_target_length = data_args.val_max_target_length
if "test" not in raw_datasets:
raise ValueError("--do_predict requires a test dataset")
predict_dataset = raw_datasets["test"]
if data_args.max_predict_samples is not None:
predict_dataset = predict_dataset.select(
range(data_args.max_predict_samples))
with training_args.main_process_first(
desc="prediction dataset map pre-processing"):
predict_dataset = predict_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on prediction dataset",
)
# Data collator
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if training_args.fp16 else None,
)
# Metric
metric = load_metric("rouge")
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [label.strip() for label in labels]
# rougeLSum expects newline after each sentence
preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds]
labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels]
return preds, labels
def compute_metrics(eval_preds):
preds, labels = eval_preds
if isinstance(preds, tuple):
preds = preds[0]
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels,
skip_special_tokens=True)
# Some simple post-processing
decoded_preds, decoded_labels = postprocess_text(
decoded_preds, decoded_labels)
result = metric.compute(predictions=decoded_preds,
references=decoded_labels,
use_stemmer=True)
# Extract a few results from ROUGE
result = {
key: value.mid.fmeasure * 100
for key, value in result.items()
}
prediction_lens = [
np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds
]
result["gen_len"] = np.mean(prediction_lens)
result = {k: round(v, 4) for k, v in result.items()}
return result
# Initialize our Trainer
trainer = Seq2SeqTrainer(
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,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics
if training_args.predict_with_generate else None,
)
# Training
if training_args.do_train:
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
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
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.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
max_length = (training_args.generation_max_length
if training_args.generation_max_length is not None else
data_args.val_max_target_length)
num_beams = data_args.num_beams if data_args.num_beams is not None else training_args.generation_num_beams
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(max_length=max_length,
num_beams=num_beams,
metric_key_prefix="eval")
max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(
eval_dataset)
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
if training_args.do_predict:
logger.info("*** Predict ***")
predict_results = trainer.predict(predict_dataset,
metric_key_prefix="predict",
max_length=max_length,
num_beams=num_beams)
metrics = predict_results.metrics
max_predict_samples = (data_args.max_predict_samples
if data_args.max_predict_samples is not None
else len(predict_dataset))
metrics["predict_samples"] = min(max_predict_samples,
len(predict_dataset))
trainer.log_metrics("predict", metrics)
trainer.save_metrics("predict", metrics)
if trainer.is_world_process_zero():
if training_args.predict_with_generate:
predictions = tokenizer.batch_decode(
predict_results.predictions,
skip_special_tokens=True,
clean_up_tokenization_spaces=True)
predictions = [pred.strip() for pred in predictions]
output_prediction_file = os.path.join(
training_args.output_dir, "generated_predictions.txt")
with open(output_prediction_file, "w") as writer:
writer.write("\n".join(predictions))
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": "summarization"
}
if data_args.dataset_name is not None:
kwargs["dataset_tags"] = data_args.dataset_name
if data_args.dataset_config_name is not None:
kwargs["dataset_args"] = data_args.dataset_config_name
kwargs[
"dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
else:
kwargs["dataset"] = data_args.dataset_name
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, 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()
# 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)],
)
log_level = training_args.get_process_log_level()
logger.setLevel(log_level)
datasets.utils.logging.set_verbosity(log_level)
transformers.utils.logging.set_verbosity(log_level)
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# 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}"
)
logger.info(f"Training/evaluation parameters {training_args}")
# 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 before initializing model.
set_seed(training_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 data_args.train_file is not None or data_args.validation_file is not None:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.train_file.split(".")[-1]
raw_datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
else:
# Downloading and loading the swag dataset from the hub.
raw_datasets = load_dataset("swag", "regular", cache_dir=model_args.cache_dir)
# 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
# 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,
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 = AutoModelForMultipleChoice.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,
)
# When using your own dataset or a different dataset from swag, you will probably need to change this.
ending_names = [f"ending{i}" for i in range(4)]
context_name = "sent1"
question_header_name = "sent2"
if data_args.max_seq_length is None:
max_seq_length = tokenizer.model_max_length
if max_seq_length > 1024:
logger.warning(
f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
"Picking 1024 instead. You can change that default value by passing --max_seq_length xxx."
)
max_seq_length = 1024
else:
if data_args.max_seq_length > tokenizer.model_max_length:
logger.warning(
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)
# Preprocessing the datasets.
def preprocess_function(examples):
first_sentences = [[context] * 4 for context in examples[context_name]]
question_headers = examples[question_header_name]
second_sentences = [
[f"{header} {examples[end][i]}" for end in ending_names] for i, header in enumerate(question_headers)
]
# Flatten out
first_sentences = sum(first_sentences, [])
second_sentences = sum(second_sentences, [])
# Tokenize
tokenized_examples = tokenizer(
first_sentences,
second_sentences,
truncation=True,
max_length=max_seq_length,
padding="max_length" if data_args.pad_to_max_length else False,
)
# Un-flatten
return {k: [v[i : i + 4] for i in range(0, len(v), 4)] for k, v in tokenized_examples.items()}
if training_args.do_train:
if "train" not in raw_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = raw_datasets["train"]
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(range(data_args.max_train_samples))
with training_args.main_process_first(desc="train dataset map pre-processing"):
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
if "validation" not in raw_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = raw_datasets["validation"]
if data_args.max_eval_samples is not None:
eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))
with training_args.main_process_first(desc="validation dataset map pre-processing"):
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
data_collator = (
default_data_collator
if data_args.pad_to_max_length
else DataCollatorForMultipleChoice(tokenizer=tokenizer, pad_to_multiple_of=8 if training_args.fp16 else None)
)
# Metric
def compute_metrics(eval_predictions):
predictions, label_ids = eval_predictions
preds = np.argmax(predictions, axis=1)
return {"accuracy": (preds == label_ids).astype(np.float32).mean().item()}
# 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,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
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
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
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.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(eval_dataset)
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
if training_args.push_to_hub:
trainer.push_to_hub(
finetuned_from=model_args.model_name_or_path,
tasks="multiple-choice",
dataset_tags="swag",
dataset_args="regular",
dataset="SWAG",
language="en",
)
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, Seq2SeqTrainingArguments))
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 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."
)
# 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()
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 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 JSON files, this script will use the `question` column for the input question and `table` column for the corresponding table.
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name,
cache_dir=model_args.cache_dir)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
datasets = load_dataset(extension,
data_files=data_files,
cache_dir=model_args.cache_dir)
# 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
#
# 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,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# IMPORTANT: the initial BART model's decoding is penalized by no_repeat_ngram_size, and thus
# we should disable it here to avoid problematic generation
config.no_repeat_ngram_size = 0
config.max_length = 1024
config.early_stopping = False
# load tapex tokenizer
tokenizer = TapexTokenizer.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,
add_prefix_space=True,
)
# load Bart based Tapex model (default tapex-large)
model = BartForConditionalGeneration.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,
)
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
if training_args.do_train:
column_names = datasets["train"].column_names
elif training_args.do_eval:
column_names = datasets["validation"].column_names
elif training_args.do_predict:
column_names = datasets["test"].column_names
else:
logger.info(
"There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`."
)
return
# Temporarily set max_target_length for training.
max_target_length = data_args.max_target_length
padding = "max_length" if data_args.pad_to_max_length else False
if training_args.label_smoothing_factor > 0 and not hasattr(
model, "prepare_decoder_input_ids_from_labels"):
logger.warning(
"label_smoothing is enabled but the `prepare_decoder_input_ids_from_labels` method is not defined for"
f"`{model.__class__.__name__}`. This will lead to loss being calculated twice and will take up more memory"
)
def preprocess_tableqa_function(examples, is_training=False):
"""
The is_training FLAG is used to identify if we could use the supervision
to truncate the table content if it is required.
"""
questions = [question.lower() for question in examples["question"]]
example_tables = examples["table"]
tables = [
pd.DataFrame.from_records(example_table["rows"],
columns=example_table["header"])
for example_table in example_tables
]
# using wikitablequestion's answer set
answers = examples["answers"]
# IMPORTANT: we cannot pass by answers during evaluation, answers passed during training are used to
# truncate large tables in the train set!
if is_training:
model_inputs = tokenizer(
table=tables,
query=questions,
answer=answers,
max_length=data_args.max_source_length,
padding=padding,
truncation=True,
)
else:
model_inputs = tokenizer(table=tables,
query=questions,
max_length=data_args.max_source_length,
padding=padding,
truncation=True)
with tokenizer.as_target_tokenizer():
labels = tokenizer(
answer=[", ".join(answer) for answer in answers],
max_length=max_target_length,
padding=padding,
truncation=True,
)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [[
(l if l != tokenizer.pad_token_id else -100) for l in label
] for label in labels["input_ids"]]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
# in training, we can use the answer as extra information to truncate large tables
preprocess_tableqa_function_training = partial(preprocess_tableqa_function,
is_training=True)
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))
train_dataset = train_dataset.map(
preprocess_tableqa_function_training,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
max_target_length = data_args.val_max_target_length
if "validation" not in datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = datasets["validation"]
if data_args.max_eval_samples is not None:
eval_dataset = eval_dataset.select(
range(data_args.max_eval_samples))
eval_dataset = eval_dataset.map(
preprocess_tableqa_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_predict:
max_target_length = data_args.val_max_target_length
if "test" not in datasets:
raise ValueError("--do_predict requires a test dataset")
predict_dataset = datasets["test"]
if data_args.max_predict_samples is not None:
predict_dataset = predict_dataset.select(
range(data_args.max_predict_samples))
predict_dataset = predict_dataset.map(
preprocess_tableqa_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if training_args.fp16 else None,
)
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [label.strip() for label in labels]
return preds, labels
def compute_metrics(eval_preds):
preds, labels = eval_preds
if isinstance(preds, tuple):
preds = preds[0]
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels,
skip_special_tokens=True)
# Some simple post-processing
decoded_preds, decoded_labels = postprocess_text(
decoded_preds, decoded_labels)
delimiter = ", "
# define example evaluation
def evaluate_example(predict_str: str, ground_str: str):
predict_spans = predict_str.split(delimiter)
ground_spans = ground_str.split(delimiter)
predict_values = defaultdict(lambda: 0)
ground_values = defaultdict(lambda: 0)
for span in predict_spans:
try:
predict_values[float(span)] += 1
except ValueError:
predict_values[span.strip()] += 1
for span in ground_spans:
try:
ground_values[float(span)] += 1
except ValueError:
ground_values[span.strip()] += 1
_is_correct = predict_values == ground_values
return _is_correct
def get_denotation_accuracy(predictions: List[str],
references: List[str]):
assert len(predictions) == len(references)
correct_num = 0
for predict_str, ground_str in zip(predictions, references):
is_correct = evaluate_example(predict_str.lower(),
ground_str.lower())
if is_correct:
correct_num += 1
return correct_num / len(predictions)
accuracy = get_denotation_accuracy(decoded_preds, decoded_labels)
result = {"denotation_accuracy": accuracy}
return result
# Initialize our Trainer
trainer = Seq2SeqTrainer(
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,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics
if training_args.predict_with_generate else None,
)
if training_args.do_train:
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
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
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.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(max_length=data_args.val_max_target_length,
num_beams=data_args.num_beams,
metric_key_prefix="eval")
max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(
eval_dataset)
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
if training_args.do_predict:
logger.info("*** Predict ***")
predict_results = trainer.predict(
predict_dataset,
metric_key_prefix="predict",
max_length=data_args.val_max_target_length,
num_beams=data_args.num_beams,
)
metrics = predict_results.metrics
max_predict_samples = (data_args.max_predict_samples
if data_args.max_predict_samples is not None
else len(predict_dataset))
metrics["predict_samples"] = min(max_predict_samples,
len(predict_dataset))
trainer.log_metrics("predict", metrics)
trainer.save_metrics("predict", metrics)
if trainer.is_world_process_zero():
if training_args.predict_with_generate:
predictions = tokenizer.batch_decode(
predict_results.predictions,
skip_special_tokens=True,
clean_up_tokenization_spaces=True)
predictions = [pred.strip() for pred in predictions]
output_prediction_file = os.path.join(training_args.output_dir,
"tapex_predictions.txt")
with open(output_prediction_file, "w") as writer:
writer.write("\n".join(predictions))
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, 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()
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty."
"Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO 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("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_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 data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.train_file.split(".")[-1]
if extension == "txt":
extension = "text"
datasets = load_dataset(extension, data_files=data_files)
# 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
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning("You are instantiating a new config instance from scratch.")
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
)
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 model_args.model_name_or_path:
model = AutoModelForCausalLM.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,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForCausalLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
def tokenize_function(examples):
return tokenizer(examples[text_column_name])
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if data_args.block_size <= 0:
block_size = tokenizer.model_max_length
else:
if data_args.block_size > tokenizer.model_max_length:
logger.warn(
f"The block_size passed ({data_args.block_size}) is larger than the maximum length for the model"
f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}."
)
block_size = min(data_args.block_size, tokenizer.model_max_length)
# Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
def group_texts(examples):
# Concatenate all texts.
concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
total_length = len(concatenated_examples[list(examples.keys())[0]])
# We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
# customize this part to your needs.
total_length = (total_length // block_size) * block_size
# Split by chunks of max_len.
result = {
k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
for k, t in concatenated_examples.items()
}
result["labels"] = result["input_ids"].copy()
return result
# Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder
# for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value might be slower
# to preprocess.
#
# To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
lm_datasets = tokenized_datasets.map(
group_texts,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=lm_datasets["train"] if training_args.do_train else None,
eval_dataset=lm_datasets["validation"] if training_args.do_eval else None,
tokenizer=tokenizer,
# Data collator will default to DataCollatorWithPadding, so we change it.
data_collator=default_data_collator,
)
# Training
if training_args.do_train:
trainer.train(
model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
)
trainer.save_model() # Saves the tokenizer too for easy upload
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["eval_loss"])
results["perplexity"] = perplexity
output_eval_file = os.path.join(training_args.output_dir, "eval_results_clm.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in results.items():
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
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(
)
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir) and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty."
"Use --overwrite_output_dir to overcome.")
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO
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()
logger.info("Training/evaluation parameters %s", 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 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 in the summarization task, this script will use the first column for the full texts and the
# second column for the summaries (unless you specify column names for this with the `text_column` and
# `summary_column` arguments).
# For translation, only JSON files are supported, with one field named "translation" containing two keys for the
# source and target languages (unless you adapt what follows).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
datasets = load_dataset(extension, data_files=data_files)
# 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
#
# 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,
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 = AutoModelForSeq2SeqLM.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,
)
# Set decoder_start_token_id
if model.config.decoder_start_token_id is None and isinstance(
tokenizer, MBartTokenizer):
model.config.decoder_start_token_id = tokenizer.lang_code_to_id[
data_args.target_lang]
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
# For translation we set the codes of our source and target languages (only useful for mBART, the others will
# ignore those attributes).
if data_args.task.startswith("translation"):
if data_args.source_lang is not None:
tokenizer.src_lang = data_args.source_lang
if data_args.target_lang is not None:
tokenizer.tgt_lang = data_args.target_lang
# To serialize preprocess_function below, each of those four variables needs to be defined (even if we won't use
# them all).
source_lang, target_lang, text_column, summary_column = None, None, None, None
if data_args.task.startswith("summarization"):
# Get the column names for input/target.
dataset_columns = summarization_name_mapping.get(
data_args.dataset_name, None)
if data_args.text_column is None:
text_column = dataset_columns[
0] if dataset_columns is not None else column_names[0]
else:
text_column = data_args.text_column
if data_args.summary_column is None:
summary_column = dataset_columns[
1] if dataset_columns is not None else column_names[1]
else:
summary_column = data_args.summary_column
else:
# Get the language codes for input/target.
lang_search = re.match("translation_([a-z]+)_to_([a-z]+)",
data_args.task)
if data_args.source_lang is not None:
source_lang = data_args.source_lang.split("_")[0]
else:
assert (
lang_search is not None
), "Provide a source language via --source_lang or rename your task 'translation_xx_to_yy'."
source_lang = lang_search.groups()[0]
if data_args.target_lang is not None:
target_lang = data_args.target_lang.split("_")[0]
else:
assert (
lang_search is not None
), "Provide a target language via --target_lang or rename your task 'translation_xx_to_yy'."
target_lang = lang_search.groups()[1]
# Temporarily set max_target_length for training.
max_target_length = data_args.max_target_length
padding = "max_length" if data_args.pad_to_max_length else False
def preprocess_function(examples):
if data_args.task.startswith("translation"):
inputs = [ex[source_lang] for ex in examples["translation"]]
targets = [ex[target_lang] for ex in examples["translation"]]
else:
inputs = examples[text_column]
targets = examples[summary_column]
model_inputs = tokenizer(inputs,
max_length=data_args.max_source_length,
padding=padding,
truncation=True)
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets,
max_length=max_target_length,
padding=padding,
truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [[
(l if l != tokenizer.pad_token_id else -100) for l in label
] for label in labels["input_ids"]]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
if training_args.do_train:
train_dataset = datasets["train"]
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
max_target_length = data_args.val_max_target_length
eval_dataset = datasets["validation"]
if data_args.max_val_samples is not None:
eval_dataset = eval_dataset.select(range(
data_args.max_val_samples))
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
if data_args.pad_to_max_length:
data_collator = default_data_collator
else:
data_collator = DataCollatorForSeq2Seq(
tokenizer, label_pad_token_id=label_pad_token_id)
# Metric
metric_name = "rouge" if data_args.task.startswith(
"summarization") else "sacrebleu"
metric = load_metric(metric_name)
def compute_metrics(eval_preds):
preds, labels = eval_preds
if isinstance(preds, tuple):
preds = preds[0]
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels,
skip_special_tokens=True)
# Some simple post-processing
decoded_preds = [pred.strip() for pred in decoded_preds]
decoded_labels = [label.strip() for label in decoded_labels]
if metric_name == "sacrebleu":
decoded_labels = [[label] for label in decoded_labels]
result = metric.compute(predictions=decoded_preds,
references=decoded_labels)
# Extract a few results from ROUGE
if metric_name == "rouge":
result = {
key: value.mid.fmeasure * 100
for key, value in result.items()
}
else:
result = {"bleu": result["score"]}
prediction_lens = [
np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds
]
result["gen_len"] = np.mean(prediction_lens)
return result
# Initialize our Trainer
trainer = Seq2SeqTrainer(
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,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics
if training_args.predict_with_generate else None,
)
# Training
if training_args.do_train:
train_result = trainer.train(
model_path=model_args.model_name_or_path if os.path.
isdir(model_args.model_name_or_path) else None)
trainer.save_model() # Saves the tokenizer too for easy upload
output_train_file = os.path.join(training_args.output_dir,
"train_results.txt")
if trainer.is_world_process_zero():
with open(output_train_file, "w") as writer:
logger.info("***** Train results *****")
for key, value in sorted(train_result.metrics.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
# Need to save the state, since Trainer.save_model saves only the tokenizer with the model
trainer.state.save_to_json(
os.path.join(training_args.output_dir, "trainer_state.json"))
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
results = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir,
"eval_results_seq2seq.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in sorted(results.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
return results
def train(cls, args_json_filename):
parser = HfArgumentParser((ClassArguments, DatasetArguments,
DataTrainingArguments, TrainingArguments))
class_args, dataset_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(args_json_filename))
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info("Start training")
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.local_rank != -1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Set seed
set_seed(training_args.seed)
summarization_trainer = cls(**dataclasses.asdict(class_args))
logger.info("Load and process dataset")
summarization_trainer.load_and_process_data(
**dataclasses.asdict(dataset_args))
summarization_trainer.cache_dataset(data_args.train_file_path,
data_args.valid_file_path)
train_dataset = torch.load(data_args.train_file_path)
valid_dataset = torch.load(data_args.valid_file_path)
logger.info("Dataset ready")
print(type(cls.data_collator.collate_batch))
# Initialize our Trainer
trainer = Trainer(
model=summarization_trainer.model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=valid_dataset,
data_collator=cls.data_collator.collate_batch,
prediction_loss_only=True,
)
# Training
if training_args.do_train:
trainer.train(model_path=class_args.model_name_or_path if os.path.
isdir(class_args.model_name_or_path) else None)
trainer.save_model()
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
if trainer.is_world_master():
summarization_trainer.tokenizer.save_pretrained(
training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval and training_args.local_rank in [-1, 0]:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir,
"eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(eval_output.keys()):
logger.info(" %s = %s", key, str(eval_output[key]))
writer.write("%s = %s\n" % (key, str(eval_output[key])))
results.update(eval_output)
return None
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()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# 1. First, let's load the dataset
raw_datasets = DatasetDict()
if training_args.do_train:
raw_datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=data_args.train_split_name,
use_auth_token=data_args.use_auth_token,
)
if data_args.audio_column_name not in raw_datasets["train"].column_names:
raise ValueError(
f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--audio_column_name` to the correct audio column - one of "
f"{', '.join(raw_datasets['train'].column_names)}."
)
if data_args.text_column_name not in raw_datasets["train"].column_names:
raise ValueError(
f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--text_column_name` to the correct text column - one of "
f"{', '.join(raw_datasets['train'].column_names)}."
)
if data_args.max_train_samples is not None:
raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
if training_args.do_eval:
raw_datasets["eval"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=data_args.eval_split_name,
use_auth_token=data_args.use_auth_token,
)
if data_args.max_eval_samples is not None:
raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
# 2. We remove some special characters from the datasets
# that make training complicated and do not help in transcribing the speech
# E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
# that could be easily picked up by the model
chars_to_ignore_regex = (
f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
)
text_column_name = data_args.text_column_name
def remove_special_characters(batch):
if chars_to_ignore_regex is not None:
batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
else:
batch["target_text"] = batch[text_column_name].lower() + " "
return batch
with training_args.main_process_first(desc="dataset map special characters removal"):
raw_datasets = raw_datasets.map(
remove_special_characters,
remove_columns=[text_column_name],
desc="remove special characters from datasets",
)
# save special tokens for tokenizer
word_delimiter_token = data_args.word_delimiter_token
unk_token = data_args.unk_token
pad_token = data_args.pad_token
# 3. Next, let's load the config as we might need it to create
# the tokenizer
# load config
config = AutoConfig.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
)
# 4. Next, if no tokenizer file is defined,
# we create the vocabulary of the model by extracting all unique characters from
# the training and evaluation datasets
# We need to make sure that only first rank saves vocabulary
# make sure all processes wait until vocab is created
tokenizer_name_or_path = model_args.tokenizer_name_or_path
tokenizer_kwargs = {}
if tokenizer_name_or_path is None:
# save vocab in training output dir
tokenizer_name_or_path = training_args.output_dir
vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
with training_args.main_process_first():
if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
os.remove(vocab_file)
with training_args.main_process_first(desc="dataset map vocabulary creation"):
if not os.path.isfile(vocab_file):
os.makedirs(tokenizer_name_or_path, exist_ok=True)
vocab_dict = create_vocabulary_from_data(
raw_datasets,
word_delimiter_token=word_delimiter_token,
unk_token=unk_token,
pad_token=pad_token,
)
# save vocab dict to be loaded into tokenizer
with open(vocab_file, "w") as file:
json.dump(vocab_dict, file)
# if tokenizer has just been created
# it is defined by `tokenizer_class` if present in config else by `model_type`
tokenizer_kwargs = {
"config": config if config.tokenizer_class is not None else None,
"tokenizer_type": config.model_type if config.tokenizer_class is None else None,
"unk_token": unk_token,
"pad_token": pad_token,
"word_delimiter_token": word_delimiter_token,
}
# 5. Now we can instantiate the feature extractor, tokenizer and model
# Note for distributed training, the .from_pretrained methods guarantee that only
# one local process can concurrently download model & vocab.
# load feature_extractor and tokenizer
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name_or_path,
use_auth_token=data_args.use_auth_token,
**tokenizer_kwargs,
)
feature_extractor = AutoFeatureExtractor.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
)
# adapt config
config.update(
{
"feat_proj_dropout": model_args.feat_proj_dropout,
"attention_dropout": model_args.attention_dropout,
"hidden_dropout": model_args.hidden_dropout,
"final_dropout": model_args.final_dropout,
"mask_time_prob": model_args.mask_time_prob,
"mask_time_length": model_args.mask_time_length,
"mask_feature_prob": model_args.mask_feature_prob,
"mask_feature_length": model_args.mask_feature_length,
"gradient_checkpointing": training_args.gradient_checkpointing,
"layerdrop": model_args.layerdrop,
"ctc_loss_reduction": model_args.ctc_loss_reduction,
"pad_token_id": tokenizer.pad_token_id,
"vocab_size": len(tokenizer),
"activation_dropout": model_args.activation_dropout,
}
)
# create model
model = AutoModelForCTC.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
config=config,
use_auth_token=data_args.use_auth_token,
)
# freeze encoder
if model_args.freeze_feature_encoder:
model.freeze_feature_encoder()
# 6. Now we preprocess the datasets including loading the audio, resampling and normalization
# Thankfully, `datasets` takes care of automatically loading and resampling the audio,
# so that we just need to set the correct target sampling rate and normalize the input
# via the `feature_extractor`
# make sure that dataset decodes audio with correct sampling rate
dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
if dataset_sampling_rate != feature_extractor.sampling_rate:
raw_datasets = raw_datasets.cast_column(
data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
)
# derive max & min input length for sample rate & max duration
max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
audio_column_name = data_args.audio_column_name
num_workers = data_args.preprocessing_num_workers
# `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
phoneme_language = data_args.phoneme_language
# Preprocessing the datasets.
# We need to read the audio files as arrays and tokenize the targets.
def prepare_dataset(batch):
# load audio
sample = batch[audio_column_name]
inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
batch["input_values"] = inputs.input_values[0]
batch["input_length"] = len(batch["input_values"])
# encode targets
additional_kwargs = {}
if phoneme_language is not None:
additional_kwargs["phonemizer_lang"] = phoneme_language
batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
return batch
with training_args.main_process_first(desc="dataset map preprocessing"):
vectorized_datasets = raw_datasets.map(
prepare_dataset,
remove_columns=next(iter(raw_datasets.values())).column_names,
num_proc=num_workers,
desc="preprocess datasets",
)
def is_audio_in_length_range(length):
return length > min_input_length and length < max_input_length
# filter data that is shorter than min_input_length
vectorized_datasets = vectorized_datasets.filter(
is_audio_in_length_range,
num_proc=num_workers,
input_columns=["input_length"],
)
# 7. Next, we can prepare the training.
# Let's use word error rate (WER) as our evaluation metric,
# instantiate a data collator and the trainer
# Define evaluation metrics during training, *i.e.* word error rate, character error rate
eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
# for large datasets it is advised to run the preprocessing on a
# single machine first with ``args.preprocessing_only`` since there will mostly likely
# be a timeout when running the script in distributed mode.
# In a second step ``args.preprocessing_only`` can then be set to `False` to load the
# cached dataset
if data_args.preprocessing_only:
logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
return
def compute_metrics(pred):
pred_logits = pred.predictions
pred_ids = np.argmax(pred_logits, axis=-1)
pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
pred_str = tokenizer.batch_decode(pred_ids)
# we do not want to group tokens when computing the metrics
label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
return metrics
# Now save everything to be able to create a single processor later
if is_main_process(training_args.local_rank):
# save feature extractor, tokenizer and config
feature_extractor.save_pretrained(training_args.output_dir)
tokenizer.save_pretrained(training_args.output_dir)
config.save_pretrained(training_args.output_dir)
try:
processor = AutoProcessor.from_pretrained(training_args.output_dir)
except (OSError, KeyError):
warnings.warn(
"Loading a processor from a feature extractor config that does not"
" include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
" attribute to your `preprocessor_config.json` file to suppress this warning: "
" `'processor_class': 'Wav2Vec2Processor'`",
FutureWarning,
)
processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
# Instantiate custom data collator
data_collator = DataCollatorCTCWithPadding(processor=processor)
# Initialize Trainer
trainer = Trainer(
model=model,
data_collator=data_collator,
args=training_args,
compute_metrics=compute_metrics,
train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
tokenizer=feature_extractor,
)
# 8. Finally, we can start training
# Training
if training_args.do_train:
# use last checkpoint if exist
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model()
metrics = train_result.metrics
max_train_samples = (
data_args.max_train_samples
if data_args.max_train_samples is not None
else len(vectorized_datasets["train"])
)
metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
max_eval_samples = (
data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
)
metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Write model card and (optionally) push to hub
config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": "speech-recognition",
"tags": ["automatic-speech-recognition", data_args.dataset_name],
"dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
"dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
}
if "common_voice" in data_args.dataset_name:
kwargs["language"] = config_name
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, 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 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."
)
# 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/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 data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name,
cache_dir=model_args.cache_dir)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
datasets = load_dataset(extension,
data_files=data_files,
field="data",
cache_dir=model_args.cache_dir)
# 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
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = XLNetConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer = XLNetTokenizerFast.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = XLNetForQuestionAnswering.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.
# Preprocessing is slighlty different for training and evaluation.
if training_args.do_train:
column_names = datasets["train"].column_names
elif training_args.do_eval:
column_names = datasets["validation"].column_names
else:
column_names = datasets["test"].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 data_args.max_seq_length > tokenizer.model_max_length:
logger.warning(
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)
# Training preprocessing
def prepare_train_features(examples):
# 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=data_args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
return_special_tokens_mask=True,
return_token_type_ids=True,
padding="max_length",
)
# 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")
# The special tokens will help us build the p_mask (which indicates the tokens that can't be in answers).
special_tokens = tokenized_examples.pop("special_tokens_mask")
# Let's label those examples!
tokenized_examples["start_positions"] = []
tokenized_examples["end_positions"] = []
tokenized_examples["is_impossible"] = []
tokenized_examples["cls_index"] = []
tokenized_examples["p_mask"] = []
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)
tokenized_examples["cls_index"].append(cls_index)
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples["token_type_ids"][i]
for k, s in enumerate(special_tokens[i]):
if s:
sequence_ids[k] = 3
context_idx = 1 if pad_on_right else 0
# Build the p_mask: non special tokens and context gets 0.0, the others get 1.0.
# The cls token gets 1.0 too (for predictions of empty answers).
tokenized_examples["p_mask"].append([
0.0 if (not special_tokens[i][k] and s == context_idx)
or k == cls_index else 1.0 for k, s in enumerate(sequence_ids)
])
# 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)
tokenized_examples["is_impossible"].append(1.0)
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] != context_idx:
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] != context_idx:
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)
tokenized_examples["is_impossible"].append(1.0)
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)
tokenized_examples["is_impossible"].append(0.0)
return tokenized_examples
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:
# Select samples from Dataset, This will help to decrease processing time
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
# Create Training Features
train_dataset = train_dataset.map(
prepare_train_features,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if data_args.max_train_samples is not None:
# Select samples from dataset again since Feature Creation might increase number of features
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
# Validation preprocessing
def prepare_validation_features(examples):
# 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=data_args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
return_special_tokens_mask=True,
return_token_type_ids=True,
padding="max_length",
)
# 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 special tokens will help us build the p_mask (which indicates the tokens that can't be in answers).
special_tokens = tokenized_examples.pop("special_tokens_mask")
# 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"] = []
# We still provide the index of the CLS token and the p_mask to the model, but not the is_impossible label.
tokenized_examples["cls_index"] = []
tokenized_examples["p_mask"] = []
for i, input_ids in enumerate(tokenized_examples["input_ids"]):
# Find the CLS token in the input ids.
cls_index = input_ids.index(tokenizer.cls_token_id)
tokenized_examples["cls_index"].append(cls_index)
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples["token_type_ids"][i]
for k, s in enumerate(special_tokens[i]):
if s:
sequence_ids[k] = 3
context_idx = 1 if pad_on_right else 0
# Build the p_mask: non special tokens and context gets 0.0, the others 1.0.
tokenized_examples["p_mask"].append([
0.0 if (not special_tokens[i][k] and s == context_idx)
or k == cls_index else 1.0 for k, s in enumerate(sequence_ids)
])
# 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_idx else None)
for k, o in enumerate(tokenized_examples["offset_mapping"][i])
]
return tokenized_examples
if training_args.do_eval:
if "validation" not in datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_examples = datasets["validation"]
if data_args.max_eval_samples is not None:
# Selecting Eval Samples from Dataset
eval_examples = eval_examples.select(
range(data_args.max_eval_samples))
# Create Features from Eval Dataset
eval_dataset = eval_examples.map(
prepare_validation_features,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if data_args.max_eval_samples is not None:
# Selecting Samples from Dataset again since Feature Creation might increase samples size
eval_dataset = eval_dataset.select(
range(data_args.max_eval_samples))
if training_args.do_predict:
if "test" not in datasets:
raise ValueError("--do_predict requires a test dataset")
predict_examples = datasets["test"]
if data_args.max_predict_samples is not None:
# We will select sample from whole data
predict_examples = predict_examples.select(
range(data_args.max_predict_samples))
# Test Feature Creation
predict_dataset = predict_examples.map(
prepare_validation_features,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if data_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(data_args.max_predict_samples))
# Data collator
# We have already padded to max length if the corresponding flag is True, otherwise we need to pad in the data
# collator.
data_collator = (default_data_collator if data_args.pad_to_max_length else
DataCollatorWithPadding(
tokenizer,
pad_to_multiple_of=8 if training_args.fp16 else None))
# 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, scores_diff_json = postprocess_qa_predictions_with_beam_search(
examples=examples,
features=features,
predictions=predictions,
version_2_with_negative=data_args.version_2_with_negative,
n_best_size=data_args.n_best_size,
max_answer_length=data_args.max_answer_length,
start_n_top=model.config.start_n_top,
end_n_top=model.config.end_n_top,
output_dir=training_args.output_dir,
is_world_process_zero=trainer.is_world_process_zero(),
prefix=stage,
)
# Format the result to the format the metric expects.
if data_args.version_2_with_negative:
formatted_predictions = [{
"id":
k,
"prediction_text":
v,
"no_answer_probability":
scores_diff_json[k]
} 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 data_args.version_2_with_negative else "squad")
def compute_metrics(p: EvalPrediction):
return metric.compute(predictions=p.predictions,
references=p.label_ids)
# Initialize our Trainer
trainer = QuestionAnsweringTrainer(
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,
eval_examples=eval_examples if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
post_process_function=post_processing_function,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
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
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
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.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(
eval_dataset)
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Prediction
if training_args.do_predict:
logger.info("*** Predict ***")
results = trainer.predict(predict_dataset, predict_examples)
metrics = results.metrics
max_predict_samples = (data_args.max_predict_samples
if data_args.max_predict_samples is not None
else len(predict_dataset))
metrics["predict_samples"] = min(max_predict_samples,
len(predict_dataset))
trainer.log_metrics("predict", metrics)
trainer.save_metrics("predict", metrics)
if training_args.push_to_hub:
trainer.push_to_hub()
def main():
# 1. Parse input arguments
# 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(
)
# 2. 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)],
)
log_level = training_args.get_process_log_level()
logger.setLevel(log_level)
transformers.utils.logging.set_verbosity(log_level)
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# 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}"
)
logger.info(f"Training/evaluation parameters {training_args}")
# 3. Detecting last checkpoint and eventualy continue from 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."
)
# 4. Load dataset
# Get the datasets: you can either provide your own CSV/JSON 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 first column for the full image path and the second column for the
# captions (unless you specify column names for this with the `image_column` and `caption_column` arguments).
#
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
dataset = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
cache_dir=model_args.cache_dir,
keep_in_memory=False,
data_dir=data_args.data_dir,
)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
dataset = load_dataset(extension,
data_files=data_files,
cache_dir=model_args.cache_dir)
# 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.
# 5. Load pretrained model, tokenizer, and feature extractor
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer)
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."
)
# Load feature_extractor, in this script we only use this to get the mean and std for normalization.
feature_extractor = AutoFeatureExtractor.from_pretrained(
model_args.feature_extractor_name or model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModel.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
config = model.config
def _freeze_params(module):
for param in module.parameters():
param.requires_grad = False
if model_args.freeze_vision_model:
_freeze_params(model.vision_model)
if model_args.freeze_text_model:
_freeze_params(model.text_model)
# set seed for torch dataloaders
set_seed(training_args.seed)
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
if training_args.do_train:
column_names = dataset["train"].column_names
elif training_args.do_eval:
column_names = dataset["validation"].column_names
elif training_args.do_predict:
column_names = dataset["test"].column_names
else:
logger.info(
"There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`."
)
return
# 6. Get the column names for input/target.
dataset_columns = dataset_name_mapping.get(data_args.dataset_name, None)
if data_args.image_column is None:
image_column = dataset_columns[
0] if dataset_columns is not None else column_names[0]
else:
image_column = data_args.image_column
if image_column not in column_names:
raise ValueError(
f"--image_column' value '{data_args.image_column}' needs to be one of: {', '.join(column_names)}"
)
if data_args.caption_column is None:
caption_column = dataset_columns[
1] if dataset_columns is not None else column_names[1]
else:
caption_column = data_args.caption_column
if caption_column not in column_names:
raise ValueError(
f"--caption_column' value '{data_args.caption_column}' needs to be one of: {', '.join(column_names)}"
)
# 7. Preprocessing the datasets.
# Initialize torchvision transforms and jit it for faster processing.
image_transformations = Transform(config.vision_config.image_size,
feature_extractor.image_mean,
feature_extractor.image_std)
image_transformations = torch.jit.script(image_transformations)
# Preprocessing the datasets.
# We need to tokenize input captions and transform the images.
def tokenize_captions(examples):
captions = [caption for caption in examples[caption_column]]
text_inputs = tokenizer(captions,
max_length=data_args.max_seq_length,
padding="max_length",
truncation=True)
examples["input_ids"] = text_inputs.input_ids
examples["attention_mask"] = text_inputs.attention_mask
return examples
def transform_images(examples):
images = [
read_image(image_file, mode=ImageReadMode.RGB)
for image_file in examples[image_column]
]
examples["pixel_values"] = [
image_transformations(image) for image in images
]
return examples
def filter_corrupt_images(examples):
"""remove problematic images"""
valid_images = []
for image_file in examples[image_column]:
try:
Image.open(image_file)
valid_images.append(True)
except Exception:
valid_images.append(False)
return valid_images
if training_args.do_train:
if "train" not in dataset:
raise ValueError("--do_train requires a train dataset")
train_dataset = dataset["train"]
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
train_dataset = train_dataset.filter(
filter_corrupt_images,
batched=True,
num_proc=data_args.preprocessing_num_workers)
train_dataset = train_dataset.map(
function=tokenize_captions,
batched=True,
remove_columns=[
col for col in column_names if col != image_column
],
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on train dataset",
)
# Transform images on the fly as doing it on the whole dataset takes too much time.
train_dataset.set_transform(transform_images)
if training_args.do_eval:
if "validation" not in dataset:
raise ValueError("--do_eval requires a train validation")
eval_dataset = dataset["validation"]
if data_args.max_eval_samples is not None:
eval_dataset = eval_dataset.select(
range(data_args.max_eval_samples))
eval_dataset = eval_dataset.filter(
filter_corrupt_images,
batched=True,
num_proc=data_args.preprocessing_num_workers)
eval_dataset = eval_dataset.map(
function=tokenize_captions,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=[
col for col in column_names if col != image_column
],
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on validation dataset",
)
# Transform images on the fly as doing it on the whole dataset takes too much time.
eval_dataset.set_transform(transform_images)
if training_args.do_predict:
if "test" not in dataset:
raise ValueError("--do_predict requires a test dataset")
test_dataset = dataset["test"]
if data_args.max_eval_samples is not None:
test_dataset = test_dataset.select(
range(data_args.max_eval_samples))
test_dataset = test_dataset.filter(
filter_corrupt_images,
batched=True,
num_proc=data_args.preprocessing_num_workers)
test_dataset = test_dataset.map(
function=tokenize_captions,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=[
col for col in column_names if col != image_column
],
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on test dataset",
)
# Transform images on the fly as doing it on the whole dataset takes too much time.
test_dataset.set_transform(transform_images)
# 8. Initalize 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,
data_collator=collate_fn,
)
# 9. Training
if training_args.do_train:
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
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model()
trainer.log_metrics("train", train_result.metrics)
trainer.save_metrics("train", train_result.metrics)
trainer.save_state()
# 10. Evaluation
if training_args.do_eval:
metrics = trainer.evaluate()
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# 11. Write Training Stats and push to hub.
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": "contrastive-image-text-modeling"
}
if data_args.dataset_name is not None:
kwargs["dataset_tags"] = data_args.dataset_name
if data_args.dataset_config_name is not None:
kwargs["dataset_args"] = data_args.dataset_config_name
kwargs[
"dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
else:
kwargs["dataset"] = data_args.dataset_name
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
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("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_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. 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.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name)
if "validation" not in datasets.keys():
datasets["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[:{data_args.validation_split_percentage}%]",
)
datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[{data_args.validation_split_percentage}%:]",
)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.train_file.split(".")[-1]
if extension == "txt":
extension = "text"
datasets = load_dataset(extension, data_files=data_files)
# 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
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config_kwargs = {
"cache_dir": model_args.cache_dir,
"revision": model_args.model_revision,
"use_auth_token": True if model_args.use_auth_token else None,
}
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name,
**config_kwargs)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path,
**config_kwargs)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
tokenizer_kwargs = {
"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,
}
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name,
**tokenizer_kwargs)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path, **tokenizer_kwargs)
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 model_args.model_name_or_path:
model = AutoModelForMaskedLM.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,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForMaskedLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
if data_args.line_by_line:
# When using line_by_line, we just tokenize each nonempty line.
padding = "max_length" if data_args.pad_to_max_length else False
def tokenize_function(examples):
# Remove empty lines
examples["text"] = [
line for line in examples["text"]
if len(line) > 0 and not line.isspace()
]
return tokenizer(
examples["text"],
padding=padding,
truncation=True,
max_length=data_args.max_seq_length,
# We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
# receives the `special_tokens_mask`.
return_special_tokens_mask=True,
)
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=[text_column_name],
load_from_cache_file=not data_args.overwrite_cache,
)
else:
# Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
# We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
# efficient when it receives the `special_tokens_mask`.
def tokenize_function(examples):
return tokenizer(examples[text_column_name],
return_special_tokens_mask=True)
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if data_args.max_seq_length is None:
max_seq_length = tokenizer.model_max_length
if max_seq_length > 1024:
logger.warn(
f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
"Picking 1024 instead. You can change that default value by passing --max_seq_length xxx."
)
max_seq_length = 1024
else:
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)
# Main data processing function that will concatenate all texts from our dataset and generate chunks of
# max_seq_length.
def group_texts(examples):
# Concatenate all texts.
concatenated_examples = {
k: sum(examples[k], [])
for k in examples.keys()
}
total_length = len(concatenated_examples[list(examples.keys())[0]])
# We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
# customize this part to your needs.
total_length = (total_length // max_seq_length) * max_seq_length
# Split by chunks of max_len.
result = {
k: [
t[i:i + max_seq_length]
for i in range(0, total_length, max_seq_length)
]
for k, t in concatenated_examples.items()
}
return result
# Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
# remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
# might be slower to preprocess.
#
# To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
tokenized_datasets = tokenized_datasets.map(
group_texts,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
# This one will take care of randomly masking the tokens.
data_collator = DataCollatorForLanguageModeling(
tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_datasets["train"]
if training_args.do_train else None,
eval_dataset=tokenized_datasets["validation"]
if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
)
# Training
if training_args.do_train:
if last_checkpoint is not None:
model_path = last_checkpoint
elif model_args.model_name_or_path is not None and os.path.isdir(
model_args.model_name_or_path):
model_path = model_args.model_name_or_path
else:
model_path = None
train_result = trainer.train(model_path=model_path)
trainer.save_model() # Saves the tokenizer too for easy upload
output_train_file = os.path.join(training_args.output_dir,
"train_results.txt")
if trainer.is_world_process_zero():
with open(output_train_file, "w") as writer:
logger.info("***** Train results *****")
for key, value in sorted(train_result.metrics.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
# Need to save the state, since Trainer.save_model saves only the tokenizer with the model
trainer.state.save_to_json(
os.path.join(training_args.output_dir, "trainer_state.json"))
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["eval_loss"])
results["perplexity"] = perplexity
output_eval_file = os.path.join(training_args.output_dir,
"eval_results_mlm.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in sorted(results.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments,
TrainingArguments, MultiLingAdapterArguments))
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, adapter_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args, adapter_args = parser.parse_args_into_dataclasses(
)
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir) and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO
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()
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)
elif data_args.train_file.endswith(".csv"):
# Loading a dataset from local csv files
datasets = load_dataset("csv",
data_files={
"train": data_args.train_file,
"validation": data_args.validation_file
})
else:
# Loading a dataset from local json files
datasets = load_dataset("json",
data_files={
"train": data_args.train_file,
"validation": data_args.validation_file
})
# See more about loading any type of standard or custom dataset at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Labels
label_list = None
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,
)
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,
)
# We use the AutoModelWithHeads class here for better adapter support.
model = AutoModelWithHeads.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,
)
model.add_classification_head(
data_args.task_name or "glue",
num_labels=num_labels,
id2label={i: v
for i, v in enumerate(label_list)}
if num_labels > 0 else None,
)
# Setup adapters
if adapter_args.train_adapter:
task_name = data_args.task_name or "glue"
# check if adapter already exists, otherwise add it
if task_name not in model.config.adapters:
# resolve the adapter config
adapter_config = AdapterConfig.load(
adapter_args.adapter_config,
non_linearity=adapter_args.adapter_non_linearity,
reduction_factor=adapter_args.adapter_reduction_factor,
)
# load a pre-trained from Hub if specified
if adapter_args.load_adapter:
model.load_adapter(
adapter_args.load_adapter,
config=adapter_config,
load_as=task_name,
)
# otherwise, add a fresh adapter
else:
model.add_adapter(task_name, config=adapter_config)
# optionally load a pre-trained language adapter
if adapter_args.load_lang_adapter:
# resolve the language adapter config
lang_adapter_config = AdapterConfig.load(
adapter_args.lang_adapter_config,
non_linearity=adapter_args.lang_adapter_non_linearity,
reduction_factor=adapter_args.lang_adapter_reduction_factor,
)
# load the language adapter from Hub
lang_adapter_name = model.load_adapter(
adapter_args.load_lang_adapter,
config=lang_adapter_config,
load_as=adapter_args.language,
)
else:
lang_adapter_name = None
# Freeze all model weights except of those of this adapter
model.train_adapter([task_name])
# Set the adapters to be used in every forward pass
if lang_adapter_name:
model.set_active_adapters(ac.Stack(lang_adapter_name, task_name))
else:
model.set_active_adapters(task_name)
else:
if adapter_args.load_adapter or adapter_args.load_lang_adapter:
raise ValueError(
"Adapters can only be loaded in adapters training mode."
"Use --train_adapter to enable adapter training")
# 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"
max_length = data_args.max_seq_length
else:
# We will pad later, dynamically at batch creation, to the max sequence length in each batch
padding = False
max_length = 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 data_args.task_name is not None and 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: 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:
label_to_id = {v: i for i, v in enumerate(label_list)}
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_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] for l in examples["label"]]
return result
datasets = datasets.map(preprocess_function,
batched=True,
load_from_cache_file=not data_args.overwrite_cache)
train_dataset = datasets["train"]
eval_dataset = datasets["validation_matched" if data_args.task_name ==
"mnli" else "validation"]
if data_args.task_name is not None:
test_dataset = datasets["test_matched" if data_args.task_name ==
"mnli" else "test"]
# 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]}.")
# 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()
}
# Initialize our Trainer
trainer_class = AdapterTrainer if adapter_args.train_adapter else Trainer
trainer = trainer_class(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset if training_args.do_eval else None,
compute_metrics=compute_metrics,
tokenizer=tokenizer,
# Data collator will default to DataCollatorWithPadding, so we change it if we already did the padding.
data_collator=default_data_collator
if data_args.pad_to_max_length else None,
)
# Training
if training_args.do_train:
trainer.train(model_path=model_args.model_name_or_path if os.path.
isdir(model_args.model_name_or_path) else None)
trainer.save_model() # Saves the tokenizer too for easy upload
# Evaluation
eval_results = {}
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):
eval_result = trainer.evaluate(eval_dataset=eval_dataset)
output_eval_file = os.path.join(training_args.output_dir,
f"eval_results_{task}.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info(f"***** Eval results {task} *****")
for key, value in eval_result.items():
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
eval_results.update(eval_result)
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")
return eval_results
def main(args_file=None):
# 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")) or args_file is not None:
# 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.
args_file_path = os.path.abspath(sys.argv[1]) if args_file is None else args_file
model_args, data_args, training_args = parser.parse_json_file(json_file=args_file_path)
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
assert model_args.model_type in list(MODEL_TYPE_TO_TOKENIZER.keys()), "model type should be 't5' or 'bart'"
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.local_rank != -1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Set seed
set_seed(training_args.seed)
# Set project name
os.environ["WANDB_PROJECT"] = "question-generation"
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
tokenizer_cls = MODEL_TYPE_TO_TOKENIZER[model_args.model_type]
tokenizer = tokenizer_cls.from_pretrained(
model_args.tokenizer_name_or_path if model_args.tokenizer_name_or_path else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
model = AutoModelForSeq2SeqLM.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
model.resize_token_embeddings(len(tokenizer))
if model_args.freeze_embeds:
logger.info("freezing embeddings of the model")
freeze_embeds(model)
assert_not_all_frozen(model)
# Get datasets
logger.info('loading dataset')
train_dataset = torch.load(data_args.train_file_path) if training_args.do_train else None
valid_dataset = torch.load(data_args.valid_file_path) if training_args.do_eval else None
logger.info('finished loading dataset')
# Initialize data_collator
data_collator = T2TDataCollator(
tokenizer=tokenizer,
model_type=model_args.model_type,
mode="training",
using_tpu=training_args.tpu_num_cores is not None
)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=valid_dataset,
data_collator=data_collator,
prediction_loss_only=True,
label_smoothing=model_args.label_smoothing
)
# disable wandb console logs
logging.getLogger('wandb.run_manager').setLevel(logging.WARNING)
# Training
if training_args.do_train:
trainer.train(
model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
)
trainer.save_model()
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
if trainer.is_world_master():
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval and training_args.local_rank in [-1, 0]:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(eval_output.keys()):
logger.info(" %s = %s", key, str(eval_output[key]))
writer.write("%s = %s\n" % (key, str(eval_output[key])))
results.update(eval_output)
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, 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",
level=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()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_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 data_args.train_file is not None or data_args.validation_file is not None:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.train_file.split(".")[-1]
datasets = load_dataset(extension, data_files=data_files)
else:
# Downloading and loading the swag dataset from the hub.
datasets = load_dataset("swag", "regular")
# 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
# 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,
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 = AutoModelForMultipleChoice.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,
)
# When using your own dataset or a different dataset from swag, you will probably need to change this.
ending_names = [f"ending{i}" for i in range(4)]
context_name = "sent1"
question_header_name = "sent2"
# Preprocessing the datasets.
def preprocess_function(examples):
first_sentences = [[context] * 4 for context in examples[context_name]]
question_headers = examples[question_header_name]
second_sentences = [[
f"{header} {examples[end][i]}" for end in ending_names
] for i, header in enumerate(question_headers)]
# Flatten out
first_sentences = sum(first_sentences, [])
second_sentences = sum(second_sentences, [])
# Tokenize
tokenized_examples = tokenizer(
first_sentences,
second_sentences,
truncation=True,
max_length=data_args.max_seq_length,
padding="max_length" if data_args.pad_to_max_length else False,
)
# Un-flatten
return {
k: [v[i:i + 4] for i in range(0, len(v), 4)]
for k, v in tokenized_examples.items()
}
tokenized_datasets = datasets.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
data_collator = (default_data_collator if data_args.pad_to_max_length else
DataCollatorForMultipleChoice(tokenizer=tokenizer))
# Metric
def compute_metrics(eval_predictions):
predictions, label_ids = eval_predictions
preds = np.argmax(predictions, axis=1)
return {
"accuracy": (preds == label_ids).astype(np.float32).mean().item()
}
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_datasets["train"]
if training_args.do_train else None,
eval_dataset=tokenized_datasets["validation"]
if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
if last_checkpoint is not None:
model_path = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
model_path = model_args.model_name_or_path
else:
model_path = None
train_result = trainer.train(model_path=model_path)
trainer.save_model() # Saves the tokenizer too for easy upload
output_train_file = os.path.join(training_args.output_dir,
"train_results.txt")
if trainer.is_world_process_zero():
with open(output_train_file, "w") as writer:
logger.info("***** Train results *****")
for key, value in sorted(train_result.metrics.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
# Need to save the state, since Trainer.save_model saves only the tokenizer with the model
trainer.state.save_to_json(
os.path.join(training_args.output_dir, "trainer_state.json"))
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
results = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir,
"eval_results_swag.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in sorted(results.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, 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(
)
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir) and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
module = import_module("tasks")
try:
token_classification_task_clazz = getattr(module, model_args.task_type)
token_classification_task: TokenClassificationTask = token_classification_task_clazz(
)
except AttributeError:
raise ValueError(
f"Task {model_args.task_type} needs to be defined as a TokenClassificationTask subclass in {module}. "
f"Available tasks classes are: {TokenClassificationTask.__subclasses__()}"
)
# Setup logging
if IS_IN_DOCKER_CONTAINER:
log_level = logging.WARNING
elif training_args.training_on_cloud:
log_level = logging.WARNING
else:
log_level = logging.INFO if (training_args.local_rank
in [-1, 0]) else logging.WARN
logging.basicConfig(
format=
"%(asctime)s - %(levelname)s - %(name)s - %(lineno)d - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=log_level)
logger.warning(f"log level: {log_level}")
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.local_rank != -1),
training_args.fp16,
)
logger.info("Training/evaluation parameters:")
for key, value in vars(training_args).items():
logger.info(" %s = %s", key, value)
# Set seed
set_seed(training_args.seed)
# Prepare CONLL-2003 task
labels = token_classification_task.get_labels(data_args.labels)
label_map: Dict[int, str] = {i: label for i, label in enumerate(labels)}
num_labels = len(labels)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
id2label=label_map,
label2id={label: i
for i, label in enumerate(labels)},
cache_dir=model_args.cache_dir,
gradient_checkpointing=training_args.gradient_checkpointing)
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,
additional_special_tokens=["[unused1]"])
model = AutoModelForTokenClassification.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,
training_args=training_args)
# Get datasets
train_dataset = (TokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.train,
training_args=training_args,
remove_label=Remove_Label) if training_args.do_train else None)
eval_dataset = (TokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.dev,
training_args=training_args,
remove_label=Remove_Label)
if training_args.do_eval or training_args.do_predict_dev
or training_args.evaluate_during_training else None)
def align_predictions(predictions: np.ndarray,
label_ids: np.ndarray,
dateset=None) -> Tuple[List[int], List[int]]:
if training_args.use_crf:
mask = []
for feature in dateset.features:
mask.append(feature.attention_mask)
if mask:
if isinstance(mask, list):
mask = np.array(mask)
mask = torch.from_numpy(mask).cuda() == 1
predictions = torch.from_numpy(predictions).cuda()
preds = model.crf.decode(predictions, mask)
else:
preds = model.crf.decode(predictions)
out_label_list = []
preds_list = []
for pred_one, label_one in zip(preds, label_ids.tolist()):
out_label = []
pred = []
for p, l in zip(pred_one, label_one):
out_label.append(label_map[l])
pred.append(label_map[p])
out_label = out_label[1:-1]
pred = pred[1:-1]
out_label_list.append(out_label)
preds_list.append(pred)
else:
preds = np.argmax(predictions, axis=2)
batch_size, seq_len = preds.shape
out_label_list = [[] for _ in range(batch_size)]
preds_list = [[] for _ in range(batch_size)]
for i in range(batch_size):
for j in range(seq_len):
if label_ids[i, j] != nn.CrossEntropyLoss().ignore_index:
out_label_list[i].append(label_map[label_ids[i][j]])
preds_list[i].append(label_map[preds[i][j]])
return preds_list, out_label_list
def precision_score(y_true, y_pred, average='micro'):
true_entities = set(y_true)
pred_entities = set(y_pred)
nb_correct = len(true_entities & pred_entities)
nb_pred = len(pred_entities)
score = nb_correct / nb_pred if nb_pred > 0 else 0
return score
def recall_score(y_true, y_pred, average='micro', suffix=False):
true_entities = set(y_true)
pred_entities = set(y_pred)
nb_correct = len(true_entities & pred_entities)
nb_true = len(true_entities)
score = nb_correct / nb_true if nb_true > 0 else 0
return score
def f_score(y_true, y_pred, average='micro', suffix=False):
true_entities = set(y_true)
pred_entities = set(y_pred)
nb_correct = len(true_entities & pred_entities)
nb_pred = len(pred_entities)
nb_true = len(true_entities)
p = nb_correct / nb_pred if nb_pred > 0 else 0
r = nb_correct / nb_true if nb_true > 0 else 0
score = 2 * p * r / (p + r) if p + r > 0 else 0
return score
def compute_metrics(p: EvalPrediction, mode, dateset=None) -> Dict:
preds_list, out_label_list = align_predictions(p.predictions,
p.label_ids, dateset)
words_list, preds_list = post_align_predictions(
dateset, preds_list, tokenizer)
if mode == "dev":
reader_file = os.path.join(data_args.data_dir, f"{mode}.txt")
elif mode == "test":
reader_file = os.path.join(data_args.test_data_dir, f"{mode}.txt")
else:
raise ValueError(f"mode is error: {mode}")
if mode == "dev":
data_dir = TRAIN_DIR
elif mode == "test":
data_dir = TEST_DIR
else:
raise ValueError(f"mode is error: {mode}")
with open(reader_file, "r", encoding="utf8") as reader:
pre_tuple, real_tuple = conver_entity_list_to_tuple(
reader, preds_list, words_list, data_dir)
# print(real_tuple[0], pre_tuple[0])
result = {
# "accuracy_score": accuracy_score(out_label_list, preds_list),
"all_precision": precision_score(real_tuple, pre_tuple),
"all_recall": recall_score(real_tuple, pre_tuple),
"all_f_score": f_score(real_tuple, pre_tuple),
}
for label in LABEL_LIST:
sub_pre_tuple = [t for t in pre_tuple[:] if t[-1] == label]
sub_real_tuple = [t for t in real_tuple[:] if t[-1] == label]
result.update({
f"{label}_precision":
precision_score(sub_real_tuple, sub_pre_tuple),
f"{label}_recall":
recall_score(sub_real_tuple, sub_pre_tuple),
f"{label}_f_score":
f_score(sub_real_tuple, sub_pre_tuple),
})
metrics_report = f"\n{'Tag':20s}\t{'Precision':9s}\t{'Recall':9s}\t{'F-Score':9s}\t\n"
for label in LABEL_LIST:
pkey = f"{label}_precision"
rkey = f"{label}_recall"
fkey = f"{label}_f_score"
p, r, f = result[pkey], result[rkey], result[fkey]
metrics_report += f"{label:20s}\t{p:9.7f}\t{r:9.7f}\t{f:9.7f}\t\n"
metrics_report += "<BLANKLINE>\n"
pkey = f"all_precision"
rkey = f"all_recall"
fkey = f"all_f_score"
p, r, f = result[pkey], result[rkey], result[fkey]
label = "ALL"
metrics_report += f"{label:20s}\t{p:9.7f}\t{r:9.7f}\t{f:9.7f}\t\n"
logger.info("--------metricd report---------")
logger.info(metrics_report)
return result
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
trainer.train(model_path=model_args.model_name_or_path if os.path.
isdir(model_args.model_name_or_path) else None)
trainer.save_model()
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
if trainer.is_world_master():
tokenizer.save_pretrained(training_args.output_dir)
select_model(training_args.output_dir, k_fold=10, select=True, rm=False)
if training_args.do_eval or training_args.do_predict or training_args.do_predict_dev:
model = AutoModelForTokenClassification.from_pretrained(
training_args.output_dir,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
training_args=training_args)
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
compute_metrics=compute_metrics,
)
logger.warning(f"load best mode from {training_args.output_dir}")
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
result = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir,
"eval_results.txt")
if trainer.is_world_master():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in result.items():
# logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
results.update(result)
# Predict
if training_args.do_predict:
test_dataset = TokenClassificationDataset(
token_classification_task=token_classification_task,
data_dir=data_args.test_data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.test,
training_args=training_args,
remove_label=Remove_Label)
# logger.warning(list(zip(tokens_list[3], test_dataset.features[3].input_ids)))
predictions, label_ids, metrics = trainer.predict(test_dataset)
preds_list, _ = align_predictions(predictions, label_ids, test_dataset)
# 数据对齐
words_list, preds_list = post_align_predictions(
test_dataset, preds_list, tokenizer)
output_test_results_file = os.path.join(training_args.output_dir,
"test_results.txt")
if trainer.is_world_master():
with open(output_test_results_file, "w") as writer:
for key, value in metrics.items():
logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
# Save predictions
output_test_predictions_file = os.path.join(training_args.output_dir,
"test_predictions.txt")
if trainer.is_world_master():
with open(output_test_predictions_file, "w") as writer:
with open(os.path.join(data_args.test_data_dir, "test.txt"),
"r") as f:
token_classification_task.write_predictions_to_file(
writer, f, preds_list, words_list)
# Predict dev
if training_args.do_predict_dev:
# logger.warning(list(zip(tokens_list[3], test_dataset.features[3].input_ids)))
predictions, label_ids, metrics = trainer.predict(
eval_dataset, description="Evaluation")
preds_list, _ = align_predictions(predictions, label_ids, eval_dataset)
# 数据对齐
words_list, preds_list = post_align_predictions(
eval_dataset, preds_list, tokenizer)
output_test_results_file = os.path.join(training_args.output_dir,
"dev_results.txt")
if trainer.is_world_master():
with open(output_test_results_file, "w") as writer:
for key, value in metrics.items():
writer.write("%s = %s\n" % (key, value))
# Save predictions
output_test_predictions_file = os.path.join(training_args.output_dir,
"dev_predictions.txt")
if trainer.is_world_master():
with open(output_test_predictions_file, "w") as writer:
with open(os.path.join(data_args.data_dir, "dev.txt"),
"r") as f:
token_classification_task.write_predictions_to_file(
writer, f, preds_list, words_list)
return results
def main():
# 在src/transformers/training_args.py中查看所有可能的参数，
# 或将 -help标志传递给此脚本。 现在，我们保留了不同的参数集，以更清晰地分离关注点。
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# 如果我们仅向脚本传递一个参数，并且它是json文件的路径，
# 让我们对其进行解析以获取参数。
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
#通过命令行传递参数
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
# 判断是不是重新训练
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir) and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
#设置日志格式
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO
if is_main_process(training_args.local_rank) else logging.WARN,
)
#打印当前设置
# 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}"
)
#打印参数
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
#随机数种子
set_seed(training_args.seed)
#num_labels 类别数量, output_mode 是任务类型，'classification'
# 加载预训练的模型和令牌生成器
# 分布式培训：
# from_pretrained方法保证只有一个本地进程可以并发下载模型和vocab。
#添加自定义参数finetuning_task, num_labels
# 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 in task_to_keys.keys():
# Downloading and loading a dataset from the hub.
datasets = load_dataset(path="data/glue.py", name=data_args.task_name)
elif data_args.task_name and data_args.task_script:
datasets = load_dataset(path=data_args.task_script,
name=data_args.task_name,
data_dir=data_args.task_dir)
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("data/csv.py", data_files=data_files)
else:
# Loading a dataset from local json files
datasets = load_dataset("data/json.py", 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, 标签设置
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)
logger.info(f"标签的总数是: {num_labels}")
# 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,
)
# .ckpt 是可以加载tensorflow的模型
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 and data_args.task_name in task_to_keys.keys(
):
# 对于sstb: sentence1_key, sentence2_key = ('sentence1', 'sentence2')
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 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: 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)}
def preprocess_function(examples):
""""
# Tokenize the texts
examples: {'idx': [0, 1], 'label': [5.0, 3.799999952316284], 'sentence1': ['A plane is taking off.', 'A man is playing a large flute.'], 'sentence2': ['An air plane is taking off.', 'A man is playing a flute.']}
"""
args = ((examples[sentence1_key], ) if sentence2_key is None else
(examples[sentence1_key], examples[sentence2_key]))
result = tokenizer(*args,
padding=padding,
max_length=data_args.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] for l in examples["label"]]
return result
datasets = datasets.map(preprocess_function,
batched=True,
load_from_cache_file=not data_args.overwrite_cache)
train_dataset = datasets["train"]
eval_dataset = datasets["validation_matched" if data_args.task_name ==
"mnli" else "validation"]
if data_args.task_name is not None or data_args.test_file is not None:
test_dataset = datasets["test_matched" if data_args.task_name ==
"mnli" else "test"]
# 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]}.")
# 设置metric的脚本
if data_args.task_name is not None and data_args.metric_script is not None:
metric = load_metric(data_args.metric_script, data_args.task_name)
else:
metric = load_metric("metric/glue.py", data_args.task_name)
# TODO: When datasets metrics include regular accuracy, make an else here and remove special branch from
# compute_metrics
# 您可以定义自定义的compute_metrics函数。 它需要一个`EvalPrediction`对象(一个带有predictions和label_ids字段的namedtuple)，并且必须返回一个值为float的字典字符串。
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()
}
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset if training_args.do_eval else None,
compute_metrics=compute_metrics,
tokenizer=tokenizer,
# Data collator will default to DataCollatorWithPadding, so we change it if we already did the padding.
data_collator=default_data_collator
if data_args.pad_to_max_length else None,
)
# Training
if training_args.do_train:
train_result = trainer.train(
model_path=model_args.model_name_or_path if os.path.
isdir(model_args.model_name_or_path) else None)
metrics = train_result.metrics
trainer.save_model() # Saves the tokenizer too for easy upload
output_train_file = os.path.join(training_args.output_dir,
"train_results.txt")
if trainer.is_world_process_zero():
with open(output_train_file, "w") as writer:
logger.info("***** Train results *****")
for key, value in sorted(metrics.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
# Need to save the state, since Trainer.save_model saves only the tokenizer with the model
trainer.state.save_to_json(
os.path.join(training_args.output_dir, "trainer_state.json"))
# Evaluation
eval_results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
# 如果是mnli, 循环以处理MNLI双重评估（匹配，不匹配)
# 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):
eval_result = trainer.evaluate(eval_dataset=eval_dataset)
output_eval_file = os.path.join(training_args.output_dir,
f"eval_results_{task}.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info(f"***** Eval results {task} *****")
for key, value in sorted(eval_result.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
eval_results.update(eval_result)
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")
return eval_results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, 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,
cache_dir=model_args.cache_dir)
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,
cache_dir=model_args.cache_dir)
else:
# Loading a dataset from local json files
datasets = load_dataset("json",
data_files=data_files,
cache_dir=model_args.cache_dir)
# 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.warning(
"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.warning(
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_eval_samples is not None:
eval_dataset = eval_dataset.select(
range(data_args.max_eval_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")
predict_dataset = datasets["test_matched" if data_args.task_name ==
"mnli" else "test"]
if data_args.max_predict_samples is not None:
predict_dataset = predict_dataset.select(
range(data_args.max_predict_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_eval_samples = (data_args.max_eval_samples
if data_args.max_eval_samples is not None else
len(eval_dataset))
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
if training_args.do_predict:
logger.info("*** Predict ***")
# Loop to handle MNLI double evaluation (matched, mis-matched)
tasks = [data_args.task_name]
predict_datasets = [predict_dataset]
if data_args.task_name == "mnli":
tasks.append("mnli-mm")
predict_datasets.append(datasets["test_mismatched"])
for predict_dataset, task in zip(predict_datasets, tasks):
# Removing the `label` columns because it contains -1 and Trainer won't like that.
predict_dataset.remove_columns_("label")
predictions = trainer.predict(
predict_dataset, metric_key_prefix="predict").predictions
predictions = np.squeeze(
predictions) if is_regression else np.argmax(predictions,
axis=1)
output_predict_file = os.path.join(training_args.output_dir,
f"predict_results_{task}.txt")
if trainer.is_world_process_zero():
with open(output_predict_file, "w") as writer:
logger.info(f"***** Predict 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")
if training_args.push_to_hub:
trainer.push_to_hub()
def main():
# region Argument parsing
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TFTrainingArguments))
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(
)
output_dir = Path(training_args.output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
# endregion
# region Logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
log_level = training_args.get_process_log_level()
logger.setLevel(log_level)
datasets.utils.logging.set_verbosity(log_level)
transformers.utils.logging.set_verbosity(log_level)
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# endregion
# region Checkpoints
checkpoint = None
if len(os.listdir(training_args.output_dir)
) > 0 and not training_args.overwrite_output_dir:
if (output_dir / CONFIG_NAME).is_file() and (
output_dir / TF2_WEIGHTS_NAME).is_file():
checkpoint = output_dir
logger.info(
f"Checkpoint detected, resuming training from checkpoint in {training_args.output_dir}. To avoid this"
" behavior, change the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
else:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to continue regardless.")
# endregion
# Set seed before initializing model.
set_seed(training_args.seed)
# region Load datasets
# 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 data_args.train_file is not None or data_args.validation_file is not None:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.train_file.split(".")[-1]
raw_datasets = load_dataset(extension,
data_files=data_files,
cache_dir=model_args.cache_dir)
else:
# Downloading and loading the swag dataset from the hub.
raw_datasets = load_dataset("swag",
"regular",
cache_dir=model_args.cache_dir)
# 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.
# When using your own dataset or a different dataset from swag, you will probably need to change this.
ending_names = [f"ending{i}" for i in range(4)]
context_name = "sent1"
question_header_name = "sent2"
# endregion
# region Load model config and tokenizer
if checkpoint is not None:
config_path = training_args.output_dir
elif model_args.config_name:
config_path = model_args.config_name
else:
config_path = model_args.model_name_or_path
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
config_path,
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,
)
# endregion
# region Dataset preprocessing
if data_args.max_seq_length is None:
max_seq_length = tokenizer.model_max_length
if max_seq_length > 1024:
logger.warning(
f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
"Picking 1024 instead. You can change that default value by passing --max_seq_length xxx."
)
max_seq_length = 1024
else:
if data_args.max_seq_length > tokenizer.model_max_length:
logger.warning(
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):
first_sentences = [[context] * 4 for context in examples[context_name]]
question_headers = examples[question_header_name]
second_sentences = [[
f"{header} {examples[end][i]}" for end in ending_names
] for i, header in enumerate(question_headers)]
# Flatten out
first_sentences = sum(first_sentences, [])
second_sentences = sum(second_sentences, [])
# Tokenize
tokenized_examples = tokenizer(first_sentences,
second_sentences,
truncation=True,
max_length=max_seq_length)
# Un-flatten
data = {
k: [v[i:i + 4] for i in range(0, len(v), 4)]
for k, v in tokenized_examples.items()
}
return data
if training_args.do_train:
if "train" not in raw_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = raw_datasets["train"]
non_label_columns = [
feature for feature in train_dataset.features
if feature not in ("label", "labels")
]
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
with training_args.main_process_first(
desc="train dataset map pre-processing"):
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
if "validation" not in raw_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = raw_datasets["validation"]
if not training_args.do_train:
non_label_columns = [
feature for feature in eval_dataset.features
if feature not in ("label", "labels")
]
if data_args.max_eval_samples is not None:
eval_dataset = eval_dataset.select(
range(data_args.max_eval_samples))
with training_args.main_process_first(
desc="validation dataset map pre-processing"):
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
# endregion
with training_args.strategy.scope():
# region Build model
if checkpoint is None:
model_path = model_args.model_name_or_path
else:
model_path = checkpoint
model = TFAutoModelForMultipleChoice.from_pretrained(
model_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,
)
num_replicas = training_args.strategy.num_replicas_in_sync
total_train_batch_size = training_args.per_device_train_batch_size * num_replicas
total_eval_batch_size = training_args.per_device_eval_batch_size * num_replicas
if training_args.do_train:
total_train_steps = (len(train_dataset) // total_train_batch_size
) * int(training_args.num_train_epochs)
optimizer, lr_schedule = create_optimizer(
init_lr=training_args.learning_rate,
num_train_steps=int(total_train_steps),
num_warmup_steps=0)
else:
optimizer = "adam" # Just put anything in here, since we're not using it anyway
model.compile(
optimizer=optimizer,
loss=tf.keras.losses.SparseCategoricalCrossentropy(
from_logits=True),
metrics=[
tf.keras.metrics.SparseCategoricalAccuracy(name="accuracy")
],
)
# endregion
# region Training
if training_args.do_train:
tf_train_dataset = convert_dataset_for_tensorflow(
train_dataset,
non_label_column_names=non_label_columns,
batch_size=total_train_batch_size)
if training_args.do_eval:
validation_data = convert_dataset_for_tensorflow(
eval_dataset,
non_label_column_names=non_label_columns,
batch_size=total_eval_batch_size)
else:
validation_data = None
model.fit(
tf_train_dataset,
validation_data=validation_data,
epochs=int(training_args.num_train_epochs),
callbacks=[
SavePretrainedCallback(output_dir=training_args.output_dir)
],
)
# endregion
# region Evaluation
if training_args.do_eval and not training_args.do_train:
# Do a standalone evaluation pass
tf_eval_dataset = convert_dataset_for_tensorflow(
eval_dataset,
non_label_column_names=non_label_columns,
batch_size=total_eval_batch_size)
model.evaluate(tf_eval_dataset)
# endregion
# region Push to hub
if training_args.push_to_hub:
model.push_to_hub(
finetuned_from=model_args.model_name_or_path,
tasks="multiple-choice",
dataset_tags="swag",
dataset_args="regular",
dataset="SWAG",
language="en",
)
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(
)
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir) and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO
if training_args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.local_rank != -1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Set seed
set_seed(training_args.seed)
# Prepare CONLL-2003 task
labels = get_labels(data_args.labels)
label_map: Dict[int, str] = {i: label for i, label in enumerate(labels)}
num_labels = len(labels)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
id2label=label_map,
label2id={label: i
for i, label in enumerate(labels)},
cache_dir=model_args.cache_dir,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast,
)
model = AutoModelForTokenClassification.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,
)
'''
model_to_save = AutoModel.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,
)
model_to_save.save_pretrained(training_args.output_dir)
tokenizer.save_pretrained(training_args.output_dir)
import pdb; pdb.set_trace()
'''
# Get datasets
train_dataset = (NerDataset(
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.train,
) if training_args.do_train else None)
eval_dataset = (NerDataset(
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.dev,
) if training_args.do_eval else None)
def align_predictions(
predictions: np.ndarray,
label_ids: np.ndarray) -> Tuple[List[int], List[int]]:
preds = np.argmax(predictions, axis=2)
batch_size, seq_len = preds.shape
out_label_list = [[] for _ in range(batch_size)]
preds_list = [[] for _ in range(batch_size)]
for i in range(batch_size):
for j in range(seq_len):
if label_ids[i, j] != nn.CrossEntropyLoss().ignore_index:
out_label_list[i].append(label_map[label_ids[i][j]])
preds_list[i].append(label_map[preds[i][j]])
return preds_list, out_label_list
def compute_metrics(p: EvalPrediction) -> Dict:
preds_list, out_label_list = align_predictions(p.predictions,
p.label_ids)
return {
"precision": precision_score(out_label_list, preds_list),
"recall": recall_score(out_label_list, preds_list),
"f1": f1_score(out_label_list, preds_list),
}
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
trainer.train(model_path=model_args.model_name_or_path if os.path.
isdir(model_args.model_name_or_path) else None)
trainer.save_model()
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
if trainer.is_world_master():
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
result = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir,
"eval_results.txt")
if trainer.is_world_master():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in result.items():
logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
results.update(result)
# Predict
if training_args.do_predict:
test_dataset = NerDataset(
data_dir=data_args.data_dir,
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args.max_seq_length,
overwrite_cache=data_args.overwrite_cache,
mode=Split.test,
)
predictions, label_ids, metrics = trainer.predict(test_dataset)
preds_list, _ = align_predictions(predictions, label_ids)
# Save predictions
output_test_results_file = os.path.join(training_args.output_dir,
"test_results.txt")
if trainer.is_world_master():
with open(output_test_results_file, "w") as writer:
logger.info("***** Test results *****")
for key, value in metrics.items():
logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
output_test_predictions_file = os.path.join(training_args.output_dir,
"test_predictions.txt")
if trainer.is_world_master():
with open(output_test_predictions_file, "w") as writer:
with open(os.path.join(data_args.data_dir, "test.txt"),
"r") as f:
example_id = 0
for line in f:
if line.startswith(
"-DOCSTART-") or line == "" or line == "\n":
writer.write(line)
if not preds_list[example_id]:
example_id += 1
elif preds_list[example_id]:
entity_label = preds_list[example_id].pop(0)
if entity_label == 'O':
output_line = line.split(
)[0] + " " + entity_label + "\n"
else:
output_line = line.split(
)[0] + " " + entity_label[0] + "\n"
# output_line = line.split()[0] + " " + preds_list[example_id].pop(0) + "\n"
writer.write(output_line)
else:
logger.warning(
"Maximum sequence length exceeded: No prediction for '%s'.",
line.split()[0])
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
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(
)
if data_args.source_prefix is None and model_args.model_name_or_path in [
"t5-small",
"t5-base",
"t5-large",
"t5-3b",
"t5-11b",
]:
logger.warning(
"You're running a t5 model but didn't provide a source prefix, which is expected, e.g. with "
"`--source_prefix 'translate English to German: ' `")
# 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()
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 JSON 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 translation, only JSON files are supported, with one field named "translation" containing two keys for the
# source and target languages (unless you adapt what follows).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
datasets = load_dataset(extension, data_files=data_files)
# 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
#
# 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,
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 = AutoModelForSeq2SeqLM.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,
)
# Set decoder_start_token_id
if model.config.decoder_start_token_id is None and isinstance(
tokenizer, (MBartTokenizer, MBartTokenizerFast)):
assert (data_args.target_lang is not None and data_args.source_lang
is not None), "mBart requires --target_lang and --source_lang"
if isinstance(tokenizer, MBartTokenizer):
model.config.decoder_start_token_id = tokenizer.lang_code_to_id[
data_args.target_lang]
else:
model.config.decoder_start_token_id = tokenizer.convert_tokens_to_ids(
data_args.target_lang)
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
prefix = data_args.source_prefix if data_args.source_prefix is not None else ""
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
if training_args.do_train:
column_names = datasets["train"].column_names
elif training_args.do_eval:
column_names = datasets["validation"].column_names
elif training_args.do_predict:
column_names = datasets["test"].column_names
else:
logger.info(
"There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`."
)
return
# For translation we set the codes of our source and target languages (only useful for mBART, the others will
# ignore those attributes).
if isinstance(tokenizer, (MBartTokenizer, MBartTokenizerFast)):
if data_args.source_lang is not None:
tokenizer.src_lang = data_args.source_lang
if data_args.target_lang is not None:
tokenizer.tgt_lang = data_args.target_lang
# Get the language codes for input/target.
source_lang = data_args.source_lang.split("_")[0]
target_lang = data_args.target_lang.split("_")[0]
# Temporarily set max_target_length for training.
max_target_length = data_args.max_target_length
padding = "max_length" if data_args.pad_to_max_length else False
if training_args.label_smoothing_factor > 0 and not hasattr(
model, "prepare_decoder_input_ids_from_labels"):
logger.warn(
"label_smoothing is enabled but the `prepare_decoder_input_ids_from_labels` method is not defined for"
f"`{model.__class__.__name__}`. This will lead to loss being calculated twice and will take up more memory"
)
def preprocess_function(examples):
inputs = [ex[source_lang] for ex in examples["translation"]]
targets = [ex[target_lang] for ex in examples["translation"]]
inputs = [prefix + inp for inp in inputs]
model_inputs = tokenizer(inputs,
max_length=data_args.max_source_length,
padding=padding,
truncation=True)
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets,
max_length=max_target_length,
padding=padding,
truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [[
(l if l != tokenizer.pad_token_id else -100) for l in label
] for label in labels["input_ids"]]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
if training_args.do_train:
train_dataset = datasets["train"]
if "train" not in datasets:
raise ValueError("--do_train requires a train dataset")
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
max_target_length = data_args.val_max_target_length
if "validation" not in datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = datasets["validation"]
if data_args.max_val_samples is not None:
eval_dataset = eval_dataset.select(range(
data_args.max_val_samples))
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_predict:
max_target_length = data_args.val_max_target_length
if "test" not in datasets:
raise ValueError("--do_predict requires a test dataset")
test_dataset = datasets["test"]
if data_args.max_test_samples is not None:
test_dataset = test_dataset.select(
range(data_args.max_test_samples))
test_dataset = test_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
if data_args.pad_to_max_length:
data_collator = default_data_collator
else:
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if training_args.fp16 else None,
)
# Metric
metric = load_metric("sacrebleu")
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [[label.strip()] for label in labels]
return preds, labels
def compute_metrics(eval_preds):
preds, labels = eval_preds
if isinstance(preds, tuple):
preds = preds[0]
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels,
skip_special_tokens=True)
# Some simple post-processing
decoded_preds, decoded_labels = postprocess_text(
decoded_preds, decoded_labels)
result = metric.compute(predictions=decoded_preds,
references=decoded_labels)
result = {"bleu": result["score"]}
prediction_lens = [
np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds
]
result["gen_len"] = np.mean(prediction_lens)
result = {k: round(v, 4) for k, v in result.items()}
return result
# Initialize our Trainer
trainer = Seq2SeqTrainer(
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,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics
if training_args.predict_with_generate else None,
)
# Training
if training_args.do_train:
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
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.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(max_length=data_args.val_max_target_length,
num_beams=data_args.num_beams,
metric_key_prefix="eval")
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 ***")
test_results = trainer.predict(
test_dataset,
metric_key_prefix="test",
max_length=data_args.val_max_target_length,
num_beams=data_args.num_beams,
)
metrics = test_results.metrics
max_test_samples = data_args.max_test_samples if data_args.max_test_samples is not None else len(
test_dataset)
metrics["test_samples"] = min(max_test_samples, len(test_dataset))
trainer.log_metrics("test", metrics)
trainer.save_metrics("test", metrics)
if trainer.is_world_process_zero():
if training_args.predict_with_generate:
test_preds = tokenizer.batch_decode(
test_results.predictions,
skip_special_tokens=True,
clean_up_tokenization_spaces=True)
test_preds = [pred.strip() for pred in test_preds]
output_test_preds_file = os.path.join(training_args.output_dir,
"test_generations.txt")
with open(output_test_preds_file, "w") as writer:
writer.write("\n".join(test_preds))
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, 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("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_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 data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.train_file.split(".")[-1]
datasets = load_dataset(extension, data_files=data_files)
# 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.
if training_args.do_train:
column_names = datasets["train"].column_names
features = datasets["train"].features
else:
column_names = datasets["validation"].column_names
features = datasets["validation"].features
text_column_name = "tokens" if "tokens" in column_names else column_names[0]
label_column_name = (
f"{data_args.task_name}_tags" if f"{data_args.task_name}_tags" in column_names else column_names[1]
)
# In the event the labels are not a `Sequence[ClassLabel]`, we will need to go through the dataset to get the
# unique labels.
def get_label_list(labels):
unique_labels = set()
for label in labels:
unique_labels = unique_labels | set(label)
label_list = list(unique_labels)
label_list.sort()
return label_list
if isinstance(features[label_column_name].feature, ClassLabel):
label_list = features[label_column_name].feature.names
# No need to convert the labels since they are already ints.
label_to_id = {i: i for i in range(len(label_list))}
else:
label_list = get_label_list(datasets["train"][label_column_name])
label_to_id = {l: i for i, l in enumerate(label_list)}
num_labels = len(label_list)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
finetuning_task=data_args.task_name,
cache_dir=model_args.cache_dir,
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=True,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForTokenClassification.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,
)
# Tokenizer check: this script requires a fast tokenizer.
if not isinstance(tokenizer, PreTrainedTokenizerFast):
raise ValueError(
"This example script only works for models that have a fast tokenizer. Checkout the big table of models "
"at https://huggingface.co/transformers/index.html#bigtable to find the model types that meet this "
"requirement"
)
# Preprocessing the dataset
# Padding strategy
padding = "max_length" if data_args.pad_to_max_length else False
# Tokenize all texts and align the labels with them.
def tokenize_and_align_labels(examples):
tokenized_inputs = tokenizer(
examples[text_column_name],
padding=padding,
truncation=True,
# We use this argument because the texts in our dataset are lists of words (with a label for each word).
is_split_into_words=True,
)
labels = []
for i, label in enumerate(examples[label_column_name]):
word_ids = tokenized_inputs.word_ids(batch_index=i)
previous_word_idx = None
label_ids = []
for word_idx in word_ids:
# Special tokens have a word id that is None. We set the label to -100 so they are automatically
# ignored in the loss function.
if word_idx is None:
label_ids.append(-100)
# We set the label for the first token of each word.
elif word_idx != previous_word_idx:
label_ids.append(label_to_id[label[word_idx]])
# For the other tokens in a word, we set the label to either the current label or -100, depending on
# the label_all_tokens flag.
else:
label_ids.append(label_to_id[label[word_idx]] if data_args.label_all_tokens else -100)
previous_word_idx = word_idx
labels.append(label_ids)
tokenized_inputs["labels"] = labels
return tokenized_inputs
tokenized_datasets = datasets.map(
tokenize_and_align_labels,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
data_collator = DataCollatorForTokenClassification(tokenizer, pad_to_multiple_of=8 if training_args.fp16 else None)
# 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)
if data_args.return_entity_level_metrics:
# Unpack nested dictionaries
final_results = {}
for key, value in results.items():
if isinstance(value, dict):
for n, v in value.items():
final_results[f"{key}_{n}"] = v
else:
final_results[key] = value
return final_results
else:
return {
"precision": results["overall_precision"],
"recall": results["overall_recall"],
"f1": results["overall_f1"],
"accuracy": results["overall_accuracy"],
}
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_datasets["train"] if training_args.do_train else None,
eval_dataset=tokenized_datasets["validation"] if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
output_train_file = os.path.join(training_args.output_dir, "train_results.txt")
if trainer.is_world_process_zero():
with open(output_train_file, "w") as writer:
logger.info("***** Train results *****")
for key, value in sorted(train_result.metrics.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
# Need to save the state, since Trainer.save_model saves only the tokenizer with the model
trainer.state.save_to_json(os.path.join(training_args.output_dir, "trainer_state.json"))
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
results = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir, "eval_results_ner.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in results.items():
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
# Predict
if training_args.do_predict:
logger.info("*** Predict ***")
test_dataset = tokenized_datasets["test"]
predictions, labels, metrics = trainer.predict(test_dataset)
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)
]
output_test_results_file = os.path.join(training_args.output_dir, "test_results.txt")
if trainer.is_world_process_zero():
with open(output_test_results_file, "w") as writer:
for key, value in sorted(metrics.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
# Save predictions
output_test_predictions_file = os.path.join(training_args.output_dir, "test_predictions.txt")
if trainer.is_world_process_zero():
with open(output_test_predictions_file, "w") as writer:
for prediction in true_predictions:
writer.write(" ".join(prediction) + "\n")
return results
async def index(request: Request):
request_content_type = request.headers["Content-Type"]
"""An input_fn that loads a pickled tensor"""
if request_content_type == "application/json":
data = await request.json()
logger.info("================ input sentences ===============")
logger.info(f"data: {data}")
logger.info(f"data: {data['review_body']}")
test_file = dict()
test_file = {
"review_id": "",
"product_id": "",
"reviewer_id": "",
"stars": "",
"review_body": data["review_body"],
"review_title": "",
"language": "en",
"product_category": "",
}
with open(
os.path.dirname(__file__) + "request_body.json",
mode="wt",
encoding="utf-8",
) as file:
json.dump(test_file, file, ensure_ascii=False, indent=2)
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
model_args, data_args, training_args = parser.parse_json_file(
json_file="./args.json")
logger.info(f"training=args: {training_args}")
# dataset = Dataset.from_dict(data)
data_files = {}
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
dataset = load_dataset(extension, data_files=data_files)
logger.info(f"dataset: {dataset}")
test_dataset = dataset["test"]
test_dataset = test_dataset.map(
_preprocess_function,
batched=True,
num_proc=None,
remove_columns=[
"review_id",
"product_id",
"reviewer_id",
"stars",
"review_body",
"review_title",
"language",
"product_category",
],
load_from_cache_file=True,
)
logger.info(f"test_dataset: {test_dataset}")
preds = predict(test_dataset, model)
return preds[0]
raise ValueError(f"Unsupported content type: {request_content_type}")
for metric_name, value in eval_metrics.items():
summary_writer.scalar(f"eval_{metric_name}", value, step)
if __name__ == "__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(
)
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir) and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty."
"Use --overwrite_output_dir to overcome.")
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
level="NOTSET",
def main():
# 1. Parse input arguments
# 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, Seq2SeqTrainingArguments))
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(
)
# 2. 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)],
)
log_level = training_args.get_process_log_level()
logger.setLevel(log_level)
datasets.utils.logging.set_verbosity(log_level)
transformers.utils.logging.set_verbosity(log_level)
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
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}"
)
logger.info(f"Training/evaluation parameters {training_args}")
# 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()
logger.info("Training/evaluation parameters %s", training_args)
# 3. Detecting last checkpoint and eventualy continue from 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 before initializing model.
set_seed(training_args.seed)
# 4. Load dataset
raw_datasets = DatasetDict()
if training_args.do_train:
raw_datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=data_args.train_split_name,
use_auth_token=True if model_args.use_auth_token else None,
)
if training_args.do_eval:
raw_datasets["eval"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=data_args.eval_split_name,
use_auth_token=True if model_args.use_auth_token else None,
)
if data_args.audio_column_name not in next(iter(
raw_datasets.values())).column_names:
raise ValueError(
f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--audio_column_name` to the correct audio column - one of "
f"{', '.join(next(iter(raw_datasets.values())).column_names)}.")
if data_args.text_column_name not in next(iter(
raw_datasets.values())).column_names:
raise ValueError(
f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--text_column_name` to the correct text column - one of "
f"{', '.join(next(iter(raw_datasets.values())).column_names)}.")
# 5. Load pretrained model, tokenizer, and feature extractor
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
feature_extractor = AutoFeatureExtractor.from_pretrained(
model_args.feature_extractor_name if model_args.feature_extractor_name
else model_args.model_name_or_path,
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 = AutoModelForSpeechSeq2Seq.from_pretrained(
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,
)
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
if model_args.freeze_feature_encoder:
model.freeze_feature_encoder()
# 6. Resample speech dataset if necassary
dataset_sampling_rate = next(iter(raw_datasets.values())).features[
data_args.audio_column_name].sampling_rate
if dataset_sampling_rate != feature_extractor.sampling_rate:
raw_datasets = raw_datasets.cast_column(
data_args.audio_column_name,
datasets.features.Audio(
sampling_rate=feature_extractor.sampling_rate))
# 7. Preprocessing the datasets.
# We need to read the audio files as arrays and tokenize the targets.
max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
audio_column_name = data_args.audio_column_name
num_workers = data_args.preprocessing_num_workers
text_column_name = data_args.text_column_name
model_input_name = feature_extractor.model_input_names[0]
do_lower_case = data_args.do_lower_case
if data_args.max_train_samples is not None:
raw_datasets["train"] = raw_datasets["train"].select(
range(data_args.max_train_samples))
if data_args.max_eval_samples is not None:
raw_datasets["eval"] = raw_datasets["eval"].select(
range(data_args.max_eval_samples))
def prepare_dataset(batch):
# process audio
sample = batch[audio_column_name]
inputs = feature_extractor(sample["array"],
sampling_rate=sample["sampling_rate"])
# process audio length
batch[model_input_name] = inputs.input_values[0]
batch["input_length"] = len(batch["input_values"])
# process targets
input_str = batch[text_column_name].lower(
) if do_lower_case else batch[text_column_name]
batch["labels"] = tokenizer(input_str).input_ids
return batch
with training_args.main_process_first(desc="dataset map pre-processing"):
vectorized_datasets = raw_datasets.map(
prepare_dataset,
remove_columns=next(iter(raw_datasets.values())).column_names,
num_proc=data_args.preprocessing_num_workers,
desc="preprocess train dataset",
)
# filter data that is shorter than min_input_length or longer than
# max_input_length
def is_audio_in_length_range(length):
return length > min_input_length and length < max_input_length
vectorized_datasets = vectorized_datasets.filter(
is_audio_in_length_range,
num_proc=num_workers,
input_columns=["input_length"],
)
# for large datasets it is advised to run the preprocessing on a
# single machine first with `args.preprocessing_only` since there will mostly likely
# be a timeout when running the script in distributed mode.
# In a second step `args.preprocessing_only` can then be set to `False` to load the
# cached dataset
if data_args.preprocessing_only:
cache = {k: v.cache_files for k, v in vectorized_datasets.items()}
logger.info(f"Data preprocessing finished. Files cached at {cache}.")
return
# 8. Load Metric
metric = load_metric("wer")
def compute_metrics(pred):
pred_ids = pred.predictions
pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
pred_str = tokenizer.batch_decode(pred_ids, skip_special_tokens=True)
# we do not want to group tokens when computing the metrics
label_str = tokenizer.batch_decode(pred.label_ids,
skip_special_tokens=True)
wer = metric.compute(predictions=pred_str, references=label_str)
return {"wer": wer}
# 9. Create a single speech processor
if is_main_process(training_args.local_rank):
# save feature extractor, tokenizer and config
feature_extractor.save_pretrained(training_args.output_dir)
tokenizer.save_pretrained(training_args.output_dir)
config.save_pretrained(training_args.output_dir)
processor = AutoProcessor.from_pretrained(training_args.output_dir)
# 10. Define data collator
data_collator = DataCollatorSpeechSeq2SeqWithPadding(
processor=processor,
decoder_start_token_id=model.config.decoder_start_token_id)
# 11. Initialize Trainer
trainer = Seq2SeqTrainer(
model=model,
args=training_args,
train_dataset=vectorized_datasets["train"]
if training_args.do_train else None,
eval_dataset=vectorized_datasets["eval"]
if training_args.do_eval else None,
tokenizer=feature_extractor,
data_collator=data_collator,
compute_metrics=compute_metrics
if training_args.predict_with_generate else None,
)
# 12. Training
if training_args.do_train:
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
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the feature extractor too for easy upload
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(vectorized_datasets["train"]))
metrics["train_samples"] = min(max_train_samples,
len(vectorized_datasets["train"]))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# 13. Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(metric_key_prefix="eval",
max_length=model.config.max_length,
num_beams=model.config.num_beams)
max_eval_samples = (data_args.max_eval_samples
if data_args.max_eval_samples is not None else len(
vectorized_datasets["eval"]))
metrics["eval_samples"] = min(max_eval_samples,
len(vectorized_datasets["eval"]))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# 14. Write Training Stats
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": "speech recognition"
}
if data_args.dataset_name is not None:
kwargs["dataset_tags"] = data_args.dataset_name
if data_args.dataset_config_name is not None:
kwargs["dataset_args"] = data_args.dataset_config_name
kwargs[
"dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
else:
kwargs["dataset"] = data_args.dataset_name
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
return results
