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)
# 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,
}
# tokenizer = GPT2TokenizerFast(
# os.path.join('../tokenizer', 'vocab.json'),
# os.path.join('../tokenizer', 'merges.txt'),
# bos_token='<s>',
# eos_token='</s>',
# sep_token='</s>',
# cls_token='<s>',
# unk_token='<unk>',
# pad_token='<pad>',
# mask_token='<mask>',
# )
tokenizer = GPT2TokenizerFast.from_pretrained('gpt2')
tokenizer.add_special_tokens({
'bos_token': '<s>',
'eos_token': '</s>',
'sep_token': '</s>',
'cls_token': '<s>',
'unk_token': '<unk>',
'pad_token': '<pad>',
'mask_token': '<mask>',
})
# tokenizer.convert_tokens_to_ids(s)
model = AutoModelForSequenceClassification.from_pretrained('../test-v1/',
num_labels=2)
# Data collator
# This one will take care of randomly masking the tokens.
data_collator = DataCollatorForLanguageModeling(
tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
train_dataset = Dataset(
paths=['../bin_data/{}_text_document'.format(i) for i in range(30)],
tokenizer=tokenizer)
eval_dataset = Dataset(paths='../bin_data/30_text_document',
tokenizer=tokenizer)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
tokenizer=tokenizer,
data_collator=data_collator,
)
# Training
if training_args.do_train:
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif model_args.model_name_or_path is not None and 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
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))
perplexity = math.exp(metrics["eval_loss"])
metrics["perplexity"] = perplexity
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
logging_dir=f"{args.output_data_dir}/logs",
learning_rate=args.learning_rate,
fp16=True,
dataloader_drop_last=True,
disable_tqdm=True,
evaluation_strategy="no",
save_strategy="no",
save_total_limit=1,
logging_strategy="epoch",
)
# create Trainer instance
trainer = Trainer(
model=model,
args=training_args,
compute_metrics=compute_metrics,
train_dataset=train_dataset,
eval_dataset=test_dataset,
tokenizer=tokenizer,
)
# train model
trainer.train()
# evaluate model
eval_result = trainer.evaluate(eval_dataset=test_dataset)
# writes eval result to file which can be accessed later in s3 ouput
with open(os.path.join(args.output_data_dir, "eval_results.txt"),
"w") as writer:
print(f"***** Eval results *****")
for key, value in sorted(eval_result.items()):
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.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
)
if "validation" not in raw_datasets.keys():
raw_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,
)
raw_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 extension == "txt":
extension = "text"
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_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 = XLNetConfig()
logger.warning("You are instantiating a new config instance from scratch.")
if model_args.config_overrides is not None:
logger.info(f"Overriding config: {model_args.config_overrides}")
config.update_from_string(model_args.config_overrides)
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 = XLNetLMHeadModel.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 = XLNetLMHeadModel.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 = raw_datasets["train"].column_names
else:
column_names = raw_datasets["validation"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
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)
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=max_seq_length)
with training_args.main_process_first(desc="dataset map tokenization"):
tokenized_datasets = raw_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,
desc="Running tokenizer on dataset line_by_line",
)
else:
# Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
def tokenize_function(examples):
return tokenizer(examples[text_column_name])
with training_args.main_process_first(desc="dataset map tokenization"):
tokenized_datasets = raw_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,
desc="Running tokenizer on every text in dataset",
)
# 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
with training_args.main_process_first(desc="grouping texts together"):
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,
desc=f"Grouping texts in chunks of {max_seq_length}",
)
if training_args.do_train:
if "train" not in tokenized_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = tokenized_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 = tokenized_datasets["validation"]
if data_args.max_eval_samples is not None:
eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))
# Data collator
data_collator = DataCollatorForPermutationLanguageModeling(
tokenizer=tokenizer,
plm_probability=data_args.plm_probability,
max_span_length=data_args.max_span_length,
)
# 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,
)
# 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))
try:
perplexity = math.exp(metrics["eval_loss"])
except OverflowError:
perplexity = float("inf")
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, "tasks": "language-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
trainer.push_to_hub(**kwargs)
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:
max_train_samples = min(len(train_dataset), data_args.max_train_samples)
train_dataset = train_dataset.select(range(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:
max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
eval_dataset = eval_dataset.select(range(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:
max_eval_samples = min(len(test_dataset), data_args.max_eval_samples)
test_dataset = test_dataset.select(range(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 test_trainer_iterable_dataset(self):
# Simulate Language Modeling with an IterableDataset, with no __len__ method
# Pick-up a tiny model, so it works on CPU
# See Issue #5990: https://github.com/huggingface/transformers/issues/5990
MODEL_ID = "sshleifer/tiny-distilbert-base-cased"
model = AutoModelForMaskedLM.from_pretrained(MODEL_ID)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
train_dataset = SampleIterableDataset(file_path=PATH_SAMPLE_TEXT,
tokenizer=tokenizer)
training_args = TrainingArguments(output_dir="./examples",
no_cuda=True,
max_steps=2)
data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer,
mlm=True,
mlm_probability=0.15)
training_args = TrainingArguments(output_dir="./examples",
no_cuda=True,
max_steps=2)
trainer = Trainer(model=model,
args=training_args,
train_dataset=train_dataset,
data_collator=data_collator)
trainer.train()
loader = trainer.get_train_dataloader()
self.assertIsInstance(loader, torch.utils.data.DataLoader)
self.assertIsInstance(
loader.sampler,
torch.utils.data.dataloader._InfiniteConstantSampler)
# Exception if giving iterable dataset and no max_steps
with self.assertRaises(ValueError):
training_args = TrainingArguments(output_dir="./examples",
no_cuda=True)
_ = Trainer(model=model,
args=training_args,
train_dataset=train_dataset,
data_collator=data_collator)
# Exception if eval_dataset is iterable in __init__
with self.assertRaises(ValueError):
training_args = TrainingArguments(output_dir="./examples",
no_cuda=True,
max_steps=2)
_ = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=train_dataset,
data_collator=data_collator,
)
# Exception if predicting with iterable dataset
with self.assertRaises(ValueError):
training_args = TrainingArguments(output_dir="./examples",
no_cuda=True)
trainer = Trainer(model=model,
args=training_args,
data_collator=data_collator)
trainer.predict(train_dataset)
# Exception if evaluating with iterable dataset
with self.assertRaises(ValueError):
training_args = TrainingArguments(output_dir="./examples",
no_cuda=True)
trainer = Trainer(model=model,
args=training_args,
data_collator=data_collator)
trainer.evaluate(train_dataset)
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)
try:
num_labels = glue_tasks_num_labels[data_args.task_name]
output_mode = glue_output_modes[data_args.task_name]
except KeyError:
raise ValueError("Task not found: %s" % (data_args.task_name))
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
finetuning_task=data_args.task_name,
cache_dir=model_args.cache_dir,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
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,
)
# Get datasets
print("###################PRINTING MODEL ARGS############################")
print(data_args)
train_dataset = (GlueDataset(
data_args, tokenizer=tokenizer, cache_dir=model_args.cache_dir)
if training_args.do_train else None)
eval_dataset = (GlueDataset(data_args,
tokenizer=tokenizer,
mode="dev",
cache_dir=model_args.cache_dir)
if training_args.do_eval else None)
test_dataset = (GlueDataset(data_args,
tokenizer=tokenizer,
mode="test",
cache_dir=model_args.cache_dir)
if training_args.do_predict else None)
def build_compute_metrics_fn(
task_name: str) -> Callable[[EvalPrediction], Dict]:
def compute_metrics_fn(p: EvalPrediction):
if output_mode == "classification":
preds = np.argmax(p.predictions, axis=1)
elif output_mode == "regression":
preds = np.squeeze(p.predictions)
return glue_compute_metrics(task_name, preds, p.label_ids)
return compute_metrics_fn
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
compute_metrics=build_compute_metrics_fn(data_args.task_name),
)
# 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
eval_results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
# Loop to handle MNLI double evaluation (matched, mis-matched)
eval_datasets = [eval_dataset]
if data_args.task_name == "mnli":
mnli_mm_data_args = dataclasses.replace(data_args,
task_name="mnli-mm")
eval_datasets.append(
GlueDataset(mnli_mm_data_args,
tokenizer=tokenizer,
mode="dev",
cache_dir=model_args.cache_dir))
for eval_dataset in eval_datasets:
print("#############DOING IT HERE###########################")
trainer.compute_metrics = build_compute_metrics_fn(
eval_dataset.args.task_name)
eval_result = trainer.evaluate(eval_dataset=eval_dataset)
output_eval_file = os.path.join(
training_args.output_dir,
f"eval_results_{eval_dataset.args.task_name}.txt")
if trainer.is_world_master():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results {} *****".format(
eval_dataset.args.task_name))
for key, value in eval_result.items():
logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
eval_results.update(eval_result)
if training_args.do_predict:
logging.info("*** Test ***")
test_datasets = [test_dataset]
if data_args.task_name == "mnli":
mnli_mm_data_args = dataclasses.replace(data_args,
task_name="mnli-mm")
test_datasets.append(
GlueDataset(mnli_mm_data_args,
tokenizer=tokenizer,
mode="test",
cache_dir=model_args.cache_dir))
counter = 0
for test_dataset in test_datasets:
predictions = trainer.predict(
test_dataset=test_dataset).predictions
print(
"####################PRINTING PREDS##############################"
)
print("shape of predictions", predictions.shape)
np.save("qnli_predicts_0_" + str(counter), predictions)
if output_mode == "classification":
predictions = np.argmax(predictions, axis=1)
output_test_file = os.path.join(
training_args.output_dir,
f"test_results_{test_dataset.args.task_name}.txt")
if trainer.is_world_master():
with open(output_test_file, "w") as writer:
logger.info("***** Test results {} *****".format(
test_dataset.args.task_name))
writer.write("index\tprediction\n")
for index, item in enumerate(predictions):
if output_mode == "regression":
writer.write("%d\t%3.3f\n" % (index, item))
else:
item = test_dataset.get_labels()[item]
writer.write("%d\t%s\n" % (index, item))
counter = counter + 1
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))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if training_args.do_train:
raise ValueError("Training not supported with this script.")
if data_args.eval_data_file is None:
raise ValueError(
"Cannot do evaluation without an evaluation data file.")
# 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)
# 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)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported, but you can do it from another script, save it,"
"and load it from here, using --tokenizer_name")
inline_meta = None
if model_args.controllable_model == 'ctrl-gpt2':
print("Loaded added tokens:", tokenizer.added_tokens_encoder)
inline_meta = TARGET_CTRL_CODE
assert inline_meta in tokenizer.get_vocab().keys()
print("Using inline meta:", inline_meta)
if model_args.model_name_or_path:
controllable_model = model_args.controllable_model
if controllable_model and controllable_model.startswith('affect-gpt2'):
logger.info("Using AffectLM for evaluation")
model = AffectGPT2LMHeadModel.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,
)
# Evaluate with non-toxic affect
beta = int(
re.search(r'beta-([0-9]+)', controllable_model).group(1))
logging.info(f"Setting AffectLM beta to {beta}")
model.affect.beta = beta
affect_label = F.one_hot(torch.LongTensor([AFFECT_LABEL_IDX]),
num_classes=AFFECT_NUM_LABELS).float().to(
training_args.device)
model.set_affect_labels(affect_label)
else:
model = AutoModelWithLMHead.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 = AutoModelWithLMHead.from_config(config)
model.resize_token_embeddings(len(tokenizer))
if config.model_type in ["bert", "roberta", "distilbert", "camembert"
] and not data_args.mlm:
raise ValueError(
"BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the --mlm "
"flag (masked language modeling).")
if data_args.block_size <= 0:
data_args.block_size = tokenizer.max_len
# Our input block size will be the max possible for the model
else:
data_args.block_size = min(data_args.block_size, tokenizer.max_len)
# Get datasets
eval_dataset = get_dataset(data_args,
tokenizer=tokenizer,
local_rank=training_args.local_rank,
inline_meta=inline_meta)
data_collator = DataCollatorForLanguageModeling(
tokenizer=tokenizer,
mlm=data_args.mlm,
mlm_probability=data_args.mlm_probability)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
data_collator=data_collator,
eval_dataset=eval_dataset,
prediction_loss_only=True,
)
# Evaluation
results = {}
if training_args.local_rank in [-1, 0]:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["loss"])
result = {"perplexity": perplexity}
output_eval_file = os.path.join(training_args.output_dir,
"eval_results_lm.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
results.update(result)
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, 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()
raw_datasets["train"] = load_dataset(data_args.dataset_name,
data_args.dataset_config_name,
split=data_args.train_split_name)
raw_datasets["eval"] = load_dataset(data_args.dataset_name,
data_args.dataset_config_name,
split=data_args.eval_split_name)
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)}.")
# prepare dataset
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))
# 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)
def remove_special_characters(batch):
if chars_to_ignore_regex is not None:
batch["target_text"] = re.sub(
chars_to_ignore_regex, "",
batch[data_args.text_column_name]).lower() + " "
else:
batch["target_text"] = batch[
data_args.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=[data_args.text_column_name],
desc="remove special characters from datasets",
)
# 3. Next, 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
vocab_file = os.path.join(training_args.output_dir, "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(training_args.output_dir, exist_ok=True)
vocab_dict = create_vocabulary_from_data(raw_datasets)
# save vocab dict to be loaded into tokenizer
with open(vocab_file, "w") as file:
json.dump(vocab_dict, file)
# 4. Now we can instantiate the configuration, 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 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)
# tokenizer is defined by `tokenizer_class` if present in config else by `model_type`
config_for_tokenizer = config if config.tokenizer_class is not None else None
tokenizer_type = config.model_type if config.tokenizer_class is None else None
# load feature_extractor, tokenizer and create processor
tokenizer = AutoTokenizer.from_pretrained(
training_args.output_dir,
config=config_for_tokenizer,
tokenizer_type=tokenizer_type,
unk_token="[UNK]",
pad_token="[PAD]",
word_delimiter_token="|",
use_auth_token=data_args.use_auth_token,
)
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)
processor = Wav2Vec2Processor(feature_extractor=feature_extractor,
tokenizer=tokenizer)
# 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,
"gradient_checkpointing": training_args.gradient_checkpointing,
"layerdrop": model_args.layerdrop,
"ctc_loss_reduction": model_args.ctc_loss_reduction,
"pad_token_id": processor.tokenizer.pad_token_id,
"vocab_size": len(processor.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_extractor:
model.freeze_feature_extractor()
# 5. 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
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 * processor.feature_extractor.sampling_rate
min_input_length = data_args.min_duration_in_seconds * processor.feature_extractor.sampling_rate
# Preprocessing the datasets.
# We need to read the audio files as arrays and tokenize the targets.
def prepare_dataset(batch):
# load audio
sample = batch[data_args.audio_column_name]
batch["input_values"] = processor(
sample["array"],
sampling_rate=sample["sampling_rate"],
truncate=True,
max_length=max_input_length).input_values[0]
batch["input_length"] = len(batch["input_values"])
# Setup the processor for targets
with processor.as_target_processor():
batch["labels"] = processor(batch["target_text"]).input_ids
return batch
with training_args.main_process_first(desc="dataset map preprocessing"):
vectorized_datasets = raw_datasets.map(
prepare_dataset,
remove_columns=raw_datasets["train"].column_names,
num_proc=data_args.preprocessing_num_workers,
desc="preprocess datasets",
)
if min_input_length > 0.0:
# filter data that is shorter than min_input_length
vectorized_datasets = vectorized_datasets.filter(
lambda x: x > min_input_length,
num_proc=data_args.preprocessing_num_workers,
input_columns=["input_length"],
)
vectorized_datasets = vectorized_datasets.remove_columns(
"input_length")
# 6. 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 Metric during training
wer_metric = load_metric("wer")
# 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] = processor.tokenizer.pad_token_id
pred_str = processor.batch_decode(pred_ids)
# we do not want to group tokens when computing the metrics
label_str = processor.batch_decode(pred.label_ids, group_tokens=False)
wer = wer_metric.compute(predictions=pred_str, references=label_str)
return {"wer": wer}
# 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=processor.feature_extractor,
)
# 7. 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
# Save the feature_extractor and the tokenizer
if is_main_process(training_args.local_rank):
processor.save_pretrained(training_args.output_dir)
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()
# 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}")
# 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."
)
# Initialize our dataset and prepare it for the 'image-classification' task.
ds = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
data_files=data_args.data_files,
cache_dir=model_args.cache_dir,
task="image-classification",
)
# Define torchvision transforms to be applied to each image.
normalize = Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
_train_transforms = Compose(
[
RandomResizedCrop(data_args.image_size),
RandomHorizontalFlip(),
ToTensor(),
normalize,
]
)
_val_transforms = Compose(
[
Resize(data_args.image_size),
CenterCrop(data_args.image_size),
ToTensor(),
normalize,
]
)
def train_transforms(example_batch):
"""Apply _train_transforms across a batch."""
example_batch["pixel_values"] = [_train_transforms(pil_loader(f)) for f in example_batch["image_file_path"]]
return example_batch
def val_transforms(example_batch):
"""Apply _val_transforms across a batch."""
example_batch["pixel_values"] = [_val_transforms(pil_loader(f)) for f in example_batch["image_file_path"]]
return example_batch
# If we don't have a validation split, split off a percentage of train as validation.
data_args.train_val_split = None if "validation" in ds.keys() else data_args.train_val_split
if isinstance(data_args.train_val_split, float) and data_args.train_val_split > 0.0:
split = ds["train"].train_test_split(data_args.train_val_split)
ds["train"] = split["train"]
ds["validation"] = split["test"]
# Prepare label mappings.
# We'll include these in the model's config to get human readable labels in the Inference API.
labels = ds["train"].features["labels"].names
label2id, id2label = dict(), dict()
for i, label in enumerate(labels):
label2id[label] = str(i)
id2label[str(i)] = label
# Load the accuracy metric from the datasets package
metric = datasets.load_metric("accuracy")
# Define our 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):
"""Computes accuracy on a batch of predictions"""
return metric.compute(predictions=np.argmax(p.predictions, axis=1), references=p.label_ids)
config = AutoConfig.from_pretrained(
model_args.config_name or model_args.model_name_or_path,
num_labels=len(labels),
label2id=label2id,
id2label=id2label,
finetuning_task="image-classification",
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForImageClassification.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,
)
# NOTE - We aren't directly using this feature extractor since we defined custom transforms above.
# We initialize this instance below and pass it to Trainer to ensure that the feature extraction
# config, preprocessor_config.json, is included in output directories.
# This way if we push a model to the hub, the inference widget will work.
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,
size=data_args.image_size,
image_mean=normalize.mean,
image_std=normalize.std,
)
if training_args.do_train:
if "train" not in ds:
raise ValueError("--do_train requires a train dataset")
if data_args.max_train_samples is not None:
ds["train"] = ds["train"].shuffle(seed=training_args.seed).select(range(data_args.max_train_samples))
# Set the training transforms
ds["train"].set_transform(train_transforms)
if training_args.do_eval:
if "validation" not in ds:
raise ValueError("--do_eval requires a validation dataset")
if data_args.max_eval_samples is not None:
ds["validation"] = (
ds["validation"].shuffle(seed=training_args.seed).select(range(data_args.max_eval_samples))
)
# Set the validation transforms
ds["validation"].set_transform(val_transforms)
# Initalize our trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=ds["train"] if training_args.do_train else None,
eval_dataset=ds["validation"] if training_args.do_eval else None,
compute_metrics=compute_metrics,
tokenizer=feature_extractor,
data_collator=collate_fn,
)
# 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()
# Evaluation
if training_args.do_eval:
metrics = trainer.evaluate()
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Write model card and (optionally) push to hub
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": "image-classification",
"dataset": data_args.dataset_name,
"tags": ["image-classification"],
}
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
def train(args):
wandb.login()
seed_everything(args.seed)
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
MODEL_NAME = args.model_name
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
# load dataset
train_dataset_dir = "/opt/ml/code2/data/train_new_tag.tsv"
train_data = load_dataset(train_dataset_dir)
train_x = list(train_data.iloc[:, 0])
train_y = list(train_data.iloc[:, -1])
valid_dataset_dir = "/opt/ml/code2/data/valid_tag.tsv"
valid_data = load_dataset(valid_dataset_dir)
val_x = list(valid_data.iloc[:, 0])
val_y = list(valid_data.iloc[:, -1])
# tokenize datasets
tokenized_train = tokenized_dataset(train_x, tokenizer)
tokenized_val = tokenized_dataset(val_x, tokenizer)
# make dataset for pytorch
RE_train_dataset = RE_Dataset(tokenized_train, train_y)
RE_valid_dataset = RE_Dataset(tokenized_val, val_y)
# instantiate pretrained language model
model = AutoModelForSequenceClassification.from_pretrained(MODEL_NAME,
num_labels=8)
model.resize_token_embeddings(len(tokenizer))
model.to(device)
# optimizer and scheduler
# optimizer = AdamW(model.parameters(), lr=args.lr, weight_decay=args.weight_decay)
# scheduler = get_linear_schedule_with_warmup(optimizer=optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=300*args.epochs)
# callbacks
early_stopping = EarlyStoppingCallback(
early_stopping_patience=args.early_stopping_patience,
early_stopping_threshold=0.00005)
training_args = TrainingArguments(
output_dir='./results',
logging_dir='./logs',
logging_steps=100,
save_total_limit=1,
evaluation_strategy='steps',
eval_steps=100,
load_best_model_at_end=True,
metric_for_best_model='accuracy',
greater_is_better=True,
dataloader_num_workers=args.num_workers,
fp16=True,
seed=args.seed,
run_name=args.run_name,
num_train_epochs=args.epochs,
per_device_train_batch_size=args.train_batch_size,
per_device_eval_batch_size=args.eval_batch_size,
label_smoothing_factor=args.label_smoothing_factor,
learning_rate=args.lr,
warmup_steps=args.warmup_steps,
weight_decay=args.weight_decay,
)
trainer = Trainer(
model=model, # the instantiated 🤗 Transformers model to be trained
args=training_args, # training arguments, defined above
train_dataset=RE_train_dataset, # training dataset
eval_dataset=RE_valid_dataset, # evaluation dataset
tokenizer=tokenizer,
compute_metrics=compute_metrics, # define metrics function
# optimizers=[optimizer, scheduler],
callbacks=[early_stopping])
# train model
trainer.train()
wandb.finish()
output_dir = 'hf_' + is_grad + f'{model_name}' + f'_{exprmt_ds}' + f'_{traning_samples}'
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
args = TrainingArguments(
output_dir=output_dir,
evaluation_strategy='epoch',
learning_rate=2e-5,
per_device_train_batch_size=batch_size,
per_device_eval_batch_size=batch_size,
num_train_epochs=10,
weight_decay=0.01,
model_parallel=True
) # model_parallel should use in conjunction with model.to('cuda')
trainer = Trainer(model.to(device),
args,
train_dataset=tokenized_datasets['train'],
eval_dataset=tokenized_datasets['train'],
data_collator=data_collator,
tokenizer=tokenizer,
compute_metrics=compute_metrics)
trainer.train()
trainer.save_model(args.output_dir)
# eval_result = trainer.evaluate()
# print(eval_result)
predictions, labels, _ = trainer.predict(tokenized_datasets['test'])
predictions = np.argmax(predictions, axis=2)
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
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,
)
# Add a number token
num_added_toks = tokenizer.add_tokens(["<NUM>"])
print("We have added", num_added_toks, "tokens")
model.resize_token_embeddings(len(tokenizer))
# 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)
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.data_dir, "test.txt"), "r") as f:
example_id = 0
for line in f:
if line.startswith("-DOCSTART-") or line == "" or line == "\n":
writer.write(line)
if not preds_list[example_id]:
example_id += 1
elif preds_list[example_id]:
output_line = (
line.split()[0]
+ " "
+ preds_list[example_id].pop(0)
+ "\n"
)
writer.write(output_line)
else:
logger.warning(
"Maximum sequence length exceeded: No prediction for '%s'.",
line.split()[0],
)
return results
def compute_metrics(p: EvalPrediction) -> Dict:
sequential = list(range(len(dataset)))
success = p.predictions.tolist(
) == sequential and p.label_ids.tolist() == sequential
if not success and training_args.local_rank == 0:
logger.warning(
"Predictions and/or labels do not match expected results:\n - predictions: "
f"{p.predictions.tolist()}\n - labels: {p.label_ids.tolist()}\n - expected: {sequential}"
)
return {"success": success}
trainer = Trainer(
model=DummyModel(),
args=training_args,
data_collator=DummyDataCollator(),
eval_dataset=dataset,
compute_metrics=compute_metrics,
)
metrics = trainer.evaluate()
logger.info(metrics)
if metrics["eval_success"] is not True:
logger.error(metrics)
exit(1)
p = trainer.predict(dataset)
logger.info(p.metrics)
if p.metrics["eval_success"] is not True:
logger.error(p.metrics)
exit(1)
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(
)
checkpoint_dir = hyperparam_path_for_initializing_evidence_selector(
model_args, data_args, training_args)
ckpt_dir = Path(checkpoint_dir)
postfix = ""
if training_args.do_train:
postfix += "_train"
elif training_args.do_eval:
postfix += "_eval"
setup_root_logger(ckpt_dir,
training_args.local_rank,
debug=False,
postfix=postfix)
training_args.output_dir = checkpoint_dir
# 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).
if data_args.dataset not in ['race', 'dream']:
raise ValueError("Dataset should be race or dream.")
else:
if data_args.dataset == 'race':
from mcmrc.data_utils.processors import prepare_features_for_reading_evidence
if data_args.dataset == 'dream':
pass
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
# 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.
data_files = {}
data_files[
'train'] = data_args.train_file if data_args.train_file is not None else None
data_files[
'validation'] = data_args.validation_file if data_args.validation_file is not None else None
data_files[
'test'] = data_args.test_file if data_args.test_file is not None else None
datasets = load_dataset(
data_args.dataload_script,
data_args.dataload_split,
data_files=data_files if data_files['train'] is not None else None,
data_dir=data_args.data_dir)
# 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,
use_fast=model_args.use_fast_tokenizer,
)
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,
)
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
all_pseudo_label = load_pseudo_label(data_args.pseudo_label_path)
if data_args.run_pseudo_label_with_options:
pseudo_logit = all_pseudo_label['options_prob_diff']
else:
pseudo_logit = all_pseudo_label['logit']
acc = all_pseudo_label['acc']
# Data collator
data_collator = 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()
}
eval_on_dev = (data_args.eval_dataset == "all" or data_args.eval_dataset
== "dev") and training_args.do_eval
eval_on_test = (data_args.eval_dataset == "all" or data_args.eval_dataset
== "test") and training_args.do_eval
train_results = {}
eval_results = {}
test_results = {}
for evidence_num in range(1, data_args.max_evidence_len + 1):
pprepare_features_for_using_pseudo_label_as_evidence = partial(
prepare_features_for_reading_evidence,
run_pseudo_label_with_options=data_args.
run_pseudo_label_with_options,
evidence_logits=pseudo_logit,
evidence_len=evidence_num,
tokenizer=tokenizer,
data_args=data_args)
tokenized_datasets = datasets.map(
pprepare_features_for_using_pseudo_label_as_evidence,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# 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,
)
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
# 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,
f"evidence_{evidence_num}_trainer_state.json"))
for key in list(train_result.metric.keys()):
train_results[
f'evidence{evidence_num}_{key}'] = train_result.metric[key]
if eval_on_dev:
logger.info("*** Evaluate ***")
results = trainer.evaluate(
eval_dataset=tokenized_datasets["validation"])
for key in list(results.keys()):
eval_results[f'evidence{evidence_num}_{key}'] = results[key]
if eval_on_test:
logger.info("*** Test ***")
results = trainer.evaluate(eval_dataset=tokenized_datasets["test"])
for key in list(results.keys()):
test_results[f'evidence{evidence_num}_{key}'] = results[key]
if eval_on_dev:
output_eval_file = os.path.join(training_args.output_dir,
"eval_results.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in sorted(eval_results.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
if eval_on_test:
output_test_file = os.path.join(training_args.output_dir,
"test_results.txt")
with open(output_test_file, "w") as writer:
logger.info("***** Test results *****")
for key, value in sorted(test_results.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
if training_args.do_train:
output_train_file = os.path.join(training_args.output_dir,
"train_results.txt")
with open(output_train_file, "w") as writer:
logger.info("***** Train results *****")
for key, value in sorted(train_results.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
def process_questions(train_csv,
val_csv,
test_csv,
train_output,
val_output,
test_output,
vocab_path=None,
fine_tune_out_path=None,
wandb_log=True,
model_name='bert-base-uncased'):
''' Encode question tokens'''
print('Loading tokenizer')
tokenizer = AutoTokenizer.from_pretrained(model_name)
print('Load data')
tokenized_datasets = load_dataset('csv',
delimiter="\t",
data_files={
'train': train_csv,
'val': val_csv,
'test': test_csv
})
print('Tokenizing questions')
tokenized_datasets = tokenized_datasets.map(tokenize_function(
tokenizer, 'question'),
batched=True,
remove_columns=["question"])
model_training_datasets = tokenized_datasets
print('Load Vocab')
if (vocab_path):
with open(vocab_path, 'r') as f:
vocab = json.load(f)
else:
vocab = create_vocab(train_annotation_json)
print('Tokenizing answers')
tokenized_datasets = tokenized_datasets.map(tokenize_answer(vocab),
batched=False,
remove_columns=["answer"])
print('Renaming fields')
if (model_name == 'bert-base-uncased'):
tokenized_datasets = tokenized_datasets.map(
lambda instance: {
'question_id': instance['id'],
'video_ids': instance['key'],
'video_name': instance['gif_name'],
'question_tokens': instance['input_ids'],
'question_attention_mask': instance['attention_mask'],
'question_token_type_ids': instance['token_type_ids']
},
batched=True,
remove_columns=[
'id', 'input_ids', 'attention_mask', 'token_type_ids',
'description', 'key', 'type', 'vid_id'
])
elif (model_name == 'roberta-base'
or model_name == 'distilbert-base-uncased'):
tokenized_datasets = tokenized_datasets.map(
lambda instance: {
'question_id': instance['id'],
'video_ids': instance['key'],
'video_name': instance['gif_name'],
'question_tokens': instance['input_ids'],
'question_attention_mask': instance['attention_mask']
},
batched=True,
remove_columns=[
'id', 'input_ids', 'attention_mask', 'description', 'key',
'type', 'vid_id'
])
print('Saving datasets')
with open(train_output, 'wb') as f:
pickle.dump(tokenized_datasets['train'], f)
with open(val_output, 'wb') as f:
pickle.dump(tokenized_datasets['val'], f)
with open(test_output, 'wb') as f:
pickle.dump(tokenized_datasets['test'], f)
if (fine_tune_out_path):
print('Finetuning Masked LM Bert model with train questions')
model = AutoModelForMaskedLM.from_pretrained(model_name)
data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer,
mlm_probability=0.15)
training_args = TrainingArguments('test-clm',
per_device_train_batch_size=64,
evaluation_strategy="epoch",
learning_rate=2e-5,
weight_decay=0.01)
trainer = Trainer(model=model,
args=training_args,
train_dataset=model_training_datasets["train"],
eval_dataset=model_training_datasets["val"],
data_collator=data_collator)
if (wandb_log):
import wandb
perplexity_pretrained_train = trainer.evaluate(
model_training_datasets["train"])['eval_loss']
wandb.log({'hf_perplexity_train': perplexity_pretrained_train})
perplexity_pretrained_val = trainer.evaluate(
model_training_datasets["val"])['eval_loss']
wandb.log({'hf_perplexity_val': perplexity_pretrained_val})
trainer.train()
eval_perplexity = trainer.evaluate()['eval_loss']
print(
f"Model finetuned with validation perpexity of {eval_perplexity}")
if (wandb_log):
perplexity_finetuned_train = trainer.evaluate(
model_training_datasets["train"])['eval_loss']
wandb.log(
{'finetuned_perplexity_train': perplexity_finetuned_train})
perplexity_finetuned_val = trainer.evaluate(
model_training_datasets["val"])['eval_loss']
wandb.log({'finetuned_perplexity_val': perplexity_finetuned_val})
print('Saving Model')
model.save_pretrained(save_directory=fine_tune_out_path)
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))
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))
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 run(self):
"""Run experiments"""
result_collector = ResultCollector(self.dataset_name,
self.experiment_type)
normalized_encoder, normalized_decoder, number_of_labels = self.encode_labels(
)
config = RobertaConfig.from_pretrained("roberta-base")
config.num_labels = number_of_labels
tokenizer, model = utils.provide_model_and_tokenizer(
self.parameter['model_name'],
self.parameter['pretrained_model_or_path'],
config=config)
tf_ds = {}
for key in self.dataset:
df_ds = self.dataset[key]
if self.test:
# load only subset of the data
df_ds = df_ds[:10]
self.logger.warning(
'Run in test mode - dataset reduced to 10 records!')
if self.parameter['description']:
texts = list(
(df_ds['title'] + ' - ' + df_ds['description']).values)
else:
texts = df_ds['title'].values
if self.parameter['preprocessing'] == True:
texts = [preprocess(value) for value in texts]
# Normalize label values
labels = [
value.replace(' ', '_') for value in df_ds['category'].values
]
tf_ds[key] = CategoryDatasetRNN(texts, labels, tokenizer,
normalized_encoder)
timestamp = time.time()
string_timestamp = datetime.fromtimestamp(timestamp).strftime(
'%Y-%m-%d_%H-%M-%S')
training_args = TrainingArguments(
output_dir='{}/models/{}/transformers/model/{}'.format(
self.data_dir, self.dataset_name,
self.parameter['experiment_name']),
# output directory
num_train_epochs=self.
parameter['epochs'], # total # of training epochs
learning_rate=self.parameter['learning_rate'],
per_device_train_batch_size=self.
parameter['per_device_train_batch_size'],
# batch size per device during training
per_device_eval_batch_size=64, # batch size for evaluation
warmup_steps=
500, # number of warmup steps for learning rate scheduler
weight_decay=self.
parameter['weight_decay'], # strength of weight decay
logging_dir='{}/models/{}/transformers/logs-{}'.format(
self.data_dir, self.dataset_name, string_timestamp),
# directory for storing logs
save_total_limit=5, # Save only the last 5 Checkpoints
metric_for_best_model=self.parameter['metric_for_best_model'],
load_best_model_at_end=True,
gradient_accumulation_steps=self.
parameter['gradient_accumulation_steps'],
seed=self.parameter['seed'],
disable_tqdm=True)
evaluator = scorer.HierarchicalScorer(
self.parameter['experiment_name'],
self.tree,
transformer_decoder=normalized_decoder)
trainer = Trainer(
model=
model, # the instantiated 🤗 Transformers model to be trained
args=training_args, # training arguments, defined above
train_dataset=tf_ds[
'train'], # tensorflow_datasets training dataset
eval_dataset=tf_ds[
'validate'], # tensorflow_datasets evaluation dataset
compute_metrics=evaluator.compute_metrics_transformers_rnn)
self.logger.info('Start training!')
trainer.train()
for split in ['train', 'validate', 'test']:
result_collector.results['{}+{}'.format(self.parameter['experiment_name'], split)] \
= trainer.evaluate(tf_ds[split])
trainer.save_model()
# Persist results
result_collector.persist_results(timestamp)
from transformers import TrainingArguments
training_args = TrainingArguments(
output_dir="./wav2vec2-base-timit-demo",
group_by_length=True,
per_device_train_batch_size=32,
evaluation_strategy="steps",
num_train_epochs=1,
fp16=True,
save_steps=500,
eval_steps=500,
logging_steps=500,
learning_rate=1e-4,
weight_decay=0.005,
warmup_steps=1000,
save_total_limit=2,
)
from transformers import Trainer
trainer = Trainer(
model=model,
data_collator=data_collator,
args=training_args,
compute_metrics=compute_metrics,
train_dataset=timit_prepared["train"],
eval_dataset=timit_prepared["test"],
tokenizer=processor.feature_extractor,
)
trainer.train()
def main():
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if data_args.eval_data_file is None and training_args.do_eval:
raise ValueError(
"Cannot do evaluation without an evaluation data file. Either supply a file to --eval_data_file "
"or remove the --do_eval argument.")
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_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(training_args.seed)
model = LongformerForMaskedLM.from_pretrained(
'allenai/longformer-base-4096')
tokenizer = LongformerTokenizer.from_pretrained(
'allenai/longformer-base-4096')
tokenizer.add_tokens(['<doc-s>'], special_tokens=True)
tokenizer.add_tokens(['</doc-s>'], special_tokens=True)
data_args.block_size = 4096
train_dataset = get_dataset(data_args,
tokenizer=tokenizer,
local_rank=training_args.local_rank)
model.resize_token_embeddings(len(tokenizer))
data_collator = DataCollatorForLanguageModeling(
tokenizer=tokenizer,
mlm=data_args.mlm,
mlm_probability=data_args.mlm_probability,
globalize_special_tokens=data_args.globalize_special_tokens)
trainer = Trainer(
model=model,
args=training_args,
data_collator=data_collator,
train_dataset=train_dataset,
# eval_dataset=eval_dataset,
prediction_loss_only=True,
)
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)
if trainer.is_world_master():
tokenizer.save_pretrained(training_args.output_dir)
results = {}
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["loss"])
result = {"perplexity": perplexity}
output_eval_file = os.path.join(training_args.output_dir,
"eval_results_lm.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
results.update(result)
return results
def treinar():
"""
Treina o modelo ("fine-tunind"), utilizando como modelo pré-treinado o BERTimbau - Portuguese BERT, da Neuralmind
(https://github.com/neuralmind-ai/portuguese-bert).
"""
unique_tags, train_dataset, val_dataset, tokenizer, id2tag, tag2id = __processar_base(
)
configuracao = AutoConfig.from_pretrained(str(
config.subdiretorio_modelo_neuralmind_bert_base),
id2label=id2tag,
label2id=tag2id)
model = BertForTokenClassification.from_pretrained(
config.subdiretorio_modelo_neuralmind_bert_base, config=configuracao)
training_args = TrainingArguments(
output_dir='./results', # output directory
num_train_epochs=3, # total number of training epochs
# per_device_train_batch_size=16, # batch size per device during training
# per_device_eval_batch_size=64, # batch size for evaluation
per_device_train_batch_size=1, # batch size per device during training
per_device_eval_batch_size=1, # batch size for evaluation
warmup_steps=500, # number of warmup steps for learning rate scheduler
weight_decay=0.01, # strength of weight decay
logging_steps=10)
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(id2tag[label_ids[i][j]])
preds_list[i].append(id2tag[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 {
"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),
}
trainer = Trainer(
model=model, # the instantiated 🤗 Transformers model to be trained
args=training_args, # training arguments, defined above
train_dataset=train_dataset, # training dataset
eval_dataset=val_dataset, # evaluation dataset,
compute_metrics=compute_metrics)
trainer.train()
logging.basicConfig(level=logging.INFO)
trainer.evaluate(val_dataset)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, GlueDataTrainingArguments, TrainingArguments))
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)
try:
num_labels = glue_tasks_num_labels[data_args.task_name]
output_mode = glue_output_modes[data_args.task_name]
except KeyError:
raise ValueError("Task not found: %s" % (data_args.task_name))
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
finetuning_task=data_args.task_name,
cache_dir=model_args.cache_dir,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
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,
)
# Get datasets
train_dataset = (
GlueDataset(data_args, tokenizer=tokenizer, local_rank=training_args.local_rank)
if training_args.do_train
else None
)
eval_dataset = (
GlueDataset(data_args, tokenizer=tokenizer, local_rank=training_args.local_rank, evaluate=True)
if training_args.do_eval
else None
)
def compute_metrics(p: EvalPrediction) -> Dict:
return glue_compute_metrics(data_args.task_name, output_mode, p.predictions, p.label_ids)
# 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 and training_args.local_rank in [-1, 0]:
logger.info("*** Evaluate ***")
# Loop to handle MNLI double evaluation (matched, mis-matched)
eval_datasets = [eval_dataset]
if data_args.task_name == "mnli":
mnli_mm_data_args = dataclasses.replace(data_args, task_name="mnli-mm")
eval_datasets.append(
GlueDataset(mnli_mm_data_args, tokenizer=tokenizer, local_rank=training_args.local_rank, evaluate=True)
)
for eval_dataset in eval_datasets:
full_result = trainer.evaluate(eval_dataset=eval_dataset)
from pandas import DataFrame
res = DataFrame()
res['scores'] = full_result.predictions[:, 1]
res['labels'] = full_result.label_ids
output_eval_file0 = os.path.join(
training_args.output_dir, f"eval_results_{eval_dataset.args.task_name}.csv"
)
res.to_csv(output_eval_file0)
result = full_result.metrics
output_eval_file = os.path.join(
training_args.output_dir, f"eval_results_{eval_dataset.args.task_name}.txt"
)
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results {} *****".format(eval_dataset.args.task_name))
for key, value in result.items():
logger.info(" {:<7} = {:.3f}".format(key, value))
writer.write("%s = %s\n" % (key, value))
results.update(result)
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, 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.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
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]
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.
if training_args.do_train:
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
label_keys = list(range(len(label_list)))
else:
label_list = get_label_list(raw_datasets["train"][label_column_name])
label_keys = 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_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,
use_fast=True,
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,
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#supported-frameworks to find the model types that meet this "
"requirement")
if model.config.label2id != PretrainedConfig(
num_labels=num_labels).label2id:
label_name_to_id = {k: v for k, v in model.config.label2id.items()}
if list(sorted(label_name_to_id.keys())) == list(sorted(label_list)):
label_to_id = {k: int(label_name_to_id[k]) for k in label_keys}
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.",
)
else:
label_to_id = {k: i for i, k in enumerate(label_keys)}
model.config.label2id = label_to_id
model.config.id2label = {i: l for l, i in label_to_id.items()}
# Map that sends B-Xxx label to its I-Xxx counterpart
b_to_i_label = []
for idx, label in enumerate(label_list):
if label.startswith("B-") and label.replace("B-", "I-") in label_list:
b_to_i_label.append(label_list.index(label.replace("B-", "I-")))
else:
b_to_i_label.append(idx)
# 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,
max_length=data_args.max_seq_length,
# 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:
if data_args.label_all_tokens:
label_ids.append(
b_to_i_label[label_to_id[label[word_idx]]])
else:
label_ids.append(-100)
previous_word_idx = word_idx
labels.append(label_ids)
tokenized_inputs["labels"] = labels
return tokenized_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(
tokenize_and_align_labels,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on train dataset",
)
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(
tokenize_and_align_labels,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on validation dataset",
)
if training_args.do_predict:
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(
tokenize_and_align_labels,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on prediction dataset",
)
# 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=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)
metrics = train_result.metrics
trainer.save_model() # Saves the tokenizer too for easy upload
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)
# Predict
if training_args.do_predict:
logger.info("*** Predict ***")
predictions, labels, metrics = trainer.predict(
predict_dataset, metric_key_prefix="predict")
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)]
trainer.log_metrics("predict", metrics)
trainer.save_metrics("predict", metrics)
# Save predictions
output_predictions_file = os.path.join(training_args.output_dir,
"predictions.txt")
if trainer.is_world_process_zero():
with open(output_predictions_file, "w") as writer:
for prediction in true_predictions:
writer.write(" ".join(prediction) + "\n")
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": "token-classification"
}
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)
file_path="data/train.txt",
block_size=128,
)
from transformers import DataCollatorForLanguageModeling
data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer,
mlm=True,
mlm_probability=0.15)
from transformers import Trainer, TrainingArguments
training_args = TrainingArguments(
output_dir="./output",
overwrite_output_dir=True,
num_train_epochs=1,
per_device_train_batch_size=64,
save_steps=10_000,
save_total_limit=2,
prediction_loss_only=True,
)
trainer = Trainer(
model=model,
args=training_args,
data_collator=data_collator,
train_dataset=dataset,
)
trainer.train()
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)],
)
logger.setLevel(logging.INFO if training_args.should_log 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 training_args.should_log:
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}")
# 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.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.")
if model_args.config_overrides is not None:
logger.info(f"Overriding config: {model_args.config_overrides}")
config.update_from_string(model_args.config_overrides)
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:
model = AutoModelForCausalLM.from_config(config)
n_params = sum(dict((p.data_ptr(), p.numel()) for p in model.parameters()).values())
logger.info(f"Training new model from scratch - Total size={n_params/2**20:.2f}M params")
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:
output = tokenizer(examples[text_column_name])
# 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()}
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))
try:
perplexity = math.exp(metrics["eval_loss"])
except OverflowError:
perplexity = float("inf")
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, "tasks": "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(
)
# 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}")
# 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."
)
# Initialize our dataset.
ds = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
data_files=data_args.data_files,
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
# If we don't have a validation split, split off a percentage of train as validation.
data_args.train_val_split = None if "validation" in ds.keys(
) else data_args.train_val_split
if isinstance(data_args.train_val_split,
float) and data_args.train_val_split > 0.0:
split = ds["train"].train_test_split(data_args.train_val_split)
ds["train"] = split["train"]
ds["validation"] = split["test"]
# Create config
# 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_or_path:
config = AutoConfig.from_pretrained(model_args.config_name_or_path,
**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.")
if model_args.config_overrides is not None:
logger.info(f"Overriding config: {model_args.config_overrides}")
config.update_from_string(model_args.config_overrides)
logger.info(f"New config: {config}")
# make sure the decoder_type is "simmim" (only relevant for BEiT)
if hasattr(config, "decoder_type"):
config.decoder_type = "simmim"
# adapt config
model_args.image_size = model_args.image_size if model_args.image_size is not None else config.image_size
model_args.patch_size = model_args.patch_size if model_args.patch_size is not None else config.patch_size
model_args.encoder_stride = (model_args.encoder_stride
if model_args.encoder_stride is not None else
config.encoder_stride)
config.update({
"image_size": model_args.image_size,
"patch_size": model_args.patch_size,
"encoder_stride": model_args.encoder_stride,
})
# create feature extractor
if model_args.feature_extractor_name:
feature_extractor = AutoFeatureExtractor.from_pretrained(
model_args.feature_extractor_name, **config_kwargs)
elif model_args.model_name_or_path:
feature_extractor = AutoFeatureExtractor.from_pretrained(
model_args.model_name_or_path, **config_kwargs)
else:
FEATURE_EXTRACTOR_TYPES = {
conf.model_type: feature_extractor_class
for conf, feature_extractor_class in
FEATURE_EXTRACTOR_MAPPING.items()
}
feature_extractor = FEATURE_EXTRACTOR_TYPES[model_args.model_type]()
# create model
if model_args.model_name_or_path:
model = AutoModelForMaskedImageModeling.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 = AutoModelForMaskedImageModeling.from_config(config)
if training_args.do_train:
column_names = ds["train"].column_names
else:
column_names = ds["validation"].column_names
if data_args.image_column_name is not None:
image_column_name = data_args.image_column_name
elif "image" in column_names:
image_column_name = "image"
elif "img" in column_names:
image_column_name = "img"
else:
image_column_name = column_names[0]
# transformations as done in original SimMIM paper
# source: https://github.com/microsoft/SimMIM/blob/main/data/data_simmim.py
transforms = Compose([
Lambda(lambda img: img.convert("RGB") if img.mode != "RGB" else img),
RandomResizedCrop(model_args.image_size,
scale=(0.67, 1.0),
ratio=(3.0 / 4.0, 4.0 / 3.0)),
RandomHorizontalFlip(),
ToTensor(),
Normalize(mean=feature_extractor.image_mean,
std=feature_extractor.image_std),
])
# create mask generator
mask_generator = MaskGenerator(
input_size=model_args.image_size,
mask_patch_size=data_args.mask_patch_size,
model_patch_size=model_args.patch_size,
mask_ratio=data_args.mask_ratio,
)
def preprocess_images(examples):
"""Preprocess a batch of images by applying transforms + creating a corresponding mask, indicating
which patches to mask."""
examples["pixel_values"] = [
transforms(image) for image in examples[image_column_name]
]
examples["mask"] = [
mask_generator() for i in range(len(examples[image_column_name]))
]
return examples
if training_args.do_train:
if "train" not in ds:
raise ValueError("--do_train requires a train dataset")
if data_args.max_train_samples is not None:
ds["train"] = ds["train"].shuffle(seed=training_args.seed).select(
range(data_args.max_train_samples))
# Set the training transforms
ds["train"].set_transform(preprocess_images)
if training_args.do_eval:
if "validation" not in ds:
raise ValueError("--do_eval requires a validation dataset")
if data_args.max_eval_samples is not None:
ds["validation"] = (ds["validation"].shuffle(
seed=training_args.seed).select(
range(data_args.max_eval_samples)))
# Set the validation transforms
ds["validation"].set_transform(preprocess_images)
# Initialize our trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=ds["train"] if training_args.do_train else None,
eval_dataset=ds["validation"] if training_args.do_eval else None,
tokenizer=feature_extractor,
data_collator=collate_fn,
)
# 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()
# Evaluation
if training_args.do_eval:
metrics = trainer.evaluate()
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Write model card and (optionally) push to hub
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": "masked-image-modeling",
"dataset": data_args.dataset_name,
"tags": ["masked-image-modeling"],
}
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(
)
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 is None:
block_size = tokenizer.model_max_length
if block_size > 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 --block_size xxx."
)
block_size = 1024
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:
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_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 train(epoch, vocab_size, train_files_path, save_path, learning_rate,
save_steps, per_gpu_train_batch_size, gradient_accumulation_steps):
"""
从头训练一个语言模型RoBERTa.
"""
from transformers import RobertaConfig
config = RobertaConfig( #可调整参数
vocab_size=vocab_size,
max_position_embeddings=514,
num_attention_heads=12,
num_hidden_layers=12,
type_vocab_size=1,
)
# 不训练，直接用BERT的分词器
from transformers import BertTokenizer
tokenizer = BertTokenizer.from_pretrained("hfl/chinese-roberta-wwm-ext",
max_len=512)
# 事实上，hfl/chinese-roberta-wwm-ext/vocab.txt == chinese_L-12_H-768_A-12/vocab.txt #词典
from transformers import RobertaForMaskedLM
"""
源码：
class RobertaForMaskedLM()
def __init__(self, config):
super().__init__(config)
if not config.is_decoder:
logger.warning("If you want to use `RobertaLMHeadModel` as a standalone, add `is_decoder=True.`")
self.roberta = RobertaModel(config, add_pooling_layer=False)
self.lm_head = RobertaLMHead(config)
self.init_weights()
可以发现，RobertaForMaskedLM包含了roberta主体模型(不含pooling层)和一个语言模型输出层
"""
# 初始化模型
model = RobertaForMaskedLM(config=config)
print("参数数量: ", model.num_parameters()) # 1亿参数
dataset = TextDataset(
tokenizer=tokenizer,
file_path=train_files_path,
block_size=512,
)
from transformers import DataCollatorForLanguageModeling
data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer,
mlm=True,
mlm_probability=0.15)
from transformers import Trainer, TrainingArguments
training_args = TrainingArguments(
output_dir=save_path,
overwrite_output_dir=True,
num_train_epochs=epoch,
per_gpu_train_batch_size=per_gpu_train_batch_size,
save_steps=save_steps,
save_total_limit=2,
gradient_accumulation_steps=
gradient_accumulation_steps, # 32*8=256 batch_size
learning_rate=learning_rate,
weight_decay=0.01,
adam_epsilon=1e-6,
warmup_steps=10000,
)
trainer = Trainer(
model=model,
args=training_args,
data_collator=data_collator,
train_dataset=dataset,
)
trainer.train()
trainer.save_model(save_path)
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)
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}")
# Set seed before initializing model.
set_seed(training_args.seed)
# 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 train from scratch."
)
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."
)
# Initialize our dataset and prepare it for the audio classification task.
raw_datasets = DatasetDict()
raw_datasets["train"] = load_dataset(
data_args.dataset_name, data_args.dataset_config_name, split=data_args.train_split_name
)
raw_datasets["eval"] = load_dataset(
data_args.dataset_name, data_args.dataset_config_name, split=data_args.eval_split_name
)
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.label_column_name not in raw_datasets["train"].column_names:
raise ValueError(
f"--label_column_name {data_args.label_column_name} not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--label_column_name` to the correct text column - one of "
f"{', '.join(raw_datasets['train'].column_names)}."
)
# Setting `return_attention_mask=True` is the way to get a correctly masked mean-pooling over
# transformer outputs in the classifier, but it doesn't always lead to better accuracy
feature_extractor = AutoFeatureExtractor.from_pretrained(
model_args.feature_extractor_name or model_args.model_name_or_path,
return_attention_mask=model_args.attention_mask,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# `datasets` takes care of automatically loading and resampling the audio,
# so we just need to set the correct target sampling rate.
raw_datasets = raw_datasets.cast_column(
data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
)
def train_transforms(batch):
"""Apply train_transforms across a batch."""
output_batch = {"input_values": []}
for audio in batch[data_args.audio_column_name]:
wav = random_subsample(
audio["array"], max_length=data_args.max_length_seconds, sample_rate=feature_extractor.sampling_rate
)
output_batch["input_values"].append(wav)
output_batch["labels"] = [label for label in batch[data_args.label_column_name]]
return output_batch
def val_transforms(batch):
"""Apply val_transforms across a batch."""
output_batch = {"input_values": []}
for audio in batch[data_args.audio_column_name]:
wav = audio["array"]
output_batch["input_values"].append(wav)
output_batch["labels"] = [label for label in batch[data_args.label_column_name]]
return output_batch
# Prepare label mappings.
# We'll include these in the model's config to get human readable labels in the Inference API.
labels = raw_datasets["train"].features[data_args.label_column_name].names
label2id, id2label = dict(), dict()
for i, label in enumerate(labels):
label2id[label] = str(i)
id2label[str(i)] = label
# Load the accuracy metric from the datasets package
metric = datasets.load_metric("accuracy")
# Define our compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with
# `predictions` and `label_ids` fields) and has to return a dictionary string to float.
def compute_metrics(eval_pred):
"""Computes accuracy on a batch of predictions"""
predictions = np.argmax(eval_pred.predictions, axis=1)
return metric.compute(predictions=predictions, references=eval_pred.label_ids)
config = AutoConfig.from_pretrained(
model_args.config_name or model_args.model_name_or_path,
num_labels=len(labels),
label2id=label2id,
id2label=id2label,
finetuning_task="audio-classification",
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForAudioClassification.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,
)
# freeze the convolutional waveform encoder
if model_args.freeze_feature_extractor:
model.freeze_feature_extractor()
if training_args.do_train:
if data_args.max_train_samples is not None:
raw_datasets["train"] = (
raw_datasets["train"].shuffle(seed=training_args.seed).select(range(data_args.max_train_samples))
)
# Set the training transforms
raw_datasets["train"].set_transform(train_transforms, output_all_columns=False)
if training_args.do_eval:
if data_args.max_eval_samples is not None:
raw_datasets["eval"] = (
raw_datasets["eval"].shuffle(seed=training_args.seed).select(range(data_args.max_eval_samples))
)
# Set the validation transforms
raw_datasets["eval"].set_transform(val_transforms, output_all_columns=False)
# Initialize our trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=raw_datasets["train"] if training_args.do_train else None,
eval_dataset=raw_datasets["eval"] if training_args.do_eval else None,
compute_metrics=compute_metrics,
tokenizer=feature_extractor,
)
# 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()
# Evaluation
if training_args.do_eval:
metrics = trainer.evaluate()
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Write model card and (optionally) push to hub
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": "audio-classification",
"dataset": data_args.dataset_name,
"tags": ["audio-classification"],
}
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",
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(f"Training/evaluation parameters {training_args}")
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
# or specify a GLUE benchmark task (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use as labels the column called 'label' and as pair of sentences the
# sentences in columns called 'sentence1' and 'sentence2' if such column exists or the first two columns not named
# label if at least two columns are provided.
#
# If the CSVs/JSONs contain only one non-label column, the script does single sentence classification on this
# single column. You can easily tweak this behavior (see below)
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.task_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset("glue", data_args.task_name)
else:
# Loading a dataset from your local files.
# CSV/JSON training and evaluation files are needed.
data_files = {
"train": data_args.train_file,
"validation": data_args.validation_file
}
# Get the test dataset: you can provide your own CSV/JSON test file (see below)
# when you use `do_predict` without specifying a GLUE benchmark task.
if training_args.do_predict:
if data_args.test_file is not None:
train_extension = data_args.train_file.split(".")[-1]
test_extension = data_args.test_file.split(".")[-1]
assert (
test_extension == train_extension
), "`test_file` should have the same extension (csv or json) as `train_file`."
data_files["test"] = data_args.test_file
else:
raise ValueError(
"Need either a GLUE task or a test file for `do_predict`.")
for key in data_files.keys():
logger.info(f"load a local file for {key}: {data_files[key]}")
if data_args.train_file.endswith(".csv"):
# Loading a dataset from local csv files
datasets = load_dataset("csv", data_files=data_files)
else:
# Loading a dataset from local json files
datasets = load_dataset("json", data_files=data_files)
# See more about loading any type of standard or custom dataset at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Labels
if data_args.task_name is not None:
is_regression = data_args.task_name == "stsb"
if not is_regression:
label_list = datasets["train"].features["label"].names
num_labels = len(label_list)
else:
num_labels = 1
else:
# Trying to have good defaults here, don't hesitate to tweak to your needs.
is_regression = datasets["train"].features["label"].dtype in [
"float32", "float64"
]
if is_regression:
num_labels = 1
else:
# A useful fast method:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.unique
label_list = datasets["train"].unique("label")
label_list.sort() # Let's sort it for determinism
num_labels = len(label_list)
# Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
finetuning_task=data_args.task_name,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForSequenceClassification.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# Preprocessing the datasets
if data_args.task_name is not None:
sentence1_key, sentence2_key = task_to_keys[data_args.task_name]
else:
# Again, we try to have some nice defaults but don't hesitate to tweak to your use case.
non_label_column_names = [
name for name in datasets["train"].column_names if name != "label"
]
if "sentence1" in non_label_column_names and "sentence2" in non_label_column_names:
sentence1_key, sentence2_key = "sentence1", "sentence2"
else:
if len(non_label_column_names) >= 2:
sentence1_key, sentence2_key = non_label_column_names[:2]
else:
sentence1_key, sentence2_key = non_label_column_names[0], None
# Padding strategy
if data_args.pad_to_max_length:
padding = "max_length"
else:
# We will pad later, dynamically at batch creation, to the max sequence length in each batch
padding = False
# Some models have set the order of the labels to use, so let's make sure we do use it.
label_to_id = None
if (model.config.label2id !=
PretrainedConfig(num_labels=num_labels).label2id
and data_args.task_name is not None and 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
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]}.")
# 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 = 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:
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)
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 ***")
# 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 train(args):
# load model and tokenizer
MODEL_NAME = args.pretrained_model
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
# load dataset
train_dataset = load_data("/opt/ml/input/data/train/train.tsv")
#dev_dataset = load_data("./dataset/train/dev.tsv")
train_label = train_dataset['label'].values
#dev_label = dev_dataset['label'].values
# tokenizing dataset
tokenized_train = tokenized_dataset(train_dataset, tokenizer)
#tokenized_dev = tokenized_dataset(dev_dataset, tokenizer)
# make dataset for pytorch.
RE_train_dataset = RE_Dataset(tokenized_train, train_label)
#RE_dev_dataset = RE_Dataset(tokenized_dev, dev_label)
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
# setting model hyperparameter
config_module = getattr(import_module("transformers"),
args.model_type + "Config")
model_config = config_module.from_pretrained(MODEL_NAME)
model_config.num_labels = 42
model_module = getattr(import_module("transformers"),
args.model_type + "ForSequenceClassification")
model = model_module.from_pretrained(MODEL_NAME, config=model_config)
model.parameters
model.to(device)
output_dir = increment_output_dir(args.output_dir)
# 사용한 option 외에도 다양한 option들이 있습니다.
# https://huggingface.co/transformers/main_classes/trainer.html#trainingarguments 참고해주세요.
training_args = TrainingArguments(
output_dir=output_dir, # output directory
save_total_limit=args.save_total_limit, # number of total save model.
save_steps=args.save_steps, # model saving step.
num_train_epochs=args.epochs, # total number of training epochs
learning_rate=args.lr, # learning_rate
per_device_train_batch_size=args.
batch_size, # batch size per device during training
#per_device_eval_batch_size=16, # batch size for evaluation
warmup_steps=500, # number of warmup steps for learning rate scheduler
weight_decay=0.01, # strength of weight decay
logging_dir='./logs', # directory for storing logs
logging_steps=100, # log saving step.
#evaluation_strategy='steps', # evaluation strategy to adopt during training
# `no`: No evaluation during training.
# `steps`: Evaluate every `eval_steps`.
# `epoch`: Evaluate every end of epoch.
#eval_steps = 500, # evaluation step.
)
trainer = Trainer(
model=model, # the instantiated 🤗 Transformers model to be trained
args=training_args, # training arguments, defined above
train_dataset=RE_train_dataset, # training dataset
#eval_dataset=RE_dev_dataset, # evaluation dataset
#compute_metrics=compute_metrics # define metrics function
)
# train model
trainer.train()
