def __init__(self, xs=[]):
self.params = ps = parse_params(xs)
self.mgr = mgr = Accelerator()
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
log.info(mgr.state)
log.setLevel(
logging.INFO if mgr.is_local_main_process else logging.ERROR)
if mgr.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
if ps.seed is not None:
set_seed(ps.seed)
if mgr.is_main_process:
if ps.push_to_hub:
if ps.hub_model_id is None:
x = get_full_repo_name(Path(ps.out_dir).name,
token=ps.hub_token)
else:
x = ps.hub_model_id
self.repo = Repository(ps.out_dir, clone_from=x)
elif ps.out_dir is not None:
os.makedirs(ps.out_dir, exist_ok=True)
mgr.wait_for_everyone()
self.padding = "max_len" if ps.pad_to_max_length else False
def main():
args = parse_args()
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
# Setup logging, we only want one process per machine to log things on the screen.
logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
if jax.process_index() == 0:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# Handle the repository creation
if args.push_to_hub:
if args.hub_model_id is None:
repo_name = get_full_repo_name(Path(args.output_dir).absolute().name, token=args.hub_token)
else:
repo_name = args.hub_model_id
repo = Repository(args.output_dir, clone_from=repo_name)
# Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
# or specify a GLUE benchmark task (the dataset will be downloaded automatically from the datasets Hub).
# For CSV/JSON files, this script will use as labels the column called 'label' and as pair of sentences the
# sentences in columns called 'sentence1' and 'sentence2' if such column exists or the first two columns not named
# label if at least two columns are provided.
# If the CSVs/JSONs contain only one non-label column, the script does single sentence classification on this
# single column. You can easily tweak this behavior (see below)
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if args.task_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset("glue", args.task_name)
else:
# Loading the dataset from local csv or json file.
data_files = {}
if args.train_file is not None:
data_files["train"] = args.train_file
if args.validation_file is not None:
data_files["validation"] = args.validation_file
extension = (args.train_file if args.train_file is not None else args.valid_file).split(".")[-1]
raw_datasets = load_dataset(extension, data_files=data_files)
# See more about loading any type of standard or custom dataset at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Labels
if args.task_name is not None:
is_regression = args.task_name == "stsb"
if not is_regression:
label_list = raw_datasets["train"].features["label"].names
num_labels = len(label_list)
else:
num_labels = 1
else:
# Trying to have good defaults here, don't hesitate to tweak to your needs.
is_regression = raw_datasets["train"].features["label"].dtype in ["float32", "float64"]
if is_regression:
num_labels = 1
else:
# A useful fast method:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.unique
label_list = raw_datasets["train"].unique("label")
label_list.sort() # Let's sort it for determinism
num_labels = len(label_list)
# Load pretrained model and tokenizer
config = AutoConfig.from_pretrained(args.model_name_or_path, num_labels=num_labels, finetuning_task=args.task_name)
tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
model = FlaxAutoModelForSequenceClassification.from_pretrained(args.model_name_or_path, config=config)
# Preprocessing the datasets
if args.task_name is not None:
sentence1_key, sentence2_key = task_to_keys[args.task_name]
else:
# Again, we try to have some nice defaults but don't hesitate to tweak to your use case.
non_label_column_names = [name for name in raw_datasets["train"].column_names if name != "label"]
if "sentence1" in non_label_column_names and "sentence2" in non_label_column_names:
sentence1_key, sentence2_key = "sentence1", "sentence2"
else:
if len(non_label_column_names) >= 2:
sentence1_key, sentence2_key = non_label_column_names[:2]
else:
sentence1_key, sentence2_key = non_label_column_names[0], None
# Some models have set the order of the labels to use, so let's make sure we do use it.
label_to_id = None
if (
model.config.label2id != PretrainedConfig(num_labels=num_labels).label2id
and args.task_name is not None
and not is_regression
):
# Some have all caps in their config, some don't.
label_name_to_id = {k.lower(): v for k, v in model.config.label2id.items()}
if list(sorted(label_name_to_id.keys())) == list(sorted(label_list)):
logger.info(
f"The configuration of the model provided the following label correspondence: {label_name_to_id}. "
"Using it!"
)
label_to_id = {i: label_name_to_id[label_list[i]] for i in range(num_labels)}
else:
logger.warning(
"Your model seems to have been trained with labels, but they don't match the dataset: ",
f"model labels: {list(sorted(label_name_to_id.keys()))}, dataset labels: {list(sorted(label_list))}."
"\nIgnoring the model labels as a result.",
)
elif args.task_name is None:
label_to_id = {v: i for i, v in enumerate(label_list)}
def preprocess_function(examples):
# Tokenize the texts
texts = (
(examples[sentence1_key],) if sentence2_key is None else (examples[sentence1_key], examples[sentence2_key])
)
result = tokenizer(*texts, padding="max_length", max_length=args.max_length, truncation=True)
if "label" in examples:
if label_to_id is not None:
# Map labels to IDs (not necessary for GLUE tasks)
result["labels"] = [label_to_id[l] for l in examples["label"]]
else:
# In all cases, rename the column to labels because the model will expect that.
result["labels"] = examples["label"]
return result
processed_datasets = raw_datasets.map(
preprocess_function, batched=True, remove_columns=raw_datasets["train"].column_names
)
train_dataset = processed_datasets["train"]
eval_dataset = processed_datasets["validation_matched" if args.task_name == "mnli" else "validation"]
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")
# Define a summary writer
has_tensorboard = is_tensorboard_available()
if has_tensorboard and jax.process_index() == 0:
try:
from flax.metrics.tensorboard import SummaryWriter
summary_writer = SummaryWriter(args.output_dir)
summary_writer.hparams(vars(args))
except ImportError as ie:
has_tensorboard = False
logger.warning(
f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
)
else:
logger.warning(
"Unable to display metrics through TensorBoard because the package is not installed: "
"Please run pip install tensorboard to enable."
)
def write_metric(train_metrics, eval_metrics, train_time, step):
summary_writer.scalar("train_time", train_time, step)
train_metrics = get_metrics(train_metrics)
for key, vals in train_metrics.items():
tag = f"train_{key}"
for i, val in enumerate(vals):
summary_writer.scalar(tag, val, step - len(vals) + i + 1)
for metric_name, value in eval_metrics.items():
summary_writer.scalar(f"eval_{metric_name}", value, step)
num_epochs = int(args.num_train_epochs)
rng = jax.random.PRNGKey(args.seed)
dropout_rngs = jax.random.split(rng, jax.local_device_count())
train_batch_size = args.per_device_train_batch_size * jax.local_device_count()
eval_batch_size = args.per_device_eval_batch_size * jax.local_device_count()
learning_rate_fn = create_learning_rate_fn(
len(train_dataset), train_batch_size, args.num_train_epochs, args.num_warmup_steps, args.learning_rate
)
state = create_train_state(
model, learning_rate_fn, is_regression, num_labels=num_labels, weight_decay=args.weight_decay
)
# define step functions
def train_step(
state: train_state.TrainState, batch: Dict[str, Array], dropout_rng: PRNGKey
) -> Tuple[train_state.TrainState, float]:
"""Trains model with an optimizer (both in `state`) on `batch`, returning a pair `(new_state, loss)`."""
dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
targets = batch.pop("labels")
def loss_fn(params):
logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
loss = state.loss_fn(logits, targets)
return loss
grad_fn = jax.value_and_grad(loss_fn)
loss, grad = grad_fn(state.params)
grad = jax.lax.pmean(grad, "batch")
new_state = state.apply_gradients(grads=grad)
metrics = jax.lax.pmean({"loss": loss, "learning_rate": learning_rate_fn(state.step)}, axis_name="batch")
return new_state, metrics, new_dropout_rng
p_train_step = jax.pmap(train_step, axis_name="batch", donate_argnums=(0,))
def eval_step(state, batch):
logits = state.apply_fn(**batch, params=state.params, train=False)[0]
return state.logits_fn(logits)
p_eval_step = jax.pmap(eval_step, axis_name="batch")
if args.task_name is not None:
metric = load_metric("glue", args.task_name)
else:
metric = load_metric("accuracy")
logger.info(f"===== Starting training ({num_epochs} epochs) =====")
train_time = 0
# make sure weights are replicated on each device
state = replicate(state)
for epoch in range(1, num_epochs + 1):
logger.info(f"Epoch {epoch}")
logger.info(" Training...")
train_start = time.time()
train_metrics = []
rng, input_rng = jax.random.split(rng)
# train
for batch in glue_train_data_collator(input_rng, train_dataset, train_batch_size):
state, metrics, dropout_rngs = p_train_step(state, batch, dropout_rngs)
train_metrics.append(metrics)
train_time += time.time() - train_start
logger.info(f" Done! Training metrics: {unreplicate(metrics)}")
logger.info(" Evaluating...")
# evaluate
for batch in glue_eval_data_collator(eval_dataset, eval_batch_size):
labels = batch.pop("labels")
predictions = p_eval_step(state, batch)
metric.add_batch(predictions=chain(*predictions), references=chain(*labels))
# evaluate also on leftover examples (not divisible by batch_size)
num_leftover_samples = len(eval_dataset) % eval_batch_size
# make sure leftover batch is evaluated on one device
if num_leftover_samples > 0 and jax.process_index() == 0:
# take leftover samples
batch = eval_dataset[-num_leftover_samples:]
batch = {k: jnp.array(v) for k, v in batch.items()}
labels = batch.pop("labels")
predictions = eval_step(unreplicate(state), batch)
metric.add_batch(predictions=predictions, references=labels)
eval_metric = metric.compute()
logger.info(f" Done! Eval metrics: {eval_metric}")
cur_step = epoch * (len(train_dataset) // train_batch_size)
# Save metrics
if has_tensorboard and jax.process_index() == 0:
write_metric(train_metrics, eval_metric, train_time, cur_step)
# save checkpoint after each epoch and push checkpoint to the hub
if jax.process_index() == 0:
params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
model.save_pretrained(args.output_dir, params=params)
tokenizer.save_pretrained(args.output_dir)
if args.push_to_hub:
repo.push_to_hub(commit_message=f"Saving weights and logs of epoch {epoch}", blocking=False)
# save the eval metrics in json
if jax.process_index() == 0:
eval_metric = {f"eval_{metric_name}": value for metric_name, value in eval_metric.items()}
path = os.path.join(args.output_dir, "eval_results.json")
with open(path, "w") as f:
json.dump(eval_metric, f, indent=4, sort_keys=True)
def main():
args = parse_args()
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
accelerator = Accelerator()
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state)
# Setup logging, we only want one process per machine to log things on the screen.
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(
logging.INFO if accelerator.is_local_main_process else logging.ERROR)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.push_to_hub:
if args.hub_model_id is None:
repo_name = get_full_repo_name(Path(args.output_dir).name,
token=args.hub_token)
else:
repo_name = args.hub_model_id
repo = Repository(args.output_dir, clone_from=repo_name)
elif args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
accelerator.wait_for_everyone()
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(args.dataset_name,
args.dataset_config_name)
else:
data_files = {}
if args.train_file is not None:
data_files["train"] = args.train_file
if args.validation_file is not None:
data_files["validation"] = args.validation_file
if args.test_file is not None:
data_files["test"] = args.test_file
extension = args.train_file.split(".")[-1]
raw_datasets = load_dataset(extension,
data_files=data_files,
field="data")
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = XLNetConfig.from_pretrained(args.model_name_or_path)
tokenizer = XLNetTokenizerFast.from_pretrained(args.model_name_or_path)
model = XLNetForQuestionAnswering.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config)
# Preprocessing the datasets.
# Preprocessing is slighlty different for training and evaluation.
column_names = raw_datasets["train"].column_names
question_column_name = "question" if "question" in column_names else column_names[
0]
context_column_name = "context" if "context" in column_names else column_names[
1]
answer_column_name = "answers" if "answers" in column_names else column_names[
2]
# Padding side determines if we do (question|context) or (context|question).
pad_on_right = tokenizer.padding_side == "right"
if args.max_seq_length > tokenizer.model_max_length:
logger.warning(
f"The max_seq_length passed ({args.max_seq_length}) is larger than the maximum length for the"
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(args.max_seq_length, tokenizer.model_max_length)
# Training preprocessing
def prepare_train_features(examples):
# Some of the questions have lots of whitespace on the left, which is not useful and will make the
# truncation of the context fail (the tokenized question will take a lots of space). So we remove that
# left whitespace
examples[question_column_name] = [
q.lstrip() for q in examples[question_column_name]
]
# Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
# in one example possible giving several features when a context is long, each of those features having a
# context that overlaps a bit the context of the previous feature.
tokenized_examples = tokenizer(
examples[
question_column_name if pad_on_right else context_column_name],
examples[
context_column_name if pad_on_right else question_column_name],
truncation="only_second" if pad_on_right else "only_first",
max_length=max_seq_length,
stride=args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
return_special_tokens_mask=True,
return_token_type_ids=True,
padding="max_length",
)
# Since one example might give us several features if it has a long context, we need a map from a feature to
# its corresponding example. This key gives us just that.
sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
# The offset mappings will give us a map from token to character position in the original context. This will
# help us compute the start_positions and end_positions.
offset_mapping = tokenized_examples.pop("offset_mapping")
# The special tokens will help us build the p_mask (which indicates the tokens that can't be in answers).
special_tokens = tokenized_examples.pop("special_tokens_mask")
# Let's label those examples!
tokenized_examples["start_positions"] = []
tokenized_examples["end_positions"] = []
tokenized_examples["is_impossible"] = []
tokenized_examples["cls_index"] = []
tokenized_examples["p_mask"] = []
for i, offsets in enumerate(offset_mapping):
# We will label impossible answers with the index of the CLS token.
input_ids = tokenized_examples["input_ids"][i]
cls_index = input_ids.index(tokenizer.cls_token_id)
tokenized_examples["cls_index"].append(cls_index)
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples["token_type_ids"][i]
for k, s in enumerate(special_tokens[i]):
if s:
sequence_ids[k] = 3
context_idx = 1 if pad_on_right else 0
# Build the p_mask: non special tokens and context gets 0.0, the others get 1.0.
# The cls token gets 1.0 too (for predictions of empty answers).
tokenized_examples["p_mask"].append([
0.0 if (not special_tokens[i][k] and s == context_idx)
or k == cls_index else 1.0 for k, s in enumerate(sequence_ids)
])
# One example can give several spans, this is the index of the example containing this span of text.
sample_index = sample_mapping[i]
answers = examples[answer_column_name][sample_index]
# If no answers are given, set the cls_index as answer.
if len(answers["answer_start"]) == 0:
tokenized_examples["start_positions"].append(cls_index)
tokenized_examples["end_positions"].append(cls_index)
tokenized_examples["is_impossible"].append(1.0)
else:
# Start/end character index of the answer in the text.
start_char = answers["answer_start"][0]
end_char = start_char + len(answers["text"][0])
# Start token index of the current span in the text.
token_start_index = 0
while sequence_ids[token_start_index] != context_idx:
token_start_index += 1
# End token index of the current span in the text.
token_end_index = len(input_ids) - 1
while sequence_ids[token_end_index] != context_idx:
token_end_index -= 1
# Detect if the answer is out of the span (in which case this feature is labeled with the CLS index).
if not (offsets[token_start_index][0] <= start_char
and offsets[token_end_index][1] >= end_char):
tokenized_examples["start_positions"].append(cls_index)
tokenized_examples["end_positions"].append(cls_index)
tokenized_examples["is_impossible"].append(1.0)
else:
# Otherwise move the token_start_index and token_end_index to the two ends of the answer.
# Note: we could go after the last offset if the answer is the last word (edge case).
while token_start_index < len(offsets) and offsets[
token_start_index][0] <= start_char:
token_start_index += 1
tokenized_examples["start_positions"].append(
token_start_index - 1)
while offsets[token_end_index][1] >= end_char:
token_end_index -= 1
tokenized_examples["end_positions"].append(
token_end_index + 1)
tokenized_examples["is_impossible"].append(0.0)
return tokenized_examples
if "train" not in raw_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = raw_datasets["train"]
if args.max_train_samples is not None:
# We will select sample from whole data if agument is specified
train_dataset = train_dataset.select(range(args.max_train_samples))
# Create train feature from dataset
with accelerator.main_process_first():
train_dataset = train_dataset.map(
prepare_train_features,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on train dataset",
)
if args.max_train_samples is not None:
# Number of samples might increase during Feature Creation, We select only specified max samples
train_dataset = train_dataset.select(range(args.max_train_samples))
# Validation preprocessing
def prepare_validation_features(examples):
# Some of the questions have lots of whitespace on the left, which is not useful and will make the
# truncation of the context fail (the tokenized question will take a lots of space). So we remove that
# left whitespace
examples[question_column_name] = [
q.lstrip() for q in examples[question_column_name]
]
# Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
# in one example possible giving several features when a context is long, each of those features having a
# context that overlaps a bit the context of the previous feature.
tokenized_examples = tokenizer(
examples[
question_column_name if pad_on_right else context_column_name],
examples[
context_column_name if pad_on_right else question_column_name],
truncation="only_second" if pad_on_right else "only_first",
max_length=max_seq_length,
stride=args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
return_special_tokens_mask=True,
return_token_type_ids=True,
padding="max_length",
)
# Since one example might give us several features if it has a long context, we need a map from a feature to
# its corresponding example. This key gives us just that.
sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
# The special tokens will help us build the p_mask (which indicates the tokens that can't be in answers).
special_tokens = tokenized_examples.pop("special_tokens_mask")
# For evaluation, we will need to convert our predictions to substrings of the context, so we keep the
# corresponding example_id and we will store the offset mappings.
tokenized_examples["example_id"] = []
# We still provide the index of the CLS token and the p_mask to the model, but not the is_impossible label.
tokenized_examples["cls_index"] = []
tokenized_examples["p_mask"] = []
for i, input_ids in enumerate(tokenized_examples["input_ids"]):
# Find the CLS token in the input ids.
cls_index = input_ids.index(tokenizer.cls_token_id)
tokenized_examples["cls_index"].append(cls_index)
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples["token_type_ids"][i]
for k, s in enumerate(special_tokens[i]):
if s:
sequence_ids[k] = 3
context_idx = 1 if pad_on_right else 0
# Build the p_mask: non special tokens and context gets 0.0, the others 1.0.
tokenized_examples["p_mask"].append([
0.0 if (not special_tokens[i][k] and s == context_idx)
or k == cls_index else 1.0 for k, s in enumerate(sequence_ids)
])
# One example can give several spans, this is the index of the example containing this span of text.
sample_index = sample_mapping[i]
tokenized_examples["example_id"].append(
examples["id"][sample_index])
# Set to None the offset_mapping that are not part of the context so it's easy to determine if a token
# position is part of the context or not.
tokenized_examples["offset_mapping"][i] = [
(o if sequence_ids[k] == context_idx else None)
for k, o in enumerate(tokenized_examples["offset_mapping"][i])
]
return tokenized_examples
if "validation" not in raw_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_examples = raw_datasets["validation"]
if args.max_eval_samples is not None:
# We will select sample from whole data
eval_examples = eval_examples.select(range(args.max_eval_samples))
# Validation Feature Creation
with accelerator.main_process_first():
eval_dataset = eval_examples.map(
prepare_validation_features,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on validation dataset",
)
if args.max_eval_samples is not None:
# During Feature creation dataset samples might increase, we will select required samples again
eval_dataset = eval_dataset.select(range(args.max_eval_samples))
if args.do_predict:
if "test" not in raw_datasets:
raise ValueError("--do_predict requires a test dataset")
predict_examples = raw_datasets["test"]
if args.max_predict_samples is not None:
# We will select sample from whole data
predict_examples = predict_examples.select(
range(args.max_predict_samples))
# Predict Feature Creation
with accelerator.main_process_first():
predict_dataset = predict_examples.map(
prepare_validation_features,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on prediction dataset",
)
if args.max_predict_samples is not None:
# During Feature creation dataset samples might increase, we will select required samples again
predict_dataset = predict_dataset.select(
range(args.max_predict_samples))
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(
f"Sample {index} of the training set: {train_dataset[index]}.")
# DataLoaders creation:
if args.pad_to_max_length:
# If padding was already done ot max length, we use the default data collator that will just convert everything
# to tensors.
data_collator = default_data_collator
else:
# Otherwise, `DataCollatorWithPadding` will apply dynamic padding for us (by padding to the maximum length of
# the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple
# of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).
data_collator = DataCollatorWithPadding(
tokenizer,
pad_to_multiple_of=(8 if accelerator.use_fp16 else None))
train_dataloader = DataLoader(train_dataset,
shuffle=True,
collate_fn=data_collator,
batch_size=args.per_device_train_batch_size)
eval_dataset_for_model = eval_dataset.remove_columns(
["example_id", "offset_mapping"])
eval_dataloader = DataLoader(eval_dataset_for_model,
collate_fn=data_collator,
batch_size=args.per_device_eval_batch_size)
if args.do_predict:
predict_dataset_for_model = predict_dataset.remove_columns(
["example_id", "offset_mapping"])
predict_dataloader = DataLoader(
predict_dataset_for_model,
collate_fn=data_collator,
batch_size=args.per_device_eval_batch_size)
# Post-processing:
def post_processing_function(examples,
features,
predictions,
stage="eval"):
# Post-processing: we match the start logits and end logits to answers in the original context.
predictions, scores_diff_json = postprocess_qa_predictions_with_beam_search(
examples=examples,
features=features,
predictions=predictions,
version_2_with_negative=args.version_2_with_negative,
n_best_size=args.n_best_size,
max_answer_length=args.max_answer_length,
start_n_top=model.config.start_n_top,
end_n_top=model.config.end_n_top,
output_dir=args.output_dir,
prefix=stage,
)
# Format the result to the format the metric expects.
if args.version_2_with_negative:
formatted_predictions = [{
"id":
k,
"prediction_text":
v,
"no_answer_probability":
scores_diff_json[k]
} for k, v in predictions.items()]
else:
formatted_predictions = [{
"id": k,
"prediction_text": v
} for k, v in predictions.items()]
references = [{
"id": ex["id"],
"answers": ex[answer_column_name]
} for ex in examples]
return EvalPrediction(predictions=formatted_predictions,
label_ids=references)
metric = load_metric(
"squad_v2" if args.version_2_with_negative else "squad")
def create_and_fill_np_array(start_or_end_logits, dataset, max_len):
"""
Create and fill numpy array of size len_of_validation_data * max_length_of_output_tensor
Args:
start_or_end_logits(:obj:`tensor`):
This is the output predictions of the model. We can only enter either start or end logits.
eval_dataset: Evaluation dataset
max_len(:obj:`int`):
The maximum length of the output tensor. ( See the model.eval() part for more details )
"""
step = 0
# create a numpy array and fill it with -100.
logits_concat = np.full((len(dataset), max_len),
-100,
dtype=np.float32)
# Now since we have create an array now we will populate it with the outputs gathered using accelerator.gather
for i, output_logit in enumerate(
start_or_end_logits): # populate columns
# We have to fill it such that we have to take the whole tensor and replace it on the newly created array
# And after every iteration we have to change the step
batch_size = output_logit.shape[0]
cols = output_logit.shape[1]
if step + batch_size < len(dataset):
logits_concat[step:step + batch_size, :cols] = output_logit
else:
logits_concat[step:, :cols] = output_logit[:len(dataset) -
step]
step += batch_size
return logits_concat
# Optimizer
# Split weights in two groups, one with weight decay and the other not.
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay":
args.weight_decay,
},
{
"params": [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0,
},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate)
# Prepare everything with our `accelerator`.
model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader)
# Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be
# shorter in multiprocess)
# Scheduler and math around the number of training steps.
num_update_steps_per_epoch = math.ceil(
len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
else:
args.num_train_epochs = math.ceil(args.max_train_steps /
num_update_steps_per_epoch)
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=args.num_warmup_steps,
num_training_steps=args.max_train_steps,
)
# Train!
total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(
f" Instantaneous batch size per device = {args.per_device_train_batch_size}"
)
logger.info(
f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}"
)
logger.info(
f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
# Only show the progress bar once on each machine.
progress_bar = tqdm(range(args.max_train_steps),
disable=not accelerator.is_local_main_process)
completed_steps = 0
for epoch in range(args.num_train_epochs):
model.train()
for step, batch in enumerate(train_dataloader):
outputs = model(**batch)
loss = outputs.loss
loss = loss / args.gradient_accumulation_steps
accelerator.backward(loss)
if step % args.gradient_accumulation_steps == 0 or step == len(
train_dataloader) - 1:
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
progress_bar.update(1)
completed_steps += 1
if completed_steps >= args.max_train_steps:
break
if args.push_to_hub and epoch < args.num_train_epochs - 1:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir,
save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
repo.push_to_hub(
commit_message=f"Training in progress epoch {epoch}",
blocking=False,
auto_lfs_prune=True)
# intialize all lists to collect the batches
all_start_top_log_probs = []
all_start_top_index = []
all_end_top_log_probs = []
all_end_top_index = []
all_cls_logits = []
for step, batch in enumerate(eval_dataloader):
with torch.no_grad():
outputs = model(**batch)
start_top_log_probs = outputs.start_top_log_probs
start_top_index = outputs.start_top_index
end_top_log_probs = outputs.end_top_log_probs
end_top_index = outputs.end_top_index
cls_logits = outputs.cls_logits
if not args.pad_to_max_length: # necessary to pad predictions and labels for being gathered
start_top_log_probs = accelerator.pad_across_processes(
start_top_log_probs, dim=1, pad_index=-100)
start_top_index = accelerator.pad_across_processes(
start_top_index, dim=1, pad_index=-100)
end_top_log_probs = accelerator.pad_across_processes(
end_top_log_probs, dim=1, pad_index=-100)
end_top_index = accelerator.pad_across_processes(
end_top_index, dim=1, pad_index=-100)
cls_logits = accelerator.pad_across_processes(cls_logits,
dim=1,
pad_index=-100)
all_start_top_log_probs.append(
accelerator.gather(start_top_log_probs).cpu().numpy())
all_start_top_index.append(
accelerator.gather(start_top_index).cpu().numpy())
all_end_top_log_probs.append(
accelerator.gather(end_top_log_probs).cpu().numpy())
all_end_top_index.append(
accelerator.gather(end_top_index).cpu().numpy())
all_cls_logits.append(accelerator.gather(cls_logits).cpu().numpy())
max_len = max([x.shape[1] for x in all_end_top_log_probs
]) # Get the max_length of the tensor
# concatenate all numpy arrays collected above
start_top_log_probs_concat = create_and_fill_np_array(
all_start_top_log_probs, eval_dataset, max_len)
start_top_index_concat = create_and_fill_np_array(all_start_top_index,
eval_dataset, max_len)
end_top_log_probs_concat = create_and_fill_np_array(
all_end_top_log_probs, eval_dataset, max_len)
end_top_index_concat = create_and_fill_np_array(all_end_top_index,
eval_dataset, max_len)
cls_logits_concat = np.concatenate(all_cls_logits, axis=0)
# delete the list of numpy arrays
del start_top_log_probs
del start_top_index
del end_top_log_probs
del end_top_index
del cls_logits
outputs_numpy = (
start_top_log_probs_concat,
start_top_index_concat,
end_top_log_probs_concat,
end_top_index_concat,
cls_logits_concat,
)
prediction = post_processing_function(eval_examples, eval_dataset,
outputs_numpy)
eval_metric = metric.compute(predictions=prediction.predictions,
references=prediction.label_ids)
logger.info(f"Evaluation metrics: {eval_metric}")
if args.do_predict:
# intialize all lists to collect the batches
all_start_top_log_probs = []
all_start_top_index = []
all_end_top_log_probs = []
all_end_top_index = []
all_cls_logits = []
for step, batch in enumerate(predict_dataloader):
with torch.no_grad():
outputs = model(**batch)
start_top_log_probs = outputs.start_top_log_probs
start_top_index = outputs.start_top_index
end_top_log_probs = outputs.end_top_log_probs
end_top_index = outputs.end_top_index
cls_logits = outputs.cls_logits
if not args.pad_to_max_length: # necessary to pad predictions and labels for being gathered
start_top_log_probs = accelerator.pad_across_processes(
start_top_log_probs, dim=1, pad_index=-100)
start_top_index = accelerator.pad_across_processes(
start_top_index, dim=1, pad_index=-100)
end_top_log_probs = accelerator.pad_across_processes(
end_top_log_probs, dim=1, pad_index=-100)
end_top_index = accelerator.pad_across_processes(
end_top_index, dim=1, pad_index=-100)
cls_logits = accelerator.pad_across_processes(
cls_logits, dim=1, pad_index=-100)
all_start_top_log_probs.append(
accelerator.gather(start_top_log_probs).cpu().numpy())
all_start_top_index.append(
accelerator.gather(start_top_index).cpu().numpy())
all_end_top_log_probs.append(
accelerator.gather(end_top_log_probs).cpu().numpy())
all_end_top_index.append(
accelerator.gather(end_top_index).cpu().numpy())
all_cls_logits.append(
accelerator.gather(cls_logits).cpu().numpy())
max_len = max([x.shape[1] for x in all_end_top_log_probs
]) # Get the max_length of the tensor
# concatenate all numpy arrays collected above
start_top_log_probs_concat = create_and_fill_np_array(
all_start_top_log_probs, predict_dataset, max_len)
start_top_index_concat = create_and_fill_np_array(
all_start_top_index, predict_dataset, max_len)
end_top_log_probs_concat = create_and_fill_np_array(
all_end_top_log_probs, predict_dataset, max_len)
end_top_index_concat = create_and_fill_np_array(
all_end_top_index, predict_dataset, max_len)
cls_logits_concat = np.concatenate(all_cls_logits, axis=0)
# delete the list of numpy arrays
del start_top_log_probs
del start_top_index
del end_top_log_probs
del end_top_index
del cls_logits
outputs_numpy = (
start_top_log_probs_concat,
start_top_index_concat,
end_top_log_probs_concat,
end_top_index_concat,
cls_logits_concat,
)
prediction = post_processing_function(predict_examples,
predict_dataset, outputs_numpy)
predict_metric = metric.compute(predictions=prediction.predictions,
references=prediction.label_ids)
logger.info(f"Predict metrics: {predict_metric}")
if args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir,
save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
if args.push_to_hub:
repo.push_to_hub(commit_message="End of training",
auto_lfs_prune=True)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir) and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty."
"Use --overwrite_output_dir to overcome.")
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
level="NOTSET",
datefmt="[%X]",
)
# Log on each process the small summary:
logger = logging.getLogger(__name__)
# Set the verbosity to info of the Transformers logger (on main process only):
logger.info(f"Training/evaluation parameters {training_args}")
# Set seed before initializing model.
set_seed(training_args.seed)
# Handle the repository creation
if training_args.push_to_hub:
if training_args.hub_model_id is None:
repo_name = get_full_repo_name(Path(
training_args.output_dir).absolute().name,
token=training_args.hub_token)
else:
repo_name = training_args.hub_model_id
repo = Repository(training_args.output_dir, clone_from=repo_name)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets 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_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 extension == "txt":
extension = "text"
datasets = load_dataset(extension,
data_files=data_files,
cache_dir=model_args.cache_dir)
if "validation" not in datasets.keys():
datasets["validation"] = load_dataset(
extension,
data_files=data_files,
split=f"train[:{data_args.validation_split_percentage}%]",
cache_dir=model_args.cache_dir,
)
datasets["train"] = load_dataset(
extension,
data_files=data_files,
split=f"train[{data_args.validation_split_percentage}%:]",
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
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.config_name:
config = T5Config.from_pretrained(model_args.config_name,
cache_dir=model_args.cache_dir,
vocab_size=len(tokenizer))
elif model_args.model_name_or_path:
config = T5Config.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.")
# 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]
max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
# Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
# Since we make sure that all sequences are of the same length, no attention_mask is needed.
def tokenize_function(examples):
return tokenizer(examples[text_column_name],
return_attention_mask=False)
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,
)
# T5-like span masked language modeling will fuse consecutively masked tokens to a single sentinel token.
# To ensure that the input length is `max_seq_length`, we need to increase the maximum length
# according to `mlm_probability` and `mean_noise_span_length`. We can also define the label length accordingly.
expanded_inputs_length, targets_length = compute_input_and_target_lengths(
inputs_length=max_seq_length,
noise_density=data_args.mlm_probability,
mean_noise_span_length=data_args.mean_noise_span_length,
)
# Main data processing function that will concatenate all texts from our dataset and generate chunks of expanded_inputs_length.
def group_texts(examples):
# Concatenate all texts.
concatenated_examples = {
k: list(chain(*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.
if total_length >= expanded_inputs_length:
total_length = (total_length //
expanded_inputs_length) * expanded_inputs_length
# Split by chunks of max_len.
result = {
k: [
t[i:i + expanded_inputs_length]
for i in range(0, total_length, expanded_inputs_length)
]
for k, t in concatenated_examples.items()
}
return result
# Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
# remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
# might be slower to preprocess.
#
# To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
tokenized_datasets = tokenized_datasets.map(
group_texts,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
# Enable tensorboard only on the master node
has_tensorboard = is_tensorboard_available()
if has_tensorboard and jax.process_index() == 0:
try:
from flax.metrics.tensorboard import SummaryWriter
summary_writer = SummaryWriter(
log_dir=Path(training_args.output_dir))
except ImportError as ie:
has_tensorboard = False
logger.warning(
f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
)
else:
logger.warning(
"Unable to display metrics through TensorBoard because the package is not installed: "
"Please run pip install tensorboard to enable.")
# Initialize our training
rng = jax.random.PRNGKey(training_args.seed)
dropout_rngs = jax.random.split(rng, jax.local_device_count())
if model_args.model_name_or_path:
model = FlaxT5ForConditionalGeneration.from_pretrained(
model_args.model_name_or_path,
config=config,
seed=training_args.seed,
dtype=getattr(jnp, model_args.dtype))
else:
config.vocab_size = len(tokenizer)
model = FlaxT5ForConditionalGeneration(config,
seed=training_args.seed,
dtype=getattr(
jnp, model_args.dtype))
# Data collator
# This one will take care of randomly masking the tokens.
data_collator = FlaxDataCollatorForT5MLM(
tokenizer=tokenizer,
noise_density=data_args.mlm_probability,
mean_noise_span_length=data_args.mean_noise_span_length,
input_length=max_seq_length,
target_length=targets_length,
pad_token_id=model.config.pad_token_id,
decoder_start_token_id=model.config.decoder_start_token_id,
)
# Store some constant
num_epochs = int(training_args.num_train_epochs)
train_batch_size = int(
training_args.per_device_train_batch_size) * jax.device_count()
eval_batch_size = int(
training_args.per_device_eval_batch_size) * jax.device_count()
num_train_steps = len(
tokenized_datasets["train"]) // train_batch_size * num_epochs
# Create learning rate schedule
warmup_fn = optax.linear_schedule(
init_value=0.0,
end_value=training_args.learning_rate,
transition_steps=training_args.warmup_steps)
decay_fn = optax.linear_schedule(
init_value=training_args.learning_rate,
end_value=0,
transition_steps=num_train_steps - training_args.warmup_steps,
)
linear_decay_lr_schedule_fn = optax.join_schedules(
schedules=[warmup_fn, decay_fn],
boundaries=[training_args.warmup_steps])
# We use Optax's "masking" functionality to not apply weight decay
# to bias and LayerNorm scale parameters. decay_mask_fn returns a
# mask boolean with the same structure as the parameters.
# The mask is True for parameters that should be decayed.
def decay_mask_fn(params):
flat_params = traverse_util.flatten_dict(params)
flat_mask = {
path: (path[-1] != "bias"
and path[-2:] not in [("layer_norm", "scale"),
("final_layer_norm", "scale")])
for path in flat_params
}
return traverse_util.unflatten_dict(flat_mask)
# create adam optimizer
if training_args.adafactor:
# We use the default parameters here to initialize adafactor,
# For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
optimizer = optax.adafactor(
learning_rate=linear_decay_lr_schedule_fn, )
else:
optimizer = optax.adamw(
learning_rate=linear_decay_lr_schedule_fn,
b1=training_args.adam_beta1,
b2=training_args.adam_beta2,
weight_decay=training_args.weight_decay,
mask=decay_mask_fn,
)
# Setup train state
state = train_state.TrainState.create(apply_fn=model.__call__,
params=model.params,
tx=optimizer)
# Define gradient update step fn
def train_step(state, batch, dropout_rng):
dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
def loss_fn(params):
labels = batch.pop("labels")
logits = state.apply_fn(**batch,
params=params,
dropout_rng=dropout_rng,
train=True)[0]
# compute loss
loss = optax.softmax_cross_entropy(
logits, onehot(labels, logits.shape[-1])).mean()
return loss
grad_fn = jax.value_and_grad(loss_fn)
loss, grad = grad_fn(state.params)
grad = jax.lax.pmean(grad, "batch")
new_state = state.apply_gradients(grads=grad)
metrics = jax.lax.pmean(
{
"loss": loss,
"learning_rate": linear_decay_lr_schedule_fn(state.step)
},
axis_name="batch")
return new_state, metrics, new_dropout_rng
# Create parallel version of the train step
p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0, ))
# Define eval fn
def eval_step(params, batch):
labels = batch.pop("labels")
logits = model(**batch, params=params, train=False)[0]
# compute loss
loss = optax.softmax_cross_entropy(logits,
onehot(labels, logits.shape[-1]))
# compute accuracy
accuracy = jnp.equal(jnp.argmax(logits, axis=-1), labels)
# summarize metrics
metrics = {"loss": loss.mean(), "accuracy": accuracy.mean()}
metrics = jax.lax.pmean(metrics, axis_name="batch")
return metrics
p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0, ))
# Replicate the train state on each device
state = jax_utils.replicate(state)
train_time = 0
epochs = tqdm(range(num_epochs), desc="Epoch ... ", position=0)
for epoch in epochs:
# ======================== Training ================================
train_start = time.time()
train_metrics = []
# Create sampling rng
rng, input_rng = jax.random.split(rng)
# Generate an epoch by shuffling sampling indices from the train dataset
num_train_samples = len(tokenized_datasets["train"])
train_samples_idx = jax.random.permutation(
input_rng, jnp.arange(num_train_samples))
train_batch_idx = generate_batch_splits(train_samples_idx,
train_batch_size)
# Gather the indexes for creating the batch and do a training step
for step, batch_idx in enumerate(
tqdm(train_batch_idx, desc="Training...", position=1)):
samples = [
tokenized_datasets["train"][int(idx)] for idx in batch_idx
]
model_inputs = data_collator(samples)
# Model forward
model_inputs = shard(model_inputs.data)
state, train_metric, dropout_rngs = p_train_step(
state, model_inputs, dropout_rngs)
train_metrics.append(train_metric)
cur_step = epoch * (num_train_samples // train_batch_size) + step
if cur_step % training_args.logging_steps == 0 and cur_step > 0:
# Save metrics
train_metric = jax_utils.unreplicate(train_metric)
train_time += time.time() - train_start
if has_tensorboard and jax.process_index() == 0:
write_train_metric(summary_writer, train_metrics,
train_time, cur_step)
epochs.write(
f"Step... ({cur_step} | Loss: {train_metric['loss'].mean()}, Learning Rate: {train_metric['learning_rate'].mean()})"
)
train_metrics = []
if cur_step % training_args.eval_steps == 0 and cur_step > 0:
# ======================== Evaluating ==============================
num_eval_samples = len(tokenized_datasets["validation"])
eval_samples_idx = jnp.arange(num_eval_samples)
eval_batch_idx = generate_batch_splits(eval_samples_idx,
eval_batch_size)
eval_metrics = []
for i, batch_idx in enumerate(
tqdm(eval_batch_idx, desc="Evaluating ...",
position=2)):
samples = [
tokenized_datasets["validation"][int(idx)]
for idx in batch_idx
]
model_inputs = data_collator(samples)
# Model forward
model_inputs = shard(model_inputs.data)
metrics = p_eval_step(state.params, model_inputs)
eval_metrics.append(metrics)
# get eval metrics
eval_metrics = get_metrics(eval_metrics)
eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
# Update progress bar
epochs.write(
f"Step... ({cur_step} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
)
# Save metrics
if has_tensorboard and jax.process_index() == 0:
write_eval_metric(summary_writer, eval_metrics, cur_step)
if cur_step % training_args.save_steps == 0 and cur_step > 0:
# save checkpoint after each epoch and push checkpoint to the hub
if jax.process_index() == 0:
params = jax.device_get(
jax.tree_map(lambda x: x[0], state.params))
model.save_pretrained(training_args.output_dir,
params=params)
tokenizer.save_pretrained(training_args.output_dir)
if training_args.push_to_hub:
repo.push_to_hub(
commit_message=
f"Saving weights and logs of step {cur_step}",
blocking=False)
# Eval after training
if training_args.do_eval:
num_eval_samples = len(tokenized_datasets["validation"])
eval_samples_idx = jnp.arange(num_eval_samples)
eval_batch_idx = generate_batch_splits(eval_samples_idx,
eval_batch_size)
eval_metrics = []
for i, batch_idx in enumerate(
tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
samples = [
tokenized_datasets["validation"][int(idx)] for idx in batch_idx
]
model_inputs = data_collator(samples)
# Model forward
model_inputs = shard(model_inputs.data)
metrics = p_eval_step(state.params, model_inputs)
eval_metrics.append(metrics)
# get eval metrics
eval_metrics = get_metrics(eval_metrics)
eval_metrics = jax.tree_map(lambda metric: jnp.mean(metric).item(),
eval_metrics)
if jax.process_index() == 0:
eval_metrics = {
f"eval_{metric_name}": value
for metric_name, value in eval_metrics.items()
}
path = os.path.join(training_args.output_dir, "eval_results.json")
with open(path, "w") as f:
json.dump(eval_metrics, f, indent=4, sort_keys=True)
def main():
args = parse_args()
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
handler = DistributedDataParallelKwargs(find_unused_parameters=True)
accelerator = Accelerator(kwargs_handlers=[handler])
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state)
# Setup logging, we only want one process per machine to log things on the screen.
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(logging.INFO if accelerator.is_local_main_process else logging.ERROR)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.push_to_hub:
if args.hub_model_id is None:
repo_name = get_full_repo_name(Path(args.output_dir).name, token=args.hub_token)
else:
repo_name = args.hub_model_id
repo = Repository(args.output_dir, clone_from=repo_name)
elif args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
accelerator.wait_for_everyone()
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets for token classification task available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'tokens' or the first column if no column called
# 'tokens' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(args.dataset_name, args.dataset_config_name)
else:
data_files = {}
if args.train_file is not None:
data_files["train"] = args.train_file
if args.validation_file is not None:
data_files["validation"] = args.validation_file
extension = args.train_file.split(".")[-1]
raw_datasets = load_dataset(extension, data_files=data_files)
# Trim a number of training examples
if args.debug:
for split in raw_datasets.keys():
raw_datasets[split] = raw_datasets[split].select(range(100))
# 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 raw_datasets["train"] is not None:
column_names = raw_datasets["train"].column_names
features = raw_datasets["train"].features
else:
column_names = raw_datasets["validation"].column_names
features = raw_datasets["validation"].features
if args.text_column_name is not None:
text_column_name = args.text_column_name
elif "tokens" in column_names:
text_column_name = "tokens"
else:
text_column_name = column_names[0]
if args.label_column_name is not None:
label_column_name = args.label_column_name
elif f"{args.task_name}_tags" in column_names:
label_column_name = f"{args.task_name}_tags"
else:
label_column_name = column_names[1]
# In the event the labels are not a `Sequence[ClassLabel]`, we will need to go through the dataset to get the
# unique labels.
def get_label_list(labels):
unique_labels = set()
for label in labels:
unique_labels = unique_labels | set(label)
label_list = list(unique_labels)
label_list.sort()
return label_list
if isinstance(features[label_column_name].feature, ClassLabel):
label_list = features[label_column_name].feature.names
# No need to convert the labels since they are already ints.
else:
label_list = get_label_list(raw_datasets["train"][label_column_name])
num_labels = len(label_list)
# 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)
# Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if args.config_name:
config = LukeConfig.from_pretrained(args.config_name, num_labels=num_labels)
elif args.model_name_or_path:
config = LukeConfig.from_pretrained(args.model_name_or_path, num_labels=num_labels)
else:
logger.warning("You are instantiating a new config instance from scratch.")
tokenizer_name_or_path = args.tokenizer_name if args.tokenizer_name else args.model_name_or_path
if not tokenizer_name_or_path:
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."
)
tokenizer = LukeTokenizer.from_pretrained(
tokenizer_name_or_path,
use_fast=False,
task="entity_span_classification",
max_entity_length=args.max_entity_length,
max_mention_length=args.max_mention_length,
)
if args.model_name_or_path:
model = LukeForEntitySpanClassification.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
)
else:
logger.info("Training new model from scratch")
model = LukeForEntitySpanClassification.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
padding = "max_length" if args.pad_to_max_length else False
def compute_sentence_boundaries_for_luke(examples):
sentence_boundaries = []
for tokens in examples[text_column_name]:
sentence_boundaries.append([0, len(tokens)])
examples["sentence_boundaries"] = sentence_boundaries
return examples
def compute_entity_spans_for_luke(examples):
all_entity_spans = []
texts = []
all_labels_entity_spans = []
all_original_entity_spans = []
for labels, tokens, sentence_boundaries in zip(
examples[label_column_name], examples[text_column_name], examples["sentence_boundaries"]
):
subword_lengths = [len(tokenizer.tokenize(token)) for token in tokens]
total_subword_length = sum(subword_lengths)
_, context_end = sentence_boundaries
if total_subword_length > args.max_length - 2:
cur_length = sum(subword_lengths[:context_end])
idx = context_end - 1
while cur_length > args.max_length - 2:
cur_length -= subword_lengths[idx]
context_end -= 1
idx -= 1
text = ""
sentence_words = tokens[:context_end]
sentence_subword_lengths = subword_lengths[:context_end]
word_start_char_positions = []
word_end_char_positions = []
labels_positions = {}
for word, label in zip(sentence_words, labels):
if word[0] == "'" or (len(word) == 1 and is_punctuation(word)):
text = text.rstrip()
word_start_char_positions.append(len(text))
text += word
word_end_char_positions.append(len(text))
text += " "
labels_positions[(word_start_char_positions[-1], word_end_char_positions[-1])] = label
text = text.rstrip()
texts.append(text)
entity_spans = []
labels_entity_spans = []
original_entity_spans = []
for word_start in range(len(sentence_words)):
for word_end in range(word_start, len(sentence_words)):
if (
sum(sentence_subword_lengths[word_start:word_end]) <= tokenizer.max_mention_length
and len(entity_spans) < tokenizer.max_entity_length
):
entity_spans.append((word_start_char_positions[word_start], word_end_char_positions[word_end]))
original_entity_spans.append((word_start, word_end + 1))
if (
word_start_char_positions[word_start],
word_end_char_positions[word_end],
) in labels_positions:
labels_entity_spans.append(
labels_positions[
(word_start_char_positions[word_start], word_end_char_positions[word_end])
]
)
else:
labels_entity_spans.append(0)
all_entity_spans.append(entity_spans)
all_labels_entity_spans.append(labels_entity_spans)
all_original_entity_spans.append(original_entity_spans)
examples["entity_spans"] = all_entity_spans
examples["text"] = texts
examples["labels_entity_spans"] = all_labels_entity_spans
examples["original_entity_spans"] = all_original_entity_spans
return examples
def tokenize_and_align_labels(examples):
entity_spans = []
for v in examples["entity_spans"]:
entity_spans.append(list(map(tuple, v)))
tokenized_inputs = tokenizer(
examples["text"],
entity_spans=entity_spans,
max_length=args.max_length,
padding=padding,
truncation=True,
)
if padding == "max_length":
tokenized_inputs["labels"] = padding_tensor(
examples["labels_entity_spans"], -100, tokenizer.padding_side, tokenizer.max_entity_length
)
tokenized_inputs["original_entity_spans"] = padding_tensor(
examples["original_entity_spans"], (-1, -1), tokenizer.padding_side, tokenizer.max_entity_length
)
tokenized_inputs[label_column_name] = padding_tensor(
examples[label_column_name], -1, tokenizer.padding_side, tokenizer.max_entity_length
)
else:
tokenized_inputs["labels"] = [ex[: tokenizer.max_entity_length] for ex in examples["labels_entity_spans"]]
tokenized_inputs["original_entity_spans"] = [
ex[: tokenizer.max_entity_length] for ex in examples["original_entity_spans"]
]
tokenized_inputs[label_column_name] = [
ex[: tokenizer.max_entity_length] for ex in examples[label_column_name]
]
return tokenized_inputs
with accelerator.main_process_first():
raw_datasets = raw_datasets.map(
compute_sentence_boundaries_for_luke,
batched=True,
desc="Adding sentence boundaries",
)
raw_datasets = raw_datasets.map(
compute_entity_spans_for_luke,
batched=True,
desc="Adding sentence spans",
)
processed_raw_datasets = raw_datasets.map(
tokenize_and_align_labels,
batched=True,
remove_columns=raw_datasets["train"].column_names,
desc="Running tokenizer on dataset",
)
train_dataset = processed_raw_datasets["train"]
eval_dataset = processed_raw_datasets["validation"]
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")
# DataLoaders creation:
if args.pad_to_max_length:
# If padding was already done ot max length, we use the default data collator that will just convert everything
# to tensors.
data_collator = default_data_collator
else:
# Otherwise, `DataCollatorForTokenClassification` will apply dynamic padding for us (by padding to the maximum length of
# the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple
# of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).
data_collator = DataCollatorForLukeTokenClassification(
tokenizer, pad_to_multiple_of=(8 if accelerator.use_fp16 else None)
)
train_dataloader = DataLoader(
train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size
)
eval_dataloader = DataLoader(eval_dataset, collate_fn=data_collator, batch_size=args.per_device_eval_batch_size)
# Optimizer
# Split weights in two groups, one with weight decay and the other not.
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
"weight_decay": args.weight_decay,
},
{
"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)],
"weight_decay": 0.0,
},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate)
# Use the device given by the `accelerator` object.
device = accelerator.device
model.to(device)
# Prepare everything with our `accelerator`.
model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader
)
# Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be
# shorter in multiprocess)
# Scheduler and math around the number of training steps.
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
else:
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=args.num_warmup_steps,
num_training_steps=args.max_train_steps,
)
# Metrics
metric = load_metric("seqeval")
def get_luke_labels(outputs, ner_tags, original_entity_spans):
true_predictions = []
true_labels = []
for output, original_spans, tags in zip(outputs.logits, original_entity_spans, ner_tags):
true_tags = [val for val in tags if val != -1]
true_original_spans = [val for val in original_spans if val != (-1, -1)]
max_indices = torch.argmax(output, axis=1)
max_logits = torch.max(output, axis=1).values
predictions = []
for logit, index, span in zip(max_logits, max_indices, true_original_spans):
if index != 0:
predictions.append((logit, span, label_list[index]))
predicted_sequence = [label_list[0]] * len(true_tags)
for _, span, label in sorted(predictions, key=lambda o: o[0], reverse=True):
if all([o == label_list[0] for o in predicted_sequence[span[0] : span[1]]]):
predicted_sequence[span[0]] = label
if span[1] - span[0] > 1:
predicted_sequence[span[0] + 1 : span[1]] = [label] * (span[1] - span[0] - 1)
true_predictions.append(predicted_sequence)
true_labels.append([label_list[tag_id] for tag_id in true_tags])
return true_predictions, true_labels
def compute_metrics():
results = metric.compute()
if 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"],
}
# Train!
total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {args.per_device_train_batch_size}")
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
# Only show the progress bar once on each machine.
progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process)
completed_steps = 0
for epoch in range(args.num_train_epochs):
model.train()
for step, batch in enumerate(train_dataloader):
_ = batch.pop("original_entity_spans")
outputs = model(**batch)
loss = outputs.loss
loss = loss / args.gradient_accumulation_steps
accelerator.backward(loss)
if step % args.gradient_accumulation_steps == 0 or step == len(train_dataloader) - 1:
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
progress_bar.update(1)
completed_steps += 1
if completed_steps >= args.max_train_steps:
break
model.eval()
for step, batch in enumerate(eval_dataloader):
original_entity_spans = batch.pop("original_entity_spans")
with torch.no_grad():
outputs = model(**batch)
preds, refs = get_luke_labels(outputs, batch[label_column_name], original_entity_spans)
metric.add_batch(
predictions=preds,
references=refs,
) # predictions and preferences are expected to be a nested list of labels, not label_ids
eval_metric = compute_metrics()
accelerator.print(f"epoch {epoch}:", eval_metric)
if args.push_to_hub and epoch < args.num_train_epochs - 1:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir, save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
repo.push_to_hub(
commit_message=f"Training in progress epoch {epoch}", blocking=False, auto_lfs_prune=True
)
if args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir, save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
if args.push_to_hub:
repo.push_to_hub(commit_message="End of training", auto_lfs_prune=True)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
# Setup logging, we only want one process per machine to log things on the screen.
logger.setLevel(logging.INFO if jax.process_index() ==
0 else logging.ERROR)
if jax.process_index() == 0:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# Handle the repository creation
if training_args.push_to_hub:
if training_args.hub_model_id is None:
repo_name = get_full_repo_name(Path(
training_args.output_dir).absolute().name,
token=training_args.hub_token)
else:
repo_name = training_args.hub_model_id
repo = Repository(training_args.output_dir, clone_from=repo_name)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets for token classification task available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'tokens' or the first column if no column called
# 'tokens' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name,
cache_dir=model_args.cache_dir)
else:
# Loading the dataset from local csv or json file.
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = (data_args.train_file if data_args.train_file is not None
else data_args.valid_file).split(".")[-1]
raw_datasets = load_dataset(extension,
data_files=data_files,
cache_dir=model_args.cache_dir)
# See more about loading any type of standard or custom dataset at
# https://huggingface.co/docs/datasets/loading_datasets.html.
if raw_datasets["train"] is not None:
column_names = raw_datasets["train"].column_names
features = raw_datasets["train"].features
else:
column_names = raw_datasets["validation"].column_names
features = raw_datasets["validation"].features
if data_args.text_column_name is not None:
text_column_name = data_args.text_column_name
elif "tokens" in column_names:
text_column_name = "tokens"
else:
text_column_name = column_names[0]
if data_args.label_column_name is not None:
label_column_name = data_args.label_column_name
elif f"{data_args.task_name}_tags" in column_names:
label_column_name = f"{data_args.task_name}_tags"
else:
label_column_name = column_names[1]
# In the event the labels are not a `Sequence[ClassLabel]`, we will need to go through the dataset to get the
# unique labels.
def get_label_list(labels):
unique_labels = set()
for label in labels:
unique_labels = unique_labels | set(label)
label_list = list(unique_labels)
label_list.sort()
return label_list
if isinstance(features[label_column_name].feature, ClassLabel):
label_list = features[label_column_name].feature.names
# No need to convert the labels since they are already ints.
label_to_id = {i: i for i in range(len(label_list))}
else:
label_list = get_label_list(raw_datasets["train"][label_column_name])
label_to_id = {l: i for i, l in enumerate(label_list)}
num_labels = len(label_list)
# Load pretrained model and tokenizer
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
label2id=label_to_id,
id2label={i: l
for l, i in label_to_id.items()},
finetuning_task=data_args.task_name,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer_name_or_path = model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path
if config.model_type in {"gpt2", "roberta"}:
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
add_prefix_space=True,
)
else:
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = FlaxAutoModelForTokenClassification.from_pretrained(
model_args.model_name_or_path,
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# Preprocessing the datasets
# Tokenize all texts and align the labels with them.
def tokenize_and_align_labels(examples):
tokenized_inputs = tokenizer(
examples[text_column_name],
max_length=data_args.max_seq_length,
padding="max_length",
truncation=True,
# We use this argument because the texts in our dataset are lists of words (with a label for each word).
is_split_into_words=True,
)
labels = []
for i, label in enumerate(examples[label_column_name]):
word_ids = tokenized_inputs.word_ids(batch_index=i)
previous_word_idx = None
label_ids = []
for word_idx in word_ids:
# Special tokens have a word id that is None. We set the label to -100 so they are automatically
# ignored in the loss function.
if word_idx is None:
label_ids.append(-100)
# We set the label for the first token of each word.
elif word_idx != previous_word_idx:
label_ids.append(label_to_id[label[word_idx]])
# For the other tokens in a word, we set the label to either the current label or -100, depending on
# the label_all_tokens flag.
else:
label_ids.append(label_to_id[label[word_idx]] if data_args.
label_all_tokens else -100)
previous_word_idx = word_idx
labels.append(label_ids)
tokenized_inputs["labels"] = labels
return tokenized_inputs
processed_raw_datasets = raw_datasets.map(
tokenize_and_align_labels,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
remove_columns=raw_datasets["train"].column_names,
desc="Running tokenizer on dataset",
)
train_dataset = processed_raw_datasets["train"]
eval_dataset = processed_raw_datasets["validation"]
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(
f"Sample {index} of the training set: {train_dataset[index]}.")
# Define a summary writer
summary_writer = tensorboard.SummaryWriter(training_args.output_dir)
summary_writer.hparams({
**training_args.to_dict(),
**vars(model_args),
**vars(data_args)
})
def write_train_metric(summary_writer, train_metrics, train_time, step):
summary_writer.scalar("train_time", train_time, step)
train_metrics = get_metrics(train_metrics)
for key, vals in train_metrics.items():
tag = f"train_{key}"
for i, val in enumerate(vals):
summary_writer.scalar(tag, val, step - len(vals) + i + 1)
def write_eval_metric(summary_writer, eval_metrics, step):
for metric_name, value in eval_metrics.items():
summary_writer.scalar(f"eval_{metric_name}", value, step)
num_epochs = int(training_args.num_train_epochs)
rng = jax.random.PRNGKey(training_args.seed)
dropout_rngs = jax.random.split(rng, jax.local_device_count())
train_batch_size = training_args.per_device_train_batch_size * jax.local_device_count(
)
eval_batch_size = training_args.per_device_eval_batch_size * jax.local_device_count(
)
learning_rate_fn = create_learning_rate_fn(
len(train_dataset),
train_batch_size,
training_args.num_train_epochs,
training_args.warmup_steps,
training_args.learning_rate,
)
state = create_train_state(model,
learning_rate_fn,
num_labels=num_labels,
training_args=training_args)
# define step functions
def train_step(
state: train_state.TrainState, batch: Dict[str, Array],
dropout_rng: PRNGKey) -> Tuple[train_state.TrainState, float]:
"""Trains model with an optimizer (both in `state`) on `batch`, returning a pair `(new_state, loss)`."""
dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
targets = batch.pop("labels")
def loss_fn(params):
logits = state.apply_fn(**batch,
params=params,
dropout_rng=dropout_rng,
train=True)[0]
loss = state.loss_fn(logits, targets)
return loss
grad_fn = jax.value_and_grad(loss_fn)
loss, grad = grad_fn(state.params)
grad = jax.lax.pmean(grad, "batch")
new_state = state.apply_gradients(grads=grad)
metrics = jax.lax.pmean(
{
"loss": loss,
"learning_rate": learning_rate_fn(state.step)
},
axis_name="batch")
return new_state, metrics, new_dropout_rng
p_train_step = jax.pmap(train_step,
axis_name="batch",
donate_argnums=(0, ))
def eval_step(state, batch):
logits = state.apply_fn(**batch, params=state.params, train=False)[0]
return state.logits_fn(logits)
p_eval_step = jax.pmap(eval_step, axis_name="batch")
metric = load_metric("seqeval")
def get_labels(y_pred, y_true):
# Transform predictions and references tensos to numpy arrays
# Remove ignored index (special tokens)
true_predictions = [[
label_list[p] for (p, l) in zip(pred, gold_label) if l != -100
] for pred, gold_label in zip(y_pred, y_true)]
true_labels = [[
label_list[l] for (p, l) in zip(pred, gold_label) if l != -100
] for pred, gold_label in zip(y_pred, y_true)]
return true_predictions, true_labels
def compute_metrics():
results = metric.compute()
if data_args.return_entity_level_metrics:
# Unpack nested dictionaries
final_results = {}
for key, value in results.items():
if isinstance(value, dict):
for n, v in value.items():
final_results[f"{key}_{n}"] = v
else:
final_results[key] = value
return final_results
else:
return {
"precision": results["overall_precision"],
"recall": results["overall_recall"],
"f1": results["overall_f1"],
"accuracy": results["overall_accuracy"],
}
logger.info(f"===== Starting training ({num_epochs} epochs) =====")
train_time = 0
# make sure weights are replicated on each device
state = replicate(state)
train_time = 0
step_per_epoch = len(train_dataset) // train_batch_size
total_steps = step_per_epoch * num_epochs
epochs = tqdm(range(num_epochs),
desc=f"Epoch ... (1/{num_epochs})",
position=0)
for epoch in epochs:
train_start = time.time()
train_metrics = []
# Create sampling rng
rng, input_rng = jax.random.split(rng)
# train
for step, batch in enumerate(
tqdm(
train_data_collator(input_rng, train_dataset,
train_batch_size),
total=step_per_epoch,
desc="Training...",
position=1,
)):
state, train_metric, dropout_rngs = p_train_step(
state, batch, dropout_rngs)
train_metrics.append(train_metric)
cur_step = (epoch * step_per_epoch) + (step + 1)
if cur_step % training_args.logging_steps == 0 and cur_step > 0:
# Save metrics
train_metric = unreplicate(train_metric)
train_time += time.time() - train_start
if jax.process_index() == 0:
write_train_metric(summary_writer, train_metrics,
train_time, cur_step)
epochs.write(
f"Step... ({cur_step}/{total_steps} | Training Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
)
train_metrics = []
if cur_step % training_args.eval_steps == 0 and cur_step > 0:
eval_metrics = {}
# evaluate
for batch in tqdm(
eval_data_collator(eval_dataset, eval_batch_size),
total=len(eval_dataset) // eval_batch_size,
desc="Evaluating ...",
position=2,
):
labels = batch.pop("labels")
predictions = p_eval_step(state, batch)
predictions = np.array(
[pred for pred in chain(*predictions)])
labels = np.array([label for label in chain(*labels)])
labels[np.array(chain(
*batch["attention_mask"])) == 0] = -100
preds, refs = get_labels(predictions, labels)
metric.add_batch(
predictions=preds,
references=refs,
)
# evaluate also on leftover examples (not divisible by batch_size)
num_leftover_samples = len(eval_dataset) % eval_batch_size
# make sure leftover batch is evaluated on one device
if num_leftover_samples > 0 and jax.process_index() == 0:
# take leftover samples
batch = eval_dataset[-num_leftover_samples:]
batch = {k: np.array(v) for k, v in batch.items()}
labels = batch.pop("labels")
predictions = eval_step(unreplicate(state), batch)
labels = np.array(labels)
labels[np.array(batch["attention_mask"]) == 0] = -100
preds, refs = get_labels(predictions, labels)
metric.add_batch(
predictions=preds,
references=refs,
)
eval_metrics = compute_metrics()
if data_args.return_entity_level_metrics:
logger.info(
f"Step... ({cur_step}/{total_steps} | Validation metrics: {eval_metrics}"
)
else:
logger.info(
f"Step... ({cur_step}/{total_steps} | Validation f1: {eval_metrics['f1']}, Validation Acc: {eval_metrics['accuracy']})"
)
if jax.process_index() == 0:
write_eval_metric(summary_writer, eval_metrics, cur_step)
if (cur_step % training_args.save_steps == 0
and cur_step > 0) or (cur_step == total_steps):
# save checkpoint after each epoch and push checkpoint to the hub
if jax.process_index() == 0:
params = jax.device_get(unreplicate(state.params))
model.save_pretrained(training_args.output_dir,
params=params)
tokenizer.save_pretrained(training_args.output_dir)
if training_args.push_to_hub:
repo.push_to_hub(
commit_message=
f"Saving weights and logs of step {cur_step}",
blocking=False)
epochs.desc = f"Epoch ... {epoch + 1}/{num_epochs}"
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.")
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
# Setup logging, we only want one process per machine to log things on the screen.
logger.setLevel(logging.INFO if jax.process_index() ==
0 else logging.ERROR)
if jax.process_index() == 0:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# Set the verbosity to info of the Transformers logger (on main process only):
logger.info(f"Training/evaluation parameters {training_args}")
# Handle the repository creation
if training_args.push_to_hub:
if training_args.hub_model_id is None:
repo_name = get_full_repo_name(Path(
training_args.output_dir).absolute().name,
token=training_args.hub_token)
else:
repo_name = training_args.hub_model_id
repo = Repository(training_args.output_dir, clone_from=repo_name)
# Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files this script will use the first column for the full texts and the second column for the
# summaries (unless you specify column names for this with the `text_column` and `summary_column` arguments).
#
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)
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.
# Load pretrained model and tokenizer
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 = FlaxAutoModelForSeq2SeqLM.from_pretrained(
model_args.model_name_or_path,
config=config,
seed=training_args.seed,
dtype=getattr(jnp, model_args.dtype))
else:
model = FlaxAutoModelForSeq2SeqLM.from_config(config,
seed=training_args.seed,
dtype=getattr(
jnp,
model_args.dtype))
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
prefix = data_args.source_prefix if data_args.source_prefix is not None else ""
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
if training_args.do_train:
column_names = 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
# Get the column names for input/target.
dataset_columns = summarization_name_mapping.get(data_args.dataset_name,
None)
if data_args.text_column is None:
text_column = dataset_columns[
0] if dataset_columns is not None else column_names[0]
else:
text_column = data_args.text_column
if text_column not in column_names:
raise ValueError(
f"--text_column' value '{data_args.text_column}' needs to be one of: {', '.join(column_names)}"
)
if data_args.summary_column is None:
summary_column = dataset_columns[
1] if dataset_columns is not None else column_names[1]
else:
summary_column = data_args.summary_column
if summary_column not in column_names:
raise ValueError(
f"--summary_column' value '{data_args.summary_column}' needs to be one of: {', '.join(column_names)}"
)
# Temporarily set max_target_length for training.
max_target_length = data_args.max_target_length
# In Flax, for seq2seq models we need to pass `decoder_input_ids`
# as the Flax models don't accept `labels`, we need to prepare the decoder_input_ids here
# for that dynamically import the `shift_tokens_right` function from the model file
model_module = __import__(model.__module__,
fromlist=["shift_tokens_tight"])
shift_tokens_right_fn = getattr(model_module, "shift_tokens_right")
# Setting padding="max_length" as we need fixed length inputs for jitted functions
def preprocess_function(examples):
inputs = examples[text_column]
targets = examples[summary_column]
inputs = [prefix + inp for inp in inputs]
model_inputs = tokenizer(inputs,
max_length=data_args.max_source_length,
padding="max_length",
truncation=True,
return_tensors="np")
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets,
max_length=max_target_length,
padding="max_length",
truncation=True,
return_tensors="np")
model_inputs["labels"] = labels["input_ids"]
decoder_input_ids = shift_tokens_right_fn(
labels["input_ids"], config.pad_token_id,
config.decoder_start_token_id)
model_inputs["decoder_input_ids"] = np.asarray(decoder_input_ids)
# We need decoder_attention_mask so we can ignore pad tokens from loss
model_inputs["decoder_attention_mask"] = labels["attention_mask"]
return model_inputs
if training_args.do_train:
if "train" not in dataset:
raise ValueError("--do_train requires a train dataset")
train_dataset = dataset["train"]
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on train dataset",
)
if training_args.do_eval:
max_target_length = data_args.val_max_target_length
if "validation" not in dataset:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = dataset["validation"]
if data_args.max_eval_samples is not None:
eval_dataset = eval_dataset.select(
range(data_args.max_eval_samples))
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on validation dataset",
)
if training_args.do_predict:
max_target_length = data_args.val_max_target_length
if "test" not in dataset:
raise ValueError("--do_predict requires a test dataset")
predict_dataset = dataset["test"]
if data_args.max_predict_samples is not None:
predict_dataset = predict_dataset.select(
range(data_args.max_predict_samples))
predict_dataset = predict_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on prediction dataset",
)
# Metric
metric = load_metric("rouge")
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [label.strip() for label in labels]
# rougeLSum expects newline after each sentence
preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds]
labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels]
return preds, labels
def compute_metrics(preds, labels):
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
decoded_labels = tokenizer.batch_decode(labels,
skip_special_tokens=True)
# Some simple post-processing
decoded_preds, decoded_labels = postprocess_text(
decoded_preds, decoded_labels)
result = metric.compute(predictions=decoded_preds,
references=decoded_labels,
use_stemmer=True)
# Extract a few results from ROUGE
result = {
key: value.mid.fmeasure * 100
for key, value in result.items()
}
prediction_lens = [
np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds
]
result["gen_len"] = np.mean(prediction_lens)
result = {k: round(v, 4) for k, v in result.items()}
return result
# Enable tensorboard only on the master node
has_tensorboard = is_tensorboard_available()
if has_tensorboard and jax.process_index() == 0:
try:
from flax.metrics.tensorboard import SummaryWriter
summary_writer = SummaryWriter(
log_dir=Path(training_args.output_dir))
except ImportError as ie:
has_tensorboard = False
logger.warning(
f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
)
else:
logger.warning(
"Unable to display metrics through TensorBoard because the package is not installed: "
"Please run pip install tensorboard to enable.")
# Initialize our training
rng = jax.random.PRNGKey(training_args.seed)
rng, dropout_rng = jax.random.split(rng)
# Store some constant
num_epochs = int(training_args.num_train_epochs)
train_batch_size = int(
training_args.per_device_train_batch_size) * jax.device_count()
eval_batch_size = int(
training_args.per_device_eval_batch_size) * jax.device_count()
steps_per_epoch = len(train_dataset) // train_batch_size
total_train_steps = steps_per_epoch * num_epochs
# Create learning rate schedule
linear_decay_lr_schedule_fn = create_learning_rate_fn(
len(train_dataset),
train_batch_size,
training_args.num_train_epochs,
training_args.warmup_steps,
training_args.learning_rate,
)
# We use Optax's "masking" functionality to not apply weight decay
# to bias and LayerNorm scale parameters. decay_mask_fn returns a
# mask boolean with the same structure as the parameters.
# The mask is True for parameters that should be decayed.
# Note that this mask is specifically adapted for FlaxBart.
# For FlaxT5, one should correct the layer norm parameter naming
# accordingly - see `run_t5_mlm_flax.py` e.g.
def decay_mask_fn(params):
flat_params = traverse_util.flatten_dict(params)
layer_norm_params = [(name, "scale") for name in [
"self_attn_layer_norm", "layernorm_embedding", "final_layer_norm"
]]
flat_mask = {
path: (path[-1] != "bias" and path[-2:] not in layer_norm_params)
for path in flat_params
}
return traverse_util.unflatten_dict(flat_mask)
# create adam optimizer
adamw = optax.adamw(
learning_rate=linear_decay_lr_schedule_fn,
b1=training_args.adam_beta1,
b2=training_args.adam_beta2,
eps=training_args.adam_epsilon,
weight_decay=training_args.weight_decay,
mask=decay_mask_fn,
)
# Setup train state
state = TrainState.create(apply_fn=model.__call__,
params=model.params,
tx=adamw,
dropout_rng=dropout_rng)
# label smoothed cross entropy
def loss_fn(logits, labels, padding_mask, label_smoothing_factor=0.0):
"""
The label smoothing implementation is adapted from Flax's official example:
https://github.com/google/flax/blob/87a211135c6a377c8f29048a1cac3840e38b9da4/examples/wmt/train.py#L104
"""
vocab_size = logits.shape[-1]
confidence = 1.0 - label_smoothing_factor
low_confidence = (1.0 - confidence) / (vocab_size - 1)
normalizing_constant = -(confidence * jnp.log(confidence) +
(vocab_size - 1) * low_confidence *
jnp.log(low_confidence + 1e-20))
soft_labels = onehot(labels,
vocab_size,
on_value=confidence,
off_value=low_confidence)
loss = optax.softmax_cross_entropy(logits, soft_labels)
loss = loss - normalizing_constant
# ignore padded tokens from loss
loss = loss * padding_mask
loss = loss.sum() / padding_mask.sum()
return loss
# Define gradient update step fn
def train_step(state, batch, label_smoothing_factor=0.0):
dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
def compute_loss(params):
labels = batch.pop("labels")
logits = state.apply_fn(**batch,
params=params,
dropout_rng=dropout_rng,
train=True)[0]
loss = loss_fn(logits, labels, batch["decoder_attention_mask"],
label_smoothing_factor)
return loss
grad_fn = jax.value_and_grad(compute_loss)
loss, grad = grad_fn(state.params)
grad = jax.lax.pmean(grad, "batch")
new_state = state.apply_gradients(grads=grad,
dropout_rng=new_dropout_rng)
metrics = {
"loss": loss,
"learning_rate": linear_decay_lr_schedule_fn(state.step)
}
metrics = jax.lax.pmean(metrics, axis_name="batch")
return new_state, metrics
# Define eval fn
def eval_step(params, batch, label_smoothing_factor=0.0):
labels = batch.pop("labels")
logits = model(**batch, params=params, train=False)[0]
loss = loss_fn(logits, labels, batch["decoder_attention_mask"],
label_smoothing_factor)
# summarize metrics
metrics = {"loss": loss}
metrics = jax.lax.pmean(metrics, axis_name="batch")
return metrics
# Define generation function
max_length = (data_args.val_max_target_length
if data_args.val_max_target_length is not None else
model.config.max_length)
num_beams = data_args.num_beams if data_args.num_beams is not None else model.config.num_beams
gen_kwargs = {"max_length": max_length, "num_beams": num_beams}
def generate_step(params, batch):
model.params = params
output_ids = model.generate(batch["input_ids"],
attention_mask=batch["attention_mask"],
**gen_kwargs)
return output_ids.sequences
# Create parallel version of the train and eval step
p_train_step = jax.pmap(partial(
train_step,
label_smoothing_factor=training_args.label_smoothing_factor),
"batch",
donate_argnums=(0, ))
p_eval_step = jax.pmap(
partial(eval_step,
label_smoothing_factor=training_args.label_smoothing_factor),
"batch")
p_generate_step = jax.pmap(generate_step, "batch")
# Replicate the train state on each device
state = state.replicate()
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {num_epochs}")
logger.info(
f" Instantaneous batch size per device = {training_args.per_device_train_batch_size}"
)
logger.info(
f" Total train batch size (w. parallel & distributed) = {train_batch_size}"
)
logger.info(f" Total optimization steps = {total_train_steps}")
train_time = 0
epochs = tqdm(range(num_epochs),
desc=f"Epoch ... (1/{num_epochs})",
position=0)
for epoch in epochs:
# ======================== Training ================================
train_start = time.time()
# Create sampling rng
rng, input_rng = jax.random.split(rng)
train_metrics = []
# Generate an epoch by shuffling sampling indices from the train dataset
train_loader = data_loader(input_rng,
train_dataset,
train_batch_size,
shuffle=True)
steps_per_epoch = len(train_dataset) // train_batch_size
# train
for _ in tqdm(range(steps_per_epoch),
desc="Training...",
position=1,
leave=False):
batch = next(train_loader)
state, train_metric = p_train_step(state, batch)
train_metrics.append(train_metric)
train_time += time.time() - train_start
train_metric = unreplicate(train_metric)
epochs.write(
f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
)
# ======================== Evaluating ==============================
eval_metrics = []
eval_preds = []
eval_labels = []
eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size)
eval_steps = len(eval_dataset) // eval_batch_size
for _ in tqdm(range(eval_steps),
desc="Evaluating...",
position=2,
leave=False):
# Model forward
batch = next(eval_loader)
labels = batch["labels"]
metrics = p_eval_step(state.params, batch)
eval_metrics.append(metrics)
# generation
if data_args.predict_with_generate:
generated_ids = p_generate_step(state.params, batch)
eval_preds.extend(
jax.device_get(
generated_ids.reshape(-1, gen_kwargs["max_length"])))
eval_labels.extend(
jax.device_get(labels.reshape(-1, labels.shape[-1])))
# normalize eval metrics
eval_metrics = get_metrics(eval_metrics)
eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
# compute ROUGE metrics
rouge_desc = ""
if data_args.predict_with_generate:
rouge_metrics = compute_metrics(eval_preds, eval_labels)
eval_metrics.update(rouge_metrics)
rouge_desc = " ".join([
f"Eval {key}: {value} |"
for key, value in rouge_metrics.items()
])
# Print metrics and update progress bar
desc = f"Epoch... ({epoch + 1}/{num_epochs} | Eval Loss: {eval_metrics['loss']} | {rouge_desc})"
epochs.write(desc)
epochs.desc = desc
# Save metrics
if has_tensorboard and jax.process_index() == 0:
cur_step = epoch * (len(train_dataset) // train_batch_size)
write_metric(summary_writer, train_metrics, eval_metrics,
train_time, cur_step)
# save checkpoint after each epoch and push checkpoint to the hub
if jax.process_index() == 0:
params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
model.save_pretrained(training_args.output_dir, params=params)
tokenizer.save_pretrained(training_args.output_dir)
if training_args.push_to_hub:
repo.push_to_hub(
commit_message=f"Saving weights and logs of epoch {epoch}",
blocking=False)
# ======================== Prediction loop ==============================
if training_args.do_predict:
logger.info("*** Predict ***")
pred_metrics = []
pred_generations = []
pred_labels = []
pred_loader = data_loader(input_rng, predict_dataset, eval_batch_size)
pred_steps = len(predict_dataset) // eval_batch_size
for _ in tqdm(range(pred_steps),
desc="Predicting...",
position=2,
leave=False):
# Model forward
batch = next(pred_loader)
labels = batch["labels"]
metrics = p_eval_step(state.params, batch)
pred_metrics.append(metrics)
# generation
if data_args.predict_with_generate:
generated_ids = p_generate_step(state.params, batch)
pred_generations.extend(
jax.device_get(
generated_ids.reshape(-1, gen_kwargs["max_length"])))
pred_labels.extend(
jax.device_get(labels.reshape(-1, labels.shape[-1])))
# normalize prediction metrics
pred_metrics = get_metrics(pred_metrics)
pred_metrics = jax.tree_map(jnp.mean, pred_metrics)
# compute ROUGE metrics
rouge_desc = ""
if data_args.predict_with_generate:
rouge_metrics = compute_metrics(pred_generations, pred_labels)
pred_metrics.update(rouge_metrics)
rouge_desc = " ".join([
f"Predict {key}: {value} |"
for key, value in rouge_metrics.items()
])
# Print metrics
desc = f"Predict Loss: {pred_metrics['loss']} | {rouge_desc})"
logger.info(desc)
# save final metrics in json
if jax.process_index() == 0:
rouge_metrics = {
f"test_{metric_name}": value
for metric_name, value in rouge_metrics.items()
}
path = os.path.join(training_args.output_dir, "test_results.json")
with open(path, "w") as f:
json.dump(rouge_metrics, f, indent=4, sort_keys=True)
def main():
args = parse_args()
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
accelerator = Accelerator()
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state)
# Setup logging, we only want one process per machine to log things on the screen.
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(logging.INFO if accelerator.is_local_main_process else logging.ERROR)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.push_to_hub:
if args.hub_model_id is None:
repo_name = get_full_repo_name(Path(args.output_dir).name, token=args.hub_token)
else:
repo_name = args.hub_model_id
repo = Repository(args.output_dir, clone_from=repo_name)
elif args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
accelerator.wait_for_everyone()
# Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
# or specify a GLUE benchmark task (the dataset will be downloaded automatically from the datasets Hub).
# For CSV/JSON files, this script will use as labels the column called 'label' and as pair of sentences the
# sentences in columns called 'sentence1' and 'sentence2' if such column exists or the first two columns not named
# label if at least two columns are provided.
# If the CSVs/JSONs contain only one non-label column, the script does single sentence classification on this
# single column. You can easily tweak this behavior (see below)
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if args.task_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset("glue", args.task_name)
else:
# Loading the dataset from local csv or json file.
data_files = {}
if args.train_file is not None:
data_files["train"] = args.train_file
if args.validation_file is not None:
data_files["validation"] = args.validation_file
extension = (args.train_file if args.train_file is not None else args.valid_file).split(".")[-1]
raw_datasets = load_dataset(extension, data_files=data_files)
# See more about loading any type of standard or custom dataset at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Labels
if args.task_name is not None:
is_regression = args.task_name == "stsb"
if not is_regression:
label_list = raw_datasets["train"].features["label"].names
num_labels = len(label_list)
else:
num_labels = 1
else:
# Trying to have good defaults here, don't hesitate to tweak to your needs.
is_regression = raw_datasets["train"].features["label"].dtype in ["float32", "float64"]
if is_regression:
num_labels = 1
else:
# A useful fast method:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.unique
label_list = raw_datasets["train"].unique("label")
label_list.sort() # Let's sort it for determinism
num_labels = len(label_list)
# Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(args.model_name_or_path, num_labels=num_labels, finetuning_task=args.task_name)
tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
model = AutoModelForSequenceClassification.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
)
# Preprocessing the datasets
if args.task_name is not None:
sentence1_key, sentence2_key = task_to_keys[args.task_name]
else:
# Again, we try to have some nice defaults but don't hesitate to tweak to your use case.
non_label_column_names = [name for name in raw_datasets["train"].column_names if name != "label"]
if "sentence1" in non_label_column_names and "sentence2" in non_label_column_names:
sentence1_key, sentence2_key = "sentence1", "sentence2"
else:
if len(non_label_column_names) >= 2:
sentence1_key, sentence2_key = non_label_column_names[:2]
else:
sentence1_key, sentence2_key = non_label_column_names[0], None
# Some models have set the order of the labels to use, so let's make sure we do use it.
label_to_id = None
if (
model.config.label2id != PretrainedConfig(num_labels=num_labels).label2id
and args.task_name is not None
and not is_regression
):
# Some have all caps in their config, some don't.
label_name_to_id = {k.lower(): v for k, v in model.config.label2id.items()}
if list(sorted(label_name_to_id.keys())) == list(sorted(label_list)):
logger.info(
f"The configuration of the model provided the following label correspondence: {label_name_to_id}. "
"Using it!"
)
label_to_id = {i: label_name_to_id[label_list[i]] for i in range(num_labels)}
else:
logger.warning(
"Your model seems to have been trained with labels, but they don't match the dataset: ",
f"model labels: {list(sorted(label_name_to_id.keys()))}, dataset labels: {list(sorted(label_list))}."
"\nIgnoring the model labels as a result.",
)
elif args.task_name is None:
label_to_id = {v: i for i, v in enumerate(label_list)}
if label_to_id is not None:
model.config.label2id = label_to_id
model.config.id2label = {id: label for label, id in config.label2id.items()}
elif args.task_name is not None and not is_regression:
model.config.label2id = {l: i for i, l in enumerate(label_list)}
model.config.id2label = {id: label for label, id in config.label2id.items()}
padding = "max_length" if args.pad_to_max_length else False
def preprocess_function(examples):
# Tokenize the texts
texts = (
(examples[sentence1_key],) if sentence2_key is None else (examples[sentence1_key], examples[sentence2_key])
)
result = tokenizer(*texts, padding=padding, max_length=args.max_length, truncation=True)
if "label" in examples:
if label_to_id is not None:
# Map labels to IDs (not necessary for GLUE tasks)
result["labels"] = [label_to_id[l] for l in examples["label"]]
else:
# In all cases, rename the column to labels because the model will expect that.
result["labels"] = examples["label"]
return result
processed_datasets = raw_datasets.map(
preprocess_function,
batched=True,
remove_columns=raw_datasets["train"].column_names,
desc="Running tokenizer on dataset",
)
train_dataset = processed_datasets["train"]
eval_dataset = processed_datasets["validation_matched" if args.task_name == "mnli" else "validation"]
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")
# DataLoaders creation:
if args.pad_to_max_length:
# If padding was already done ot max length, we use the default data collator that will just convert everything
# to tensors.
data_collator = default_data_collator
else:
# Otherwise, `DataCollatorWithPadding` will apply dynamic padding for us (by padding to the maximum length of
# the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple
# of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).
data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=(8 if accelerator.use_fp16 else None))
train_dataloader = DataLoader(
train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size
)
eval_dataloader = DataLoader(eval_dataset, collate_fn=data_collator, batch_size=args.per_device_eval_batch_size)
# Optimizer
# Split weights in two groups, one with weight decay and the other not.
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
"weight_decay": args.weight_decay,
},
{
"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)],
"weight_decay": 0.0,
},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate)
# Prepare everything with our `accelerator`.
model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader
)
# Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be
# shorter in multiprocess)
# Scheduler and math around the number of training steps.
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
else:
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=args.num_warmup_steps,
num_training_steps=args.max_train_steps,
)
# Get the metric function
if args.task_name is not None:
metric = load_metric("glue", args.task_name)
else:
metric = load_metric("accuracy")
# Train!
total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {args.per_device_train_batch_size}")
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
# Only show the progress bar once on each machine.
progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process)
completed_steps = 0
for epoch in range(args.num_train_epochs):
model.train()
for step, batch in enumerate(train_dataloader):
outputs = model(**batch)
loss = outputs.loss
loss = loss / args.gradient_accumulation_steps
accelerator.backward(loss)
if step % args.gradient_accumulation_steps == 0 or step == len(train_dataloader) - 1:
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
progress_bar.update(1)
completed_steps += 1
if completed_steps >= args.max_train_steps:
break
model.eval()
for step, batch in enumerate(eval_dataloader):
outputs = model(**batch)
predictions = outputs.logits.argmax(dim=-1) if not is_regression else outputs.logits.squeeze()
metric.add_batch(
predictions=accelerator.gather(predictions),
references=accelerator.gather(batch["labels"]),
)
eval_metric = metric.compute()
logger.info(f"epoch {epoch}: {eval_metric}")
if args.push_to_hub and epoch < args.num_train_epochs - 1:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir, save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
repo.push_to_hub(commit_message=f"Training in progress epoch {epoch}", blocking=False)
if args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir, save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
if args.push_to_hub:
repo.push_to_hub(commit_message="End of training")
if args.task_name == "mnli":
# Final evaluation on mismatched validation set
eval_dataset = processed_datasets["validation_mismatched"]
eval_dataloader = DataLoader(
eval_dataset, collate_fn=data_collator, batch_size=args.per_device_eval_batch_size
)
eval_dataloader = accelerator.prepare(eval_dataloader)
model.eval()
for step, batch in enumerate(eval_dataloader):
outputs = model(**batch)
predictions = outputs.logits.argmax(dim=-1)
metric.add_batch(
predictions=accelerator.gather(predictions),
references=accelerator.gather(batch["labels"]),
)
eval_metric = metric.compute()
logger.info(f"mnli-mm: {eval_metric}")
def main():
args = parse_args()
if args.source_prefix is None and args.model_name_or_path in [
"t5-small",
"t5-base",
"t5-large",
"t5-3b",
"t5-11b",
]:
logger.warning(
"You're running a t5 model but didn't provide a source prefix, which is the expected, e.g. with "
"`--source_prefix 'summarize: ' `")
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
accelerator = Accelerator()
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state)
# Setup logging, we only want one process per machine to log things on the screen.
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(
logging.INFO if accelerator.is_local_main_process else logging.ERROR)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.push_to_hub:
if args.hub_model_id is None:
repo_name = get_full_repo_name(Path(args.output_dir).name,
token=args.hub_token)
else:
repo_name = args.hub_model_id
repo = Repository(args.output_dir, clone_from=repo_name)
elif args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
accelerator.wait_for_everyone()
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(args.dataset_name,
args.dataset_config_name)
else:
data_files = {}
if args.train_file is not None:
data_files["train"] = args.train_file
if args.validation_file is not None:
data_files["validation"] = args.validation_file
extension = args.train_file.split(".")[-1]
raw_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
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if args.config_name:
config = AutoConfig.from_pretrained(args.config_name)
elif args.model_name_or_path:
config = AutoConfig.from_pretrained(args.model_name_or_path)
else:
config = CONFIG_MAPPING[args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
if args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(
args.tokenizer_name, use_fast=not args.use_slow_tokenizer)
elif args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
args.model_name_or_path, use_fast=not args.use_slow_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 args.model_name_or_path:
model = AutoModelForSeq2SeqLM.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForSeq2SeqLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
prefix = args.source_prefix if args.source_prefix is not None else ""
# Preprocessing the datasets.
# First we tokenize all the texts.
column_names = raw_datasets["train"].column_names
# Get the column names for input/target.
dataset_columns = summarization_name_mapping.get(args.dataset_name, None)
if args.text_column is None:
text_column = dataset_columns[
0] if dataset_columns is not None else column_names[0]
else:
text_column = args.text_column
if text_column not in column_names:
raise ValueError(
f"--text_column' value '{args.text_column}' needs to be one of: {', '.join(column_names)}"
)
if args.summary_column is None:
summary_column = dataset_columns[
1] if dataset_columns is not None else column_names[1]
else:
summary_column = args.summary_column
if summary_column not in column_names:
raise ValueError(
f"--summary_column' value '{args.summary_column}' needs to be one of: {', '.join(column_names)}"
)
# Temporarily set max_target_length for training.
max_target_length = args.max_target_length
padding = "max_length" if args.pad_to_max_length else False
def preprocess_function(examples):
inputs = examples[text_column]
targets = examples[summary_column]
inputs = [prefix + inp for inp in inputs]
model_inputs = tokenizer(inputs,
max_length=args.max_source_length,
padding=padding,
truncation=True)
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets,
max_length=max_target_length,
padding=padding,
truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and args.ignore_pad_token_for_loss:
labels["input_ids"] = [[
(l if l != tokenizer.pad_token_id else -100) for l in label
] for label in labels["input_ids"]]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
processed_datasets = raw_datasets.map(
preprocess_function,
batched=True,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on dataset",
)
train_dataset = processed_datasets["train"]
eval_dataset = processed_datasets["validation"]
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 1):
logger.info(
f"Sample {index} of the training set: {train_dataset[index]}.")
label_pad_token_id = -100 if args.ignore_pad_token_for_loss else tokenizer.pad_token_id
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if accelerator.use_fp16 else None,
)
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [label.strip() for label in labels]
# rougeLSum expects newline after each sentence
preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds]
labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels]
return preds, labels
train_dataloader = DataLoader(train_dataset,
shuffle=True,
collate_fn=data_collator,
batch_size=args.per_device_train_batch_size)
eval_dataloader = DataLoader(eval_dataset,
collate_fn=data_collator,
batch_size=args.per_device_eval_batch_size)
# Optimizer
# Split weights in two groups, one with weight decay and the other not.
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay":
args.weight_decay,
},
{
"params": [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0,
},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate)
# Prepare everything with our `accelerator`.
model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader)
# Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be
# shorter in multiprocess)
# Scheduler and math around the number of training steps.
num_update_steps_per_epoch = math.ceil(
len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
else:
args.num_train_epochs = math.ceil(args.max_train_steps /
num_update_steps_per_epoch)
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=args.num_warmup_steps,
num_training_steps=args.max_train_steps,
)
# Metric
metric = load_metric("rouge")
# Train!
total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(
f" Instantaneous batch size per device = {args.per_device_train_batch_size}"
)
logger.info(
f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}"
)
logger.info(
f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
# Only show the progress bar once on each machine.
progress_bar = tqdm(range(args.max_train_steps),
disable=not accelerator.is_local_main_process)
completed_steps = 0
for epoch in range(args.num_train_epochs):
model.train()
for step, batch in enumerate(train_dataloader):
outputs = model(**batch)
loss = outputs.loss
loss = loss / args.gradient_accumulation_steps
accelerator.backward(loss)
if step % args.gradient_accumulation_steps == 0 or step == len(
train_dataloader) - 1:
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
progress_bar.update(1)
completed_steps += 1
if completed_steps >= args.max_train_steps:
break
model.eval()
if args.val_max_target_length is None:
args.val_max_target_length = args.max_target_length
gen_kwargs = {
"max_length":
args.val_max_target_length
if args is not None else config.max_length,
"num_beams":
args.num_beams,
}
for step, batch in enumerate(eval_dataloader):
with torch.no_grad():
generated_tokens = accelerator.unwrap_model(model).generate(
batch["input_ids"],
attention_mask=batch["attention_mask"],
**gen_kwargs,
)
generated_tokens = accelerator.pad_across_processes(
generated_tokens, dim=1, pad_index=tokenizer.pad_token_id)
labels = batch["labels"]
if not args.pad_to_max_length:
# If we did not pad to max length, we need to pad the labels too
labels = accelerator.pad_across_processes(
batch["labels"],
dim=1,
pad_index=tokenizer.pad_token_id)
generated_tokens = accelerator.gather(
generated_tokens).cpu().numpy()
labels = accelerator.gather(labels).cpu().numpy()
if args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels,
tokenizer.pad_token_id)
if isinstance(generated_tokens, tuple):
generated_tokens = generated_tokens[0]
decoded_preds = tokenizer.batch_decode(
generated_tokens, skip_special_tokens=True)
decoded_labels = tokenizer.batch_decode(
labels, skip_special_tokens=True)
decoded_preds, decoded_labels = postprocess_text(
decoded_preds, decoded_labels)
metric.add_batch(predictions=decoded_preds,
references=decoded_labels)
result = metric.compute(use_stemmer=True)
# Extract a few results from ROUGE
result = {
key: value.mid.fmeasure * 100
for key, value in result.items()
}
result = {k: round(v, 4) for k, v in result.items()}
logger.info(result)
if args.push_to_hub and epoch < args.num_train_epochs - 1:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir,
save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
repo.push_to_hub(
commit_message=f"Training in progress epoch {epoch}",
blocking=False)
if args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir,
save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
if args.push_to_hub:
repo.push_to_hub(commit_message="End of training")
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.")
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
# Setup logging, we only want one process per machine to log things on the screen.
logger.setLevel(logging.INFO if jax.process_index() ==
0 else logging.ERROR)
if jax.process_index() == 0:
transformers.utils.logging.set_verbosity_info()
else:
transformers.utils.logging.set_verbosity_error()
# Set the verbosity to info of the Transformers logger (on main process only):
logger.info(f"Training/evaluation parameters {training_args}")
# set seed for random transforms and torch dataloaders
set_seed(training_args.seed)
# Handle the repository creation
if training_args.push_to_hub:
if training_args.hub_model_id is None:
repo_name = get_full_repo_name(Path(
training_args.output_dir).absolute().name,
token=training_args.hub_token)
else:
repo_name = training_args.hub_model_id
repo = Repository(training_args.output_dir, clone_from=repo_name)
# Initialize datasets and pre-processing transforms
# We use torchvision here for faster pre-processing
# Note that here we are using some default pre-processing, for maximum accuray
# one should tune this part and carefully select what transformations to use.
normalize = transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
train_dataset = torchvision.datasets.ImageFolder(
data_args.train_dir,
transforms.Compose([
transforms.RandomResizedCrop(data_args.image_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
]),
)
eval_dataset = torchvision.datasets.ImageFolder(
data_args.validation_dir,
transforms.Compose([
transforms.Resize(data_args.image_size),
transforms.CenterCrop(data_args.image_size),
transforms.ToTensor(),
normalize,
]),
)
# Load pretrained model and tokenizer
if model_args.config_name:
config = AutoConfig.from_pretrained(
model_args.config_name,
num_labels=len(train_dataset.classes),
image_size=data_args.image_size,
cache_dir=model_args.cache_dir,
)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(
model_args.model_name_or_path,
num_labels=len(train_dataset.classes),
image_size=data_args.image_size,
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.model_name_or_path:
model = FlaxAutoModelForImageClassification.from_pretrained(
model_args.model_name_or_path,
config=config,
seed=training_args.seed,
dtype=getattr(jnp, model_args.dtype))
else:
model = FlaxAutoModelForImageClassification.from_config(
config,
seed=training_args.seed,
dtype=getattr(jnp, model_args.dtype))
# Store some constant
num_epochs = int(training_args.num_train_epochs)
train_batch_size = int(
training_args.per_device_train_batch_size) * jax.device_count()
eval_batch_size = int(
training_args.per_device_eval_batch_size) * jax.device_count()
steps_per_epoch = len(train_dataset) // train_batch_size
total_train_steps = steps_per_epoch * num_epochs
def collate_fn(examples):
pixel_values = torch.stack([example[0] for example in examples])
labels = torch.tensor([example[1] for example in examples])
batch = {"pixel_values": pixel_values, "labels": labels}
batch = {k: v.numpy() for k, v in batch.items()}
return batch
# Create data loaders
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=train_batch_size,
shuffle=True,
num_workers=data_args.preprocessing_num_workers,
persistent_workers=True,
drop_last=True,
collate_fn=collate_fn,
)
eval_loader = torch.utils.data.DataLoader(
eval_dataset,
batch_size=eval_batch_size,
shuffle=False,
num_workers=data_args.preprocessing_num_workers,
persistent_workers=True,
drop_last=True,
collate_fn=collate_fn,
)
# Enable tensorboard only on the master node
has_tensorboard = is_tensorboard_available()
if has_tensorboard and jax.process_index() == 0:
try:
from flax.metrics.tensorboard import SummaryWriter
summary_writer = SummaryWriter(
log_dir=Path(training_args.output_dir))
except ImportError as ie:
has_tensorboard = False
logger.warning(
f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
)
else:
logger.warning(
"Unable to display metrics through TensorBoard because the package is not installed: "
"Please run pip install tensorboard to enable.")
# Initialize our training
rng = jax.random.PRNGKey(training_args.seed)
rng, dropout_rng = jax.random.split(rng)
# Create learning rate schedule
linear_decay_lr_schedule_fn = create_learning_rate_fn(
len(train_dataset),
train_batch_size,
training_args.num_train_epochs,
training_args.warmup_steps,
training_args.learning_rate,
)
# create adam optimizer
adamw = optax.adamw(
learning_rate=linear_decay_lr_schedule_fn,
b1=training_args.adam_beta1,
b2=training_args.adam_beta2,
eps=training_args.adam_epsilon,
weight_decay=training_args.weight_decay,
)
# Setup train state
state = TrainState.create(apply_fn=model.__call__,
params=model.params,
tx=adamw,
dropout_rng=dropout_rng)
def loss_fn(logits, labels):
loss = optax.softmax_cross_entropy(logits,
onehot(labels, logits.shape[-1]))
return loss.mean()
# Define gradient update step fn
def train_step(state, batch):
dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
def compute_loss(params):
labels = batch.pop("labels")
logits = state.apply_fn(**batch,
params=params,
dropout_rng=dropout_rng,
train=True)[0]
loss = loss_fn(logits, labels)
return loss
grad_fn = jax.value_and_grad(compute_loss)
loss, grad = grad_fn(state.params)
grad = jax.lax.pmean(grad, "batch")
new_state = state.apply_gradients(grads=grad,
dropout_rng=new_dropout_rng)
metrics = {
"loss": loss,
"learning_rate": linear_decay_lr_schedule_fn(state.step)
}
metrics = jax.lax.pmean(metrics, axis_name="batch")
return new_state, metrics
# Define eval fn
def eval_step(params, batch):
labels = batch.pop("labels")
logits = model(**batch, params=params, train=False)[0]
loss = loss_fn(logits, labels)
# summarize metrics
accuracy = (jnp.argmax(logits, axis=-1) == labels).mean()
metrics = {"loss": loss, "accuracy": accuracy}
metrics = jax.lax.pmean(metrics, axis_name="batch")
return metrics
# Create parallel version of the train and eval step
p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0, ))
p_eval_step = jax.pmap(eval_step, "batch")
# Replicate the train state on each device
state = state.replicate()
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {num_epochs}")
logger.info(
f" Instantaneous batch size per device = {training_args.per_device_train_batch_size}"
)
logger.info(
f" Total train batch size (w. parallel & distributed) = {train_batch_size}"
)
logger.info(f" Total optimization steps = {total_train_steps}")
train_time = 0
epochs = tqdm(range(num_epochs),
desc=f"Epoch ... (1/{num_epochs})",
position=0)
for epoch in epochs:
# ======================== Training ================================
train_start = time.time()
# Create sampling rng
rng, input_rng = jax.random.split(rng)
train_metrics = []
steps_per_epoch = len(train_dataset) // train_batch_size
train_step_progress_bar = tqdm(total=steps_per_epoch,
desc="Training...",
position=1,
leave=False)
# train
for batch in train_loader:
batch = shard(batch)
state, train_metric = p_train_step(state, batch)
train_metrics.append(train_metric)
train_step_progress_bar.update(1)
train_time += time.time() - train_start
train_metric = unreplicate(train_metric)
train_step_progress_bar.close()
epochs.write(
f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
)
# ======================== Evaluating ==============================
eval_metrics = []
eval_steps = len(eval_dataset) // eval_batch_size
eval_step_progress_bar = tqdm(total=eval_steps,
desc="Evaluating...",
position=2,
leave=False)
for batch in eval_loader:
# Model forward
batch = shard(batch)
metrics = p_eval_step(state.params, batch)
eval_metrics.append(metrics)
eval_step_progress_bar.update(1)
# normalize eval metrics
eval_metrics = get_metrics(eval_metrics)
eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
# Print metrics and update progress bar
eval_step_progress_bar.close()
desc = (
f"Epoch... ({epoch + 1}/{num_epochs} | Eval Loss: {round(eval_metrics['loss'].item(), 4)} | "
f"Eval Accuracy: {round(eval_metrics['accuracy'].item(), 4)})")
epochs.write(desc)
epochs.desc = desc
# Save metrics
if has_tensorboard and jax.process_index() == 0:
cur_step = epoch * (len(train_dataset) // train_batch_size)
write_metric(summary_writer, train_metrics, eval_metrics,
train_time, cur_step)
# save checkpoint after each epoch and push checkpoint to the hub
if jax.process_index() == 0:
params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
model.save_pretrained(training_args.output_dir, params=params)
if training_args.push_to_hub:
repo.push_to_hub(
commit_message=f"Saving weights and logs of epoch {epoch}",
blocking=False)
def main():
args = parse_args()
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
accelerator = Accelerator()
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state)
# Setup logging, we only want one process per machine to log things on the screen.
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(
logging.INFO if accelerator.is_local_main_process else logging.ERROR)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.push_to_hub:
if args.hub_model_id is None:
repo_name = get_full_repo_name(Path(args.output_dir).name,
token=args.hub_token)
else:
repo_name = args.hub_model_id
repo = Repository(args.output_dir, clone_from=repo_name)
elif args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
accelerator.wait_for_everyone()
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets for token classification task available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'tokens' or the first column if no column called
# 'tokens' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(args.dataset_name,
args.dataset_config_name)
else:
data_files = {}
if args.train_file is not None:
data_files["train"] = args.train_file
if args.validation_file is not None:
data_files["validation"] = args.validation_file
extension = args.train_file.split(".")[-1]
raw_datasets = load_dataset(extension, data_files=data_files)
# Trim a number of training examples
if args.debug:
for split in raw_datasets.keys():
raw_datasets[split] = raw_datasets[split].select(range(100))
# 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 raw_datasets["train"] is not None:
column_names = raw_datasets["train"].column_names
features = raw_datasets["train"].features
else:
column_names = raw_datasets["validation"].column_names
features = raw_datasets["validation"].features
if args.text_column_name is not None:
text_column_name = args.text_column_name
elif "tokens" in column_names:
text_column_name = "tokens"
else:
text_column_name = column_names[0]
if args.label_column_name is not None:
label_column_name = args.label_column_name
elif f"{args.task_name}_tags" in column_names:
label_column_name = f"{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
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if args.config_name:
config = AutoConfig.from_pretrained(args.config_name,
num_labels=num_labels)
elif args.model_name_or_path:
config = AutoConfig.from_pretrained(args.model_name_or_path,
num_labels=num_labels)
else:
config = CONFIG_MAPPING[args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
tokenizer_name_or_path = args.tokenizer_name if args.tokenizer_name else args.model_name_or_path
if not tokenizer_name_or_path:
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 config.model_type in {"gpt2", "roberta"}:
tokenizer = AutoTokenizer.from_pretrained(tokenizer_name_or_path,
use_fast=True,
add_prefix_space=True)
else:
tokenizer = AutoTokenizer.from_pretrained(tokenizer_name_or_path,
use_fast=True)
if args.model_name_or_path:
model = AutoModelForTokenClassification.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForTokenClassification.from_config(config)
model.resize_token_embeddings(len(tokenizer))
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 datasets.
# First we tokenize all the texts.
padding = "max_length" if 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],
max_length=args.max_length,
padding=padding,
truncation=True,
# We use this argument because the texts in our dataset are lists of words (with a label for each word).
is_split_into_words=True,
)
labels = []
for i, label in enumerate(examples[label_column_name]):
word_ids = tokenized_inputs.word_ids(batch_index=i)
previous_word_idx = None
label_ids = []
for word_idx in word_ids:
# Special tokens have a word id that is None. We set the label to -100 so they are automatically
# ignored in the loss function.
if word_idx is None:
label_ids.append(-100)
# We set the label for the first token of each word.
elif word_idx != previous_word_idx:
label_ids.append(label_to_id[label[word_idx]])
# For the other tokens in a word, we set the label to either the current label or -100, depending on
# the label_all_tokens flag.
else:
if 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
with accelerator.main_process_first():
processed_raw_datasets = raw_datasets.map(
tokenize_and_align_labels,
batched=True,
remove_columns=raw_datasets["train"].column_names,
desc="Running tokenizer on dataset",
)
train_dataset = processed_raw_datasets["train"]
eval_dataset = processed_raw_datasets["validation"]
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(
f"Sample {index} of the training set: {train_dataset[index]}.")
# DataLoaders creation:
if args.pad_to_max_length:
# If padding was already done ot max length, we use the default data collator that will just convert everything
# to tensors.
data_collator = default_data_collator
else:
# Otherwise, `DataCollatorForTokenClassification` will apply dynamic padding for us (by padding to the maximum length of
# the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple
# of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).
data_collator = DataCollatorForTokenClassification(
tokenizer,
pad_to_multiple_of=(8 if accelerator.use_fp16 else None))
train_dataloader = DataLoader(train_dataset,
shuffle=True,
collate_fn=data_collator,
batch_size=args.per_device_train_batch_size)
eval_dataloader = DataLoader(eval_dataset,
collate_fn=data_collator,
batch_size=args.per_device_eval_batch_size)
# Optimizer
# Split weights in two groups, one with weight decay and the other not.
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay":
args.weight_decay,
},
{
"params": [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0,
},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate)
# Use the device given by the `accelerator` object.
device = accelerator.device
model.to(device)
# Prepare everything with our `accelerator`.
model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader)
# Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be
# shorter in multiprocess)
# Scheduler and math around the number of training steps.
num_update_steps_per_epoch = math.ceil(
len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
else:
args.num_train_epochs = math.ceil(args.max_train_steps /
num_update_steps_per_epoch)
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=args.num_warmup_steps,
num_training_steps=args.max_train_steps,
)
# Metrics
metric = load_metric("seqeval")
def get_labels(predictions, references):
# Transform predictions and references tensos to numpy arrays
if device.type == "cpu":
y_pred = predictions.detach().clone().numpy()
y_true = references.detach().clone().numpy()
else:
y_pred = predictions.detach().cpu().clone().numpy()
y_true = references.detach().cpu().clone().numpy()
# Remove ignored index (special tokens)
true_predictions = [[
label_list[p] for (p, l) in zip(pred, gold_label) if l != -100
] for pred, gold_label in zip(y_pred, y_true)]
true_labels = [[
label_list[l] for (p, l) in zip(pred, gold_label) if l != -100
] for pred, gold_label in zip(y_pred, y_true)]
return true_predictions, true_labels
def compute_metrics():
results = metric.compute()
if 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"],
}
# Train!
total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(
f" Instantaneous batch size per device = {args.per_device_train_batch_size}"
)
logger.info(
f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}"
)
logger.info(
f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
# Only show the progress bar once on each machine.
progress_bar = tqdm(range(args.max_train_steps),
disable=not accelerator.is_local_main_process)
completed_steps = 0
for epoch in range(args.num_train_epochs):
model.train()
for step, batch in enumerate(train_dataloader):
outputs = model(**batch)
loss = outputs.loss
loss = loss / args.gradient_accumulation_steps
accelerator.backward(loss)
if step % args.gradient_accumulation_steps == 0 or step == len(
train_dataloader) - 1:
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
progress_bar.update(1)
completed_steps += 1
if completed_steps >= args.max_train_steps:
break
model.eval()
for step, batch in enumerate(eval_dataloader):
with torch.no_grad():
outputs = model(**batch)
predictions = outputs.logits.argmax(dim=-1)
labels = batch["labels"]
if not args.pad_to_max_length: # necessary to pad predictions and labels for being gathered
predictions = accelerator.pad_across_processes(predictions,
dim=1,
pad_index=-100)
labels = accelerator.pad_across_processes(labels,
dim=1,
pad_index=-100)
predictions_gathered = accelerator.gather(predictions)
labels_gathered = accelerator.gather(labels)
preds, refs = get_labels(predictions_gathered, labels_gathered)
metric.add_batch(
predictions=preds,
references=refs,
) # predictions and preferences are expected to be a nested list of labels, not label_ids
# eval_metric = metric.compute()
eval_metric = compute_metrics()
accelerator.print(f"epoch {epoch}:", eval_metric)
if args.push_to_hub and epoch < args.num_train_epochs - 1:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir,
save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
repo.push_to_hub(
commit_message=f"Training in progress epoch {epoch}",
blocking=False,
auto_lfs_prune=True)
if args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir,
save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
if args.push_to_hub:
repo.push_to_hub(commit_message="End of training",
auto_lfs_prune=True)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
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."
)
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
# Setup logging, we only want one process per machine to log things on the screen.
logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
if jax.process_index() == 0:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# Set the verbosity to info of the Transformers logger (on main process only):
logger.info(f"Training/evaluation parameters {training_args}")
# Set seed before initializing model.
set_seed(training_args.seed)
# Handle the repository creation
if training_args.push_to_hub:
if training_args.hub_model_id is None:
repo_name = get_full_repo_name(
Path(training_args.output_dir).absolute().name, token=training_args.hub_token
)
else:
repo_name = training_args.hub_model_id
repo = Repository(training_args.output_dir, clone_from=repo_name)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets 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 guarantees 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.
dataset = load_dataset(
data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir, keep_in_memory=False
)
if "validation" not in dataset.keys():
dataset["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,
)
dataset["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 = {}
dataset_args = {}
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"
dataset_args["keep_linebreaks"] = data_args.keep_linebreaks
dataset = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir, **dataset_args)
if "validation" not in dataset.keys():
dataset["validation"] = load_dataset(
extension,
data_files=data_files,
split=f"train[:{data_args.validation_split_percentage}%]",
cache_dir=model_args.cache_dir,
**dataset_args,
)
dataset["train"] = load_dataset(
extension,
data_files=data_files,
split=f"train[{data_args.validation_split_percentage}%:]",
cache_dir=model_args.cache_dir,
**dataset_args,
)
# 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 = FlaxAutoModelForCausalLM.from_pretrained(
model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
)
else:
model = FlaxAutoModelForCausalLM.from_config(
config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
)
# Preprocessing the datasets.
# First we tokenize all the texts.
if training_args.do_train:
column_names = dataset["train"].column_names
else:
column_names = dataset["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 = dataset.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 > config.max_position_embeddings:
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: list(chain(*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.
if total_length >= block_size:
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))
# Enable tensorboard only on the master node
has_tensorboard = is_tensorboard_available()
if has_tensorboard and jax.process_index() == 0:
try:
from flax.metrics.tensorboard import SummaryWriter
summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
except ImportError as ie:
has_tensorboard = False
logger.warning(
f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
)
else:
logger.warning(
"Unable to display metrics through TensorBoard because the package is not installed: "
"Please run pip install tensorboard to enable."
)
# Initialize our training
rng = jax.random.PRNGKey(training_args.seed)
rng, dropout_rng = jax.random.split(rng)
# Store some constant
num_epochs = int(training_args.num_train_epochs)
train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
steps_per_epoch = len(train_dataset) // train_batch_size
total_train_steps = steps_per_epoch * num_epochs
# Create learning rate schedule
linear_decay_lr_schedule_fn = create_learning_rate_fn(
len(train_dataset),
train_batch_size,
training_args.num_train_epochs,
training_args.warmup_steps,
training_args.learning_rate,
)
# We use Optax's "masking" functionality to not apply weight decay
# to bias and LayerNorm scale parameters. decay_mask_fn returns a
# mask boolean with the same structure as the parameters.
# The mask is True for parameters that should be decayed.
# Note that this mask is specifically adapted for FlaxGPT2.
# For other models, one should correct the layer norm parameter naming
# accordingly.
def decay_mask_fn(params):
flat_params = traverse_util.flatten_dict(params)
flat_mask = {
path: (path[-1] != "bias" and path[-2:] not in [("ln_1", "scale"), ("ln_2", "scale"), ("ln_f", "scale")])
for path in flat_params
}
return traverse_util.unflatten_dict(flat_mask)
# create adam optimizer
if training_args.adafactor:
# We use the default parameters here to initialize adafactor,
# For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
optimizer = optax.adafactor(
learning_rate=linear_decay_lr_schedule_fn,
)
else:
optimizer = optax.adamw(
learning_rate=linear_decay_lr_schedule_fn,
b1=training_args.adam_beta1,
b2=training_args.adam_beta2,
eps=training_args.adam_epsilon,
weight_decay=training_args.weight_decay,
mask=decay_mask_fn,
)
# Setup train state
state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer, dropout_rng=dropout_rng)
def loss_fn(logits, labels):
shift_logits = logits[..., :-1, :]
shift_labels = labels[..., 1:]
loss = optax.softmax_cross_entropy(shift_logits, onehot(shift_labels, shift_logits.shape[-1]))
return loss.mean()
# Define gradient update step fn
def train_step(state, batch):
dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
def compute_loss(params):
labels = batch.pop("labels")
logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
loss = loss_fn(logits, labels)
return loss
grad_fn = jax.value_and_grad(compute_loss)
loss, grad = grad_fn(state.params)
grad = jax.lax.pmean(grad, "batch")
new_state = state.apply_gradients(grads=grad, dropout_rng=new_dropout_rng)
metrics = {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}
metrics = jax.lax.pmean(metrics, axis_name="batch")
return new_state, metrics
# Define eval fn
def eval_step(params, batch):
labels = batch.pop("labels")
logits = model(**batch, params=params, train=False)[0]
loss = loss_fn(logits, labels)
# summarize metrics
metrics = {"loss": loss}
metrics = jax.lax.pmean(metrics, axis_name="batch")
return metrics
# Create parallel version of the train and eval step
p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
p_eval_step = jax.pmap(eval_step, "batch")
# Replicate the train state on each device
state = state.replicate()
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {num_epochs}")
logger.info(f" Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
logger.info(f" Total train batch size (w. parallel & distributed) = {train_batch_size}")
logger.info(f" Total optimization steps = {total_train_steps}")
train_time = 0
train_metrics = []
epochs = tqdm(range(num_epochs), desc="Epoch ... ", position=0)
for epoch in epochs:
# ======================== Training ================================
train_start = time.time()
# Create sampling rng
rng, input_rng = jax.random.split(rng)
# Generate an epoch by shuffling sampling indices from the train dataset
train_loader = data_loader(input_rng, train_dataset, train_batch_size, shuffle=True)
steps_per_epoch = len(train_dataset) // train_batch_size
# train
for step in tqdm(range(steps_per_epoch), desc="Training...", position=1, leave=False):
batch = next(train_loader)
batch = shard(batch)
state, train_metric = p_train_step(state, batch)
train_metrics.append(train_metric)
cur_step = epoch * (len(train_dataset) // train_batch_size) + step
if cur_step % training_args.logging_steps == 0 and cur_step > 0:
# Save metrics
train_metric = unreplicate(train_metric)
train_time += time.time() - train_start
if has_tensorboard and jax.process_index() == 0:
write_train_metric(summary_writer, train_metrics, train_time, cur_step)
epochs.write(
f"Step... ({cur_step} | Loss: {train_metric['loss'].mean()}, Learning Rate: {train_metric['learning_rate'].mean()})"
)
train_metrics = []
if cur_step % training_args.eval_steps == 0 and cur_step > 0:
# ======================== Evaluating ==============================
eval_metrics = []
eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size)
eval_steps = len(eval_dataset) // eval_batch_size
for _ in tqdm(range(eval_steps), desc="Evaluating...", position=2, leave=False):
# Model forward
batch = next(eval_loader)
batch = shard(batch)
metrics = p_eval_step(state.params, batch)
eval_metrics.append(metrics)
# normalize eval metrics
eval_metrics = get_metrics(eval_metrics)
eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
try:
eval_metrics["perplexity"] = math.exp(eval_metrics["loss"])
except OverflowError:
eval_metrics["perplexity"] = float("inf")
# Print metrics and update progress bar
desc = f"Step... ({cur_step} | Eval Loss: {eval_metrics['loss']} | Eval Perplexity: {eval_metrics['perplexity']})"
epochs.write(desc)
epochs.desc = desc
# Save metrics
if has_tensorboard and jax.process_index() == 0:
write_eval_metric(summary_writer, eval_metrics, cur_step)
if cur_step % training_args.save_steps == 0 and cur_step > 0:
# save checkpoint after each epoch and push checkpoint to the hub
if jax.process_index() == 0:
params = jax.device_get(unreplicate(state.params))
model.save_pretrained(training_args.output_dir, params=params)
tokenizer.save_pretrained(training_args.output_dir)
if training_args.push_to_hub:
repo.push_to_hub(commit_message=f"Saving weights and logs of step {cur_step}", blocking=False)
# Eval after training
if training_args.do_eval:
eval_metrics = []
eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size)
eval_steps = len(eval_dataset) // eval_batch_size
for _ in tqdm(range(eval_steps), desc="Evaluating...", position=2, leave=False):
# Model forward
batch = shard(next(eval_loader))
metrics = p_eval_step(state.params, batch)
eval_metrics.append(metrics)
# normalize eval metrics
eval_metrics = get_metrics(eval_metrics)
eval_metrics = jax.tree_map(lambda x: jnp.mean(x).item(), eval_metrics)
try:
eval_metrics["perplexity"] = math.exp(eval_metrics["loss"])
except OverflowError:
eval_metrics["perplexity"] = float("inf")
if jax.process_index() == 0:
eval_metrics = {f"eval_{metric_name}": value for metric_name, value in eval_metrics.items()}
path = os.path.join(training_args.output_dir, "eval_results.json")
with open(path, "w") as f:
json.dump(eval_metrics, f, indent=4, sort_keys=True)
def main():
# region Argument parsing
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, 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(
)
# endregion
# region Logging
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
# Setup logging, we only want one process per machine to log things on the screen.
logger.setLevel(logging.INFO if jax.process_index() ==
0 else logging.ERROR)
if jax.process_index() == 0:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# endregion
# Handle the repository creation
if training_args.push_to_hub:
if training_args.hub_model_id is None:
repo_name = get_full_repo_name(Path(
training_args.output_dir).absolute().name,
token=training_args.hub_token)
else:
repo_name = training_args.hub_model_id
repo = Repository(training_args.output_dir, clone_from=repo_name)
# region Load Data
# 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:
# Loading the dataset from local csv or json file.
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
raw_datasets = load_dataset(extension,
data_files=data_files,
field="data",
cache_dir=model_args.cache_dir)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# endregion
# region Load pretrained model and tokenizer
#
# Load pretrained model and tokenizer
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=True,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# endregion
# region 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")
# endregion
# region Preprocessing the datasets
# Preprocessing is slightly different for training and evaluation.
if training_args.do_train:
column_names = raw_datasets["train"].column_names
elif training_args.do_eval:
column_names = raw_datasets["validation"].column_names
else:
column_names = raw_datasets["test"].column_names
question_column_name = "question" if "question" in column_names else column_names[
0]
context_column_name = "context" if "context" in column_names else column_names[
1]
answer_column_name = "answers" if "answers" in column_names else column_names[
2]
# Padding side determines if we do (question|context) or (context|question).
pad_on_right = tokenizer.padding_side == "right"
if data_args.max_seq_length > tokenizer.model_max_length:
logger.warning(
f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the"
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
# Training preprocessing
def prepare_train_features(examples):
# Some of the questions have lots of whitespace on the left, which is not useful and will make the
# truncation of the context fail (the tokenized question will take a lots of space). So we remove that
# left whitespace
examples[question_column_name] = [
q.lstrip() for q in examples[question_column_name]
]
# Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
# in one example possible giving several features when a context is long, each of those features having a
# context that overlaps a bit the context of the previous feature.
tokenized_examples = tokenizer(
examples[
question_column_name if pad_on_right else context_column_name],
examples[
context_column_name if pad_on_right else question_column_name],
truncation="only_second" if pad_on_right else "only_first",
max_length=max_seq_length,
stride=data_args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
padding="max_length",
)
# Since one example might give us several features if it has a long context, we need a map from a feature to
# its corresponding example. This key gives us just that.
sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
# The offset mappings will give us a map from token to character position in the original context. This will
# help us compute the start_positions and end_positions.
offset_mapping = tokenized_examples.pop("offset_mapping")
# Let's label those examples!
tokenized_examples["start_positions"] = []
tokenized_examples["end_positions"] = []
for i, offsets in enumerate(offset_mapping):
# We will label impossible answers with the index of the CLS token.
input_ids = tokenized_examples["input_ids"][i]
cls_index = input_ids.index(tokenizer.cls_token_id)
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples.sequence_ids(i)
# One example can give several spans, this is the index of the example containing this span of text.
sample_index = sample_mapping[i]
answers = examples[answer_column_name][sample_index]
# If no answers are given, set the cls_index as answer.
if len(answers["answer_start"]) == 0:
tokenized_examples["start_positions"].append(cls_index)
tokenized_examples["end_positions"].append(cls_index)
else:
# Start/end character index of the answer in the text.
start_char = answers["answer_start"][0]
end_char = start_char + len(answers["text"][0])
# Start token index of the current span in the text.
token_start_index = 0
while sequence_ids[token_start_index] != (1 if pad_on_right
else 0):
token_start_index += 1
# End token index of the current span in the text.
token_end_index = len(input_ids) - 1
while sequence_ids[token_end_index] != (1 if pad_on_right else
0):
token_end_index -= 1
# Detect if the answer is out of the span (in which case this feature is labeled with the CLS index).
if not (offsets[token_start_index][0] <= start_char
and offsets[token_end_index][1] >= end_char):
tokenized_examples["start_positions"].append(cls_index)
tokenized_examples["end_positions"].append(cls_index)
else:
# Otherwise move the token_start_index and token_end_index to the two ends of the answer.
# Note: we could go after the last offset if the answer is the last word (edge case).
while token_start_index < len(offsets) and offsets[
token_start_index][0] <= start_char:
token_start_index += 1
tokenized_examples["start_positions"].append(
token_start_index - 1)
while offsets[token_end_index][1] >= end_char:
token_end_index -= 1
tokenized_examples["end_positions"].append(
token_end_index + 1)
return tokenized_examples
processed_raw_datasets = dict()
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:
# We will select sample from whole data if agument is specified
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
# Create train feature from dataset
train_dataset = train_dataset.map(
prepare_train_features,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if data_args.max_train_samples is not None:
# Number of samples might increase during Feature Creation, We select only specified max samples
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
processed_raw_datasets["train"] = train_dataset
# Validation preprocessing
def prepare_validation_features(examples):
# Some of the questions have lots of whitespace on the left, which is not useful and will make the
# truncation of the context fail (the tokenized question will take a lots of space). So we remove that
# left whitespace
examples[question_column_name] = [
q.lstrip() for q in examples[question_column_name]
]
# Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
# in one example possible giving several features when a context is long, each of those features having a
# context that overlaps a bit the context of the previous feature.
tokenized_examples = tokenizer(
examples[
question_column_name if pad_on_right else context_column_name],
examples[
context_column_name if pad_on_right else question_column_name],
truncation="only_second" if pad_on_right else "only_first",
max_length=max_seq_length,
stride=data_args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
padding="max_length",
)
# Since one example might give us several features if it has a long context, we need a map from a feature to
# its corresponding example. This key gives us just that.
sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
# For evaluation, we will need to convert our predictions to substrings of the context, so we keep the
# corresponding example_id and we will store the offset mappings.
tokenized_examples["example_id"] = []
for i in range(len(tokenized_examples["input_ids"])):
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples.sequence_ids(i)
context_index = 1 if pad_on_right else 0
# One example can give several spans, this is the index of the example containing this span of text.
sample_index = sample_mapping[i]
tokenized_examples["example_id"].append(
examples["id"][sample_index])
# Set to None the offset_mapping that are not part of the context so it's easy to determine if a token
# position is part of the context or not.
tokenized_examples["offset_mapping"][i] = [
(o if sequence_ids[k] == context_index else None)
for k, o in enumerate(tokenized_examples["offset_mapping"][i])
]
return tokenized_examples
if training_args.do_eval:
if "validation" not in raw_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_examples = raw_datasets["validation"]
if data_args.max_eval_samples is not None:
# We will select sample from whole data
eval_examples = eval_examples.select(
range(data_args.max_eval_samples))
# Validation Feature Creation
eval_dataset = eval_examples.map(
prepare_validation_features,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if data_args.max_eval_samples is not None:
# During Feature creation dataset samples might increase, we will select required samples again
eval_dataset = eval_dataset.select(
range(data_args.max_eval_samples))
processed_raw_datasets["validation"] = eval_dataset
if training_args.do_predict:
if "test" not in raw_datasets:
raise ValueError("--do_predict requires a test dataset")
predict_examples = raw_datasets["test"]
if data_args.max_predict_samples is not None:
# We will select sample from whole data
predict_examples = predict_examples.select(
range(data_args.max_predict_samples))
# Predict Feature Creation
predict_dataset = predict_examples.map(
prepare_validation_features,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if data_args.max_predict_samples is not None:
# During Feature creation dataset samples might increase, we will select required samples again
predict_dataset = predict_dataset.select(
range(data_args.max_predict_samples))
processed_raw_datasets["test"] = predict_dataset
# endregion
# region Metrics and Post-processing:
def post_processing_function(examples,
features,
predictions,
stage="eval"):
# Post-processing: we match the start logits and end logits to answers in the original context.
predictions = postprocess_qa_predictions(
examples=examples,
features=features,
predictions=predictions,
version_2_with_negative=data_args.version_2_with_negative,
n_best_size=data_args.n_best_size,
max_answer_length=data_args.max_answer_length,
null_score_diff_threshold=data_args.null_score_diff_threshold,
output_dir=training_args.output_dir,
prefix=stage,
)
# Format the result to the format the metric expects.
if data_args.version_2_with_negative:
formatted_predictions = [{
"id": k,
"prediction_text": v,
"no_answer_probability": 0.0
} for k, v in predictions.items()]
else:
formatted_predictions = [{
"id": k,
"prediction_text": v
} for k, v in predictions.items()]
references = [{
"id": ex["id"],
"answers": ex[answer_column_name]
} for ex in examples]
return EvalPrediction(predictions=formatted_predictions,
label_ids=references)
metric = load_metric(
"squad_v2" if data_args.version_2_with_negative else "squad")
def compute_metrics(p: EvalPrediction):
return metric.compute(predictions=p.predictions,
references=p.label_ids)
# Create and fill numpy array of size len_of_validation_data * max_length_of_output_tensor
def create_and_fill_np_array(start_or_end_logits, dataset, max_len):
"""
Create and fill numpy array of size len_of_validation_data * max_length_of_output_tensor
Args:
start_or_end_logits(:obj:`tensor`):
This is the output predictions of the model. We can only enter either start or end logits.
eval_dataset: Evaluation dataset
max_len(:obj:`int`):
The maximum length of the output tensor. ( See the model.eval() part for more details )
"""
step = 0
# create a numpy array and fill it with -100.
logits_concat = np.full((len(dataset), max_len),
-100,
dtype=np.float64)
# Now since we have create an array now we will populate it with the outputs of the model.
for i, output_logit in enumerate(
start_or_end_logits): # populate columns
# We have to fill it such that we have to take the whole tensor and replace it on the newly created array
# And after every iteration we have to change the step
batch_size = output_logit.shape[0]
cols = output_logit.shape[1]
if step + batch_size < len(dataset):
logits_concat[step:step + batch_size, :cols] = output_logit
else:
logits_concat[step:, :cols] = output_logit[:len(dataset) -
step]
step += batch_size
return logits_concat
# endregion
# region Training steps and logging init
train_dataset = processed_raw_datasets["train"]
eval_dataset = processed_raw_datasets["validation"]
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(
f"Sample {index} of the training set: {train_dataset[index]}.")
# Define a summary writer
summary_writer = tensorboard.SummaryWriter(training_args.output_dir)
summary_writer.hparams({
**training_args.to_dict(),
**vars(model_args),
**vars(data_args)
})
def write_train_metric(summary_writer, train_metrics, train_time, step):
summary_writer.scalar("train_time", train_time, step)
train_metrics = get_metrics(train_metrics)
for key, vals in train_metrics.items():
tag = f"train_{key}"
for i, val in enumerate(vals):
summary_writer.scalar(tag, val, step - len(vals) + i + 1)
def write_eval_metric(summary_writer, eval_metrics, step):
for metric_name, value in eval_metrics.items():
summary_writer.scalar(f"eval_{metric_name}", value, step)
num_epochs = int(training_args.num_train_epochs)
rng = jax.random.PRNGKey(training_args.seed)
dropout_rngs = jax.random.split(rng, jax.local_device_count())
train_batch_size = training_args.per_device_train_batch_size * jax.local_device_count(
)
eval_batch_size = training_args.per_device_eval_batch_size * jax.local_device_count(
)
# endregion
# region Load model
model = FlaxAutoModelForQuestionAnswering.from_pretrained(
model_args.model_name_or_path,
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
seed=training_args.seed,
dtype=getattr(jnp, model_args.dtype),
)
learning_rate_fn = create_learning_rate_fn(
len(train_dataset),
train_batch_size,
training_args.num_train_epochs,
training_args.warmup_steps,
training_args.learning_rate,
)
state = create_train_state(model,
learning_rate_fn,
num_labels=max_seq_length,
training_args=training_args)
# endregion
# region Define train step functions
def train_step(
state: train_state.TrainState, batch: Dict[str, Array],
dropout_rng: PRNGKey) -> Tuple[train_state.TrainState, float]:
"""Trains model with an optimizer (both in `state`) on `batch`, returning a pair `(new_state, loss)`."""
dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
start_positions = batch.pop("start_positions")
end_positions = batch.pop("end_positions")
targets = (start_positions, end_positions)
def loss_fn(params):
logits = state.apply_fn(**batch,
params=params,
dropout_rng=dropout_rng,
train=True)
loss = state.loss_fn(logits, targets)
return loss
grad_fn = jax.value_and_grad(loss_fn)
loss, grad = grad_fn(state.params)
grad = jax.lax.pmean(grad, "batch")
new_state = state.apply_gradients(grads=grad)
metrics = jax.lax.pmean(
{
"loss": loss,
"learning_rate": learning_rate_fn(state.step)
},
axis_name="batch")
return new_state, metrics, new_dropout_rng
p_train_step = jax.pmap(train_step,
axis_name="batch",
donate_argnums=(0, ))
# endregion
# region Define eval step functions
def eval_step(state, batch):
logits = state.apply_fn(**batch, params=state.params, train=False)
return state.logits_fn(logits)
p_eval_step = jax.pmap(eval_step, axis_name="batch")
# endregion
# region Define train and eval loop
logger.info(f"===== Starting training ({num_epochs} epochs) =====")
train_time = 0
# make sure weights are replicated on each device
state = replicate(state)
train_time = 0
step_per_epoch = len(train_dataset) // train_batch_size
total_steps = step_per_epoch * num_epochs
epochs = tqdm(range(num_epochs),
desc=f"Epoch ... (1/{num_epochs})",
position=0)
for epoch in epochs:
train_start = time.time()
train_metrics = []
# Create sampling rng
rng, input_rng = jax.random.split(rng)
# train
for step, batch in enumerate(
tqdm(
train_data_collator(input_rng, train_dataset,
train_batch_size),
total=step_per_epoch,
desc="Training...",
position=1,
),
1,
):
state, train_metric, dropout_rngs = p_train_step(
state, batch, dropout_rngs)
train_metrics.append(train_metric)
cur_step = epoch * step_per_epoch + step
if cur_step % training_args.logging_steps == 0 and cur_step > 0:
# Save metrics
train_metric = unreplicate(train_metric)
train_time += time.time() - train_start
if jax.process_index() == 0:
write_train_metric(summary_writer, train_metrics,
train_time, cur_step)
epochs.write(
f"Step... ({cur_step}/{total_steps} | Training Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
)
train_metrics = []
if (training_args.do_eval
and (cur_step % training_args.eval_steps == 0
or cur_step % step_per_epoch == 0) and cur_step > 0):
eval_metrics = {}
all_start_logits = []
all_end_logits = []
# evaluate
for batch in tqdm(
eval_data_collator(eval_dataset, eval_batch_size),
total=len(eval_dataset) // eval_batch_size,
desc="Evaluating ...",
position=2,
):
_ = batch.pop("example_id")
_ = batch.pop("offset_mapping")
predictions = p_eval_step(state, batch)
start_logits = np.array(
[pred for pred in chain(*predictions[0])])
end_logits = np.array(
[pred for pred in chain(*predictions[1])])
all_start_logits.append(start_logits)
all_end_logits.append(end_logits)
# evaluate also on leftover examples (not divisible by batch_size)
num_leftover_samples = len(eval_dataset) % eval_batch_size
# make sure leftover batch is evaluated on one device
if num_leftover_samples > 0 and jax.process_index() == 0:
# take leftover samples
batch = eval_dataset[-num_leftover_samples:]
batch = {k: np.array(v) for k, v in batch.items()}
_ = batch.pop("example_id")
_ = batch.pop("offset_mapping")
predictions = eval_step(unreplicate(state), batch)
start_logits = np.array([pred for pred in predictions[0]])
end_logits = np.array([pred for pred in predictions[1]])
all_start_logits.append(start_logits)
all_end_logits.append(end_logits)
max_len = max([x.shape[1] for x in all_start_logits
]) # Get the max_length of the tensor
# concatenate the numpy array
start_logits_concat = create_and_fill_np_array(
all_start_logits, eval_dataset, max_len)
end_logits_concat = create_and_fill_np_array(
all_end_logits, eval_dataset, max_len)
# delete the list of numpy arrays
del all_start_logits
del all_end_logits
outputs_numpy = (start_logits_concat, end_logits_concat)
prediction = post_processing_function(eval_examples,
eval_dataset,
outputs_numpy)
eval_metrics = compute_metrics(prediction)
logger.info(
f"Step... ({cur_step}/{total_steps} | Evaluation metrics: {eval_metrics})"
)
if jax.process_index() == 0:
write_eval_metric(summary_writer, eval_metrics, cur_step)
if (cur_step % training_args.save_steps == 0
and cur_step > 0) or (cur_step == total_steps):
# save checkpoint after each epoch and push checkpoint to the hub
if jax.process_index() == 0:
params = jax.device_get(unreplicate(state.params))
model.save_pretrained(training_args.output_dir,
params=params)
tokenizer.save_pretrained(training_args.output_dir)
if training_args.push_to_hub:
repo.push_to_hub(
commit_message=
f"Saving weights and logs of step {cur_step}",
blocking=False)
epochs.desc = f"Epoch ... {epoch + 1}/{num_epochs}"
)
# Log on each process the small summary:
logger = logging.getLogger(__name__)
# Set the verbosity to info of the Transformers logger (on main process only):
logger.info(f"Training/evaluation parameters {training_args}")
# Set seed before initializing model.
set_seed(training_args.seed)
# Handle the repository creation
if training_args.push_to_hub:
if training_args.hub_model_id is None:
repo_name = get_full_repo_name(
Path(training_args.output_dir).absolute().name, token=training_args.hub_token
)
else:
repo_name = training_args.hub_model_id
repo = Repository(training_args.output_dir, clone_from=repo_name)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets 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 guarantees that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
def main():
args = parse_args()
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
accelerator = Accelerator()
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state)
# Setup logging, we only want one process per machine to log things on the screen.
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(
logging.INFO if accelerator.is_local_main_process else logging.ERROR)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.push_to_hub:
if args.hub_model_id is None:
repo_name = get_full_repo_name(Path(args.output_dir).name,
token=args.hub_token)
else:
repo_name = args.hub_model_id
repo = Repository(args.output_dir, clone_from=repo_name)
elif args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
accelerator.wait_for_everyone()
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(args.dataset_name,
args.dataset_config_name)
else:
data_files = {}
if args.train_file is not None:
data_files["train"] = args.train_file
if args.validation_file is not None:
data_files["validation"] = args.validation_file
extension = args.train_file.split(".")[-1]
raw_datasets = load_dataset(extension, data_files=data_files)
# Trim a number of training examples
if args.debug:
for split in raw_datasets.keys():
raw_datasets[split] = raw_datasets[split].select(range(100))
# 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 raw_datasets["train"] is not None:
column_names = raw_datasets["train"].column_names
else:
column_names = raw_datasets["validation"].column_names
# 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"
label_column_name = "label" if "label" in column_names else "labels"
# Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if args.config_name:
config = AutoConfig.from_pretrained(args.model_name_or_path)
elif args.model_name_or_path:
config = AutoConfig.from_pretrained(args.model_name_or_path)
else:
config = CONFIG_MAPPING[args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
if args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(
args.tokenizer_name, use_fast=not args.use_slow_tokenizer)
elif args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
args.model_name_or_path, use_fast=not args.use_slow_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 args.model_name_or_path:
model = AutoModelForMultipleChoice.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForMultipleChoice.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
padding = "max_length" if args.pad_to_max_length else False
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)]
labels = examples[label_column_name]
# Flatten out
first_sentences = sum(first_sentences, [])
second_sentences = sum(second_sentences, [])
# Tokenize
tokenized_examples = tokenizer(
first_sentences,
second_sentences,
max_length=args.max_length,
padding=padding,
truncation=True,
)
# Un-flatten
tokenized_inputs = {
k: [v[i:i + 4] for i in range(0, len(v), 4)]
for k, v in tokenized_examples.items()
}
tokenized_inputs["labels"] = labels
return tokenized_inputs
processed_datasets = raw_datasets.map(
preprocess_function,
batched=True,
remove_columns=raw_datasets["train"].column_names)
train_dataset = processed_datasets["train"]
eval_dataset = processed_datasets["validation"]
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(
f"Sample {index} of the training set: {train_dataset[index]}.")
# DataLoaders creation:
if args.pad_to_max_length:
# If padding was already done ot max length, we use the default data collator that will just convert everything
# to tensors.
data_collator = default_data_collator
else:
# Otherwise, `DataCollatorWithPadding` will apply dynamic padding for us (by padding to the maximum length of
# the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple
# of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).
data_collator = DataCollatorForMultipleChoice(
tokenizer,
pad_to_multiple_of=(8 if accelerator.use_fp16 else None))
train_dataloader = DataLoader(train_dataset,
shuffle=True,
collate_fn=data_collator,
batch_size=args.per_device_train_batch_size)
eval_dataloader = DataLoader(eval_dataset,
collate_fn=data_collator,
batch_size=args.per_device_eval_batch_size)
# Optimizer
# Split weights in two groups, one with weight decay and the other not.
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay":
args.weight_decay,
},
{
"params": [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0,
},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate)
# Use the device given by the `accelerator` object.
device = accelerator.device
model.to(device)
# Prepare everything with our `accelerator`.
model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader)
# Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be
# shorter in multiprocess)
# Scheduler and math around the number of training steps.
num_update_steps_per_epoch = math.ceil(
len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
else:
args.num_train_epochs = math.ceil(args.max_train_steps /
num_update_steps_per_epoch)
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=args.num_warmup_steps,
num_training_steps=args.max_train_steps,
)
# Metrics
metric = load_metric("accuracy")
# Train!
total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(
f" Instantaneous batch size per device = {args.per_device_train_batch_size}"
)
logger.info(
f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}"
)
logger.info(
f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
# Only show the progress bar once on each machine.
progress_bar = tqdm(range(args.max_train_steps),
disable=not accelerator.is_local_main_process)
completed_steps = 0
for epoch in range(args.num_train_epochs):
model.train()
for step, batch in enumerate(train_dataloader):
outputs = model(**batch)
loss = outputs.loss
loss = loss / args.gradient_accumulation_steps
accelerator.backward(loss)
if step % args.gradient_accumulation_steps == 0 or step == len(
train_dataloader) - 1:
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
progress_bar.update(1)
completed_steps += 1
if completed_steps >= args.max_train_steps:
break
model.eval()
for step, batch in enumerate(eval_dataloader):
with torch.no_grad():
outputs = model(**batch)
predictions = outputs.logits.argmax(dim=-1)
metric.add_batch(
predictions=accelerator.gather(predictions),
references=accelerator.gather(batch["labels"]),
)
eval_metric = metric.compute()
accelerator.print(f"epoch {epoch}: {eval_metric}")
if args.push_to_hub and epoch < args.num_train_epochs - 1:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir,
save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
repo.push_to_hub(
commit_message=f"Training in progress epoch {epoch}",
blocking=False)
if args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir,
save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
if args.push_to_hub:
repo.push_to_hub(commit_message="End of training")
def main():
args = parse_args()
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
accelerator = Accelerator()
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state)
# Setup logging, we only want one process per machine to log things on the screen.
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(
logging.INFO if accelerator.is_local_main_process else logging.ERROR)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.push_to_hub:
if args.hub_model_id is None:
repo_name = get_full_repo_name(Path(args.output_dir).name,
token=args.hub_token)
else:
repo_name = args.hub_model_id
repo = Repository(args.output_dir, clone_from=repo_name)
elif args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
accelerator.wait_for_everyone()
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(args.dataset_name,
args.dataset_config_name)
if "validation" not in raw_datasets.keys():
raw_datasets["validation"] = load_dataset(
args.dataset_name,
args.dataset_config_name,
split=f"train[:{args.validation_split_percentage}%]",
)
raw_datasets["train"] = load_dataset(
args.dataset_name,
args.dataset_config_name,
split=f"train[{args.validation_split_percentage}%:]",
)
else:
data_files = {}
if args.train_file is not None:
data_files["train"] = args.train_file
if args.validation_file is not None:
data_files["validation"] = args.validation_file
extension = args.train_file.split(".")[-1]
if extension == "txt":
extension = "text"
raw_datasets = load_dataset(extension, data_files=data_files)
# If no validation data is there, validation_split_percentage will be used to divide the dataset.
if "validation" not in raw_datasets.keys():
raw_datasets["validation"] = load_dataset(
extension,
data_files=data_files,
split=f"train[:{args.validation_split_percentage}%]",
)
raw_datasets["train"] = load_dataset(
extension,
data_files=data_files,
split=f"train[{args.validation_split_percentage}%:]",
)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if args.config_name:
config = AutoConfig.from_pretrained(args.config_name)
elif args.model_name_or_path:
config = AutoConfig.from_pretrained(args.model_name_or_path)
else:
config = CONFIG_MAPPING[args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
if args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(
args.tokenizer_name, use_fast=not args.use_slow_tokenizer)
elif args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
args.model_name_or_path, use_fast=not args.use_slow_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 args.model_name_or_path:
model = AutoModelForMaskedLM.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForMaskedLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
column_names = raw_datasets["train"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
if 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 args.max_seq_length > tokenizer.model_max_length:
logger.warning(
f"The max_seq_length passed ({args.max_seq_length}) is larger than the maximum length for the"
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(args.max_seq_length, tokenizer.model_max_length)
if args.line_by_line:
# When using line_by_line, we just tokenize each nonempty line.
padding = "max_length" if args.pad_to_max_length else False
def tokenize_function(examples):
# Remove empty lines
examples[text_column_name] = [
line for line in examples[text_column_name]
if len(line) > 0 and not line.isspace()
]
return tokenizer(
examples[text_column_name],
padding=padding,
truncation=True,
max_length=max_seq_length,
# We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
# receives the `special_tokens_mask`.
return_special_tokens_mask=True,
)
tokenized_datasets = raw_datasets.map(
tokenize_function,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=[text_column_name],
load_from_cache_file=not 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.
# We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
# efficient when it receives the `special_tokens_mask`.
def tokenize_function(examples):
return tokenizer(examples[text_column_name],
return_special_tokens_mask=True)
tokenized_datasets = raw_datasets.map(
tokenize_function,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not 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.
if total_length >= max_seq_length:
total_length = (total_length //
max_seq_length) * max_seq_length
# Split by chunks of max_len.
result = {
k: [
t[i:i + max_seq_length]
for i in range(0, total_length, max_seq_length)
]
for k, t in concatenated_examples.items()
}
return result
# Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
# remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
# might be slower to preprocess.
#
# To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
tokenized_datasets = tokenized_datasets.map(
group_texts,
batched=True,
num_proc=args.preprocessing_num_workers,
load_from_cache_file=not args.overwrite_cache,
desc=f"Grouping texts in chunks of {max_seq_length}",
)
train_dataset = tokenized_datasets["train"]
eval_dataset = tokenized_datasets["validation"]
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(
f"Sample {index} of the training set: {train_dataset[index]}.")
# Data collator
# This one will take care of randomly masking the tokens.
data_collator = DataCollatorForLanguageModeling(
tokenizer=tokenizer, mlm_probability=args.mlm_probability)
# DataLoaders creation:
train_dataloader = DataLoader(train_dataset,
shuffle=True,
collate_fn=data_collator,
batch_size=args.per_device_train_batch_size)
eval_dataloader = DataLoader(eval_dataset,
collate_fn=data_collator,
batch_size=args.per_device_eval_batch_size)
# Optimizer
# Split weights in two groups, one with weight decay and the other not.
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay":
args.weight_decay,
},
{
"params": [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0,
},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate)
# Prepare everything with our `accelerator`.
model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader)
# On TPU, the tie weights in our model have been disconnected, so we need to restore the ties.
if accelerator.distributed_type == DistributedType.TPU:
model.tie_weights()
# Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be
# shorter in multiprocess)
# Scheduler and math around the number of training steps.
num_update_steps_per_epoch = math.ceil(
len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
else:
args.num_train_epochs = math.ceil(args.max_train_steps /
num_update_steps_per_epoch)
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=args.num_warmup_steps,
num_training_steps=args.max_train_steps,
)
# Train!
total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(
f" Instantaneous batch size per device = {args.per_device_train_batch_size}"
)
logger.info(
f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}"
)
logger.info(
f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
# Only show the progress bar once on each machine.
progress_bar = tqdm(range(args.max_train_steps),
disable=not accelerator.is_local_main_process)
completed_steps = 0
for epoch in range(args.num_train_epochs):
model.train()
for step, batch in enumerate(train_dataloader):
outputs = model(**batch)
loss = outputs.loss
loss = loss / args.gradient_accumulation_steps
accelerator.backward(loss)
if step % args.gradient_accumulation_steps == 0 or step == len(
train_dataloader) - 1:
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
progress_bar.update(1)
completed_steps += 1
if completed_steps >= args.max_train_steps:
break
model.eval()
losses = []
for step, batch in enumerate(eval_dataloader):
with torch.no_grad():
outputs = model(**batch)
loss = outputs.loss
losses.append(
accelerator.gather(loss.repeat(
args.per_device_eval_batch_size)))
losses = torch.cat(losses)
losses = losses[:len(eval_dataset)]
try:
perplexity = math.exp(torch.mean(losses))
except OverflowError:
perplexity = float("inf")
logger.info(f"epoch {epoch}: perplexity: {perplexity}")
if args.push_to_hub and epoch < args.num_train_epochs - 1:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir,
save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
repo.push_to_hub(
commit_message=f"Training in progress epoch {epoch}",
blocking=False)
if args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir,
save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
if args.push_to_hub:
repo.push_to_hub(commit_message="End of training")
def main():
# See all possible arguments in src/transformers/args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
args = parse_args()
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
accelerator = Accelerator()
logger.info(accelerator.state)
# Setup logging, we only want one process per machine to log things on the screen.
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(logging.INFO if accelerator.is_local_main_process else logging.ERROR)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
# set up weights and biases if available
if is_wandb_available():
import wandb
wandb.init(project=args.output_dir.split("/")[-1])
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.push_to_hub and not args.preprocessing_only:
if args.hub_model_id is None:
repo_name = get_full_repo_name(Path(args.output_dir).name, token=args.hub_token)
else:
repo_name = args.hub_model_id
repo = Repository(args.output_dir, clone_from=repo_name)
elif args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
accelerator.wait_for_everyone()
# 1. Download and create train, validation dataset
# We load all dataset configuration and datset split pairs passed in
# ``args.dataset_config_names`` and ``args.dataset_split_names``
datasets_splits = []
for dataset_config_name, train_split_name in zip(args.dataset_config_names, args.dataset_split_names):
# load dataset
dataset_split = load_dataset(
args.dataset_name, dataset_config_name, split=train_split_name, cache_dir=args.cache_dir
)
datasets_splits.append(dataset_split)
# Next, we concatenate all configurations and splits into a single training dataset
raw_datasets = DatasetDict()
if len(datasets_splits) > 1:
raw_datasets["train"] = concatenate_datasets(datasets_splits).shuffle(seed=args.seed)
else:
raw_datasets["train"] = datasets_splits[0]
# Take ``args.validation_split_percentage`` from the training dataset for the validation_split_percentage
num_validation_samples = raw_datasets["train"].num_rows * args.validation_split_percentage // 100
if num_validation_samples == 0:
raise ValueError(
"`args.validation_split_percentage` is less than a single sample "
f"for {len(raw_datasets['train'])} training samples. Increase "
"`args.num_validation_split_percentage`. "
)
raw_datasets["validation"] = raw_datasets["train"].select(range(num_validation_samples))
raw_datasets["train"] = raw_datasets["train"].select(range(num_validation_samples, raw_datasets["train"].num_rows))
# 2. Preprocess audio: load, resample, normalize and truncate
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(args.model_name_or_path)
# only normalized-inputs-training is supported
if not feature_extractor.do_normalize:
raise ValueError(
"Training is only supported for normalized inputs. " "Make sure ``feature_extractor.do_normalize == True``"
)
# set max & min audio length in number of samples
max_length = int(args.max_duration_in_seconds * feature_extractor.sampling_rate)
min_length = int(args.min_duration_in_seconds * feature_extractor.sampling_rate)
resampler = None
if raw_datasets["train"][args.audio_column_name][0].split(".")[-1] == "mp3":
# TODO(PVP) - remove hard-coded 48_000 after audio feature is merged
resampler = torchaudio.transforms.Resample(48_000, feature_extractor.sampling_rate)
def prepare_dataset(batch):
speech_array, sampling_rate = torchaudio.load(batch[args.audio_column_name])
speech_array = speech_array.squeeze()
# if necessary resample audio
if resampler is not None:
# TODO(PVP) - remove hard-coded 48_000 after audio feature is merged
speech_array = resampler(speech_array)
sampling_rate = resampler.new_freq
speech_array = speech_array.numpy()
inputs = feature_extractor(speech_array, sampling_rate=sampling_rate, max_length=max_length, truncation=True)
batch["input_values"] = inputs.input_values[0]
return batch
# load audio files into numpy arrays
with accelerator.main_process_first():
vectorized_datasets = raw_datasets.map(
prepare_dataset,
num_proc=args.preprocessing_num_workers,
remove_columns=raw_datasets["train"].column_names,
load_from_cache_file=not args.overwrite_cache,
)
vectorized_datasets = vectorized_datasets.filter(
lambda x: len(x["input_values"]) > min_length, load_from_cache_file=not args.overwrite_cache
)
# 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 args.preprocessing_only:
return
# 3. Load model
config = Wav2Vec2Config.from_pretrained(args.model_name_or_path)
# pretraining is only supported for "newer" stable layer norm architecture
# apply_spec_augment has to be True, mask_feature_prob has to be 0.0
if not config.do_stable_layer_norm or config.feat_extract_norm != "layer":
raise ValueError(
"PreTraining is only supported for ``config.do_stable_layer_norm=True`` and ``config.feat_extract_norm='layer'"
)
# initialize random model
model = Wav2Vec2ForPreTraining(config)
# Activate gradient checkpointing if needed
if args.gradient_checkpointing:
model.gradient_checkpointing_enable()
# 4. Define data collator, optimizer and scheduler
data_collator = DataCollatorForWav2Vec2Pretraining(
model=model, feature_extractor=feature_extractor, pad_to_multiple_of=args.pad_to_multiple_of
)
train_dataloader = DataLoader(
vectorized_datasets["train"],
shuffle=True,
collate_fn=data_collator,
batch_size=args.per_device_train_batch_size,
)
eval_dataloader = DataLoader(
vectorized_datasets["validation"], collate_fn=data_collator, batch_size=args.per_device_eval_batch_size
)
# Optimizer
optimizer = AdamW(
list(model.parameters()),
lr=args.learning_rate,
betas=[args.adam_beta1, args.adam_beta2],
eps=args.adam_epsilon,
)
# Prepare everything with our `accelerator`.
model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader
)
# Scheduler and math around the number of training steps.
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
else:
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=args.num_warmup_steps,
num_training_steps=args.max_train_steps,
)
# 5. Train
total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(vectorized_datasets['train'])}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {args.per_device_train_batch_size}")
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
completed_steps = 0
# Only show the progress bar once on each machine.
progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process)
completed_steps = 0
for epoch in range(args.num_train_epochs):
model.train()
for step, batch in enumerate(train_dataloader):
# compute num of losses
num_losses = batch["mask_time_indices"].sum()
sub_attention_mask = batch.pop("sub_attention_mask", None)
sub_attention_mask = (
sub_attention_mask if sub_attention_mask is not None else torch.ones_like(batch["mask_time_indices"])
)
percent_masked = num_losses / sub_attention_mask.sum()
# forward
outputs = model(**batch)
# divide loss by gradient accumulation steps since gradients
# are accumulated for multiple backward passes in PyTorch
loss = outputs.loss / args.gradient_accumulation_steps
accelerator.backward(loss)
# make sure that `num_losses` is summed for distributed training
# and average gradients over losses of all devices
if accelerator.state.num_processes > 1:
num_losses = accelerator.gather(num_losses).sum()
gradient_multiplier = accelerator.state.num_processes / num_losses
multiply_grads(model.module.parameters(), gradient_multiplier)
else:
multiply_grads(model.parameters(), 1 / num_losses)
# update step
if (step + 1) % args.gradient_accumulation_steps == 0 or step == len(train_dataloader) - 1:
# compute grad norm for monitoring
scale = (
accelerator.scaler._scale.item()
if hasattr(accelerator, "scaler") and accelerator.scaler is not None
else 1
)
if accelerator.state.num_processes > 1:
grad_norm = get_grad_norm(model.module.parameters(), scale)
else:
grad_norm = get_grad_norm(model.parameters(), scale)
# update parameters
optimizer.step()
optimizer.zero_grad()
if not accelerator.optimizer_step_was_skipped:
lr_scheduler.step()
elif accelerator.is_local_main_process:
progress_bar.write(
"Gradients have overflown - skipping update step... " f"Updating gradient scale to {scale}..."
)
# update gumbel temperature
gumbel_temperature = max(
args.max_gumbel_temperature * args.gumbel_temperature_decay ** completed_steps,
args.min_gumbel_temperature,
)
if hasattr(model, "module"):
model.module.set_gumbel_temperature(gumbel_temperature)
else:
model.set_gumbel_temperature(gumbel_temperature)
progress_bar.update(1)
completed_steps += 1
# 6. Log all results
if (step + 1) % (args.gradient_accumulation_steps * args.logging_steps) == 0:
loss.detach()
outputs.contrastive_loss.detach()
outputs.diversity_loss.detach()
if accelerator.state.num_processes > 1:
loss = accelerator.gather(loss).sum()
outputs.contrastive_loss = accelerator.gather(outputs.contrastive_loss).sum()
outputs.diversity_loss = accelerator.gather(outputs.diversity_loss).sum()
percent_masked = accelerator.gather(percent_masked).sum()
train_logs = {
"loss": (loss * args.gradient_accumulation_steps) / num_losses,
"constrast_loss": outputs.contrastive_loss / num_losses,
"div_loss": outputs.diversity_loss / num_losses,
"%_mask_idx": percent_masked / accelerator.num_processes,
"ppl": outputs.codevector_perplexity,
"lr": torch.tensor(optimizer.param_groups[0]["lr"]),
"temp": torch.tensor(gumbel_temperature),
"grad_norm": torch.tensor(grad_norm),
}
log_str = ""
for k, v in train_logs.items():
log_str += "| {}: {:.3e}".format(k, v.item())
if accelerator.is_local_main_process:
progress_bar.write(log_str)
if is_wandb_available():
wandb.log(train_logs)
# save model every `args.saving_steps` steps
if (step + 1) % (args.gradient_accumulation_steps * args.saving_steps) == 0:
if (args.push_to_hub and epoch < args.num_train_epochs - 1) or args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir, save_function=accelerator.save)
if (args.push_to_hub and epoch < args.num_train_epochs - 1) and accelerator.is_main_process:
repo.push_to_hub(commit_message=f"Training in progress step {completed_steps}", blocking=False)
# if completed steps > `args.max_train_steps` stop
if completed_steps >= args.max_train_steps:
break
# 7. Validate!
model.eval()
# init logs
val_logs = {
"val_loss": 0,
"val_contrastive_loss": 0,
"val_diversity_loss": 0,
"val_num_losses": 0,
}
for step, batch in enumerate(eval_dataloader):
with torch.no_grad():
batch.pop("sub_attention_mask", None)
outputs = model(**batch)
val_logs["val_loss"] += outputs.loss
val_logs["val_contrastive_loss"] += outputs.contrastive_loss
val_logs["val_diversity_loss"] += outputs.diversity_loss
val_logs["val_num_losses"] += batch["mask_time_indices"].sum()
# sum over devices in multi-processing
if accelerator.num_processes > 1:
val_logs = {k: accelerator.gather(v).sum() for k, v in val_logs.items()}
val_logs = {k: v / val_logs["val_num_losses"] for k, v in val_logs.items()}
log_str = ""
for k, v in val_logs.items():
log_str += "| {}: {:.3e}".format(k, v.item())
if accelerator.is_local_main_process:
progress_bar.write(log_str)
if is_wandb_available():
wandb.log(val_logs)
if args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir, save_function=accelerator.save)
if accelerator.is_main_process:
if args.push_to_hub:
repo.push_to_hub(commit_message="End of training")