def create_learning_rate_fn(config,):
"""Create learning rate schedule."""
# Linear warmup
warmup_fn = optax.linear_schedule(
init_value=0.,
end_value=config.train.lr_init,
transition_steps=config.train.warmup_steps)
if config.train.scheduler == "linear":
decay_fn = optax.linear_schedule(
init_value=config.train.lr_init,
end_value=0.,
transition_steps=config.train.max_steps - config.train.warmup_steps)
elif config.train.scheduler == "cosine":
cosine_steps = max(config.train.max_steps - config.train.warmup_steps, 1)
decay_fn = optax.cosine_decay_schedule(
init_value=config.train.lr_init, decay_steps=cosine_steps)
elif config.train.scheduler == "step":
step_steps = max(config.train.max_steps - config.train.warmup_steps, 1) # pylint: disable=unused-variable
def schedule(count):
return config.train.lr_init * (0.5**(count // 50000))
decay_fn = schedule
else:
raise NotImplementedError
schedule_fn = optax.join_schedules(
schedules=[warmup_fn, decay_fn], boundaries=[config.train.warmup_steps])
return schedule_fn
def get_learning_rate_schedule(
total_batch_size, steps_per_epoch, total_steps, optimizer_config):
"""Build the learning rate schedule function."""
base_lr = _get_batch_scaled_lr(total_batch_size, optimizer_config.base_lr,
optimizer_config.scale_by_batch)
schedule_type = optimizer_config.schedule_type
if schedule_type == 'steps':
boundaries = optimizer_config.step_decay_kwargs.decay_boundaries
boundaries.sort()
decay_rate = optimizer_config.step_decay_kwargs.decay_rate
boundaries_and_scales = {
int(boundary * total_steps): decay_rate for boundary in boundaries}
schedule_fn = optax.piecewise_constant_schedule(
init_value=base_lr, boundaries_and_scales=boundaries_and_scales)
elif schedule_type == 'cosine':
warmup_steps = (optimizer_config.cosine_decay_kwargs.warmup_epochs
* steps_per_epoch)
# Batch scale the other lr values as well:
init_value = _get_batch_scaled_lr(
total_batch_size,
optimizer_config.cosine_decay_kwargs.init_value,
optimizer_config.scale_by_batch)
end_value = _get_batch_scaled_lr(
total_batch_size,
optimizer_config.cosine_decay_kwargs.end_value,
optimizer_config.scale_by_batch)
schedule_fn = optax.warmup_cosine_decay_schedule(
init_value=init_value,
peak_value=base_lr,
warmup_steps=warmup_steps,
decay_steps=total_steps,
end_value=end_value)
elif schedule_type == 'constant_cosine':
# Convert end_value to alpha, used by cosine_decay_schedule.
alpha = optimizer_config.constant_cosine_decay_kwargs.end_value / base_lr
# Number of steps spent in constant phase.
constant_steps = int(
optimizer_config.constant_cosine_decay_kwargs.constant_fraction
* total_steps)
decay_steps = total_steps - constant_steps
constant_phase = optax.constant_schedule(value=base_lr)
decay_phase = optax.cosine_decay_schedule(
init_value=base_lr,
decay_steps=decay_steps,
alpha=alpha)
schedule_fn = optax.join_schedules(
schedules=[constant_phase, decay_phase],
boundaries=[constant_steps])
else:
raise ValueError(f'Unknown learning rate schedule: {schedule_type}')
return schedule_fn
def create_learning_rate_schedule(learning_rate: float, warmup_steps: int):
"""Creates a rsqrt schedule with linear warmup."""
return optax.join_schedules([
optax.linear_schedule(init_value=0,
end_value=learning_rate,
transition_steps=warmup_steps),
rsqrt_schedule(init_value=learning_rate, shift=warmup_steps),
],
boundaries=[warmup_steps])
def scheduler_fn(lr, init_lr, warmup_steps, num_train_steps):
decay_steps = num_train_steps - warmup_steps
warmup_fn = optax.linear_schedule(init_value=init_lr,
end_value=lr,
transition_steps=warmup_steps)
decay_fn = optax.linear_schedule(init_value=lr,
end_value=1e-7,
transition_steps=decay_steps)
lr = optax.join_schedules(schedules=[warmup_fn, decay_fn],
boundaries=[warmup_steps])
return lr
def create_learning_rate_fn(
train_ds_size: int, train_batch_size: int, num_train_epochs: int, num_warmup_steps: int, learning_rate: float
) -> Callable[[int], jnp.array]:
"""Returns a linear warmup, linear_decay learning rate function."""
steps_per_epoch = train_ds_size // train_batch_size
num_train_steps = steps_per_epoch * num_train_epochs
warmup_fn = optax.linear_schedule(init_value=0.0, end_value=learning_rate, transition_steps=num_warmup_steps)
decay_fn = optax.linear_schedule(
init_value=learning_rate, end_value=0, transition_steps=num_train_steps - num_warmup_steps
)
schedule_fn = optax.join_schedules(schedules=[warmup_fn, decay_fn], boundaries=[num_warmup_steps])
return schedule_fn
def create_learning_rate_fn(workload: spec.Workload,
hparams: spec.Hyperparameters):
"""Create learning rate schedule."""
warmup_fn = optax.linear_schedule(init_value=0.,
end_value=hparams.learning_rate,
transition_steps=hparams.warmup_steps)
cosine_fn = optax.cosine_decay_schedule(init_value=hparams.learning_rate,
decay_steps=(workload.step_hint -
hparams.warmup_steps))
schedule_fn = optax.join_schedules(schedules=[warmup_fn, cosine_fn],
boundaries=[hparams.warmup_steps])
return schedule_fn
def get_cosine_schedule(
max_learning_rate: float,
total_steps: int,
warmup_steps: int = 0) -> optax.Schedule:
"""Builds a cosine decay schedule with initial warm-up."""
if total_steps < warmup_steps:
return optax.linear_schedule(init_value=0., end_value=max_learning_rate,
transition_steps=warmup_steps)
return optax.join_schedules([
optax.linear_schedule(init_value=0., end_value=max_learning_rate,
transition_steps=warmup_steps),
optax.cosine_decay_schedule(init_value=max_learning_rate,
decay_steps=total_steps - warmup_steps),
], [warmup_steps])
def create_learning_rate_fn(hparams: spec.Hyperparamters, steps_per_epoch: int):
"""Create learning rate schedule."""
base_learning_rate = hparams.learning_rate * get_batch_size('imagenet') / 256.
warmup_fn = optax.linear_schedule(
init_value=0.,
end_value=base_learning_rate,
transition_steps=hparams.warmup_epochs * steps_per_epoch)
cosine_epochs = max(hparams.num_epochs - hparams.warmup_epochs, 1)
cosine_fn = optax.cosine_decay_schedule(
init_value=base_learning_rate,
decay_steps=cosine_epochs * steps_per_epoch)
schedule_fn = optax.join_schedules(
schedules=[warmup_fn, cosine_fn],
boundaries=[hparams.warmup_epochs * steps_per_epoch])
return schedule_fn
def get_step_schedule(
max_learning_rate: float,
total_steps: int,
warmup_steps: int = 0) -> optax.Schedule:
"""Builds a step schedule with initial warm-up."""
if total_steps < warmup_steps:
return optax.linear_schedule(init_value=0., end_value=max_learning_rate,
transition_steps=warmup_steps)
return optax.join_schedules([
optax.linear_schedule(init_value=0., end_value=max_learning_rate,
transition_steps=warmup_steps),
optax.piecewise_constant_schedule(
init_value=max_learning_rate,
boundaries_and_scales={total_steps * 2 // 3: .1}),
], [warmup_steps])
def create_learning_rate_fn(config: ml_collections.ConfigDict,
base_learning_rate: float, steps_per_epoch: int):
"""Create learning rate schedule."""
warmup_fn = optax.linear_schedule(init_value=0.,
end_value=base_learning_rate,
transition_steps=config.warmup_epochs *
steps_per_epoch)
cosine_epochs = max(config.num_epochs - config.warmup_epochs, 1)
cosine_fn = optax.cosine_decay_schedule(init_value=base_learning_rate,
decay_steps=cosine_epochs *
steps_per_epoch)
schedule_fn = optax.join_schedules(
schedules=[warmup_fn, cosine_fn],
boundaries=[config.warmup_epochs * steps_per_epoch])
return schedule_fn
def create_learning_rate_fn(config,):
"""Create learning rate schedule."""
# Linear warmup
warmup_fn = optax.linear_schedule(
init_value=0.,
end_value=config.train.lr_init,
transition_steps=config.train.warmup_steps)
decay_fn = optax.linear_schedule(
init_value=config.train.lr_init,
end_value=0.,
transition_steps=config.train.max_steps - config.train.warmup_steps)
#cosine_steps = max(config.train.max_steps - config.train.warmup_steps, 1)
#decay_fn = optax.cosine_decay_schedule(
# init_value=config.train.lr_init,
# decay_steps=cosine_steps)
schedule_fn = optax.join_schedules(
schedules=[warmup_fn, decay_fn], boundaries=[config.train.warmup_steps])
return schedule_fn
def value(self, count: JTensor) -> JTensor:
p = self.params
return jnp.array(
optax.join_schedules([s.value for s in self._schedules],
p.boundaries)(count), jnp.float32)
def value(self, value: JTensor) -> JTensor:
p = self.params
return jnp.array(
optax.join_schedules([self.p0.value, self.p1.value],
p.boundaries[:1])(value), jnp.float32)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
configure_logger(model_args, training_args)
# 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():
# make sure only "validation" and "train" keys remain"
datasets = DatasetDict()
datasets["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"{data_args.train_split_name}[:{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"{data_args.train_split_name}[{data_args.validation_split_percentage}%:]",
cache_dir=model_args.cache_dir,
)
else:
# make sure only "validation" and "train" keys remain"
datasets = DatasetDict()
datasets["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split="validation",
cache_dir=model_args.cache_dir,
)
datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"{data_args.train_split_name}",
cache_dir=model_args.cache_dir,
)
# only normalized-inputs-training is supported
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir, do_normalize=True
)
def prepare_dataset(batch):
# check that all files have the correct sampling rate
batch["speech"], _ = librosa.load(batch[data_args.speech_file_column], sr=feature_extractor.sampling_rate)
return batch
# load audio files into numpy arrays
vectorized_datasets = datasets.map(
prepare_dataset, num_proc=data_args.preprocessing_num_workers, remove_columns=datasets["train"].column_names
)
# filter audio files that are too long
vectorized_datasets = vectorized_datasets.filter(
lambda data: len(data["speech"]) < int(data_args.max_duration_in_seconds * feature_extractor.sampling_rate)
)
def normalize(batch):
return feature_extractor(batch["speech"], sampling_rate=feature_extractor.sampling_rate)
# normalize and transform to `BatchFeatures`
vectorized_datasets = vectorized_datasets.map(
normalize,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
remove_columns=vectorized_datasets["train"].column_names,
)
# 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
config = Wav2Vec2Config.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
gradient_checkpointing=model_args.gradient_checkpointing,
)
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'"
)
model = FlaxWav2Vec2ForPreTraining(
config, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
)
data_collator = FlaxDataCollatorForWav2Vec2Pretraining(
model=model, feature_extractor=feature_extractor, pad_to_multiple_of=data_args.pad_to_multiple_of
)
# 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())
gumbel_rngs = jax.random.split(rng, jax.local_device_count())
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(vectorized_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
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 and define training hyper-parameters
state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=adamw)
num_negatives = model.config.num_negatives
contrastive_logits_temperature = model.config.contrastive_logits_temperature
num_codevectors = model.config.num_codevectors_per_group * model.config.num_codevector_groups
diversity_loss_weight = model.config.diversity_loss_weight
# Define gradient update step fn
def train_step(state, batch, dropout_rng, gumbel_rng):
dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
gumbel_rng, new_gumbel_rng = jax.random.split(gumbel_rng)
def loss_fn(params):
negative_indices = batch.pop("sampled_negative_indices")
gumbel_temperature = jnp.clip(
model_args.max_gumbel_temperature * model_args.gumbel_temperature_decay ** state.step,
a_min=model_args.min_gumbel_temperature,
)
outputs = state.apply_fn(
**batch,
gumbel_temperature=gumbel_temperature,
params=params,
dropout_rng=dropout_rng,
gumbel_rng=gumbel_rng,
train=True,
)
contrastive_loss = compute_contrastive_loss(
outputs.projected_quantized_states,
outputs.projected_states,
negative_indices,
batch["mask_time_indices"],
contrastive_logits_temperature,
num_negatives,
)
diversity_loss = (num_codevectors - outputs.codevector_perplexity) / num_codevectors
loss = contrastive_loss + diversity_loss_weight * diversity_loss
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, new_gumbel_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):
negative_indices = batch.pop("sampled_negative_indices")
outputs = model(**batch, params=params, train=False)
contrastive_loss = compute_contrastive_loss(
outputs.projected_quantized_states,
outputs.projected_states,
negative_indices,
batch["mask_time_indices"],
contrastive_logits_temperature,
num_negatives,
)
diversity_loss = (num_codevectors - outputs.codevector_perplexity) / num_codevectors
loss = contrastive_loss + diversity_loss_weight * diversity_loss
# summarize metrics
metrics = {"loss": loss.mean(), "codevector_perplexity": outputs.codevector_perplexity}
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=f"Epoch ... (1/{num_epochs})", 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(vectorized_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 i, batch_idx in enumerate(tqdm(train_batch_idx, desc="Training...", position=1)):
samples = [vectorized_datasets["train"][int(idx)] for idx in batch_idx]
model_inputs = data_collator(samples)
model_inputs = shard(model_inputs.data)
# Model forward
state, train_metric, dropout_rngs, gumbel_rngs = p_train_step(
state, model_inputs, dropout_rngs, gumbel_rngs
)
train_metrics.append(train_metric)
train_time += time.time() - train_start
epochs.write(
f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {train_metric['loss'].mean()}, Learning Rate: {train_metric['learning_rate'].mean()})"
)
# ======================== Evaluating ==============================
num_eval_samples = len(vectorized_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 = [vectorized_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"Epoch... ({epoch + 1}/{num_epochs} | Loss: {eval_metrics['loss']}, Perplexity: {eval_metrics['codevector_perplexity']})"
)
# Save metrics
if has_tensorboard and jax.process_index() == 0:
cur_step = epoch * (len(vectorized_datasets["train"]) // 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, push_to_hub=training_args.push_to_hub)
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.
# Note that this mask is specifically adapted for FlaxBERT-like models.
# 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:] != ("LayerNorm", "scale")) for path in flat_params}
return traverse_util.unflatten_dict(flat_mask)
# create adam optimizer
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(
)
# Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
# information sent is the one passed as arguments along with your Python/PyTorch versions.
send_example_telemetry("run_t5_mlm",
model_args,
data_args,
framework="flax")
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=logging.INFO,
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,
use_auth_token=True if model_args.use_auth_token else None,
)
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,
use_auth_token=True if model_args.use_auth_token else None,
)
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,
use_auth_token=True if model_args.use_auth_token else None,
)
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,
use_auth_token=True if model_args.use_auth_token else None,
)
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,
use_auth_token=True if model_args.use_auth_token else None,
)
datasets["train"] = load_dataset(
extension,
data_files=data_files,
split=f"train[{data_args.validation_split_percentage}%:]",
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
# See more about loading any type of standard or custom dataset (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,
use_auth_token=True if model_args.use_auth_token else None,
)
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,
use_auth_token=True if model_args.use_auth_token else None,
)
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),
use_auth_token=True if model_args.use_auth_token else None,
)
elif model_args.model_name_or_path:
config = T5Config.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
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),
use_auth_token=True if model_args.use_auth_token else None,
)
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()
per_device_eval_batch_size = int(training_args.per_device_eval_batch_size)
eval_batch_size = per_device_eval_batch_size * jax.device_count()
num_train_steps = len(
tokenized_datasets["train"]) // train_batch_size * num_epochs
num_of_hosts = jax.process_count()
current_host_idx = jax.process_index()
# 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)
# find out all LayerNorm parameters
layer_norm_candidates = ["layernorm", "layer_norm", "ln"]
layer_norm_named_params = set([
layer[-2:] for layer_norm_name in layer_norm_candidates
for layer in flat_params.keys()
if layer_norm_name in "".join(layer).lower()
])
flat_mask = {
path: (path[-1] != "bias"
and path[-2:] not in layer_norm_named_params)
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"])
# Avoid using jax.numpy here in case of TPU training
train_samples_idx = np.random.permutation(np.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)
local_host_model_inputs = {
key: np.split(model_inputs.data[key], num_of_hosts,
axis=0)[current_host_idx]
for key, value in model_inputs.data.items()
}
# Model forward
model_inputs = shard(local_host_model_inputs)
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:"
f" {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"])
# Avoid using jax.numpy here in case of TPU training
eval_samples_idx = np.arange(num_eval_samples)
eval_batch_idx = generate_batch_splits(eval_samples_idx,
eval_batch_size,
drop_last=False)
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
metrics = pad_shard_unpad(p_eval_step, static_return=True)(
state.params,
model_inputs.data,
min_device_batch=per_device_eval_batch_size)
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"])
# Avoid using jax.numpy here in case of TPU training
eval_samples_idx = np.arange(num_eval_samples)
eval_batch_idx = generate_batch_splits(eval_samples_idx,
eval_batch_size,
drop_last=False)
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
metrics = pad_shard_unpad(p_eval_step, static_return=True)(
state.params,
model_inputs.data,
min_device_batch=per_device_eval_batch_size)
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 value(self, value: JTensor) -> JTensor:
return jnp.array(
optax.join_schedules([s.value for s in self._schedules],
self._boundaries)(value), jnp.float32)