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()
training_args.do_train = True
# training_args.do_eval = True
data_args.mlm = True
if data_args.eval_data_file is None and training_args.do_eval:
raise ValueError(
"Cannot do evaluation without an evaluation data file. Either supply a file to --eval_data_file "
"or remove the --do_eval argument.")
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir) and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
training_args.output_dir = os.environ["SM_OUTPUT_DATA_DIR"]
input_path = os.environ["SM_CHANNEL_TRAINING"]
folder = "wikitext-2-raw-v1"
cmd = "tar xzf %s -C %s" % (input_path + "/" + folder + ".tar.gz",
input_path)
assert os.system(cmd) == 0
data_folder = f"{input_path}/{folder}"
data_args.train_data_file = f"{data_folder}/wikitext-2-raw/{data_args.train_data_file}"
data_args.eval_data_file = f"{data_folder}/wikitext-2-raw/{data_args.eval_data_file}"
checkpoint_path = training_args.output_dir + "/checkpoints"
project_name = "language-modeling"
os.environ["WANDB_PROJECT"] = project_name
wandb.init(
project=project_name
) # TODO(tilo): is this really necessary? should be done by ML-library ( here transformers)
assert wandb.api.api_key is not None
assert is_wandb_available(
) # TODO(tilo): somehow I had issues, which seem to have resolved themselves
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO
if training_args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.local_rank != -1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Set seed
set_seed(training_args.seed)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name,
cache_dir=model_args.cache_dir)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path,
cache_dir=model_args.cache_dir)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name, cache_dir=model_args.cache_dir)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported, but you can do it from another script, save it,"
"and load it from here, using --tokenizer_name")
if model_args.model_name_or_path:
model = AutoModelWithLMHead.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
else:
logger.info("Training new model from scratch")
model = AutoModelWithLMHead.from_config(config)
model.resize_token_embeddings(len(tokenizer))
if config.model_type in ["bert", "roberta", "distilbert", "camembert"
] and not data_args.mlm:
raise ValueError(
"BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the"
"--mlm flag (masked language modeling).")
if data_args.block_size <= 0:
data_args.block_size = tokenizer.max_len
# Our input block size will be the max possible for the model
else:
data_args.block_size = min(data_args.block_size, tokenizer.max_len)
# Get datasets
train_dataset = (get_dataset(
data_args, tokenizer=tokenizer, cache_dir=model_args.cache_dir)
if training_args.do_train else None)
eval_dataset = (get_dataset(data_args,
tokenizer=tokenizer,
evaluate=True,
cache_dir=model_args.cache_dir)
if training_args.do_eval else None)
if config.model_type == "xlnet":
data_collator = DataCollatorForPermutationLanguageModeling(
tokenizer=tokenizer,
plm_probability=data_args.plm_probability,
max_span_length=data_args.max_span_length,
)
else:
data_collator = DataCollatorForLanguageModeling(
tokenizer=tokenizer,
mlm=data_args.mlm,
mlm_probability=data_args.mlm_probability)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
data_collator=data_collator,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
prediction_loss_only=True,
)
# Training
if training_args.do_train:
model_path = (model_args.model_name_or_path
if model_args.model_name_or_path is not None
and os.path.isdir(model_args.model_name_or_path) else
None)
trainer.train(model_path=model_path)
trainer.save_model()
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
if trainer.is_world_master():
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["eval_loss"])
result = {"perplexity": perplexity}
output_eval_file = os.path.join(training_args.output_dir,
"eval_results_lm.txt")
if trainer.is_world_master():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
results.update(result)
return results
def main():
# See all possible arguments in src/transformers/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, main_process_only=False)
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. Now we preprocess the datasets including loading the audio, resampling and normalization
# Thankfully, `datasets` takes care of automatically loading and resampling the audio,
# so that we just need to set the correct target sampling rate and normalize the input
# via the `feature_extractor`
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(args.model_name_or_path)
# make sure that dataset decodes audio with correct sampling rate
raw_datasets = raw_datasets.cast_column(
args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
)
# 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)
def prepare_dataset(batch):
sample = batch[args.audio_column_name]
inputs = feature_extractor(
sample["array"], sampling_rate=sample["sampling_rate"], max_length=max_length, truncation=True
)
batch["input_values"] = inputs.input_values[0]
batch["input_length"] = len(inputs.input_values[0])
return batch
# load via mapped files via path
cache_file_names = None
if args.train_cache_file_name is not None:
cache_file_names = {"train": args.train_cache_file_name, "validation": args.validation_cache_file_name}
# 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,
cache_file_names=cache_file_names,
)
if min_length > 0.0:
vectorized_datasets = vectorized_datasets.filter(
lambda x: x > min_length,
num_proc=args.preprocessing_num_workers,
input_columns=["input_length"],
)
vectorized_datasets = vectorized_datasets.remove_columns("input_length")
# for large datasets it is advised to run the preprocessing on a
# single machine first with ``args.preprocessing_only`` since there will mostly likely
# be a timeout when running the script in distributed mode.
# In a second step ``args.preprocessing_only`` can then be set to `False` to load the
# cached dataset
if 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
starting_epoch = 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
starting_epoch = 0
for epoch in range(starting_epoch, 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(
f"Gradients have overflown - skipping update step... 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, is_main_process=accelerator.is_main_process, 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,
auto_lfs_prune=True,
)
# 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, is_main_process=accelerator.is_main_process, save_function=accelerator.save
)
if accelerator.is_main_process:
if args.push_to_hub:
repo.push_to_hub(commit_message="End of training", auto_lfs_prune=True)
def main():
if is_wandb_available():
import wandb
parser = HfArgumentParser(
(ModelArguments, DatasetArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir) and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty."
"Use --overwrite_output_dir to overcome.")
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO
if is_main_process(training_args.local_rank) else logging.WARN,
)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name)
else:
datasets = load_from_disk(data_args.dataset_path)
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if model_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name,
cache_dir=model_args.cache_dir)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path,
cache_dir=model_args.cache_dir)
else:
config = CONFIG_MAPPING[model_args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported by this script."
"You can do it from another script, save it, and load it from here, using --tokenizer_name."
)
if model_args.model_name_or_path:
model = AutoModelForCausalLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForCausalLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
def tokenize_function(examples):
return tokenizer(examples[text_column_name])
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
num_proc=None,
remove_columns=column_names,
load_from_cache_file=not False,
)
block_size = tokenizer.model_max_length
if block_size > 1024:
logger.warn(
f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
"Picking 1024 instead. You can change that default value by passing --block_size xxx."
)
block_size = 1024
# Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
def group_texts(examples):
# Concatenate all texts.
concatenated_examples = {
k: sum(examples[k], [])
for k in examples.keys()
}
total_length = len(concatenated_examples[list(examples.keys())[0]])
# We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
# customize this part to your needs.
total_length = (total_length // block_size) * block_size
# Split by chunks of max_len.
result = {
k:
[t[i:i + block_size] for i in range(0, total_length, block_size)]
for k, t in concatenated_examples.items()
}
result["labels"] = result["input_ids"].copy()
return result
# Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder
# for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value might be slower
# to preprocess.
#
# To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
lm_datasets = tokenized_datasets.map(
group_texts,
batched=True,
num_proc=None,
load_from_cache_file=not False,
)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=lm_datasets["train"] if training_args.do_train else None,
eval_dataset=lm_datasets["validation"]
if training_args.do_eval else None,
tokenizer=tokenizer,
# Data collator will default to DataCollatorWithPadding, so we change it.
data_collator=default_data_collator,
)
# Add Callbacks
savvi_callback = SavviCallback()
trainer.add_callback(savvi_callback)
# Training
if training_args.do_train:
model_path = (model_args.model_name_or_path if
(model_args.model_name_or_path is not None
and os.path.isdir(model_args.model_name_or_path)) else
None)
trainer.train(model_path=model_path)
trainer.save_model() # Saves the tokenizer too for easy upload
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
perplexity = math.exp(eval_output["eval_loss"])
results["perplexity"] = perplexity
output_eval_file = os.path.join(training_args.output_dir,
"eval_results_clm.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in results.items():
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
savvihub.log(step=0, row={'perplexity': results["perplexity"]})
return results