def load_pet_configs(
args) -> Tuple[WrapperConfig, pet.TrainConfig, pet.EvalConfig]:
"""
Load the model, training and evaluation configs for PET from the given command line arguments.
"""
model_cfg = WrapperConfig(
model_type=args.model_type,
model_name_or_path=args.model_name_or_path,
wrapper_type=args.wrapper_type,
task_name=args.task_name,
label_list=args.label_list,
max_seq_length=args.pet_max_seq_length,
verbalizer_file=args.verbalizer_file,
cache_dir=args.cache_dir,
)
train_cfg = pet.TrainConfig(
device=args.device,
per_gpu_train_batch_size=args.pet_per_gpu_train_batch_size,
per_gpu_unlabeled_batch_size=args.pet_per_gpu_unlabeled_batch_size,
n_gpu=args.n_gpu,
num_train_epochs=args.pet_num_train_epochs,
max_steps=args.pet_max_steps,
min_steps=args.pet_min_steps,
gradient_accumulation_steps=args.pet_gradient_accumulation_steps,
weight_decay=args.weight_decay,
learning_rate=args.learning_rate,
adam_epsilon=args.adam_epsilon,
warmup_steps=args.warmup_steps,
max_grad_norm=args.max_grad_norm,
lm_training=args.lm_training,
logging_steps=args.logging_steps,
logging_number=args.logging_number,
alpha=args.alpha,
local_rank=args.local_rank,
)
eval_cfg = pet.EvalConfig(
device=args.device,
n_gpu=args.n_gpu,
metrics=args.metrics,
per_gpu_eval_batch_size=args.pet_per_gpu_eval_batch_size,
decoding_strategy=args.decoding_strategy,
priming=args.priming,
local_rank=args.local_rank,
)
return model_cfg, train_cfg, eval_cfg
def load_sequence_classifier_configs(
args) -> Tuple[WrapperConfig, pet.TrainConfig, pet.EvalConfig]:
"""
Load the model, training and evaluation configs for a regular sequence classifier from the given command line
arguments. This classifier can either be used as a standalone model or as the final classifier for PET/iPET.
"""
model_cfg = WrapperConfig(
model_type=args.model_type,
model_name_or_path=args.model_name_or_path,
wrapper_type=SEQUENCE_CLASSIFIER_WRAPPER,
task_name=args.task_name,
label_list=args.label_list,
max_seq_length=args.sc_max_seq_length,
verbalizer_file=args.verbalizer_file,
cache_dir=args.cache_dir,
)
train_cfg = pet.TrainConfig(
device=args.device,
per_gpu_train_batch_size=args.sc_per_gpu_train_batch_size,
per_gpu_unlabeled_batch_size=args.sc_per_gpu_unlabeled_batch_size,
n_gpu=args.n_gpu,
num_train_epochs=args.sc_num_train_epochs,
max_steps=args.sc_max_steps,
min_steps=args.sc_min_steps,
temperature=args.temperature,
gradient_accumulation_steps=args.sc_gradient_accumulation_steps,
weight_decay=args.weight_decay,
learning_rate=args.learning_rate,
adam_epsilon=args.adam_epsilon,
warmup_steps=args.warmup_steps,
logging_steps=args.logging_steps,
logging_number=args.logging_number,
max_grad_norm=args.max_grad_norm,
use_logits=args.method != "sequence_classifier",
local_rank=args.local_rank,
)
eval_cfg = pet.EvalConfig(
device=args.device,
n_gpu=args.n_gpu,
metrics=args.metrics,
per_gpu_eval_batch_size=args.sc_per_gpu_eval_batch_size,
local_rank=args.local_rank,
)
return model_cfg, train_cfg, eval_cfg