Code in benchmarking/examples/fine_tuning_transformer_glue
Selecting pre-trained transformer model from Hugging Face zoo and fine-tuning it to a GLUE task. This is in fact a whole family of benchmarks:
f"finetune_transformer_glue_{dataset}"
: Tune number of hyperparameters for fixed pre-trained model, selected by--model_type
f"finetune_transformer_glue_modsel_{dataset}"
: Tune the same hyperparameters and select the best pre-trained model from a list of 9 choices
Here, dataset
selects the GLUE document classification task (values are
“cola”, “mnli”, “mrpc”, “qnli”, “qqp”, “rte”, “sst2”, “stsb”, “wnli”).
from syne_tune.experiments.default_baselines import (
BayesianOptimization,
MOBSTER,
)
class Methods:
BO = "BO"
MOBSTER = "MOBSTER"
methods = {
Methods.BO: lambda method_arguments: BayesianOptimization(method_arguments),
Methods.MOBSTER: lambda method_arguments: MOBSTER(method_arguments),
}
from typing import Dict, Any
from benchmarking.examples.fine_tuning_transformer_glue.baselines import methods
from benchmarking.benchmark_definitions import (
real_benchmark_definitions as benchmark_definitions,
)
from benchmarking.benchmark_definitions.finetune_transformer_glue import (
PRETRAINED_MODELS,
MAX_RESOURCE_ATTR,
MODEL_TYPE_ATTR,
)
from syne_tune.config_space import Domain
from syne_tune.experiments.launchers.hpo_main_local import main
extra_args = [
dict(
name="num_train_epochs",
type=int,
default=3,
help="Maximum number of training epochs",
),
dict(
name="model_type",
type=str,
default="bert-base-cased",
choices=tuple(PRETRAINED_MODELS),
help="Pre-trained model to start fine-tuning from",
),
]
def map_method_args(args, method: str, method_kwargs: Dict[str, Any]) -> Dict[str, Any]:
# We need to change ``method_kwargs.config_space``, based on ``extra_args``
new_method_kwargs = method_kwargs.copy()
new_config_space = new_method_kwargs["config_space"].copy()
choose_model = isinstance(["model_name_or_path"], Domain)
new_config_space[MAX_RESOURCE_ATTR] = args.num_train_epochs
if not choose_model:
new_config_space[MODEL_TYPE_ATTR] = args.model_type
else:
# Need to change ``points_to_evaluate``
default_configuration = new_method_kwargs["scheduler_kwargs"][
"points_to_evaluate"
][0]
new_default_configuration = {
**default_configuration,
MODEL_TYPE_ATTR: args.model_type,
}
new_method_kwargs["scheduler_kwargs"]["points_to_evaluate"] = [
new_default_configuration
]
new_method_kwargs["config_space"] = new_config_space
return new_method_kwargs
if __name__ == "__main__":
main(methods, benchmark_definitions, extra_args, map_method_args)
from pathlib import Path
import benchmarking
from benchmarking.benchmark_definitions import (
real_benchmark_definitions as benchmark_definitions,
)
from benchmarking.examples.fine_tuning_transformer_glue.baselines import methods
from benchmarking.examples.fine_tuning_transformer_glue.hpo_main import extra_args
from syne_tune.experiments.launchers.launch_remote_local import launch_remote
if __name__ == "__main__":
entry_point = Path(__file__).parent / "hpo_main.py"
launch_remote(
entry_point=entry_point,
methods=methods,
benchmark_definitions=benchmark_definitions,
source_dependencies=benchmarking.__path__,
extra_args=extra_args,
)
syne-tune[gpsearchers,aws]
tqdm