Unverified Commit a59bcefb authored by Sylvain Gugger's avatar Sylvain Gugger Committed by GitHub
Browse files

Split hp search methods (#6857)

* Split the run_hp_search by backend

* Unused import
parent 23f9611c
...@@ -3,7 +3,7 @@ import os ...@@ -3,7 +3,7 @@ import os
import numpy as np import numpy as np
from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, BestRun, HPSearchBackend from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, BestRun
from transformers.utils import logging from transformers.utils import logging
...@@ -83,7 +83,7 @@ def default_hp_search_backend(): ...@@ -83,7 +83,7 @@ def default_hp_search_backend():
return "ray" return "ray"
def run_hp_search(trainer, n_trials, direction, kwargs): def run_hp_search_optuna(trainer, n_trials: int, direction: str, **kwargs) -> BestRun:
def _objective(trial, checkpoint_dir=None): def _objective(trial, checkpoint_dir=None):
model_path = None model_path = None
if checkpoint_dir: if checkpoint_dir:
...@@ -96,80 +96,88 @@ def run_hp_search(trainer, n_trials, direction, kwargs): ...@@ -96,80 +96,88 @@ def run_hp_search(trainer, n_trials, direction, kwargs):
if getattr(trainer, "objective", None) is None: if getattr(trainer, "objective", None) is None:
metrics = trainer.evaluate() metrics = trainer.evaluate()
trainer.objective = trainer.compute_objective(metrics) trainer.objective = trainer.compute_objective(metrics)
if trainer.hp_search_backend == HPSearchBackend.RAY:
trainer._tune_save_checkpoint()
ray.tune.report(objective=trainer.objective)
return trainer.objective return trainer.objective
if trainer.hp_search_backend == HPSearchBackend.OPTUNA: timeout = kwargs.pop("timeout", None)
timeout = kwargs.pop("timeout", None) n_jobs = kwargs.pop("n_jobs", 1)
n_jobs = kwargs.pop("n_jobs", 1) study = optuna.create_study(direction=direction, **kwargs)
study = optuna.create_study(direction=direction, **kwargs) study.optimize(_objective, n_trials=n_trials, timeout=timeout, n_jobs=n_jobs)
study.optimize(_objective, n_trials=n_trials, timeout=timeout, n_jobs=n_jobs) best_trial = study.best_trial
best_trial = study.best_trial return BestRun(str(best_trial.number), best_trial.value, best_trial.params)
best_run = BestRun(str(best_trial.number), best_trial.value, best_trial.params)
elif trainer.hp_search_backend == HPSearchBackend.RAY:
# The model and TensorBoard writer do not pickle so we have to remove them (if they exists) def run_hp_search_ray(trainer, n_trials: int, direction: str, **kwargs) -> BestRun:
# while doing the ray hp search. def _objective(trial, checkpoint_dir=None):
_tb_writer = trainer.tb_writer model_path = None
trainer.tb_writer = None if checkpoint_dir:
trainer.model = None for subdir in os.listdir(checkpoint_dir):
# Setup default `resources_per_trial` and `reporter`. if subdir.startswith(PREFIX_CHECKPOINT_DIR):
if "resources_per_trial" not in kwargs and trainer.args.n_gpu > 0: model_path = os.path.join(checkpoint_dir, subdir)
# `args.n_gpu` is considered the total number of GPUs that will be split trainer.objective = None
# among the `n_jobs` trainer.train(model_path=model_path, trial=trial)
n_jobs = int(kwargs.pop("n_jobs", 1)) # If there hasn't been any evaluation during the training loop.
num_gpus_per_trial = trainer.args.n_gpu if getattr(trainer, "objective", None) is None:
if num_gpus_per_trial / n_jobs >= 1: metrics = trainer.evaluate()
num_gpus_per_trial = int(np.ceil(num_gpus_per_trial / n_jobs)) trainer.objective = trainer.compute_objective(metrics)
kwargs["resources_per_trial"] = {"gpu": num_gpus_per_trial} trainer._tune_save_checkpoint()
ray.tune.report(objective=trainer.objective)
if "reporter" not in kwargs: return trainer.objective
from ray.tune import CLIReporter
# The model and TensorBoard writer do not pickle so we have to remove them (if they exists)
kwargs["progress_reporter"] = CLIReporter(metric_columns=["objective"]) # while doing the ray hp search.
if "keep_checkpoints_num" in kwargs and kwargs["keep_checkpoints_num"] > 0: _tb_writer = trainer.tb_writer
# `keep_checkpoints_num=0` would disabled checkpointing trainer.tb_writer = None
trainer.use_tune_checkpoints = True trainer.model = None
if kwargs["keep_checkpoints_num"] > 1: # Setup default `resources_per_trial` and `reporter`.
logger.warning( if "resources_per_trial" not in kwargs and trainer.args.n_gpu > 0:
"Currently keeping {} checkpoints for each trial. Checkpoints are usually huge, " # `args.n_gpu` is considered the total number of GPUs that will be split
"consider setting `keep_checkpoints_num=1`." # among the `n_jobs`
) n_jobs = int(kwargs.pop("n_jobs", 1))
if "scheduler" in kwargs: num_gpus_per_trial = trainer.args.n_gpu
from ray.tune.schedulers import ( if num_gpus_per_trial / n_jobs >= 1:
ASHAScheduler, num_gpus_per_trial = int(np.ceil(num_gpus_per_trial / n_jobs))
HyperBandForBOHB, kwargs["resources_per_trial"] = {"gpu": num_gpus_per_trial}
MedianStoppingRule,
PopulationBasedTraining, if "reporter" not in kwargs:
from ray.tune import CLIReporter
kwargs["progress_reporter"] = CLIReporter(metric_columns=["objective"])
if "keep_checkpoints_num" in kwargs and kwargs["keep_checkpoints_num"] > 0:
# `keep_checkpoints_num=0` would disabled checkpointing
trainer.use_tune_checkpoints = True
if kwargs["keep_checkpoints_num"] > 1:
logger.warning(
"Currently keeping {} checkpoints for each trial. Checkpoints are usually huge, "
"consider setting `keep_checkpoints_num=1`."
) )
if "scheduler" in kwargs:
from ray.tune.schedulers import ASHAScheduler, HyperBandForBOHB, MedianStoppingRule, PopulationBasedTraining
# Check if checkpointing is enabled for PopulationBasedTraining # Check if checkpointing is enabled for PopulationBasedTraining
if isinstance(kwargs["scheduler"], PopulationBasedTraining): if isinstance(kwargs["scheduler"], PopulationBasedTraining):
if not trainer.use_tune_checkpoints: if not trainer.use_tune_checkpoints:
logger.warning( logger.warning(
"You are using PopulationBasedTraining but you haven't enabled checkpointing. " "You are using PopulationBasedTraining but you haven't enabled checkpointing. "
"This means your trials will train from scratch everytime they are exploiting " "This means your trials will train from scratch everytime they are exploiting "
"new configurations. Consider enabling checkpointing by passing " "new configurations. Consider enabling checkpointing by passing "
"`keep_checkpoints_num=1` as an additional argument to `Trainer.hyperparameter_search`." "`keep_checkpoints_num=1` as an additional argument to `Trainer.hyperparameter_search`."
)
# Check for `do_eval` and `eval_during_training` for schedulers that require intermediate reporting.
if isinstance(
kwargs["scheduler"], (ASHAScheduler, MedianStoppingRule, HyperBandForBOHB, PopulationBasedTraining)
) and (not trainer.args.do_eval or not trainer.args.evaluate_during_training):
raise RuntimeError(
"You are using {cls} as a scheduler but you haven't enabled evaluation during training. "
"This means your trials will not report intermediate results to Ray Tune, and "
"can thus not be stopped early or used to exploit other trials parameters. "
"If this is what you want, do not use {cls}. If you would like to use {cls}, "
"make sure you pass `do_eval=True` and `evaluate_during_training=True` in the "
"Trainer `args`.".format(cls=type(kwargs["scheduler"]).__name__)
) )
analysis = ray.tune.run(_objective, config=trainer.hp_space(None), num_samples=n_trials, **kwargs) # Check for `do_eval` and `eval_during_training` for schedulers that require intermediate reporting.
best_trial = analysis.get_best_trial(metric="objective", mode=direction[:3]) if isinstance(
best_run = BestRun(best_trial.trial_id, best_trial.last_result["objective"], best_trial.config) kwargs["scheduler"], (ASHAScheduler, MedianStoppingRule, HyperBandForBOHB, PopulationBasedTraining)
trainer.tb_writer = _tb_writer ) and (not trainer.args.do_eval or not trainer.args.evaluate_during_training):
raise RuntimeError(
"You are using {cls} as a scheduler but you haven't enabled evaluation during training. "
"This means your trials will not report intermediate results to Ray Tune, and "
"can thus not be stopped early or used to exploit other trials parameters. "
"If this is what you want, do not use {cls}. If you would like to use {cls}, "
"make sure you pass `do_eval=True` and `evaluate_during_training=True` in the "
"Trainer `args`.".format(cls=type(kwargs["scheduler"]).__name__)
)
analysis = ray.tune.run(_objective, config=trainer.hp_space(None), num_samples=n_trials, **kwargs)
best_trial = analysis.get_best_trial(metric="objective", mode=direction[:3])
best_run = BestRun(best_trial.trial_id, best_trial.last_result["objective"], best_trial.config)
trainer.tb_writer = _tb_writer
return best_run return best_run
...@@ -27,7 +27,8 @@ from .integrations import ( ...@@ -27,7 +27,8 @@ from .integrations import (
is_ray_available, is_ray_available,
is_tensorboard_available, is_tensorboard_available,
is_wandb_available, is_wandb_available,
run_hp_search, run_hp_search_optuna,
run_hp_search_ray,
) )
from .modeling_utils import PreTrainedModel from .modeling_utils import PreTrainedModel
from .optimization import AdamW, get_linear_schedule_with_warmup from .optimization import AdamW, get_linear_schedule_with_warmup
...@@ -884,7 +885,8 @@ class Trainer: ...@@ -884,7 +885,8 @@ class Trainer:
self.hp_space = default_hp_space[backend] if hp_space is None else hp_space self.hp_space = default_hp_space[backend] if hp_space is None else hp_space
self.compute_objective = default_compute_objective if compute_objective is None else compute_objective self.compute_objective = default_compute_objective if compute_objective is None else compute_objective
best_run = run_hp_search(self, n_trials, direction, kwargs) run_hp_search = run_hp_search_optuna if backend == HPSearchBackend.OPTUNA else run_hp_search_ray
best_run = run_hp_search(self, n_trials, direction, **kwargs)
self.hp_search_backend = None self.hp_search_backend = None
return best_run return best_run
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment