def _exploit(
self,
trial_executor: "trial_runner.RayTrialExecutor",
trial: Trial,
trial_to_clone: Trial,
):
"""Transfers perturbed state from trial_to_clone -> trial.
If specified, also logs the updated hyperparam state.
"""
trial_state = self._trial_state[trial]
new_state = self._trial_state[trial_to_clone]
logger.info("[exploit] transferring weights from trial "
"{} (score {}) -> {} (score {})".format(
trial_to_clone, new_state.last_score, trial,
trial_state.last_score))
new_config = self._get_new_config(trial, trial_to_clone)
# Only log mutated hyperparameters and not entire config.
old_hparams = {
k: v
for k, v in trial_to_clone.config.items()
if k in self._hyperparam_mutations
}
new_hparams = {
k: v
for k, v in new_config.items() if k in self._hyperparam_mutations
}
logger.info("[explore] perturbed config from {} -> {}".format(
old_hparams, new_hparams))
if self._log_config:
self._log_config_on_step(trial_state, new_state, trial,
trial_to_clone, new_config)
new_tag = _make_experiment_tag(trial_state.orig_tag, new_config,
self._hyperparam_mutations)
if trial.status == Trial.PAUSED:
# If trial is paused we update it with a new checkpoint.
# When the trial is started again, the new checkpoint is used.
if not self._synch:
raise TuneError("Trials should be paused here only if in "
"synchronous mode. If you encounter this error"
" please raise an issue on Ray Github.")
else:
trial_executor.stop_trial(trial)
trial_executor.set_status(trial, Trial.PAUSED)
trial.set_experiment_tag(new_tag)
trial.set_config(new_config)
trial.on_checkpoint(new_state.last_checkpoint)
self._num_perturbations += 1
# Transfer over the last perturbation time as well
trial_state.last_perturbation_time = new_state.last_perturbation_time
trial_state.last_train_time = new_state.last_train_time
def on_trial_result(self, trial_runner: "trial_runner.TrialRunner",
trial: Trial, result: Dict) -> str:
if TRAINING_ITERATION not in result:
# No time reported
return TrialScheduler.CONTINUE
if not self._next_policy:
# No more changes in the config
return TrialScheduler.CONTINUE
step = result[TRAINING_ITERATION]
self._current_step = step
change_at, new_config = self._next_policy
if step < change_at:
# Don't change the policy just yet
return TrialScheduler.CONTINUE
logger.info("Population Based Training replay is now at step {}. "
"Configuration will be changed to {}.".format(
step, new_config))
checkpoint = trial_runner.trial_executor.save(trial,
Checkpoint.MEMORY,
result=result)
new_tag = make_experiment_tag(self.experiment_tag, new_config,
new_config)
trial_executor = trial_runner.trial_executor
reset_successful = trial_executor.reset_trial(trial, new_config,
new_tag)
if reset_successful:
trial_executor.restore(trial, checkpoint, block=True)
else:
trial_executor.stop_trial(trial,
destroy_pg_if_cannot_replace=False)
trial.set_experiment_tag(new_tag)
trial.set_config(new_config)
trial_executor.start_trial(trial, checkpoint, train=False)
self.current_config = new_config
self._num_perturbations += 1
self._next_policy = next(self._policy_iter, None)
return TrialScheduler.CONTINUE
def on_trial_result(
self, trial_runner: "trial_runner.TrialRunner", trial: Trial, result: Dict
) -> str:
if TRAINING_ITERATION not in result:
# No time reported
return TrialScheduler.CONTINUE
if not self._next_policy:
# No more changes in the config
return TrialScheduler.CONTINUE
step = result[TRAINING_ITERATION]
self._current_step = step
change_at, new_config = self._next_policy
if step < change_at:
# Don't change the policy just yet
return TrialScheduler.CONTINUE
logger.info(
"Population Based Training replay is now at step {}. "
"Configuration will be changed to {}.".format(step, new_config)
)
checkpoint = trial_runner.trial_executor.save(
trial, _TuneCheckpoint.MEMORY, result=result
)
new_tag = make_experiment_tag(self.experiment_tag, new_config, new_config)
trial_executor = trial_runner.trial_executor
trial_executor.stop_trial(trial)
trial_executor.set_status(trial, Trial.PAUSED)
trial.set_experiment_tag(new_tag)
trial.set_config(new_config)
trial.on_checkpoint(checkpoint)
self.current_config = new_config
self._num_perturbations += 1
self._next_policy = next(self._policy_iter, None)
return TrialScheduler.NOOP
def reset_trial(
self,
trial: Trial,
new_config: Dict,
new_experiment_tag: str,
logger_creator: Optional[Callable[[Dict], "ray.tune.Logger"]] = None,
) -> bool:
"""Tries to invoke `Trainable.reset()` to reset trial.
Args:
trial: Trial to be reset.
new_config: New configuration for Trial trainable.
new_experiment_tag: New experiment name for trial.
logger_creator: Function that instantiates a logger on the
actor process.
Returns:
True if `reset_config` is successful else False.
"""
trial.set_experiment_tag(new_experiment_tag)
trial.set_config(new_config)
trainable = trial.runner
# Pass magic variables
extra_config = copy.deepcopy(new_config)
extra_config[TRIAL_INFO] = TrialInfo(trial)
stdout_file, stderr_file = trial.log_to_file
extra_config[STDOUT_FILE] = stdout_file
extra_config[STDERR_FILE] = stderr_file
with self._change_working_directory(trial):
with warn_if_slow("reset"):
try:
reset_val = ray.get(
trainable.reset.remote(extra_config, logger_creator),
timeout=DEFAULT_GET_TIMEOUT,
)
except GetTimeoutError:
logger.exception("Trial %s: reset timed out.", trial)
return False
return reset_val
def _exploit(self, trial_executor: "trial_executor.TrialExecutor",
trial: Trial, trial_to_clone: Trial):
"""Transfers perturbed state from trial_to_clone -> trial.
If specified, also logs the updated hyperparam state.
"""
trial_state = self._trial_state[trial]
new_state = self._trial_state[trial_to_clone]
logger.info("[exploit] transferring weights from trial "
"{} (score {}) -> {} (score {})".format(
trial_to_clone, new_state.last_score, trial,
trial_state.last_score))
new_config = self._get_new_config(trial, trial_to_clone)
# Only log mutated hyperparameters and not entire config.
old_hparams = {
k: v
for k, v in trial_to_clone.config.items()
if k in self._hyperparam_mutations
}
new_hparams = {
k: v
for k, v in new_config.items() if k in self._hyperparam_mutations
}
logger.info("[explore] perturbed config from {} -> {}".format(
old_hparams, new_hparams))
if self._log_config:
self._log_config_on_step(trial_state, new_state, trial,
trial_to_clone, new_config)
new_tag = make_experiment_tag(trial_state.orig_tag, new_config,
self._hyperparam_mutations)
if trial.status == Trial.PAUSED:
# If trial is paused we update it with a new checkpoint.
# When the trial is started again, the new checkpoint is used.
if not self._synch:
raise TuneError("Trials should be paused here only if in "
"synchronous mode. If you encounter this error"
" please raise an issue on Ray Github.")
trial.set_experiment_tag(new_tag)
trial.set_config(new_config)
trial.on_checkpoint(new_state.last_checkpoint)
else:
# If trial is running, we first try to reset it.
# If that is unsuccessful, then we have to stop it and start it
# again with a new checkpoint.
reset_successful = trial_executor.reset_trial(
trial, new_config, new_tag)
# TODO(ujvl): Refactor Scheduler abstraction to abstract
# mechanism for trial restart away. We block on restore
# and suppress train on start as a stop-gap fix to
# https://github.com/ray-project/ray/issues/7258.
if reset_successful:
trial_executor.restore(trial,
new_state.last_checkpoint,
block=True)
else:
trial_executor.stop_trial(trial)
trial.set_experiment_tag(new_tag)
trial.set_config(new_config)
trial_executor.start_trial(trial,
new_state.last_checkpoint,
train=False)
self._num_perturbations += 1
# Transfer over the last perturbation time as well
trial_state.last_perturbation_time = new_state.last_perturbation_time
trial_state.last_train_time = new_state.last_train_time