def save(self, trial, storage=Checkpoint.PERSISTENT, result=None):
"""Saves the trial's state to a checkpoint asynchronously.
Args:
trial (Trial): The trial to be saved.
storage (str): Where to store the checkpoint. Defaults to
PERSISTENT.
result (dict): The state of this trial as a dictionary to be saved.
If result is None, the trial's last result will be used.
Returns:
Checkpoint object, or None if an Exception occurs.
"""
result = result or trial.last_result
with self._change_working_directory(trial):
if storage == Checkpoint.MEMORY:
value = trial.runner.save_to_object.remote()
checkpoint = Checkpoint(storage, value, result)
trial.on_checkpoint(checkpoint)
else:
value = trial.runner.save.remote()
checkpoint = Checkpoint(storage, value, result)
trial.saving_to = checkpoint
self._running[value] = trial
return checkpoint
def save(
self, trial, storage=Checkpoint.PERSISTENT, result: Optional[Dict] = None
) -> Checkpoint:
"""Saves the trial's state to a checkpoint asynchronously.
Args:
trial (Trial): The trial to be saved.
storage (str): Where to store the checkpoint. Defaults to
PERSISTENT.
result (dict): The state of this trial as a dictionary to be saved.
If result is None, the trial's last result will be used.
Returns:
Checkpoint object, or None if an Exception occurs.
"""
logger.info(f"saving trial {trial}")
result = result or trial.last_result
with self._change_working_directory(trial):
if storage == Checkpoint.MEMORY:
value = trial.runner.save_to_object.remote()
checkpoint = Checkpoint(storage, value, result)
trial.on_checkpoint(checkpoint)
else:
value = trial.runner.save.remote()
checkpoint = Checkpoint(storage, value, result)
trial.saving_to = checkpoint
self._futures[value] = (ExecutorEventType.SAVING_RESULT, trial)
return checkpoint
def save(self, trial, storage=Checkpoint.DISK, result=None):
"""Saves the trial's state to a checkpoint."""
result = result or trial.last_result
if storage == Checkpoint.MEMORY:
value = trial.runner.save_to_object.remote()
checkpoint = Checkpoint(storage, value, result)
else:
with warn_if_slow("save_checkpoint_to_disk"):
value = ray.get(trial.runner.save.remote())
checkpoint = Checkpoint(storage, value, result)
with warn_if_slow("on_checkpoint", DEFAULT_GET_TIMEOUT) as profile:
try:
trial.on_checkpoint(checkpoint)
except Exception:
logger.exception("Trial %s: Error handling checkpoint %s",
trial, checkpoint.value)
return None
if profile.too_slow and trial.sync_on_checkpoint:
logger.warning(
"Consider turning off forced head-worker trial checkpoint "
"syncs by setting sync_on_checkpoint=False. Note that this "
"might result in faulty trial restoration for some worker "
"failure modes.")
return checkpoint.value
def _exploit_trial(self, trial_executor: 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._trials_states_dict[trial]
new_state = self._trials_states_dict[trial_to_clone]
if not new_state.last_checkpoint:
logger.info(
"[pbt]: no checkpoint for trial. Skip exploit for Trial {}".
format(trial))
return
new_config = explore(trial_to_clone.config, self._hyperparam_mutations,
self._hyperparam_mutate_probability,
self._explore_func)
logger.info(
"[exploit] transferring weights from trial {} (score {}) -> {} (score {})"
.format(trial_to_clone, new_state.last_score, trial,
trial_state.last_score))
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)
reset_successful = trial_executor.reset_trial(trial, new_config,
new_tag)
if reset_successful:
trial_executor.restore(
trial, Checkpoint.from_object(new_state.last_checkpoint))
else:
trial_executor.stop_trial(trial, stop_logger=False)
trial.config = new_config
trial.experiment_tag = new_tag
trial_executor.start_trial(
trial, Checkpoint.from_object(new_state.last_checkpoint))
# TODO: move to Exploiter
new_state.num_steps = 0
trial_state.num_steps = 0
new_state.num_explorations = 0
trial_state.num_explorations += 1
self._num_explorations += 1
# Transfer over the last perturbation time as well
trial_state.last_perturbation_time = new_state.last_perturbation_time
def _exploit(self, trial_executor, trial, trial_to_clone):
"""Transfers perturbed state from trial_to_clone -> trial."""
trial_state = self._trial_state[trial]
new_state = self._trial_state[trial_to_clone]
if not new_state.last_checkpoint:
print("[pbt] warn: no checkpoint for trial, skip exploit", trial)
return
new_config = explore(trial_to_clone.config, self._hyperparam_mutations,
self._resample_probability,
self._custom_explore_fn)
print("[exploit] transferring weights from trial "
"{} (score {}) -> {} (score {})".format(trial_to_clone,
new_state.last_score,
trial,
trial_state.last_score))
# TODO(ekl) restarting the trial is expensive. We should implement a
# lighter way reset() method that can alter the trial config.
trial_executor.stop_trial(trial, stop_logger=False)
trial.config = new_config
trial.experiment_tag = make_experiment_tag(trial_state.orig_tag,
new_config,
self._hyperparam_mutations)
trial_executor.start_trial(
trial, Checkpoint.from_object(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
def _exploit(self, trial_executor, trial, trial_to_clone):
"""Transfers perturbed state from trial_to_clone -> trial."""
trial_state = self._trial_state[trial]
new_state = self._trial_state[trial_to_clone]
if not new_state.last_checkpoint:
logger.warning("[pbt]: no checkpoint for trial."
" Skip exploit for Trial {}".format(trial))
return
new_config = explore(trial_to_clone.config, self._hyperparam_mutations,
self._resample_probability,
self._custom_explore_fn)
logger.warning("[exploit] transferring weights from trial "
"{} (score {}) -> {} (score {})".format(
trial_to_clone, new_state.last_score, trial,
trial_state.last_score))
# TODO(ekl) restarting the trial is expensive. We should implement a
# lighter way reset() method that can alter the trial config.
new_tag = make_experiment_tag(trial_state.orig_tag, new_config,
self._hyperparam_mutations)
reset_successful = trial_executor.reset_trial(trial, new_config,
new_tag)
if not reset_successful:
trial_executor.stop_trial(trial, stop_logger=False)
trial.config = new_config
trial.experiment_tag = new_tag
trial_executor.start_trial(
trial, Checkpoint.from_object(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
def save(self, trial, storage=Checkpoint.PERSISTENT, result=None):
"""Saves the trial's state to a checkpoint.
Args:
trial (Trial): The state of this trial to be saved.
storage (str): Where to store the checkpoint. Defaults to
PERSISTENT.
result (dict): The state of this trial as a dictionary to be saved.
If result is None, the trial's last result will be used.
Returns:
Checkpoint future, or None if an Exception occurs.
"""
result = result or trial.last_result
with self._change_working_directory(trial):
if storage == Checkpoint.MEMORY:
value = trial.runner.save_to_object.remote()
checkpoint = Checkpoint(storage, value, result)
else:
with warn_if_slow("save_checkpoint_to_storage"):
# TODO(ujvl): Make this asynchronous.
value = ray.get(trial.runner.save.remote())
checkpoint = Checkpoint(storage, value, result)
with warn_if_slow("on_checkpoint", DEFAULT_GET_TIMEOUT) as profile:
try:
trial.on_checkpoint(checkpoint)
except Exception:
logger.exception("Trial %s: Error handling checkpoint %s",
trial, checkpoint.value)
return None
if profile.too_slow and trial.sync_on_checkpoint:
logger.warning(
"Consider turning off forced head-worker trial checkpoint "
"syncs by setting sync_on_checkpoint=False. Note that this "
"might result in faulty trial restoration for some worker "
"failure modes.")
return checkpoint.value
def checkpoint(self):
return Checkpoint(Checkpoint.MEMORY, "None", {})
def save(self, trial, type=Checkpoint.PERSISTENT, result=None):
return Checkpoint(Checkpoint.PERSISTENT, trial.trainable_name, result)
def _exploit(self, trial_executor, trial, trial_to_clone):
"""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]
if not new_state.last_checkpoint:
logger.info("[pbt]: no checkpoint for trial."
" Skip exploit for Trial {}".format(trial))
return
# if we are at a new timestep, we dont want to penalise for trials still going
if self.data['T'].max() > self.latest:
self.current = None
print("\n\n\n\n Copying: \n{} \n with:{} \n\n".format(
str(trial), str(trial_to_clone)))
new_config, lengthscale, mindist, meandist, data = explore(
self.data, self.bounds, self.current, trial_to_clone, trial,
trial_to_clone.config, self._hyperparam_mutations,
self._resample_probability)
# important to replace the old values, since we are copying across
self.data = data.copy()
# if the current guy youre selecting is at a point youve already done,
# then append the data to the "current" which is the points in the current batch
new = []
for key in self._hyperparam_mutations.keys():
new.append(new_config[key])
new = np.array(new)
new = new.reshape(1, new.size)
if self.data['T'].max() > self.latest:
self.latest = self.data['T'].max()
self.current = new.copy()
else:
self.current = np.concatenate((self.current, new), axis=0)
print("\n\n\n\n\n Currently Evaluating \n\n\n\n\n")
print(self.current)
print("\n\n\n\n\n")
# log the lengthscale
self.meta['timesteps'].append(self.data['T'].values[-1])
self.meta['lengthscales'].append(lengthscale)
self.meta['closest'].append(mindist)
self.meta['meandist'].append(meandist)
meta = pd.DataFrame({
'timesteps': self.meta['timesteps'],
'lengthscales': self.meta['lengthscales'],
'closest': self.meta['closest'],
'meandist': self.meta['meandist']
})
meta.to_csv('meta_data.csv')
logger.info("[exploit] transferring weights from trial "
"{} (score {}) -> {} (score {})".format(
trial_to_clone, new_state.last_score, trial,
trial_state.last_score))
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)
reset_successful = trial_executor.reset_trial(trial, new_config,
new_tag)
if reset_successful:
trial_executor.restore(
trial, Checkpoint.from_object(new_state.last_checkpoint))
else:
trial_executor.stop_trial(trial, stop_logger=False)
trial.config = new_config
trial.experiment_tag = new_tag
trial_executor.start_trial(
trial, Checkpoint.from_object(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