def get_reduced_configs(self):
"""Reduce the experiments to restart."""
iteration_config = self.experiment_group.iteration_config
if iteration_config is None:
logger.error(
'Experiment group `%s` attempt to update iteration, but has no iteration',
self.experiment_group.id,
extra={'stack': True})
return
search_manager = self.experiment_group.search_manager
# Get the number of experiments to keep
n_configs_to_keep = search_manager.get_n_config_to_keep_for_iteration(
iteration=iteration_config.iteration,
bracket_iteration=iteration_config.bracket_iteration)
# Get the last group's experiments metrics
experiments_metrics = self.experiment_group.iteration_config.experiments_metrics
# Order the experiments
reverse = Optimization.maximize(self.experiment_group.hptuning_config.
hyperband.metric.optimization)
experiments_metrics = sorted(experiments_metrics,
key=lambda x: x[1],
reverse=reverse)
# Keep n experiments
return [xp[0] for xp in experiments_metrics[:n_configs_to_keep]]
def get_ordered_experiments_by_metric(self, experiment_ids: List[int],
metric: str, optimization: str):
query = self.get_annotated_experiments_with_metric(
metric=metric, experiment_ids=experiment_ids)
metric_order_by = '{}{}'.format(
'-' if Optimization.maximize(optimization) else '', metric)
return query.order_by(metric_order_by)
def should_stop_early(self) -> bool:
filters = []
for early_stopping_metric in self.early_stopping:
comparison = (
'gte' if Optimization.maximize(early_stopping_metric.optimization) else 'lte')
metric_filter = 'last_metric__{}__{}'.format(
early_stopping_metric.metric, comparison)
filters.append({metric_filter: early_stopping_metric.value})
if filters:
return self.experiments.filter(functools.reduce(OR, [Q(**f) for f in filters])).exists()
return False
def parse_y(self, metrics):
if not metrics:
return metrics
y_values = []
for value in metrics:
if Optimization.maximize(self.hptuning_config.bo.metric.optimization):
y_values.append(float(value))
else:
y_values.append(-float(value))
return np.array(y_values)
