def hp_search(self):
if not self.remote:
if self.opt_model.max_instances_at_once > torch.cuda.device_count():
raise Exception(''' 'max_instances_at_once' must be smaller or equal to the number of available gpus''')
if not hasattr(self.opt_model, 'name'):
logger.info("no 'update_optimal_model' method, checking for model.txt file . . . ")
self.update_optimal_model()
# initialize tuner and gun i.e.
ongoing_trials = OngoingTrials()
tuner = Tuner(self.opt_model, ongoing_trials)
gun = Launcher(self.opt_model, ongoing_trials, remote=self.remote)
logger.info('commencing hyper-parameter search . . . ')
tuner.search_hp()
gun.launch_trials()
tuner.end_trial()
# starting second set of trials
tuner.search_hp()
while ongoing_trials.status is not 'STOPPED':
gun.launch_trials()
tuner.end_trial()
# starting next set of trials
tuner.search_hp()
best_trial = tuner.get_best_trial()
logger.info('best trial: ', json.dumps(best_trial))
if os.path.exists(self.path_to_best_trial):
logger.info('overwriting best_trial.json . . .')
os.remove(self.path_to_best_trial)
with open(self.path_to_best_trial, 'w') as fp:
json.dump(best_trial, fp)
logger.info('results saved to best_trial.json')
def hp_search(self):
if not self.remote:
if self.opt_model.max_instances_at_once > torch.cuda.device_count(
):
raise Exception(
''' 'max_instances_at_once' must be smaller or equal to the number of available gpus'''
)
if not hasattr(self.opt_model, 'name'):
logger.info(
"no 'update_optimal_model' method, checking for model.txt file . . . "
)
self.update_optimal_model()
# initialize tuner and gun i.e.
ongoing_trials = OngoingTrials()
tuner = Tuner(self.opt_model, ongoing_trials)
gun = Launcher(self.opt_model, ongoing_trials, remote=self.remote)
logger.info('commencing hyper-parameter search . . . ')
tuner.search_hp()
gun.launch_trials()
tuner.end_trial()
# starting second set of trials
tuner.search_hp()
while ongoing_trials.status is not 'STOPPED':
gun.launch_trials()
tuner.end_trial()
# starting next set of trials
tuner.search_hp()
trials = tuner.get_trials()
sorted_trial_ids = tuner.get_sorted_trial_ids()
string1 = self.path_to_best_checkpoint.split('.')[0]
for i in range(len(sorted_trial_ids)):
save_checkpoint_location = string1 + str(i) + '.pt'
if os.path.exists(save_checkpoint_location):
logger.info('overwriting checkpoint . . .')
os.remove(save_checkpoint_location)
logger.info('trial ' + sorted_trial_ids[i] + '\tval: ' +
str(trials[sorted_trial_ids[i]]['metrics']))
torch.save(trials[sorted_trial_ids[i]]['checkpoint'],
save_checkpoint_location)
logger.info('best trial: ' +
str(trials[sorted_trial_ids[0]]['hp_values']) +
'\nbest value: ' +
str(trials[sorted_trial_ids[0]]['metrics']))
best_trial = trials[sorted_trial_ids[0]]['hp_values']
if os.path.exists(self.path_to_best_trial):
logger.info('overwriting best_trial.json . . .')
os.remove(self.path_to_best_trial)
with open(self.path_to_best_trial, 'w') as fp:
json.dump(best_trial, fp)
logger.info('results saved to best_trial.json')