def main(dataset_info, mode, seed, params, mode_args=None):
"""This command line interface has three different operation modes:
* CV: useful for the Tweakathon
* 1/3 test split: useful to evaluate a configuration
* cv on 2/3 train split: useful to optimize hyperparameters in a training
mode before testing a configuration on the 1/3 test split.
It must by no means be used for the Auto part of the competition!
"""
if mode != "test":
num_run = get_new_run_num()
for key in params:
try:
params[key] = int(params[key])
except:
try:
params[key] = float(params[key])
except:
pass
if seed is not None:
seed = int(float(seed))
else:
seed = 1
output_dir = os.getcwd()
D = store_and_or_load_data(dataset_info=dataset_info, outputdir=output_dir)
cs = get_configuration_space(D.info)
configuration = configuration_space.Configuration(cs, params)
metric = D.info['metric']
global evaluator
# Train/test split
if mode == 'holdout':
evaluator = HoldoutEvaluator(D,
configuration,
with_predictions=True,
all_scoring_functions=True,
output_y_test=True,
seed=seed,
num_run=num_run)
evaluator.fit()
signal.signal(15, empty_signal_handler)
evaluator.finish_up()
model_directory = os.path.join(os.getcwd(), "models_%d" % seed)
if os.path.exists(model_directory):
model_filename = os.path.join(model_directory,
"%s.model" % num_run)
with open(model_filename, "w") as fh:
pickle.dump(evaluator.model, fh, -1)
elif mode == 'test':
evaluator = TestEvaluator(D,
configuration,
all_scoring_functions=True,
seed=seed)
evaluator.fit()
scores = evaluator.predict()
duration = time.time() - evaluator.starttime
score = scores[metric]
additional_run_info = ";".join(
["%s: %s" % (m_, value) for m_, value in scores.items()])
additional_run_info += ";" + "duration: " + str(duration)
print "Result for ParamILS: %s, %f, 1, %f, %d, %s" % (
"SAT", abs(duration), score, evaluator.seed, additional_run_info)
# CV on the whole training set
elif mode == 'cv':
evaluator = CVEvaluator(D,
configuration,
with_predictions=True,
all_scoring_functions=True,
output_y_test=True,
cv_folds=mode_args['folds'],
seed=seed,
num_run=num_run)
evaluator.fit()
signal.signal(15, empty_signal_handler)
evaluator.finish_up()
elif mode == 'partial_cv':
evaluator = CVEvaluator(D,
configuration,
all_scoring_functions=True,
cv_folds=mode_args['folds'],
seed=seed,
num_run=num_run)
evaluator.partial_fit(mode_args['fold'])
scores = evaluator.predict()
duration = time.time() - evaluator.starttime
score = scores[metric]
additional_run_info = ";".join(
["%s: %s" % (m_, value) for m_, value in scores.items()])
additional_run_info += ";" + "duration: " + str(duration)
print "Result for ParamILS: %s, %f, 1, %f, %d, %s" % (
"SAT", abs(duration), score, evaluator.seed, additional_run_info)
elif mode == 'nested-cv':
evaluator = NestedCVEvaluator(D,
configuration,
with_predictions=True,
inner_cv_folds=mode_args['inner_folds'],
outer_cv_folds=mode_args['outer_folds'],
all_scoring_functions=True,
output_y_test=True,
seed=seed,
num_run=num_run)
evaluator.fit()
signal.signal(15, empty_signal_handler)
evaluator.finish_up()
else:
raise ValueError("Must choose a legal mode.")
