def write_training_runs(args):
filer = ExperimentFiler(args.experiment_dir)
csv_file = filer.experiment_file("training_runs.csv")
with open(csv_file, 'wb') as my_file:
my_file.write('Generation, %s, Fitness\n' % resolve_alt_objective(args))
results_files = get_results_files(args)
generation_filer = GenerationFiler(args.experiment_dir)
for results_file in results_files:
generation = generation_filer.get_generation_from_path(results_file)
persistence = ResultsDictPersistence(args.experiment_dir, generation,
logger=None)
results_dict = persistence.restore()
if len(results_dict) == 0:
# File not found
continue
for key in results_dict.keys():
result = results_dict[key]
try:
# XXX Use FitnessObjectives prepared from config?
alt_objective = result['metrics'][args.alt_objective]
fitness = result['metrics']['fitness'] # XXX not kosher
except Exception:
try:
alt_objective = result['metrics'][args.alt_objective]
fitness = result['fitness'] # XXX not kosher
except Exception as exception:
if args.alt_objective == "num_params":
fitness = result['fitness'] # XXX not kosher
# XXX What generates this params file?
cache_file = generation_filer.get_generation_file(
"candidate_{0}.params".format(key))
if os.path.exists(cache_file):
with open(cache_file, 'rb') as my_file:
alt_objective = my_file.read()
else:
undefined = 0
k = undefined # XXX
print("Extracting num params from network {}".format(k))
model = get_model(args.experiment_dir, key, generation)
alt_objective = str(model.count_params())
with open(cache_file, 'wb') as my_file:
my_file.write(alt_objective)
else:
raise exception
if args.alt_objective == 'training_time':
alt_objective = str(float(alt_objective) / 3600.0)
with open(csv_file, 'ab') as my_file:
line = '%s %s %s\n' % (generation, alt_objective, fitness)
my_file.write(line)
return csv_file
def draw_best_candidate_results(self,
best_candidate,
generation=None,
suffix=''):
"""
:param best_candidate: A candidate object comprising the best of a
generation.
:param generation: Default value is None
:param suffix: Default value is an empty string
"""
experiment_config = self.master_config.get('experiment_config')
if not experiment_config.get('visualize'):
return
best_id = self.candidate_util.get_candidate_id(best_candidate)
best_fitness = self.candidate_util.get_candidate_fitness(
best_candidate)
fitness = best_fitness if best_fitness is None else \
round(best_fitness, 4)
# Determine the output file name basis
# XXX Use fitness for now.
# Later on can address multi-objective goals.
metric_name = "fitness"
if generation is not None:
# Put the file in the gen_NN directory.
# Call it best_candidate to match the best_candidate.json
# that gets put there
base_name = "best_{0}_candidate".format(metric_name)
filer = GenerationFiler(self.experiment_dir, generation)
base_path = filer.get_generation_file(base_name)
else:
# We do not have a generation that we know about so write out
# the old-school file name.
# XXX Not entirely sure when this path would be taken
base_name = "F{0}_ID-{1}_{2}best_{3}".format(
fitness, best_id, suffix, metric_name)
filer = ExperimentFiler(self.experiment_dir)
base_path = filer.experiment_file(base_name)
# NetworkVisualizers use the build_training_model() which requires
# a data_dict of file keys -> file paths to exist. Domains that
# wish to visualize their networks that use the data_dict will
# need to deal with a None value for data dict in the visualization
# case.
data_dict = None
visualizer = NetworkMultiVisualizer(self.master_config,
data_dict,
base_path,
logger=self.logger)
visualizer.visualize(best_candidate)