regimes = args.regimes.split(',')
pupil_names = get_pupil_names_from_eval_dir(eval_dir)
if args.pupil_names is not None:
pupil_names = args.pupil_names.split(',')
for_plotting = get_optimizer_evaluation_results(eval_dir, hp_names,
AVERAGING_NUMBER)
if args.target_metrics is not None:
metrics = args.target_metrics.split(',')
best = get_best(for_plotting, model, value_filter=args.value_filter)
indents = [4, 8, 12]
print(best)
if model == 'pupil':
for dataset_name in dataset_names:
print('dataset:', dataset_name)
for metric in metrics:
b = best[dataset_name][metric]
print(' ' * indents[0] + metric + ':', b[1])
print_hps(hp_names, b[0], indents[1])
else:
for pupil_name in pupil_names:
print('pupil name:', pupil_name)
for metric in metrics:
print(' ' * indents[0] + metric + ':')
for regime in regimes:
b = best[pupil_name][metric][regime]
print(' ' * indents[1] + regime + ' result:', regime)
print(' ' * indents[2] + 'result:', b[1])
print_hps(hp_names, b[0], indents[2])
other_hyperparameters=other_hyperparameters,
initial_experiment_counter_value=initial_experiment_counter_value,
**launch_kwargs
)
hp_names = get_hp_names_from_conf_file(parameter_set_file_name)
for_plotting = get_optimizer_evaluation_results(save_path, hp_names, AVERAGING_NUMBER)
best = get_best(for_plotting, 'optimizer')
metric_res = best['adam_prep']['loss']
best_on_valid = metric_res['validation']
print(' ' * 2 + 'loss' + ':', best_on_valid[1])
print_hps(hp_names, best_on_valid[0], 4)
best_conf = dict(list(zip(hp_names, best_on_valid[0])))
env.build_pupil(
batch_size=BATCH_SIZE,
**LSTM_SIZE,
regime='training_with_meta_optimizer',
additional_metrics=add_metrics,
going_to_limit_memory=True,
)
env.build_optimizer(
**OPTIMIZER_PARAMETERS,
optimizer_init_parameter=best_conf['optimizer_init_parameter'],
)
hp_names = get_hp_names_from_conf_file(parameter_set_file_name)
for_plotting = get_pupil_evaluation_results(save_path, hp_names)
best = get_best(for_plotting, 'pupil')
env.build_pupil(batch_size=BATCH_SIZE,
num_layers=2,
num_hidden_nodes=[1000],
input_shape=[3072],
num_classes=10,
additional_metrics=add_metrics,
optimizer=opt)
for dataset_name, dataset_res in best.items():
print('dataset:', dataset_name)
for metric, b in dataset_res.items():
print(' ' * 2 + metric + ':', b[1])
print_hps(hp_names, b[0], 4)
best_conf = dict(list(zip(hp_names, b[0])))
training_path = os.path.join(base, metric + '_best', 'test',
'training')
env.train(
allow_growth=True,
# save_path='debug_grid_search',
result_types=['loss', 'bpc', 'perplexity', 'accuracy'],
additions_to_feed_dict=train_add_feed,
# pupil_restore_paths=['debug_empty_meta_optimizer/not_learning_issue_es20_nn20/checkpoints/0'],
# stop=stop_specs,
save_path=training_path,
restore_path=RESTORE_PATH,
stop=1000,
results_collect_interval=1000,
