import sys
sys.path.insert(0, '../../filling_level/vggish/') # nopep8
from main import Config, run_kfold, get_cmd_args
from time import localtime, strftime
# import argparse
if __name__ == "__main__":
# Reproduce the best experiment
# if True, will use the pre-trained model and make predictions, if False, will train the model
use_pretrained = True
exp_name = 200903163404
cfg = Config()
cfg.load_from(path=f'./predictions/{exp_name}/cfg.txt')
# replacing the time with the old_time + current_time such that there is no collision
if use_pretrained:
cfg.init_time = exp_name
else:
cfg.init_time = f'{cfg.init_time}_{strftime("%y%m%d%H%M%S", localtime())}'
# Expected average of Best Metrics on Each Valid Set: 0.912957 @ 200903163404
run_kfold(cfg, use_pretrained, get_cmd_args().predict_on_private)
# Experiment with other parameters
# cfg = Config()
# cfg.assign_variable('task', 'ftype')
# cfg.assign_variable('output_dim', 4)
# cfg.assign_variable('model_type', 'GRU')
# cfg.assign_variable('bi_dir', False)
# cfg.assign_variable('device', 'cuda:0')
# cfg.assign_variable('data_root', '../../filling_level/vggish/vggish_features')
# cfg.assign_variable('batch_size', 64)
# cfg.assign_variable('input_dim', 128)
