parser.add_argument('--model_type', type=str,
choices=['decision_level_max_pooling',
'decision_level_average_pooling',
'decision_level_single_attention',
'decision_level_multi_attention',
'feature_level_single_attention'])
parser.add_argument('--learning_rate', type=float, default=1e-3)
subparsers = parser.add_subparsers(dest='mode')
parser_train = subparsers.add_parser('train')
parser_get_avg_stats = subparsers.add_parser('get_avg_stats')
args = parser.parse_args()
args.filename = utilities.get_filename(__file__)
# Logs
sub_dir = os.path.join(args.filename,
'balance_type={}'.format(args.balance_type),
'model_type={}'.format(args.model_type))
logs_dir = os.path.join(args.workspace, 'logs', sub_dir)
utilities.create_folder(logs_dir)
logging = utilities.create_logging(logs_dir, filemode='w')
logging.info(os.path.abspath(__file__))
logging.info(args)
totest = 0
# parser_train.add_argument('--cuda', action='store_true', default=True)
# append from main.py
parser_train = subparsers.add_parser('train')
parser_train.add_argument('--workspace', type=str, required=True)
parser_train.add_argument('--sample_rate', type=int, default=32000)
parser_train.add_argument('--window_size', type=int, default=1024)
parser_train.add_argument('--hop_size', type=int, default=320)
parser_train.add_argument('--mel_bins', type=int, default=64)
parser_train.add_argument('--fmin', type=int, default=50)
parser_train.add_argument('--fmax', type=int, default=14000)
parser_train.add_argument('--model_type', type=str, required=True)
parser_train.add_argument('--loss_type', type=str, default='clip_bce', choices=['clip_bce'])
parser_train.add_argument('--balanced', type=str, default='balanced', choices=['none', 'balanced', 'alternate'])
parser_train.add_argument('--augmentation', type=str, default='mixup', choices=['none', 'mixup'])
parser_train.add_argument('--batch_size', type=int, default=32)
parser_train.add_argument('--learning_rate', type=float, default=1e-3)
parser_train.add_argument('--resume_iteration', type=int, default=0)
parser_train.add_argument('--early_stop', type=int, default=20000) # early_stop * batch_size / num_trn_samples = epoch
parser_train.add_argument('--pretrained_checkpoint_path', type=str)
# parser_train.add_argument('--freeze_base', action='store_true', default=False)
parser_train.add_argument('--freeze_base_num', type=int, default=0)
# Parse arguments
args = parser.parse_args()
args.filename = get_filename(__file__)
if args.mode == 'train':
train(args)
else:
raise Exception('Error argument!')