parser.add_argument('--save', help="save values to destination", type=str)
args = parser.parse_args()
input_folder = args.input_folder
target_folder = args.target_folder
# save_folder = 'results/baseline/raw'
frame = []
note = []
out_key = []
results = {}
if args.save is not None:
safe_mkdir(args.save)
for fn in os.listdir(input_folder):
if fn.endswith(
'_pr.csv') and not fn.startswith('.') and not 'chpn-e01' in fn:
filename_input = os.path.join(input_folder, fn)
filename_target = os.path.join(target_folder,
fn).replace('_pr.csv', '.mid')
print(filename_input)
data = DataMaps()
data.make_from_file(filename_target,
'time', [0, 30],
acoustic_model='kelz')
input_roll = np.loadtxt(filename_input)
parser.add_argument("--features", help="Use features in the x data points.", action="store_true")
parser.add_argument("--weight", help="This model will output weights directly.", action="store_true")
parser.add_argument("--epochs", help="The number of epochs to train for. Defaults to 100.",
type=int, default=100)
parser.add_argument("--lr", help="The learning rate of the Adam optimizer. Defaults to 0.001.",
type=float, default=0.001)
parser.add_argument("--no_lstm", help="Do not use the LSTM prior in the data.", action="store_true")
parser.add_argument("--out", help="The directory to save the model to. Defaults to '.' (current directory)",
default=".")
args = parser.parse_args()
ac_pitches = args.a_pitches
ac_pitches.append(0)
ac_pitches = sorted(list(set(ac_pitches)))
la_pitches = args.l_pitches
la_pitches.append(0)
la_pitches = sorted(list(set(la_pitches)))
safe_mkdir(args.out)
train_model_full(args.data, history=args.history, ac_pitch_window=ac_pitches,
la_pitch_window=la_pitches, min_diff=args.min_diff, features=args.features,
out=args.out, is_weight=args.weight, epochs=args.epochs, no_lstm=args.no_lstm,
lr=args.lr)
model_param['n_hidden'] = n_hidden
model_param['learning_rate'] = learning_rate
model_param['chunks'] = max_len
model_param['scheduled_sampling'] = args.sched_sampl
model_param['sampl_mix_weight'] = args.sampl_mix_weight
model_param['grad_clip'] = args.grad_clip
if args.pitchwise is None:
model_param['pitchwise'] = False
else:
model_param['pitchwise'] = True
model_param['n_notes'] = 2 * args.pitchwise + 1
train_param['batch_size'] = 500
save_path = args.save_path
log_path = os.path.join("ckpt", save_path)
safe_mkdir(log_path)
# f= open(os.path.join(log_path,"log.txt"), 'w')
# sys.stdout = f
model = make_model_from_dataset(data, model_param)
model.print_params()
if args.resume:
safe_mkdir(os.path.join(save_path, 'resume'))
model.resume_training(save_path, data, os.path.join(save_path, 'resume'),
train_param)
else:
if args.sched_sampl_load is not None:
model.resume_training(args.sched_sampl_load, data, save_path,
train_param)
else: