def preprocess_midi(path):
note_seq = NoteSeq.from_drum_midi_file(path)#get note
note_seq.adjust_time(-note_seq.notes[0].start)# change offset of note seq
event_seq = EventSeq.from_note_seq(note_seq)
control_seq = ControlSeq.from_event_seq(event_seq)
return event_seq.to_array(), control_seq.to_compressed_array()
def transposition(events, controls, offset=0):
# events [steps, batch_size, event_dim]
# return events, controls
events = np.array(events, dtype=np.int64)
controls = np.array(controls, dtype=np.float32)
event_feat_ranges = EventSeq.feat_ranges()
on = event_feat_ranges['note_on']
off = event_feat_ranges['note_off']
if offset > 0:
indeces0 = (((on.start <= events) & (events < on.stop - offset)) |
((off.start <= events) & (events < off.stop - offset)))
indeces1 = (((on.stop - offset <= events) & (events < on.stop)) |
((off.stop - offset <= events) & (events < off.stop)))
events[indeces0] += offset
events[indeces1] += offset - 12
elif offset < 0:
indeces0 = (((on.start - offset <= events) & (events < on.stop)) |
((off.start - offset <= events) & (events < off.stop)))
indeces1 = (((on.start <= events) & (events < on.start - offset)) |
((off.start <= events) & (events < off.start - offset)))
events[indeces0] += offset
events[indeces1] += offset + 12
assert ((0 <= events) & (events < EventSeq.dim())).all()
histr = ControlSeq.feat_ranges()['pitch_histogram']
controls[:, :, histr.start:histr.stop] = np.roll(
controls[:, :, histr.start:histr.stop], offset, -1)
return events, controls
def main(_):
model_path = os.path.join('./model/', FLAGS.name)
if os.path.exists(model_path) is False:
os.makedirs(model_path)
if os.path.isdir(model_path):
checkpoint_path =\
tf.train.latest_checkpoint(model_path)
dataset = load_dataset()
batch_gen = dataset.batches(FLAGS.num_seqs, FLAGS.num_steps, 10)
print(EventSeq.dim())
model = CharRNN(EventSeq.dim(),
ControlSeq.dim(),
num_seqs=FLAGS.num_seqs,
num_steps=FLAGS.num_steps,
lstm_size=FLAGS.lstm_size,
num_layers=FLAGS.num_layers,
learning_rate=FLAGS.learning_rate,
train_keep_prob=FLAGS.train_keep_prob,
use_embedding=FLAGS.use_embedding,
embedding_size=FLAGS.embedding_size)
model.sess.run(tf.global_variables_initializer())
model.load(checkpoint_path)
model.train(
batch_gen,
FLAGS.max_steps,
model_path,
FLAGS.save_every_n,
FLAGS.log_every_n,
)
def main(_):
if os.path.isfile(FLAGS.control) or os.path.isdir(FLAGS.control):
if os.path.isdir(FLAGS.control):
files = list(utils.find_files_by_extensions(FLAGS.control))
assert len(files) > 0, 'no file in "{control}"'.format(
control=FLAGS.control)
control = np.random.choice(files)
events, compressed_controls = torch.load(FLAGS.control)
controls = ControlSeq.recover_compressed_array(compressed_controls)
max_len = FLAGS.max_length
if FLAGS.max_length == 0:
max_len = controls.shape[0]
control = np.expand_dims(controls, 1).repeat(1, 1)
control = 'control sequence from "{control}"'.format(control=control)
assert max_len > 0, 'either max length or control sequence length should be given'
#FLAGS.start_string = FLAGS.start_string.decode('utf-8')
if os.path.isdir(FLAGS.checkpoint_path):
FLAGS.checkpoint_path =\
tf.train.latest_checkpoint(FLAGS.checkpoint_path)
model = CharRNN(EventSeq.dim(),
ControlSeq.dim(),
sampling=True,
lstm_size=FLAGS.lstm_size,
num_layers=FLAGS.num_layers,
use_embedding=FLAGS.use_embedding,
embedding_size=FLAGS.embedding_size)
model.sess.run(tf.global_variables_initializer())
model.load(FLAGS.checkpoint_path)
outputs = model.sample(1000,
prime=events[0:100],
vocab_size=EventSeq.dim())
outputs = outputs.reshape([-1, 1])
print(outputs)
name = 'output-{i:03d}.mid'.format(i=0)
path = os.path.join("output/", name)
n_notes = utils.event_indeces_to_midi_file(outputs[:, 0], path)
print('===> {path} ({n_notes} notes)'.format(path=path, n_notes=n_notes))
def __init__(self, root, verbose=False):
assert os.path.isdir(root), root
paths = utils.find_files_by_extensions(root, ['.data'])
self.root = root
self.samples = []
self.seqlens = []
if verbose:
paths = Bar(root).iter(list(paths))
for path in paths:
eventseq, controlseq = torch.load(path)
controlseq = ControlSeq.recover_compressed_array(controlseq)
assert len(eventseq) == len(controlseq)
self.samples.append((eventseq, controlseq))
self.seqlens.append(len(eventseq))
self.avglen = np.mean(self.seqlens)
def preprocess_midi(path):
note_seq = NoteSeq.from_midi_file(path)
note_seq.adjust_time(-note_seq.notes[0].start)
event_seq = EventSeq.from_note_seq(note_seq)
control_seq = ControlSeq.from_event_seq(event_seq)
return event_seq.to_array(), control_seq.to_compressed_array()
sess_path = options.sess_path
data_path = options.data_path
saving_interval = options.saving_interval
learning_rate = options.learning_rate
batch_size = options.batch_size
window_size = options.window_size
stride_size = options.stride_size
use_transposition = options.use_transposition
control_ratio = options.control_ratio
teacher_forcing_ratio = options.teacher_forcing_ratio
reset_optimizer = options.reset_optimizer
enable_logging = options.enable_logging
event_dim = EventSeq.dim()
control_dim = ControlSeq.dim()
model_config = config.model
model_params = utils.params2dict(options.model_params)
model_config.update(model_params)
device = config.device
print('-' * 70)
print('Session path:', sess_path)
print('Dataset path:', data_path)
print('Saving interval:', saving_interval)
print('-' * 70)
print('Hyperparameters:', utils.dict2params(model_config))
print('Learning rate:', learning_rate)
print('Batch size:', batch_size)
if use_beam_search:
greedy_ratio = 'DISABLED'
else:
beam_size = 'DISABLED'
assert os.path.isfile(sess_path), f'"{sess_path}" is not a file'
if control is not None:
if os.path.isfile(control) or os.path.isdir(control):
if os.path.isdir(control):
files = list(utils.find_files_by_extensions(control))
assert len(files) > 0, f'no file in "{control}"'
control = np.random.choice(files)
_, compressed_controls = torch.load(control)
controls = ControlSeq.recover_compressed_array(compressed_controls)
if max_len == 0:
max_len = controls.shape[0]
controls = torch.tensor(controls, dtype=torch.float32)
controls = controls.unsqueeze(1).repeat(1, batch_size, 1).to(device)
control = f'control sequence from "{control}"'
else:
pitch_histogram, note_density = control.split(';')
pitch_histogram = list(filter(len, pitch_histogram.split(',')))
if len(pitch_histogram) == 0:
pitch_histogram = np.ones(12) / 12
else:
pitch_histogram = np.array(list(map(float, pitch_histogram)))
assert pitch_histogram.size == 12
assert np.all(pitch_histogram >= 0)
import torch
from sequence import EventSeq, ControlSeq
#pylint: disable=E1101
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = {
'init_dim': 32,
'event_dim': EventSeq.dim(),
'control_dim': ControlSeq.dim(),
'hidden_dim': 512,
'gru_layers': 3,
'gru_dropout': 0.3,
}
train = {
'learning_rate': 0.001,
'batch_size': 64,
'window_size': 200,
'stride_size': 10,
'use_transposition': False,
'control_ratio': 1.0,
'teacher_forcing_ratio': 1.0
}
