def _ExampleToInputs(self,
ex,
truncated_length=0,
crop_training_sequence_to_notes=False):
hparams = copy.deepcopy(constants.DEFAULT_HPARAMS)
hparams.crop_training_sequence_to_notes = crop_training_sequence_to_notes
filename = ex.features.feature['id'].bytes_list.value[0]
sequence, crop_beginning_seconds = data.preprocess_sequence(
ex.features.feature['sequence'].bytes_list.value[0], hparams)
wav_data = ex.features.feature['audio'].bytes_list.value[0]
if crop_training_sequence_to_notes:
wav_data = audio_io.crop_wav_data(wav_data, hparams.sample_rate,
crop_beginning_seconds,
sequence.total_time)
spec = data.wav_to_spec(wav_data, hparams=hparams)
roll = sequences_lib.sequence_to_pianoroll(
sequence,
frames_per_second=data.hparams_frames_per_second(hparams),
min_pitch=constants.MIN_MIDI_PITCH,
max_pitch=constants.MAX_MIDI_PITCH,
min_frame_occupancy_for_label=0.0,
onset_mode='length_ms',
onset_length_ms=32.,
onset_delay_ms=0.)
length = data.wav_to_num_frames(
wav_data, frames_per_second=data.hparams_frames_per_second(hparams))
return self._DataToInputs(spec, roll.active, roll.weights, length, filename,
truncated_length)
def _ExampleToInputs(self,
ex,
truncated_length=0,
crop_training_sequence_to_notes=False):
hparams = copy.deepcopy(constants.DEFAULT_HPARAMS)
hparams.crop_training_sequence_to_notes = crop_training_sequence_to_notes
filename = ex.features.feature['id'].bytes_list.value[0]
sequence, crop_beginning_seconds = data.preprocess_sequence(
ex.features.feature['sequence'].bytes_list.value[0], hparams)
wav_data = ex.features.feature['audio'].bytes_list.value[0]
if crop_training_sequence_to_notes:
wav_data = audio_io.crop_wav_data(wav_data, hparams.sample_rate,
crop_beginning_seconds,
sequence.total_time)
spec = data.wav_to_spec(wav_data, hparams=hparams)
roll = sequences_lib.sequence_to_pianoroll(
sequence,
frames_per_second=data.hparams_frames_per_second(hparams),
min_pitch=constants.MIN_MIDI_PITCH,
max_pitch=constants.MAX_MIDI_PITCH,
min_frame_occupancy_for_label=0.0,
onset_mode='length_ms',
onset_length_ms=32.,
onset_delay_ms=0.)
length = data.wav_to_num_frames(
wav_data,
frames_per_second=data.hparams_frames_per_second(hparams))
return self._DataToInputs(spec, roll.active, roll.weights, length,
filename, truncated_length)
def generate_train_set():
"""Generate the train TFRecord."""
train_file_pairs = []
for directory in train_dirs:
path = os.path.join(FLAGS.input_dir, directory)
path = os.path.join(path, '*.wav')
wav_files = glob.glob(path)
# find matching mid files
for wav_file in wav_files:
base_name_root, _ = os.path.splitext(wav_file)
mid_file = base_name_root + '.mid'
train_file_pairs.append((wav_file, mid_file))
train_output_name = os.path.join(FLAGS.output_dir,
'maps_config2_train.tfrecord')
with tf.python_io.TFRecordWriter(train_output_name) as writer:
for pair in train_file_pairs:
print(pair)
# load the wav data
wav_data = tf.gfile.Open(pair[0]).read()
samples = audio_io.wav_data_to_samples(wav_data, FLAGS.sample_rate)
# load the midi data and convert to a notesequence
midi_data = tf.gfile.Open(pair[1]).read()
ns = midi_io.midi_to_sequence_proto(midi_data)
splits = find_split_points(ns, samples, FLAGS.sample_rate,
FLAGS.min_length, FLAGS.max_length)
for start, end in zip(splits[:-1], splits[1:]):
if end - start < FLAGS.min_length:
continue
new_ns = sequences_lib.extract_subsequence(ns, start, end)
new_wav_data = audio_io.crop_wav_data(wav_data, FLAGS.sample_rate,
start, end - start)
example = tf.train.Example(features=tf.train.Features(feature={
'id':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[pair[0]]
)),
'sequence':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[new_ns.SerializeToString()]
)),
'audio':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[new_wav_data]
))
}))
writer.write(example.SerializeToString())
def generate_train_set():
"""Generate the train TFRecord."""
train_file_pairs = []
for directory in train_dirs:
path = os.path.join(FLAGS.input_dir, directory)
path = os.path.join(path, '*.wav')
wav_files = glob.glob(path)
# find matching mid files
for wav_file in wav_files:
base_name_root, _ = os.path.splitext(wav_file)
mid_file = base_name_root + '.mid'
train_file_pairs.append((wav_file, mid_file))
train_output_name = os.path.join(FLAGS.output_dir,
'maps_config2_train.tfrecord')
with tf.python_io.TFRecordWriter(train_output_name) as writer:
for pair in train_file_pairs:
print(pair)
# load the wav data
wav_data = tf.gfile.Open(pair[0]).read()
samples = audio_io.wav_data_to_samples(wav_data, FLAGS.sample_rate)
# load the midi data and convert to a notesequence
midi_data = tf.gfile.Open(pair[1]).read()
ns = midi_io.midi_to_sequence_proto(midi_data)
splits = find_split_points(ns, samples, FLAGS.sample_rate,
FLAGS.min_length, FLAGS.max_length)
for start, end in zip(splits[:-1], splits[1:]):
if end - start < FLAGS.min_length:
continue
new_ns = sequences_lib.extract_subsequence(ns, start, end)
new_wav_data = audio_io.crop_wav_data(wav_data,
FLAGS.sample_rate, start,
end - start)
example = tf.train.Example(features=tf.train.Features(
feature={
'id':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[pair[0]])),
'sequence':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[new_ns.SerializeToString()])),
'audio':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[new_wav_data]))
}))
writer.write(example.SerializeToString())
def transform_wav_data(wav_data, sequence_tensor):
"""Transforms with sox."""
sequence, cropped_beginning_seconds = preprocess_sequence(
sequence_tensor, hparams)
# Only do audio transformations during training.
if is_training:
wav_data = audio_io.jitter_wav_data(wav_data, hparams.sample_rate,
jitter_amount_sec)
wav_data = audio_transform.transform_wav_audio(wav_data, hparams)
# If requested, crop wav.
if hparams.crop_training_sequence_to_notes:
wav_data = audio_io.crop_wav_data(wav_data, hparams.sample_rate,
cropped_beginning_seconds,
sequence.total_time)
# Normalize.
if hparams.normalize_audio:
wav_data = audio_io.normalize_wav_data(wav_data, hparams.sample_rate)
return [wav_data]
def transform_wav_data(wav_data, sequence_tensor):
"""Transforms with sox."""
sequence, cropped_beginning_seconds = preprocess_sequence(
sequence_tensor, hparams)
# Only do audio transformations during training.
if is_training:
wav_data = audio_io.jitter_wav_data(wav_data, hparams.sample_rate,
jitter_amount_sec)
wav_data = audio_transform.transform_wav_audio(wav_data, hparams)
# If requested, crop wav.
if hparams.crop_training_sequence_to_notes:
wav_data = audio_io.crop_wav_data(wav_data, hparams.sample_rate,
cropped_beginning_seconds,
sequence.total_time)
# Normalize.
if hparams.normalize_audio:
wav_data = audio_io.normalize_wav_data(wav_data, hparams.sample_rate)
return [wav_data]
def generate_train_set(exclude_ids):
"""Generate the train TFRecord."""
train_file_pairs = []
for directory in train_dirs:
path = os.path.join(FLAGS.input_dir, directory)
path = os.path.join(path, '*.wav')
wav_files = glob.glob(path)
# find matching mid files
for wav_file in wav_files:
base_name_root, _ = os.path.splitext(wav_file)
mid_file = base_name_root + '.mid'
if filename_to_id(wav_file) not in exclude_ids:
train_file_pairs.append((wav_file, mid_file))
train_output_name = os.path.join(FLAGS.output_dir,
'maps_config2_train.tfrecord')
with tf.python_io.TFRecordWriter(train_output_name) as writer:
for pair in train_file_pairs:
print(pair)
# load the wav data
wav_data = tf.gfile.Open(pair[0], 'rb').read()
samples = audio_io.wav_data_to_samples(wav_data, FLAGS.sample_rate)
samples = librosa.util.normalize(samples, norm=np.inf)
# load the midi data and convert to a notesequence
ns = midi_io.midi_file_to_note_sequence(pair[1])
splits = find_split_points(ns, samples, FLAGS.sample_rate,
FLAGS.min_length, FLAGS.max_length)
velocities = [note.velocity for note in ns.notes]
velocity_max = np.max(velocities)
velocity_min = np.min(velocities)
new_velocity_tuple = music_pb2.VelocityRange(
min=velocity_min, max=velocity_max)
for start, end in zip(splits[:-1], splits[1:]):
if end - start < FLAGS.min_length:
continue
new_ns = sequences_lib.extract_subsequence(ns, start, end)
new_wav_data = audio_io.crop_wav_data(wav_data, FLAGS.sample_rate,
start, end - start)
example = tf.train.Example(features=tf.train.Features(feature={
'id':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[pair[0]]
)),
'sequence':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[new_ns.SerializeToString()]
)),
'audio':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[new_wav_data]
)),
'velocity_range':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[new_velocity_tuple.SerializeToString()]
)),
}))
writer.write(example.SerializeToString())
def generate_train_set(exclude_ids):
"""Generate the train TFRecord."""
train_file_pairs = []
for directory in train_dirs:
path = os.path.join(FLAGS.input_dir, directory)
path = os.path.join(path, '*.wav')
wav_files = glob.glob(path)
# find matching mid files
for wav_file in wav_files:
base_name_root, _ = os.path.splitext(wav_file)
mid_file = base_name_root + '.mid'
if filename_to_id(wav_file) not in exclude_ids:
train_file_pairs.append((wav_file, mid_file))
train_output_name = os.path.join(FLAGS.output_dir,
'maps_config2_train.tfrecord')
with tf.python_io.TFRecordWriter(train_output_name) as writer:
for pair in train_file_pairs:
print(pair)
# load the wav data
wav_data = tf.gfile.Open(pair[0], 'rb').read()
samples = audio_io.wav_data_to_samples(wav_data, FLAGS.sample_rate)
samples = librosa.util.normalize(samples, norm=np.inf)
# load the midi data and convert to a notesequence
ns = midi_io.midi_file_to_note_sequence(pair[1])
splits = find_split_points(ns, samples, FLAGS.sample_rate,
FLAGS.min_length, FLAGS.max_length)
velocities = [note.velocity for note in ns.notes]
velocity_max = np.max(velocities)
velocity_min = np.min(velocities)
new_velocity_tuple = music_pb2.VelocityRange(min=velocity_min,
max=velocity_max)
for start, end in zip(splits[:-1], splits[1:]):
if end - start < FLAGS.min_length:
continue
new_ns = sequences_lib.extract_subsequence(ns, start, end)
new_wav_data = audio_io.crop_wav_data(wav_data,
FLAGS.sample_rate, start,
end - start)
example = tf.train.Example(features=tf.train.Features(
feature={
'id':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[pair[0]])),
'sequence':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[new_ns.SerializeToString()])),
'audio':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[new_wav_data])),
'velocity_range':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[new_velocity_tuple.SerializeToString()])),
}))
writer.write(example.SerializeToString())
