def process_record(wav_data,
ns,
example_id,
min_length=5,
max_length=20,
sample_rate=16000,
allow_empty_notesequence=False):
"""Split a record into chunks and create an example proto.
To use the full length audio and notesequence, set min_length=0 and
max_length=-1.
Args:
wav_data: audio data in WAV format.
ns: corresponding NoteSequence.
example_id: id for the example proto
min_length: minimum length in seconds for audio chunks.
max_length: maximum length in seconds for audio chunks.
sample_rate: desired audio sample rate.
allow_empty_notesequence: whether an empty NoteSequence is allowed.
Yields:
Example protos.
"""
samples = audio_io.wav_data_to_samples(wav_data, sample_rate)
samples = librosa.util.normalize(samples, norm=np.inf)
if max_length == min_length:
splits = np.arange(0, ns.total_time, max_length)
elif max_length > 0:
splits = find_split_points(ns, samples, sample_rate, min_length, max_length)
else:
splits = [0, ns.total_time]
velocities = [note.velocity for note in ns.notes]
velocity_max = np.max(velocities) if velocities else 0
velocity_min = np.min(velocities) if velocities else 0
velocity_range = music_pb2.VelocityRange(min=velocity_min, max=velocity_max)
for start, end in zip(splits[:-1], splits[1:]):
if end - start < min_length:
continue
if start == 0 and end == ns.total_time:
new_ns = ns
else:
new_ns = sequences_lib.extract_subsequence(ns, start, end)
if not new_ns.notes and not allow_empty_notesequence:
tf.logging.warning('skipping empty sequence')
continue
if start == 0 and end == ns.total_time:
new_samples = samples
else:
# the resampling that happen in crop_wav_data is really slow
# and we've already done it once, avoid doing it twice
new_samples = audio_io.crop_samples(samples, sample_rate, start,
end - start)
new_wav_data = audio_io.samples_to_wav_data(new_samples, sample_rate)
yield create_example(
example_id, new_ns, new_wav_data, velocity_range=velocity_range)
def split2batch(audio, sequence):
from magenta.models.onsets_frames_transcription.audio_label_data_utils import find_split_points
pad_num = int(math.ceil(
sequence.total_time * cfg.SAMPLE_RATE)) - audio.shape[0]
if pad_num > 0:
audio = np.concatenate((audio, np.zeros((pad_num), dtype=audio.dtype)))
splits = [0, sequence.total_time] if cfg.MAX_SPLIT_LENGTH == 0 else \
find_split_points(sequence, audio, cfg.SAMPLE_RATE, cfg.MIN_SPLIT_LENGTH, cfg.MAX_SPLIT_LENGTH)
samples = []
for start, end in zip(splits[:-1], splits[1:]):
if end - start < cfg.MIN_SPLIT_LENGTH:
continue
split_audio, split_seq = audio, sequence
if not (start == 0 and end == sequence.total_time):
split_seq = sequences_lib.extract_subsequence(sequence, start, end)
split_audio = audio_io.crop_samples(audio, cfg.SAMPLE_RATE, start,
end - start)
pad_num = int(math.ceil(
cfg.MAX_SPLIT_LENGTH * cfg.SAMPLE_RATE)) - split_audio.shape[0]
if pad_num > 0:
split_audio = np.concatenate(
(split_audio, np.zeros((pad_num), dtype=split_audio.dtype)))
samples.append((split_audio, split_seq))
return samples
def process_record(wav_data,
ns,
example_id,
min_length=5,
max_length=20,
sample_rate=16000,
allow_empty_notesequence=False,
load_audio_with_librosa=False):
"""Split a record into chunks and create an example proto.
To use the full length audio and notesequence, set min_length=0 and
max_length=-1.
Args:
wav_data: audio data in WAV format.
ns: corresponding NoteSequence.
example_id: id for the example proto
min_length: minimum length in seconds for audio chunks.
max_length: maximum length in seconds for audio chunks.
sample_rate: desired audio sample rate.
allow_empty_notesequence: whether an empty NoteSequence is allowed.
load_audio_with_librosa: Use librosa for sampling. Works with 24-bit wavs.
Yields:
Example protos.
"""
try:
if load_audio_with_librosa:
samples = audio_io.wav_data_to_samples_librosa(
wav_data, sample_rate)
else:
samples = audio_io.wav_data_to_samples(wav_data, sample_rate)
except audio_io.AudioIOReadError as e:
print('Exception %s', e)
return
samples = librosa.util.normalize(samples, norm=np.inf)
# Add padding to samples if notesequence is longer.
pad_to_samples = int(math.ceil(ns.total_time * sample_rate))
padding_needed = pad_to_samples - samples.shape[0]
if padding_needed > 5 * sample_rate:
raise ValueError(
'Would have padded {} more than 5 seconds to match note sequence total '
'time. ({} original samples, {} sample rate, {} sample seconds, '
'{} sequence seconds) This likely indicates a problem with the source '
'data.'.format(example_id, samples.shape[0], sample_rate,
samples.shape[0] / sample_rate, ns.total_time))
samples = np.pad(samples, (0, max(0, padding_needed)), 'constant')
if max_length == min_length:
splits = np.arange(0, ns.total_time, max_length)
elif max_length > 0:
splits = find_split_points(ns, samples, sample_rate, min_length,
max_length)
else:
splits = [0, ns.total_time]
velocities = [note.velocity for note in ns.notes]
velocity_max = np.max(velocities) if velocities else 0
velocity_min = np.min(velocities) if velocities else 0
velocity_range = music_pb2.VelocityRange(min=velocity_min,
max=velocity_max)
for start, end in zip(splits[:-1], splits[1:]):
if end - start < min_length:
continue
if start == 0 and end == ns.total_time:
new_ns = ns
else:
new_ns = sequences_lib.extract_subsequence(ns, start, end)
if not new_ns.notes and not allow_empty_notesequence:
tf.logging.warning('skipping empty sequence')
continue
if start == 0 and end == ns.total_time:
new_samples = samples
else:
# the resampling that happen in crop_wav_data is really slow
# and we've already done it once, avoid doing it twice
new_samples = audio_io.crop_samples(samples, sample_rate, start,
end - start)
new_wav_data = audio_io.samples_to_wav_data(new_samples, sample_rate)
yield create_example(example_id,
new_ns,
new_wav_data,
velocity_range=velocity_range)