def _CreateExamplesAndExpectedInputs(self, truncated_length, lengths, expected_num_inputs): hparams = copy.deepcopy(configs.DEFAULT_HPARAMS) examples = [] expected_inputs = [] for i, length in enumerate(lengths): wav_samples = np.zeros( (np.int((length / data.hparams_frames_per_second(hparams)) * hparams.sample_rate), 1), np.float32) wav_data = audio_io.samples_to_wav_data(wav_samples, hparams.sample_rate) num_frames = data.wav_to_num_frames( wav_data, frames_per_second=data.hparams_frames_per_second(hparams)) seq = self._SyntheticSequence( num_frames / data.hparams_frames_per_second(hparams), i + constants.MIN_MIDI_PITCH) examples.append(self._FillExample(seq, wav_data, 'ex%d' % i)) expected_inputs += self._ExampleToInputs(examples[-1], truncated_length) self.assertEqual(expected_num_inputs, len(expected_inputs)) return examples, expected_inputs
def generate_test_set(): """Generate the test TFRecord.""" test_file_pairs = [] for directory in test_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' test_file_pairs.append((wav_file, mid_file)) test_output_name = os.path.join(FLAGS.output_dir, 'maps_config2_test.tfrecord') with tf.python_io.TFRecordWriter(test_output_name) as writer: for idx, pair in enumerate(test_file_pairs): print('{} of {}: {}'.format(idx, len(test_file_pairs), pair[0])) # load the wav data and resample it. samples = audio_io.load_audio(pair[0], FLAGS.sample_rate) wav_data = audio_io.samples_to_wav_data(samples, FLAGS.sample_rate) # load the midi data and convert to a notesequence ns = midi_io.midi_file_to_note_sequence(pair[1]) example = audio_label_data_utils.create_example( pair[0], ns, wav_data) writer.write(example.SerializeToString()) return [filename_to_id(wav) for wav, _ in test_file_pairs]
def mix_examples(mixid_exs, sample_rate, load_audio_with_librosa): """Mix several Examples together to create a new example.""" mixid, exs = mixid_exs del mixid example_samples = [] example_sequences = [] for ex_str in exs: ex = tf.train.Example.FromString(ex_str) wav_data = ex.features.feature['audio'].bytes_list.value[0] 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) example_samples.append(samples) ns = music_pb2.NoteSequence.FromString( ex.features.feature['sequence'].bytes_list.value[0]) example_sequences.append(ns) mixed_samples, mixed_sequence = audio_label_data_utils.mix_sequences( individual_samples=example_samples, sample_rate=sample_rate, individual_sequences=example_sequences) mixed_wav_data = audio_io.samples_to_wav_data(mixed_samples, sample_rate) mixed_id = '::'.join(['mixed'] + [ns.id for ns in example_sequences]) mixed_sequence.id = mixed_id mixed_filename = '::'.join(['mixed'] + [ns.filename for ns in example_sequences]) mixed_sequence.filename = mixed_filename examples = list( audio_label_data_utils.process_record(mixed_wav_data, mixed_sequence, mixed_id, min_length=0, max_length=-1, sample_rate=sample_rate)) assert len(examples) == 1 return examples[0]
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] velocity_range = velocity_range_from_sequence(ns) 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 _CreateSyntheticExample(self): sequence = self._CreateSyntheticSequence() wav_samples = np.zeros(9 * SAMPLE_RATE, np.float32) wav_data = audio_io.samples_to_wav_data(wav_samples, SAMPLE_RATE) return wav_data, sequence