def get_onsets_and_pitch_labels(midifile):
pattern = midi.MIDIFile(midifile)
intervals = []
labels = []
for onset, _pitch, duration, velocity, _channel in pattern.sustained_notes:
label = int(_pitch) # do not subtract 21; mir_eval needs pitches strictly >= 0 anyways
intervals.append([onset, onset + duration])
labels.append(label)
return np.array(intervals), np.array(labels)
def load_performance(path,
piece,
spectrogram_params,
coords,
coord2onset,
sf_path,
tempo_factor=1.,
real_perf=False,
transpose=0):
if real_perf:
wav_path = os.path.join(path, 'performance',
piece + f'_{tempo_factor}.wav')
midi_path = os.path.join(path, 'performance', piece + '.mid')
else:
if tempo_factor == -1:
# flag to indicate no tempo factor
midi_path = os.path.join(path, 'performance', piece + '.mid')
else:
midi_path = os.path.join(path, 'performance',
piece + f'_tempo_{tempo_factor}.mid')
midi = mm_midi.MIDIFile(midi_path)
if transpose != 0:
notes = midi.notes
notes[:, 1] += transpose
midi = mm_midi.MIDIFile.from_notes(notes)
if real_perf and tempo_factor != -1:
spec = wav_to_spec_otf(wav_path, spectrogram_params)
else:
spec = midi_to_spec_otf(midi,
spectrogram_params,
sound_font_path=sf_path)
spec = np.pad(spec, ((0, 0), (spectrogram_params['pad'], 0)),
mode='constant')
onsets = (midi.notes[:, 0] * spectrogram_params['fps']).astype(int)
onsets, coords_new = merge_onsets(onsets, copy.deepcopy(coords),
coord2onset[0])
interpol_fnc = interpolate.interp1d(onsets,
coords_new.T,
kind='previous',
bounds_error=False,
fill_value=(coords_new[0, :],
coords_new[-1, :]))
return spec, onsets, coords_new, interpol_fnc
def get_y_from_file(midifile, n_frames, dt):
pattern = midi.MIDIFile(midifile)
y = np.zeros((n_frames, 88)).astype(np.float32)
for onset, _pitch, duration, velocity, _channel in pattern.notes:
pitch = int(_pitch)
frame_start = int(np.round(onset / dt))
frame_end = int(np.round((onset + duration) / dt))
# even if the event was too short, always produce a label!
if frame_start == frame_end:
frame_end += 1
label = pitch - 21
y[frame_start:frame_end, label] = 1
return y
def from_midi(annot_path, uid, hop_samples=256, unique_mf0=True):
check_dir(CACHED_FILES_PATH)
# Check if there is a cached numpy binary
cached_path = os.path.join(CACHED_FILES_PATH, "{}_{}.npz".format(uid, hop_samples))
if os.path.isfile(cached_path):
times, freqs, notes, voicing = np.load(cached_path).values()
return Annotation(times, freqs, notes, voicing)
else:
pattern = midi.MIDIFile(annot_path)
tree = IntervalTree()
for onset, _pitch, duration, velocity, _channel in pattern.notes:
pitch = int(_pitch)
frame_start = onset
frame_end = onset + duration
tree[frame_start:frame_end] = pitch
max_time = max(tree)[1]
times = np.arange(0, max_time, hop_samples/44100)
notes_mf0 = []
for t in times:
notes_at_t = []
for note in tree[t]:
notes_at_t.append(note.data)
if unique_mf0:
# remove duplicate notes
notes_at_t = list(set(notes_at_t))
notes_mf0.append(notes_at_t)
max_polyphony = np.max([len(frame) for frame in notes_mf0])
freqs = np.zeros((len(times), max_polyphony))
notes = np.zeros((len(times), max_polyphony))
voicing = np.zeros((len(times),), dtype=np.int32)
for i, notes_at_i in enumerate(notes_mf0):
for j, note in enumerate(notes_at_i):
notes[i, j] = note
freqs[i, j] = mir_eval.util.midi_to_hz(note)
voicing[i] += 1
annot = Annotation(times, freqs, notes, voicing)
np.savez(cached_path, annot.times, annot.freqs, annot.notes, annot.voicing)
return annot
def get_y_from_file(midifile, n_frames, audio_options):
pattern = midi.MIDIFile(midifile)
dt = float(audio_options['hop_size']) / float(audio_options['sample_rate'])
y_frames = np.zeros((n_frames, 88)).astype(np.float32)
y_velocity = np.zeros((n_frames, 88)).astype(np.float32)
for onset, _pitch, duration, velocity, _channel in pattern.sustained_notes:
pitch = int(_pitch)
label = pitch - 21
note_start = int(np.round(onset / dt))
note_end = int(np.round((onset + duration) / dt))
y_frames[note_start:note_end + 1,
label] = curve(note_start, note_end + 1)
y_velocity[note_start:note_end + 1, label] = velocity / 127.
return y_frames, y_velocity
def get_y_from_file(midifile, n_frames, audio_options):
with warnings.catch_warnings():
warnings.simplefilter('ignore')
pattern = midi.MIDIFile(midifile)
dt = float(audio_options['hop_size']) / float(
audio_options['sample_rate'])
y_onsets = np.zeros((n_frames, 88), dtype=np.uint8)
y_frames = np.zeros((n_frames, 88), dtype=np.uint8)
y_offsets = np.zeros((n_frames, 88), dtype=np.uint8)
for onset, _pitch, duration, velocity, _channel in pattern.sustained_notes:
pitch = int(_pitch)
label = pitch - 21
note_start = int(np.round(onset / dt))
note_end = int(np.round((onset + duration) / dt))
# some of the midi-files have onsets/offsets larger
# than n_frames. they were manually checked, and it's
# actually not an issue at all.
# see data-preparation/maestro-inconsistencies/* for
# scripts that perform visual inspection!
if note_start < n_frames:
if note_end >= n_frames:
# print('weird_offset', midifile)
note_end = n_frames - 1
y_onsets[note_start, label] = 1
y_frames[note_start:note_end + 1, label] = 1
y_offsets[note_end, label] = 1
else:
# print('weird_onset', midifile)
pass
return y_onsets, y_frames, y_offsets
def test_midi(self):
run_single(self.bin, stereo_sample_file, tmp_result, args=['--midi'])
result = midi.MIDIFile(tmp_result).notes
self.assertTrue(np.allclose(result[:, :2], self.result, atol=1e-3))
