def play_midi(self, filename, track=1, transpose=6):
midi = MidiFile(filename)
tune = midi.read_track(track)
self.play_tune(midi.tempo,
tune,
transpose=transpose,
name=filename)
def test_03_read_track(self):
m = MidiFile(self.sample_file)
notes = m.read_track(1)
for n, expected in zip(notes[:4], (('r', 0.4343891402714933),
('c2', 1.116279069767441),
('r', 48.00000000000171),
('f1', 1.499999999999998))):
self.assertEqual(n[0], expected[0])
self.assertTrue(isclose(n[1], n[1]))
def __init__(self, infile):
self.my_tracks = [] #read from file
self.my_midiMap = [] # matches MidiSndObj.note_event() with NoteEvents
self.next_time = 0.0
self.tempo = 120.0
self.tick_counter = 0
self.each_time = (60.0 / self.tempo) / 96.0
self.m = MidiFile()
self.m.open(infile)
self.m.read()
self.m.close()
class ViziMidiFile(object):
def AddNote(self, e):
print e
def __init__(self, infile):
self.my_tracks = [] #read from file
self.my_midiMap = [] # matches MidiSndObj.note_event() with NoteEvents
self.next_time = 0.0
self.tempo = 120.0
self.tick_counter = 0
self.each_time = (60.0 / self.tempo) / 96.0
self.m = MidiFile()
self.m.open(infile)
self.m.read()
self.m.close()
def __repr__(self):
return `self.__dict__`
def Print(self):
print self.m
def Restart(self, restart_tick = 0):
for my_track in self.my_tracks:
my_track.Restart(restart_tick)
self.tick_counter = restart_tick
self.next_time = restart_tick * self.each_time
def DoProcess(self, this_time): # increments the tick counter when the time is right
if self.next_time >= this_time:
return True
else:
ret_value = 0
self.next_time += self.each_time
self.tick_counter += 1
for this_track in self.my_tracks:
ret_value += this_track.DoProcess(self.tick_counter)
return ret_value
def test_02_init(self):
m = MidiFile(self.sample_file)
def test_01_init_no_file(self):
with self.assertRaises(OSError):
m = MidiFile('nonexist')
# Ick, ick, ick. Use a real timer
if trackPosition < len(masterTrack):
tickCount = masterTrack[trackPosition].delta * 300
if __name__ == "__main__":
logging.basicConfig(level=logging.ERROR,
format='%(asctime)s %(levelname)s: %(message)s',
datefmt='%Y-%m-%d %H:%M:%S')
masterTrack = None
if len(sys.argv) > 1:
fname = sys.argv[1]
if fname.endswith(".mid"):
(time, tracks) = MidiFile.parseFile(fname)
else:
tracks = MusicParser.processMusic(sys.argv[1])
masterTrack = MidiTools.mergeTracks(tracks)
tickCount = 0
trackPosition = 0
# MidiTools.printTrack(masterTrack)
pygame.init()
pygame.midi.init()
# MIDI keyboard
inp = None
def preprocess(filename, timidity, latency, truncate, pad=1, get_raw=False):
"""
Preprocess an audio file ands its MIDI counterpart. Computes transforms and labels.
:param filename: audio filename
:param timidity: set to True if the files was rendered with timidity
:param latency: in seconds
:param truncate: in seconds (0 for no truncation)
:param pad: in seconds, will be added at the start and end before spectral transforms
:param get_raw: set to True to return raw computed spectrograms (e.g. for visualization)
:return:
"""
filename_midi = filename.rsplit('.')[0] + '.mid'
dname = filename.replace('/', '_').replace('\\', '_')
# Load files
ipad = int(pad * 44100)
audio_pad = (ipad, ipad
) # add one blank second at the beginning and at the end
if truncate > 0:
audio = AudioFile(filename,
truncate=int(truncate * 44100),
pad=audio_pad)
else:
audio = AudioFile(filename, pad=audio_pad)
mid = MidiFile(filename_midi)
step = 0.02 # seconds
latency = int(round(latency / step, 0))
# Compute spectrograms
spectrograms = ComputeSpectrograms(audio, step=step)
# Compute filtered spectrograms
melgrams = ComputeMelLayers(spectrograms, step, audio.Fs, latency)
# Build the input tensor
cnn_window = 15
tensor_mel = BuildTensor(melgrams[:, 2], cnn_window)
# Compute CQT
FreqAxisLog, time, cqgram = ComputeCqt(audio,
200.,
4000.,
step,
latency,
r=3)
tensor_cqt = BuildTensor([
cqgram,
], cnn_window)
# Global data length
max_len = min(tensor_mel.shape[0], tensor_cqt.shape[0])
# Compute output labels
notes = mid.getNotes(timidity)
notes_onset = np.array(notes)[:, 0] # get only the note timing
notes_value = np.array(notes, dtype=np.int)[:, 1] # get only the key value
onset_labels = np.zeros(max_len)
onset_caracs = np.zeros((max_len, 5))
onset_caracs[:, 2] = np.arange(max_len)
note_low = 21 # lowest midi note on a keyboard
note_high = 108 # highest midi note on a keyboard
notes_labels = np.zeros((max_len, note_high - note_low + 1))
notes_caracs = np.zeros((max_len, note_high - note_low + 1))
for i in range(len(notes_onset)):
t_win = int(np.floor(
(notes_onset[i] + audio_pad[0] / audio.Fs) / step))
if t_win >= len(onset_labels):
break
if t_win >= 0:
onset_labels[t_win] = 1
onset_caracs[t_win][0] += 1 # nb_notes
onset_caracs[t_win][1] = max(onset_caracs[t_win][1],
notes[i][2]) # max volume
if t_win + 1 < len(onset_labels):
onset_caracs[t_win + 1:, 2] -= onset_caracs[t_win + 1][
2] # nb of blank windows since the last onset
n = notes_value[i] - note_low
notes_labels[t_win][n] = 1
notes_caracs[t_win][n] = notes[i][2] # volume
counter = 0
for i in range(len(onset_labels) - 1, -1, -1):
onset_caracs[i][3] = counter
if onset_labels[i] == 1:
counter = 0
else:
counter += 1
onset_caracs[:, 4] = np.minimum(onset_caracs[:, 2], onset_caracs[:, 3])
# Extract useful CQT
select = [i for i in range(max_len) if onset_labels[i] > 0]
tensor_cqt_select = np.take(tensor_cqt, select, axis=0)
notes_labels_select = np.take(notes_labels, select, axis=0)
notes_caracs_select = np.take(notes_caracs, select, axis=0)
if not get_raw:
return (tensor_mel[:max_len, ...], tensor_cqt_select, onset_labels,
onset_caracs, notes_labels_select, notes_caracs_select, dname)
else:
return (melgrams, tensor_mel, onset_labels, cqgram, tensor_cqt, time,
FreqAxisLog, max_len, step)
