print "error: " + str(new_error)
if abs(error - new_error) < 0.1: break
error = new_error
# save the network
print "Saving neural network..."
NetworkWriter.writeToFile(net, os.path.basename(dirname) + 'net')
if __name__ == '__main__':
dirname = os.path.normpath(sys.argv[1])
# wave_reader.extractFeatures(track)
trainNetwork(dirname)
net = NetworkReader.readFrom(os.path.basename(dirname) + 'net')
# predict on some of the training examples
print "Predicting on training set"
data = numpy.genfromtxt(os.path.join(dirname, 'train09_seg.csv'), delimiter=",")
labels = numpy.genfromtxt(os.path.join(dirname, 'train09REF.txt'), delimiter='\t')[0::10,1]
## for i in range(200):
## print net.activate(data[i]), labels[i]
cdata = numpy.array([])
for feature in data:
freq = max(0, net.activate(feature))
sample = wave_gen.saw(freq, 0.1, 44100)
cdata = numpy.concatenate([cdata, sample])
wave_gen.saveAudioBuffer('test.wav', cdata)
## for freq in labels:
## sample = wave_gen.saw(freq, 0.1, 44100)
## cdata = numpy.concatenate([cdata, sample])
## wave_gen.saveAudioBuffer('test_ref.wav', cdata)
dirname = os.path.normpath(sys.argv[1])
seq = sequencer.Sequencer() #create a sequencer to handle midi input
tracks = glob.glob(os.path.join(dirname, '*.mid')) #match any midi file
for t in tracks:
track = os.path.splitext(t)[0]
# sequence MIDI to saw + sine melody
print "Sequencing %s..." %track
wavData = seq.parseMidiFile("%s.mid" %track)
sawData = wavData[0]
melodyData = wavData[1]
print "Saving %s... wav and melody" %track
wave_gen.saveAudioBuffer("%s.wav" %track, sawData)
wave_gen.saveAudioBuffer("%smelody.wav" %track, melodyData)
print "Generating feature data..."
data = wave_reader.wavToFeatures("%s.wav" %track)
melody = wave_reader.wavToFeatures("%smelody.wav" %track)
labels = melody.argmax(axis=1) #axis=1, max frequency across a sample
numData = min(data.shape[0],labels.shape[0])
data = data[0:numData,]
labels = labels[0:numData]
matrix = numpy.concatenate([data, numpy.array([labels]).T], axis=1) #horizontal concatenate
saveFile('%sdata.csv' %track, matrix)
seq.resetSequencer()
if type(event) is midi.events.NoteOnEvent:
self.generateAudio(event.tick, event.channel)
note = event.pitch
self.noteActive[note] = True
elif type(event) is midi.events.NoteOffEvent:
self.generateAudio(event.tick, event.channel)
note = event.pitch
self.noteActive[note] = False
self.tracks.pop() #removes the extra added track (the first track is just metadata)
track_len = min([len(track) for track in self.tracks])
self.tracks = [track[0:track_len] for track in self.tracks] #trim all tracks to the exact same size
self.output[0] = numpy.sum(self.tracks, axis=0) #combine tracks
return self.output
def resetSequencer(self):
self.noteActive = [False] * 128
self.sampleRate = 44100
self.output = [numpy.empty(0),numpy.empty(0)]
self.tracks = []
self.bpm = 120 #default tempo
self.ppq = 96 #default resolution
if __name__ == '__main__':
seq = Sequencer()
samples = seq.parseMidiFile('000106b_.mid')
wave_gen.saveAudioBuffer('000106b_.wav', samples)
#samples = seq.parseMidiFile('mary.mid')
#wave_gen.saveAudioBuffer('mary.wav', samples)
#produces a result wav so you can listen to the prediction
if __name__ == '__main__':
dirname = os.path.normpath(sys.argv[1])
#track = os.path.join(dirname, '000106b_.wav')
print "Reading wav into input data..."
data = wave_reader.wavToFeatures('000106b_.wav')
numData = data.shape[0]
labels = numpy.zeros(numData)
print "Reloading neural network..."
net = NetworkReader.readFrom(os.path.basename(dirname) + 'designnet')
print "Activating neural network..."
for i in range(numData):
labels[i] = net.activate(data[i])
print "Generating result wav..."
cdata = numpy.array([])
for label in labels:
#if(freq > 0):
# freq = midi_util.frequencyToNoteFrequency(label)
label = round(label)
freq = midi_util.midiToFrequency(label)
sample = wave_gen.saw(freq, 0.25, 44100)
sample = wave_gen.saw(freq, 0.25, 44100)
cdata = numpy.concatenate([cdata, sample])
print "Saving result wav..."
wave_gen.saveAudioBuffer('000106b_predict.wav', cdata)
