def learn(self,
*training_set: Union[List[Note], ChordProgression],
epochs=None,
callback=None):
self.maxtsbq = max(n.ticks_since_beat_quantised
for notes, _ in nwise_disjoint(training_set, 2)
for n in notes)
self.maxdq = max(n.duration_quantised
for notes, changes in nwise_disjoint(training_set, 2)
for n in notes)
self.maxbeatdiff = max(m.beat - n.beat
for notes, _ in nwise_disjoint(training_set, 2)
for n, m in nwise(notes, 2))
self.model = self._build_net()
self.pitch_model = keras.models.Model(inputs=self.model.inputs,
outputs=self.model.outputs[0])
self.tsbq_model = keras.models.Model(inputs=self.model.inputs,
outputs=self.model.outputs[1])
self.dq_model = keras.models.Model(inputs=self.model.inputs,
outputs=self.model.outputs[2])
x, y = self._all_training_data(training_set)
stateful = any(
isinstance(layer, keras.layers.Recurrent) and layer.stateful
for layer in self.model.layers)
kwargs = {'batch_size': 1, 'shuffle': False} if stateful else {}
if callback:
kwargs.update({'callbacks': [callback], 'verbose': 0})
self.model.fit(x, y, epochs=epochs or self.epochs, **kwargs)
return not self.model.stop_training
def _all_training_data(self, training_set: Iterable[Union[List[Note], ChordProgression]]):
msg_callback = self.progressbar.set_text \
if isinstance(self.progressbar, gui.viewmodel.KerasProgressbar) \
else print
msg_callback('Processing training data...')
len_data = sum(len(notes) - self.order + 1 for notes in training_set[::2])
if self.progressbar is None:
self.progressbar = keras.utils.Progbar(len_data)
else:
self.progressbar.set_params({'samples': len_data})
x, p, t, d, o, b = [], [], [], [], [], []
for notes, changes in nwise_disjoint(training_set, 2):
for v in nwise(notes, self.order + 1):
self.progressbar.add(1)
xx = self._encode_network_input(v[:self.order],
changes.unique(v[-1].beat - 1, self.chord_radius),
changes)
if not x:
x = [[] for _ in xx]
for xi, xxi in zip(x, xx):
xi.append(xxi)
p.append(encode_int(v[-1].abcnote.value, 12))
t.append(encode_int(v[-1].ticks_since_beat_quantised, self.maxtsbq + 1))
d.append(encode_int(v[-1].duration_quantised, self.maxdq + 1))
o.append(encode_int(v[-1].octave, NUM_OCTAVES))
b.append(encode_int(v[-1].beat - v[-2].beat, self.maxbeatdiff + 1))
self.progressbar.update(self.progressbar.target, force=True)
msg_callback('Starting training...')
return [np.array(xi, dtype=bool) for xi in x],\
[np.array(arr, dtype=bool) for arr in [p, t, d, o, b]]
def _all_training_pairs(self, notes: List[Note]) -> Tuple[numpy.ndarray, numpy.ndarray]:
""" 1-of-N binary encoding of all pairs of inputs (past notes and current chord) and outputs (next note)
on the training set """
x, y = [], []
for v in nwise(notes, self.order + 1):
x.append(self._encode_network_input(v[:self.order], self.changes[v[-1].beat - 1]))
y.append(v[-1].pitch)
return numpy.array(x), numpy.array(y)
def _all_training_pairs(
self, notes: List[Note]) -> Tuple[numpy.ndarray, numpy.ndarray]:
""" 1-of-N binary encoding of all pairs of inputs (past notes and current chord) and outputs (next note)
on the training set """
x, y = [], []
for v in nwise(notes, self.order + 1):
x.append(
self._encode_network_input(v[:self.order],
self.changes[v[-1].beat - 1]))
y.append(v[-1].pitch)
return numpy.array(x), numpy.array(y)
def learn(self, seq: Sequence[State]):
""" High-level interface to train the Markov-chain.
:param seq: The complete sequence on which to train.
"""
seq = self._ensure_seq(seq)
self.training_prep(seq)
for np1gram in nwise(seq, self.order + 1):
self.training_step(np1gram)
self.training_finish_normalize()
def learn(self, notes: List[Note], *_, **__):
self.markovs_by_chord = {
chord: Markov(self.order)
for chord in self.changes
}
for markov in self.markovs_by_chord.values():
markov.training_prep([n.pitch for n in notes])
for np1gram in nwise(notes, self.order + 1):
chord = self.changes[np1gram[-1].beat - 1]
self.markovs_by_chord[chord].training_step(
[n.pitch for n in np1gram])
for markov in self.markovs_by_chord.values():
markov.training_finish_normalize()
self.past = [n.pitch for n in notes[:self.order]]
def _all_training_data(self, training_set: Iterable[Union[List[Note], ChordProgression]]):
""" 1-of-N binary encoding of all pairs of inputs (past notes and current chord) and outputs (next note)
on the training set """
x, p, t, d = [], [], [], []
for notes, changes in nwise_disjoint(training_set, 2):
for v in nwise(notes, self.order + 1):
xx = self._encode_network_input(v[:self.order],
changes.unique(v[-1].beat - 1, self.chord_radius),
changes)
if not x:
x = [[] for _ in xx]
for xi, xxi in zip(x, xx):
xi.append(xxi)
p.append(encode_int(v[-1].pitch, 127))
t.append(encode_int(v[-1].ticks_since_beat_quantised, self.maxtsbq + 1))
d.append(encode_int(v[-1].duration_quantised, self.maxdq + 1))
return [np.array(xi, dtype=bool) for xi in x], \
[np.array(p, dtype=bool), np.array(t, dtype=bool), np.array(d, dtype=bool)]
def _all_training_data(self,
training_set: Iterable[Union[List[Note],
ChordProgression]]):
""" 1-of-N binary encoding of all pairs of inputs (past notes and current chord) and outputs (next note)
on the training set """
x, p, t, d = [], [], [], []
for notes, changes in nwise_disjoint(training_set, 2):
for v in nwise(notes, self.order + 1):
xx = self._encode_network_input(
v[:self.order],
changes.unique(v[-1].beat - 1, self.chord_radius), changes)
if not x:
x = [[] for _ in xx]
for xi, xxi in zip(x, xx):
xi.append(xxi)
p.append(encode_int(v[-1].pitch, 127))
t.append(
encode_int(v[-1].ticks_since_beat_quantised,
self.maxtsbq + 1))
d.append(encode_int(v[-1].duration_quantised, self.maxdq + 1))
return [np.array(xi, dtype=bool) for xi in x], \
[np.array(p, dtype=bool), np.array(t, dtype=bool), np.array(d, dtype=bool)]
def _all_training_data(self,
training_set: Iterable[Union[List[Note],
ChordProgression]]):
msg_callback = self.progressbar.set_text \
if isinstance(self.progressbar, gui.viewmodel.KerasProgressbar) \
else print
msg_callback('Processing training data...')
len_data = sum(
len(notes) - self.order + 1 for notes in training_set[::2])
if self.progressbar is None:
self.progressbar = keras.utils.Progbar(len_data)
else:
self.progressbar.set_params({'samples': len_data})
x, p, t, d, o, b = [], [], [], [], [], []
for notes, changes in nwise_disjoint(training_set, 2):
for v in nwise(notes, self.order + 1):
self.progressbar.add(1)
xx = self._encode_network_input(
v[:self.order],
changes.unique(v[-1].beat - 1, self.chord_radius), changes)
if not x:
x = [[] for _ in xx]
for xi, xxi in zip(x, xx):
xi.append(xxi)
p.append(encode_int(v[-1].abcnote.value, 12))
t.append(
encode_int(v[-1].ticks_since_beat_quantised,
self.maxtsbq + 1))
d.append(encode_int(v[-1].duration_quantised, self.maxdq + 1))
o.append(encode_int(v[-1].octave, NUM_OCTAVES))
b.append(
encode_int(v[-1].beat - v[-2].beat, self.maxbeatdiff + 1))
self.progressbar.update(self.progressbar.target, force=True)
msg_callback('Starting training...')
return [np.array(xi, dtype=bool) for xi in x],\
[np.array(arr, dtype=bool) for arr in [p, t, d, o, b]]
def learn(self, *training_set: Union[List[Note], ChordProgression], epochs=None, callback=None):
self.maxtsbq = max(n.ticks_since_beat_quantised for notes, _ in nwise_disjoint(training_set, 2) for n in notes)
self.maxdq = max(n.duration_quantised for notes, changes in nwise_disjoint(training_set, 2) for n in notes)
self.maxbeatdiff = max(m.beat - n.beat
for notes, _ in nwise_disjoint(training_set, 2) for n, m in nwise(notes, 2))
self.model = self._build_net()
self.pitch_model = keras.models.Model(inputs=self.model.inputs, outputs=self.model.outputs[0])
self.tsbq_model = keras.models.Model(inputs=self.model.inputs, outputs=self.model.outputs[1])
self.dq_model = keras.models.Model(inputs=self.model.inputs, outputs=self.model.outputs[2])
x, y = self._all_training_data(training_set)
stateful = any(isinstance(layer, keras.layers.Recurrent) and layer.stateful for layer in self.model.layers)
kwargs = {'batch_size': 1, 'shuffle': False} if stateful else {}
if callback:
kwargs.update({'callbacks': [callback], 'verbose': 0})
self.model.fit(x, y, epochs=epochs or self.epochs, **kwargs)
return not self.model.stop_training
def back_to_the_future():
""" Find out how common it is that a later note has a smaller tick_abs """
song = 'Eb_therewill'
notes = notes_from_file(r"input/{}.mid".format(song))
print(sum(n.tick_abs > m.tick_abs for n, m in nwise(notes, 2)))
def back_to_the_future():
""" Find out how common it is that a later note has a smaller tick_abs """
song = 'Eb_therewill'
notes = notes_from_file(r"input/{}.mid".format(song))
print(sum(n.tick_abs > m.tick_abs for n, m in nwise(notes, 2)))