def _on_off_melo(self, tensor_01):
""" Convert one-song tensor to one Melo.
Tensor has to have only 0 or 1 only.
Melo returned is sorted.
Args:
tensor_01 (np.ndarray) - shape (num_of_time, num_of_features)
Output:
output_melo (Melo)
"""
output_melo = meloclass.Melo()
output_track = meloclass.Track()
output_melo.tracks.append(output_track)
for time in range(tensor_01.shape[0]):
for pitch in range(tensor_01.shape[1]):
if tensor_01[time][pitch] == 1:
if pitch % 2 == 0: # Note on event
output_track.notes.append(meloclass.NoteEvent("note_on",pitch/2+self.args.prepare_normalize_lowest,time))
else: # Note off event
output_track.notes.append(meloclass.NoteEvent("note_off",(pitch-1)/2+self.args.prepare_normalize_lowest,time))
elif tensor_01[time][pitch] != 0:
print(tensor_01[time][pitch])
raise ValueError("The tensor of tensor_of_melo cannot contain value other than 0 or 1.")
return self.melo_processor.sort_melo(output_melo)
def _convert_melo_to_relative_time(self, abs_melo):
""" Normally, Melo only uses absolute time.
However this is inconvenient to convert to MidiFile
that uses relative time. This function changes all the
NoteEvent's time to relative time (but still use smallest
unit defined by resolution instead of tick as unit of time)
NOTE that a Melo that uses relative time instead of
absolute time should never ever appear outside midimeloconverter!
Args:
abs_melo (Melo)
Output:
rel_melo (Melo) DO NOT USE THIS MELO OUTSIDE THIS CLASS.
"""
rel_melo = meloclass.Melo()
for abs_melo_track in abs_melo.tracks:
prev_time = 0
rel_melo_track = meloclass.Track()
for abs_note in abs_melo_track.notes:
rel_melo_track.notes.append(
meloclass.NoteEvent(abs_note.type, abs_note.pitch,
abs_note.time - prev_time))
prev_time = abs_note.time
rel_melo.tracks.append(rel_melo_track)
return rel_melo
def midi_to_melo(self, midifile):
""" Convert Midi to Melo.
Args:
midifile (MidiFile)
Output:
output_melo (Melo)
"""
output_melo = meloclass.Melo()
# Meaning of resolution: (1/(4*resolution))-th note.
# For example, if resolution = 1, the smallest unit of Melo
# will be quarter note. If resolution = 4, the smallest unit of
# Melo will be 16-th note.
ticks_per_smallest_unit = midifile.ticks_per_beat // self.args.prepare_resolution
for midi_track in midifile.tracks:
abs_time = 0
melo_track = meloclass.Track()
for msg in midi_track:
abs_time += msg.time # It is necessary to add MetaMessage's time too even if it is not added to Melo
if msg.type == "time_signature":
output_melo.ts_numerator = msg.numerator
output_melo.ts_denominator = msg.denominator
if not msg.is_meta and msg.channel + 1 == DRUM_CHANNEL: # ignore drums. Add 1 because mido's channel range starts from 0 while Midi standard starts from 1.
# Notice that because meta message has no attribute "channel"
# The short-circuit evaluation is needed to prevent AttributeError
continue
if msg.type == "note_on" or msg.type == "note_off":
melo_track.notes.append(
meloclass.NoteEvent(
str(msg.type), msg.note,
abs_time // ticks_per_smallest_unit))
output_melo.tracks.append(melo_track)
return output_melo
def normalize_melo(self, bound_melo):
""" Transpose one-track Melo so that its lowest note
will be around args.normalize_lowest. The transpose
will try to make the new Melo as close to C major or
A minor as possible by minimizing number of black keys.
Any note higher than args.normalize_lowest + args.normalize_range
will be removed (not to extract a melody, which was done
by raw_melo_to_melody_melo() already, but to restrict
the number of features of input of neural network)
Args:
bound_melo (Melo)
Return:
normalized_melo (Melo)
args.normalize_lowest
args.normalize_range
"""
# TODO: This is not a good transposing algorithm
# Using advanced key detection algorithm can
# help to perform better normalization.
if len(bound_melo.tracks) != 1:
raise ValueError(
"Do not call normalize_melo on Melo that does not have exactly one track. This Melo has %i tracks."
% len(bound_melo.tracks))
original_pitch_set = set([n.pitch for n in bound_melo.tracks[0].notes])
start_trans = min(
original_pitch_set) - self.args.prepare_normalize_lowest
start_pitch_set = set([p - start_trans for p in original_pitch_set])
min_black_keys = self._count_black_keys(start_pitch_set)
min_pitch_set = start_pitch_set
for i in range(1, 12):
pitch_set = set([p + i for p in start_pitch_set])
black_keys = self._count_black_keys(pitch_set)
if black_keys < min_black_keys:
min_black_keys = black_keys
min_pitch_set = pitch_set
final_trans = min(original_pitch_set) - min(min_pitch_set)
normalized_melo = meloclass.Melo()
normalized_track = meloclass.Track()
normalized_melo.tracks.append(normalized_track)
normalized_track.notes = [
meloclass.NoteEvent(n.type, n.pitch - final_trans, n.time)
for n in bound_melo.tracks[0].notes
if self.args.prepare_normalize_lowest <= n.pitch -
final_trans <= self.args.prepare_normalize_lowest +
self.args.prepare_normalize_range
]
return normalized_melo
def _hold_on_melo(self, tensor_01):
""" Convert one-song tensor to one Melo.
Tensor has to ahve only 0 or 1 only.
Melo returned is sorted.
Args:
tensor_01 (np.ndarray) - shape (num_of_time, num_of_features)
Output:
output_melo (Melo)
"""
output_melo = meloclass.Melo()
output_track = meloclass.Track()
output_melo.tracks.append(output_track)
for time in range(tensor_01.shape[0]):
for pitch in range(tensor_01.shape[1]):
if pitch % 2 == 0 and tensor_01[time][pitch] == 1:
on_note = meloclass.NoteEvent("note_on",pitch/2+self.args.prepare_normalize_lowest,time)
output_track.notes.append(on_note)
# Find note_off time
off_note_found = False
off_time = None
j = time + 1
while not off_note_found and j < tensor_01.shape[0]:
if tensor_01[j][pitch+1] == 0:
off_time = j
off_note_found = True
j += 1
# when examining a subsong, it is possile that
# an off_note cannot be found within range
# in that case, simply set j to be the last time
if j == tensor_01.shape[0]:
j = tensor_01.shape[0] - 1
off_time = j
off_note = meloclass.NoteEvent("note_off",pitch//2+self.args.prepare_normalize_lowest,off_time)
output_track.notes.append(off_note)
elif tensor_01[time][pitch] != 0 and tensor_01[time][pitch] != 1:
raise ValueError("The tensor of tensor_of_melo cannot contain value other than 0 or 1.")
return self.melo_processor.sort_melo(output_melo)