def _quantized_subsequence_to_tensors(self, quantized_subsequence):
# Reject sequences with out-of-range pitches.
if any(note.pitch < self._min_pitch or note.pitch > self._max_pitch
for note in quantized_subsequence.notes):
return [], []
# Extract all instruments.
tracks, _ = mm.extract_performances(
quantized_subsequence,
max_steps_truncate=self._max_steps_truncate,
num_velocity_bins=self._num_velocity_bins,
split_instruments=True)
# Reject sequences with too few instruments.
if not (self._min_num_instruments <= len(tracks) <=
self._max_num_instruments):
return [], []
# Sort tracks by program, with drums at the end.
tracks = sorted(tracks, key=lambda t: (t.is_drum, t.program))
chunk_size_steps = self._steps_per_bar * self._chunk_size_bars
chunks = [[] for _ in range(self._max_num_chunks)]
total_length = 0
for track in tracks:
# Make sure the track is the proper number of time steps.
track.set_length(self._max_steps_truncate)
# Split this track into chunks.
def new_performance(quantized_sequence, start_step, track=track):
return performance_lib.MetricPerformance(
quantized_sequence=quantized_sequence,
steps_per_quarter=(self._steps_per_quarter if
quantized_sequence is None else None),
start_step=start_step,
num_velocity_bins=self._num_velocity_bins,
program=track.program,
is_drum=track.is_drum)
track_chunks = split_performance(track,
chunk_size_steps,
new_performance,
clip_tied_notes=True)
assert len(track_chunks) == self._max_num_chunks
track_chunk_lengths = [
len(track_chunk) for track_chunk in track_chunks
]
# Each track chunk needs room for program token and end token.
if not all(l <= self._max_events_per_instrument - 2
for l in track_chunk_lengths):
return [], []
if not all(mm.MIN_MIDI_PROGRAM <= t.program <= mm.MAX_MIDI_PROGRAM
for t in track_chunks if not t.is_drum):
return [], []
total_length += sum(track_chunk_lengths)
# Aggregate by chunk.
for i, track_chunk in enumerate(track_chunks):
chunks[i].append(track_chunk)
# Reject sequences that are too short (in events).
if total_length < self._min_total_events:
return [], []
num_programs = mm.MAX_MIDI_PROGRAM - mm.MIN_MIDI_PROGRAM + 1
chunk_tensors = []
chunk_chord_tensors = []
for chunk_tracks in chunks:
track_tensors = []
for track in chunk_tracks:
# Add a special token for program at the beginning of each track.
track_tokens = [
self._performance_encoding.num_classes +
(num_programs if track.is_drum else track.program)
]
# Then encode the performance events.
for event in track:
track_tokens.append(
self._performance_encoding.encode_event(event))
# Then add the end token.
track_tokens.append(self.end_token)
encoded_track = data.np_onehot(track_tokens, self.output_depth,
self.output_dtype)
track_tensors.append(encoded_track)
if self._chord_encoding:
# Extract corresponding chords for each track. The chord sequences may
# be different for different tracks even though the underlying chords
# are the same, as the performance event times will generally be
# different.
try:
track_chords = chords_lib.event_list_chords(
quantized_subsequence, chunk_tracks)
except chords_lib.CoincidentChordsException:
return [], []
track_chord_tensors = []
try:
# Chord encoding for all tracks is inside this try block. If any
# track fails we need to skip the whole subsequence.
for chords in track_chords:
# Start with a pad token corresponding to the track program token.
track_chord_tokens = [self._control_pad_token]
# Then encode the chords.
for chord in chords:
track_chord_tokens.append(
self._chord_encoding.encode_event(chord))
# Then repeat the final chord for the track end token.
track_chord_tokens.append(track_chord_tokens[-1])
encoded_track_chords = data.np_onehot(
track_chord_tokens, self.control_depth,
self.control_dtype)
track_chord_tensors.append(encoded_track_chords)
except (mm.ChordSymbolException, mm.ChordEncodingException):
return [], []
chunk_chord_tensors.append(track_chord_tensors)
chunk_tensors.append(track_tensors)
return chunk_tensors, chunk_chord_tensors
def _generate(self, input_sequence, generator_options):
if len(generator_options.input_sections) > 1:
raise mm.SequenceGeneratorException(
'This model supports at most one input_sections message, but got %s'
% len(generator_options.input_sections))
if len(generator_options.generate_sections) != 1:
raise mm.SequenceGeneratorException(
'This model supports only 1 generate_sections message, but got %s'
% len(generator_options.generate_sections))
generate_section = generator_options.generate_sections[0]
if generator_options.input_sections:
input_section = generator_options.input_sections[0]
primer_sequence = mm.trim_note_sequence(input_sequence,
input_section.start_time,
input_section.end_time)
input_start_step = mm.quantize_to_step(input_section.start_time,
self.steps_per_second,
quantize_cutoff=0.0)
else:
primer_sequence = input_sequence
input_start_step = 0
last_end_time = (max(
n.end_time
for n in primer_sequence.notes) if primer_sequence.notes else 0)
if last_end_time > generate_section.start_time:
raise mm.SequenceGeneratorException(
'Got GenerateSection request for section that is before or equal to '
'the end of the NoteSequence. This model can only extend sequences. '
'Requested start time: %s, Final note end time: %s' %
(generate_section.start_time, last_end_time))
# Quantize the priming sequence.
quantized_primer_sequence = mm.quantize_note_sequence_absolute(
primer_sequence, self.steps_per_second)
extracted_perfs, _ = mm.extract_performances(
quantized_primer_sequence,
start_step=input_start_step,
num_velocity_bins=self.num_velocity_bins)
assert len(extracted_perfs) <= 1
generate_start_step = mm.quantize_to_step(generate_section.start_time,
self.steps_per_second,
quantize_cutoff=0.0)
# Note that when quantizing end_step, we set quantize_cutoff to 1.0 so it
# always rounds down. This avoids generating a sequence that ends at 5.0
# seconds when the requested end time is 4.99.
generate_end_step = mm.quantize_to_step(generate_section.end_time,
self.steps_per_second,
quantize_cutoff=1.0)
if extracted_perfs and extracted_perfs[0]:
performance = extracted_perfs[0]
else:
# If no track could be extracted, create an empty track that starts at the
# requested generate_start_step.
performance = mm.Performance(
steps_per_second=(quantized_primer_sequence.quantization_info.
steps_per_second),
start_step=generate_start_step,
num_velocity_bins=self.num_velocity_bins)
# Ensure that the track extends up to the step we want to start generating.
performance.set_length(generate_start_step - performance.start_step)
# Extract generation arguments from generator options.
arg_types = {
'note_density': lambda arg: ast.literal_eval(arg.string_value),
'pitch_histogram': lambda arg: ast.literal_eval(arg.string_value),
'disable_conditioning':
lambda arg: ast.literal_eval(arg.string_value),
'temperature': lambda arg: arg.float_value,
'beam_size': lambda arg: arg.int_value,
'branch_factor': lambda arg: arg.int_value,
'steps_per_iteration': lambda arg: arg.int_value
}
args = dict((name, value_fn(generator_options.args[name]))
for name, value_fn in arg_types.items()
if name in generator_options.args)
# Make sure note density is present when conditioning on it and not present
# otherwise.
if not self.note_density_conditioning and 'note_density' in args:
tf.logging.warning(
'Not conditioning on note density, ignoring requested density.'
)
del args['note_density']
if self.note_density_conditioning and 'note_density' not in args:
tf.logging.warning(
'Conditioning on note density but none requested, using default.'
)
args['note_density'] = [DEFAULT_NOTE_DENSITY]
# Make sure pitch class histogram is present when conditioning on it and not
# present otherwise.
if not self.pitch_histogram_conditioning and 'pitch_histogram' in args:
tf.logging.warning(
'Not conditioning on pitch histogram, ignoring requested histogram.'
)
del args['pitch_histogram']
if self.pitch_histogram_conditioning and 'pitch_histogram' not in args:
tf.logging.warning(
'Conditioning on pitch histogram but none requested, using default.'
)
args['pitch_histogram'] = [DEFAULT_PITCH_HISTOGRAM]
# Make sure disable conditioning flag is present when conditioning is
# optional and not present otherwise.
if not self.optional_conditioning and 'disable_conditioning' in args:
tf.logging.warning(
'No optional conditioning, ignoring disable conditioning flag.'
)
del args['disable_conditioning']
if self.optional_conditioning and 'disable_conditioning' not in args:
args['disable_conditioning'] = [False]
# If a single note density, pitch class histogram, or disable flag is
# present, convert to list to simplify further processing.
if (self.note_density_conditioning
and not isinstance(args['note_density'], list)):
args['note_density'] = [args['note_density']]
if (self.pitch_histogram_conditioning
and not isinstance(args['pitch_histogram'][0], list)):
args['pitch_histogram'] = [args['pitch_histogram']]
if (self.optional_conditioning
and not isinstance(args['disable_conditioning'], list)):
args['disable_conditioning'] = [args['disable_conditioning']]
# Make sure each pitch class histogram sums to one.
if self.pitch_histogram_conditioning:
for i in range(len(args['pitch_histogram'])):
total = sum(args['pitch_histogram'][i])
if total > 0:
args['pitch_histogram'][i] = [
float(count) / total
for count in args['pitch_histogram'][i]
]
else:
tf.logging.warning(
'Pitch histogram is empty, using default.')
args['pitch_histogram'][i] = DEFAULT_PITCH_HISTOGRAM
total_steps = performance.num_steps + (generate_end_step -
generate_start_step)
# Set up functions that map generation step to note density, pitch
# histogram, and disable conditioning flag.
mean_note_density = DEFAULT_NOTE_DENSITY
if self.note_density_conditioning:
args['note_density_fn'] = partial(
_step_to_note_density,
num_steps=total_steps,
note_densities=args['note_density'])
mean_note_density = sum(args['note_density']) / len(
args['note_density'])
del args['note_density']
if self.pitch_histogram_conditioning:
args['pitch_histogram_fn'] = partial(
_step_to_pitch_histogram,
num_steps=total_steps,
pitch_histograms=args['pitch_histogram'])
del args['pitch_histogram']
if self.optional_conditioning:
args['disable_conditioning_fn'] = partial(
_step_to_disable_conditioning,
num_steps=total_steps,
disable_conditioning_flags=args['disable_conditioning'])
del args['disable_conditioning']
if not performance:
# Primer is empty; let's just start with silence.
performance.set_length(
min(performance.max_shift_steps, total_steps))
while performance.num_steps < total_steps:
# Assume the average specified (or default) note density and 4 RNN steps
# per note. Can't know for sure until generation is finished because the
# number of notes per quantized step is variable.
note_density = max(1.0, mean_note_density)
steps_to_gen = total_steps - performance.num_steps
rnn_steps_to_gen = int(
math.ceil(4.0 * note_density * steps_to_gen /
self.steps_per_second))
tf.logging.info(
'Need to generate %d more steps for this sequence, will try asking '
'for %d RNN steps' % (steps_to_gen, rnn_steps_to_gen))
performance = self._model.generate_performance(
len(performance) + rnn_steps_to_gen, performance, **args)
if not self.fill_generate_section:
# In the interest of speed just go through this loop once, which may not
# entirely fill the generate section.
break
performance.set_length(total_steps)
generated_sequence = performance.to_sequence(
max_note_duration=self.max_note_duration)
assert (generated_sequence.total_time -
generate_section.end_time) <= 1e-5
return generated_sequence
def _generate(self, input_sequence, generator_options):
if len(generator_options.input_sections) > 1:
raise mm.SequenceGeneratorError(
'This model supports at most one input_sections message, but got %s' %
len(generator_options.input_sections))
if len(generator_options.generate_sections) != 1:
raise mm.SequenceGeneratorError(
'This model supports only 1 generate_sections message, but got %s' %
len(generator_options.generate_sections))
generate_section = generator_options.generate_sections[0]
if generator_options.input_sections:
input_section = generator_options.input_sections[0]
primer_sequence = mm.trim_note_sequence(
input_sequence, input_section.start_time, input_section.end_time)
input_start_step = mm.quantize_to_step(
input_section.start_time, self.steps_per_second, quantize_cutoff=0.0)
else:
primer_sequence = input_sequence
input_start_step = 0
if primer_sequence.notes:
last_end_time = max(n.end_time for n in primer_sequence.notes)
else:
last_end_time = 0
if last_end_time > generate_section.start_time:
raise mm.SequenceGeneratorError(
'Got GenerateSection request for section that is before or equal to '
'the end of the NoteSequence. This model can only extend sequences. '
'Requested start time: %s, Final note end time: %s' %
(generate_section.start_time, last_end_time))
# Quantize the priming sequence.
quantized_primer_sequence = mm.quantize_note_sequence_absolute(
primer_sequence, self.steps_per_second)
extracted_perfs, _ = mm.extract_performances(
quantized_primer_sequence, start_step=input_start_step,
num_velocity_bins=self.num_velocity_bins,
note_performance=self._note_performance)
assert len(extracted_perfs) <= 1
generate_start_step = mm.quantize_to_step(
generate_section.start_time, self.steps_per_second, quantize_cutoff=0.0)
# Note that when quantizing end_step, we set quantize_cutoff to 1.0 so it
# always rounds down. This avoids generating a sequence that ends at 5.0
# seconds when the requested end time is 4.99.
generate_end_step = mm.quantize_to_step(
generate_section.end_time, self.steps_per_second, quantize_cutoff=1.0)
if extracted_perfs and extracted_perfs[0]:
performance = extracted_perfs[0]
else:
# If no track could be extracted, create an empty track that starts at the
# requested generate_start_step.
performance = mm.Performance(
steps_per_second=(
quantized_primer_sequence.quantization_info.steps_per_second),
start_step=generate_start_step,
num_velocity_bins=self.num_velocity_bins)
# Ensure that the track extends up to the step we want to start generating.
performance.set_length(generate_start_step - performance.start_step)
# Extract generation arguments from generator options.
arg_types = {
'disable_conditioning': lambda arg: ast.literal_eval(arg.string_value),
'temperature': lambda arg: arg.float_value,
'beam_size': lambda arg: arg.int_value,
'branch_factor': lambda arg: arg.int_value,
'steps_per_iteration': lambda arg: arg.int_value
}
if self.control_signals:
for control in self.control_signals:
arg_types[control.name] = lambda arg: ast.literal_eval(arg.string_value)
args = dict((name, value_fn(generator_options.args[name]))
for name, value_fn in arg_types.items()
if name in generator_options.args)
# Make sure control signals are present and convert to lists if necessary.
if self.control_signals:
for control in self.control_signals:
if control.name not in args:
tf.logging.warning(
'Control value not specified, using default: %s = %s',
control.name, control.default_value)
args[control.name] = [control.default_value]
elif control.validate(args[control.name]):
args[control.name] = [args[control.name]]
else:
if not isinstance(args[control.name], list) or not all(
control.validate(value) for value in args[control.name]):
tf.logging.fatal(
'Invalid control value: %s = %s',
control.name, args[control.name])
# Make sure disable conditioning flag is present when conditioning is
# optional and convert to list if necessary.
if self.optional_conditioning:
if 'disable_conditioning' not in args:
args['disable_conditioning'] = [False]
elif isinstance(args['disable_conditioning'], bool):
args['disable_conditioning'] = [args['disable_conditioning']]
else:
if not isinstance(args['disable_conditioning'], list) or not all(
isinstance(value, bool) for value in args['disable_conditioning']):
tf.logging.fatal(
'Invalid disable_conditioning value: %s',
args['disable_conditioning'])
total_steps = performance.num_steps + (
generate_end_step - generate_start_step)
if 'notes_per_second' in args:
mean_note_density = (
sum(args['notes_per_second']) / len(args['notes_per_second']))
else:
mean_note_density = DEFAULT_NOTE_DENSITY
# Set up functions that map generation step to control signal values and
# disable conditioning flag.
if self.control_signals:
control_signal_fns = []
for control in self.control_signals:
control_signal_fns.append(functools.partial(
_step_to_value,
num_steps=total_steps,
values=args[control.name]))
del args[control.name]
args['control_signal_fns'] = control_signal_fns
if self.optional_conditioning:
args['disable_conditioning_fn'] = functools.partial(
_step_to_value,
num_steps=total_steps,
values=args['disable_conditioning'])
del args['disable_conditioning']
if not performance:
# Primer is empty; let's just start with silence.
performance.set_length(min(performance.max_shift_steps, total_steps))
while performance.num_steps < total_steps:
# Assume the average specified (or default) note density and 4 RNN steps
# per note. Can't know for sure until generation is finished because the
# number of notes per quantized step is variable.
note_density = max(1.0, mean_note_density)
steps_to_gen = total_steps - performance.num_steps
rnn_steps_to_gen = int(math.ceil(
4.0 * note_density * steps_to_gen / self.steps_per_second))
tf.logging.info(
'Need to generate %d more steps for this sequence, will try asking '
'for %d RNN steps' % (steps_to_gen, rnn_steps_to_gen))
performance = self._model.generate_performance(
len(performance) + rnn_steps_to_gen, performance, **args)
if not self.fill_generate_section:
# In the interest of speed just go through this loop once, which may not
# entirely fill the generate section.
break
performance.set_length(total_steps)
generated_sequence = performance.to_sequence(
max_note_duration=self.max_note_duration)
assert (generated_sequence.total_time - generate_section.end_time) <= 1e-5
return generated_sequence
def _quantized_subsequence_to_tensors(self, quantized_subsequence):
# Reject sequences with out-of-range pitches.
if any(note.pitch < self._min_pitch or note.pitch > self._max_pitch
for note in quantized_subsequence.notes):
return [], []
# Extract all instruments.
tracks, _ = mm.extract_performances(
quantized_subsequence,
max_steps_truncate=self._max_steps_truncate,
num_velocity_bins=self._num_velocity_bins,
split_instruments=True)
# Reject sequences with too few instruments.
if not (self._min_num_instruments <= len(tracks) <=
self._max_num_instruments):
return [], []
# Sort tracks by program, with drums at the end.
tracks = sorted(tracks, key=lambda t: (t.is_drum, t.program))
chunk_size_steps = self._steps_per_bar * self._chunk_size_bars
chunks = [[] for _ in range(self._max_num_chunks)]
total_length = 0
for track in tracks:
# Make sure the track is the proper number of time steps.
track.set_length(self._max_steps_truncate)
# Split this track into chunks.
def new_performance(quantized_sequence, start_step, track=track):
steps_per_quarter = (
self._steps_per_quarter if quantized_sequence is None else None)
return performance_lib.MetricPerformance(
quantized_sequence=quantized_sequence,
steps_per_quarter=steps_per_quarter,
start_step=start_step,
num_velocity_bins=self._num_velocity_bins,
program=track.program, is_drum=track.is_drum)
track_chunks = split_performance(
track, chunk_size_steps, new_performance, clip_tied_notes=True)
assert len(track_chunks) == self._max_num_chunks
track_chunk_lengths = [len(track_chunk) for track_chunk in track_chunks]
# Each track chunk needs room for program token and end token.
if not all(l <= self._max_events_per_instrument - 2
for l in track_chunk_lengths):
return [], []
if not all(mm.MIN_MIDI_PROGRAM <= t.program <= mm.MAX_MIDI_PROGRAM
for t in track_chunks if not t.is_drum):
return [], []
total_length += sum(track_chunk_lengths)
# Aggregate by chunk.
for i, track_chunk in enumerate(track_chunks):
chunks[i].append(track_chunk)
# Reject sequences that are too short (in events).
if total_length < self._min_total_events:
return [], []
num_programs = mm.MAX_MIDI_PROGRAM - mm.MIN_MIDI_PROGRAM + 1
chunk_tensors = []
chunk_chord_tensors = []
for chunk_tracks in chunks:
track_tensors = []
for track in chunk_tracks:
# Add a special token for program at the beginning of each track.
track_tokens = [self._performance_encoding.num_classes + (
num_programs if track.is_drum else track.program)]
# Then encode the performance events.
for event in track:
track_tokens.append(self._performance_encoding.encode_event(event))
# Then add the end token.
track_tokens.append(self.end_token)
encoded_track = data.np_onehot(
track_tokens, self.output_depth, self.output_dtype)
track_tensors.append(encoded_track)
if self._chord_encoding:
# Extract corresponding chords for each track. The chord sequences may
# be different for different tracks even though the underlying chords
# are the same, as the performance event times will generally be
# different.
try:
track_chords = chords_lib.event_list_chords(
quantized_subsequence, chunk_tracks)
except chords_lib.CoincidentChordsError:
return [], []
track_chord_tensors = []
try:
# Chord encoding for all tracks is inside this try block. If any
# track fails we need to skip the whole subsequence.
for chords in track_chords:
# Start with a pad token corresponding to the track program token.
track_chord_tokens = [self._control_pad_token]
# Then encode the chords.
for chord in chords:
track_chord_tokens.append(
self._chord_encoding.encode_event(chord))
# Then repeat the final chord for the track end token.
track_chord_tokens.append(track_chord_tokens[-1])
encoded_track_chords = data.np_onehot(
track_chord_tokens, self.control_depth, self.control_dtype)
track_chord_tensors.append(encoded_track_chords)
except (mm.ChordSymbolError, mm.ChordEncodingError):
return [], []
chunk_chord_tensors.append(track_chord_tensors)
chunk_tensors.append(track_tensors)
return chunk_tensors, chunk_chord_tensors