def decode_to_note_sequence(self, ids, strip_extraneous=False):
"""Transform a sequence of event indices into a performance NoteSequence.
Args:
ids: List of performance event indices.
strip_extraneous: Whether to strip EOS and padding from the end of `ids`.
Returns:
A NoteSequence.
"""
if strip_extraneous:
ids = text_encoder.strip_ids(ids,
list(range(self.num_reserved_ids)))
# Decode indices corresponding to event n-grams back into the n-grams.
event_ids = []
for i in ids:
if i >= self.unigram_vocab_size:
event_ids += self._ngrams[i - self.unigram_vocab_size]
else:
event_ids.append(i)
performance = note_seq.Performance(
quantized_sequence=None,
steps_per_second=self._steps_per_second,
num_velocity_bins=self._num_velocity_bins)
for i in event_ids:
performance.append(
self._encoding.decode_event(i - self.num_reserved_ids))
ns = performance.to_sequence()
return ns
def encode_note_sequence(self, ns):
performance = note_seq.Performance(
note_seq.quantize_note_sequence_absolute(ns, self.midi_encoder._steps_per_second),
num_velocity_bins=self.midi_encoder._num_velocity_bins)
event_ids = [self.midi_encoder._encoding.encode_event(event) +
self.midi_encoder.num_reserved_ids
for event in performance]
# Greedily encode performance event n-grams as new indices.
ids = []
j = 0
while j < len(event_ids):
ngram = ()
best_ngram = None
for i in range(j, len(event_ids)):
ngram += (event_ids[i],)
if self.midi_encoder._ngrams_trie.has_key(ngram):
best_ngram = ngram
if not self.midi_encoder._ngrams_trie.has_subtrie(ngram):
break
if best_ngram is not None:
ids.append(self.midi_encoder._ngrams_trie[best_ngram])
j += len(best_ngram)
else:
j += 1
if self.midi_encoder._add_eos:
ids.append(text_encoder.EOS_ID)
return ids
def encode_note_sequence(self, ns):
"""Transform a NoteSequence into a list of performance event indices.
Args:
ns: NoteSequence proto containing the performance to encode.
Returns:
ids: List of performance event indices.
"""
performance = note_seq.Performance(
note_seq.quantize_note_sequence_absolute(ns, self._steps_per_second),
num_velocity_bins=self._num_velocity_bins)
event_ids = [self._encoding.encode_event(event) + self.num_reserved_ids
for event in performance]
# Greedily encode performance event n-grams as new indices.
ids = []
j = 0
while j < len(event_ids):
ngram = ()
for i in range(j, len(event_ids)):
ngram += (event_ids[i],)
if self._ngrams_trie.has_key(ngram):
best_ngram = ngram
if not self._ngrams_trie.has_subtrie(ngram):
break
ids.append(self._ngrams_trie[best_ngram])
j += len(best_ngram)
if self._add_eos:
ids.append(text_encoder.EOS_ID)
return ids
def decode(self, ids, strip_extraneous=False):
"""Transform a sequence of event indices into a performance MIDI file.
Args:
ids: List of performance event indices.
strip_extraneous: Whether to strip EOS and padding from the end of `ids`.
Returns:
Path to the temporary file where the MIDI was saved.
"""
if strip_extraneous:
ids = text_encoder.strip_ids(ids, list(range(self.num_reserved_ids)))
# Decode indices corresponding to event n-grams back into the n-grams.
event_ids = []
for i in ids:
if i >= self.unigram_vocab_size:
event_ids += self._ngrams[i - self.unigram_vocab_size]
else:
event_ids.append(i)
performance = note_seq.Performance(
quantized_sequence=None,
steps_per_second=self._steps_per_second,
num_velocity_bins=self._num_velocity_bins)
for i in event_ids:
performance.append(self._encoding.decode_event(i - self.num_reserved_ids))
ns = performance.to_sequence()
_, tmp_file_path = tempfile.mkstemp('_decode.mid')
note_seq.sequence_proto_to_midi_file(ns, tmp_file_path)
return tmp_file_path
def decode_ids(self, ids):
event_ids = []
for i in ids:
if i >= self.midi_encoder.unigram_vocab_size:
event_ids += self.midi_encoder._ngrams[i - self.midi_encoder.unigram_vocab_size]
else:
event_ids.append(i)
performance = note_seq.Performance(
quantized_sequence=None,
steps_per_second=self.midi_encoder._steps_per_second,
num_velocity_bins=self.midi_encoder._num_velocity_bins
)
for i in event_ids:
if i >= self.midi_encoder.num_reserved_ids:
performance.append(self.midi_encoder._encoding.decode_event(i - self.midi_encoder.num_reserved_ids))
return performance.to_sequence()
def _generate(self, input_sequence, generator_options):
if len(generator_options.input_sections) > 1:
raise sequence_generator.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 sequence_generator.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 = note_seq.trim_note_sequence(
input_sequence, input_section.start_time,
input_section.end_time)
input_start_step = note_seq.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 sequence_generator.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 = note_seq.quantize_note_sequence_absolute(
primer_sequence, self.steps_per_second)
extracted_perfs, _ = performance_pipeline.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 = note_seq.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 = note_seq.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 = note_seq.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