def _to_tensors(self, note_sequence): """Converts NoteSequence to unique, one-hot tensor sequences.""" try: if self._steps_per_quarter: quantized_sequence = mm.quantize_note_sequence( note_sequence, self._steps_per_quarter) if (mm.steps_per_bar_in_quantized_sequence(quantized_sequence) != self._steps_per_bar): return [], [] else: quantized_sequence = mm.quantize_note_sequence_absolute( note_sequence, self._steps_per_second) except (mm.BadTimeSignatureException, mm.NonIntegerStepsPerBarException, mm.NegativeTimeException) as e: return [], [] event_lists, unused_stats = self._event_extractor_fn( quantized_sequence) if self._pad_to_total_time: for e in event_lists: e.set_length(len(e) + e.start_step, from_left=True) e.set_length(quantized_sequence.total_quantized_steps) if self._slice_steps: sliced_event_tuples = [] for l in event_lists: for i in range(self._slice_steps, len(l) + 1, self._steps_per_bar): sliced_event_tuples.append( tuple(l[i - self._slice_steps:i])) else: sliced_event_tuples = [tuple(l) for l in event_lists] # TODO(adarob): Consider handling the fact that different event lists can # be mapped to identical tensors by the encoder_decoder (e.g., Drums). unique_event_tuples = list(set(sliced_event_tuples)) unique_event_tuples = self._maybe_sample_outputs(unique_event_tuples) seqs = [] for t in unique_event_tuples: seqs.append( np_onehot( [self._legacy_encoder_decoder.encode_event(e) for e in t] + ([] if self.end_token is None else [self.end_token]), self.output_depth, self.output_dtype)) return seqs, seqs
def _to_tensors(self, note_sequence): """Converts NoteSequence to unique, one-hot tensor sequences.""" try: if self._steps_per_quarter: quantized_sequence = mm.quantize_note_sequence( note_sequence, self._steps_per_quarter) if (mm.steps_per_bar_in_quantized_sequence(quantized_sequence) != self._steps_per_bar): return [], [] else: quantized_sequence = mm.quantize_note_sequence_absolute( note_sequence, self._steps_per_second) except (mm.BadTimeSignatureException, mm.NonIntegerStepsPerBarException, mm.NegativeTimeException) as e: return [], [] event_lists, unused_stats = self._event_extractor_fn(quantized_sequence) if self._pad_to_total_time: for e in event_lists: e.set_length(len(e) + e.start_step, from_left=True) e.set_length(quantized_sequence.total_quantized_steps) if self._slice_steps: sliced_event_tuples = [] for l in event_lists: for i in range(self._slice_steps, len(l) + 1, self._steps_per_bar): sliced_event_tuples.append(tuple(l[i - self._slice_steps: i])) else: sliced_event_tuples = [tuple(l) for l in event_lists] # TODO(adarob): Consider handling the fact that different event lists can # be mapped to identical tensors by the encoder_decoder (e.g., Drums). unique_event_tuples = list(set(sliced_event_tuples)) unique_event_tuples = self._maybe_sample_outputs(unique_event_tuples) seqs = [] for t in unique_event_tuples: seqs.append(np_onehot( [self._legacy_encoder_decoder.encode_event(e) for e in t] + ([] if self.end_token is None else [self.end_token]), self.output_depth, self.output_dtype)) return seqs, seqs
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)) qpm = (input_sequence.tempos[0].qpm if input_sequence and input_sequence.tempos else mm.DEFAULT_QUARTERS_PER_MINUTE) steps_per_second = mm.steps_per_quarter_to_steps_per_second(self.steps_per_quarter, qpm) 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, steps_per_second, quantize_cutoff=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('start time: %s, Final note end time: %s' % (generate_section.start_time, last_end_time)) quantized_sequence = mm.quantize_note_sequence(primer_sequence, self.steps_per_quarter) extracted_melodies, _ = mm.extract_melodies(quantized_sequence, search_start_step=input_start_step, min_bars=0,min_unique_pitches=1, gap_bars=float('inf'),ignore_polyphonic_notes=True) assert len(extracted_melodies) <= 1 start_step = mm.quantize_to_step( generate_section.start_time, steps_per_second, quantize_cutoff=0) end_step = mm.quantize_to_step(generate_section.end_time, steps_per_second, quantize_cutoff=1.0) if extracted_melodies and extracted_melodies[0]: melody = extracted_melodies[0] else: steps_per_bar = int(mm.steps_per_bar_in_quantized_sequence(quantized_sequence)) melody = mm.Melody([],start_step=max(0, start_step - 1),steps_per_bar=steps_per_bar,steps_per_quarter=self.steps_per_quarter) melody.set_length(start_step - melody.start_step) arg_types = { '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) generated_melody = self._model.generate_melody(end_step - melody.start_step, melody, **args) generated_sequence = generated_melody.to_sequence(qpm=qpm) assert (generated_sequence.total_time - generate_section.end_time) <= 1e-5 return generated_sequence
def process_dataset(self, dataset, conf): """ Augment, transform and tokenize each sample in <dataset> Return: list of tokenised sequences """ # Abstract to ARGS at some point quantize = conf['quantize'] steps_per_quarter = conf['steps_per_quarter'] filter_4_4 = conf['filter_4_4'] # maybe we dont want this? # To midi note sequence using magent dev_sequences = [mm.midi_to_note_sequence(features["midi"]) for features in dataset] if self.augment_stretch: augmented = self._augment_stretch(dev_sequences) # Tripling the total number of sequences dev_sequences = dev_sequences + augmented if quantize: dev_sequences = [self._quantize(d, steps_per_quarter) for d in dev_sequences] # Filter out those that are not in 4/4 and do not have any notes dev_sequences = [ s for s in dev_sequences if self._is_4_4(s) and len(s.notes) > 0 and s.notes[-1].quantized_end_step > mm.steps_per_bar_in_quantized_sequence(s) ] # note sequence -> [(pitch, vel_bucket, start timestep)] tokens = [self._tokenize(d, steps_per_quarter, quantize) for d in dev_sequences] if self.shuffle: np.random.shuffle(tokens) stream = self._join_token_list(tokens, n=1) return torch.tensor(stream)
def _to_tensors(self, note_sequence): """Converts NoteSequence to unique sequences.""" try: quantized_sequence = mm.quantize_note_sequence( note_sequence, self._steps_per_quarter) if (mm.steps_per_bar_in_quantized_sequence(quantized_sequence) != self._steps_per_bar): return [], [] except (mm.BadTimeSignatureException, mm.NonIntegerStepsPerBarException, mm.NegativeTimeException) as e: return [], [] new_notes = [] for n in quantized_sequence.notes: if not n.is_drum: continue if n.pitch not in self._pitch_class_map: continue n.pitch = self._pitch_class_map[n.pitch] new_notes.append(n) del quantized_sequence.notes[:] quantized_sequence.notes.extend(new_notes) event_lists, unused_stats = self._drums_extractor_fn(quantized_sequence) if self._pad_to_total_time: for e in event_lists: e.set_length(len(e) + e.start_step, from_left=True) e.set_length(quantized_sequence.total_quantized_steps) if self._slice_steps: sliced_event_tuples = [] for l in event_lists: for i in range(self._slice_steps, len(l) + 1, self._steps_per_bar): sliced_event_tuples.append(tuple(l[i - self._slice_steps: i])) else: sliced_event_tuples = [tuple(l) for l in event_lists] unique_event_tuples = list(set(sliced_event_tuples)) unique_event_tuples = self._maybe_sample_outputs(unique_event_tuples) rolls = [] oh_vecs = [] for t in unique_event_tuples: if self._roll_input or self._roll_output: if self.end_token is not None: t_roll = list(t) + [(self._pr_encoder_decoder.input_size - 1,)] else: t_roll = t rolls.append(np.vstack([ self._pr_encoder_decoder.events_to_input(t_roll, i).astype(np.bool) for i in range(len(t_roll))])) if not (self._roll_input and self._roll_output): labels = [self._oh_encoder_decoder.encode_event(e) for e in t] if self.end_token is not None: labels += [self._oh_encoder_decoder.num_classes] oh_vecs.append(np_onehot( labels, self._oh_encoder_decoder.num_classes + (self.end_token is not None), np.bool)) if self._roll_input: input_seqs = [ np.append(roll, np.expand_dims(np.all(roll == 0, axis=1), axis=1), axis=1) for roll in rolls] else: input_seqs = oh_vecs output_seqs = rolls if self._roll_output else oh_vecs return input_seqs, output_seqs
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)) qpm = (input_sequence.tempos[0].qpm if input_sequence and input_sequence.tempos else mm.DEFAULT_QUARTERS_PER_MINUTE) generate_section = generator_options.generate_sections[0] if generator_options.input_sections: # Use primer melody from input section only. Take backing chords from # beginning of input section through end of generate section. input_section = generator_options.input_sections[0] primer_sequence = mm.trim_note_sequence(input_sequence, input_section.start_time, input_section.end_time) backing_sequence = mm.trim_note_sequence(input_sequence, input_section.start_time, generate_section.end_time) input_start_step = self.seconds_to_steps(input_section.start_time, qpm) else: # No input section. Take primer melody from the beginning of the sequence # up until the start of the generate section. primer_sequence = mm.trim_note_sequence( input_sequence, 0.0, generate_section.start_time) backing_sequence = mm.trim_note_sequence(input_sequence, 0.0, generate_section.end_time) 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 input section. This model can only extend melodies. ' 'Requested start time: %s, Final note end time: %s' % (generate_section.start_time, last_end_time)) # Quantize the priming and backing sequences. quantized_primer_sequence = mm.quantize_note_sequence( primer_sequence, self._steps_per_quarter) quantized_backing_sequence = mm.quantize_note_sequence( backing_sequence, self._steps_per_quarter) # Setting gap_bars to infinite ensures that the entire input will be used. extracted_melodies, _ = mm.extract_melodies( quantized_primer_sequence, search_start_step=input_start_step, min_bars=0, min_unique_pitches=1, gap_bars=float('inf'), ignore_polyphonic_notes=True) assert len(extracted_melodies) <= 1 start_step = self.seconds_to_steps(generate_section.start_time, qpm) end_step = self.seconds_to_steps(generate_section.end_time, qpm) if extracted_melodies and extracted_melodies[0]: melody = extracted_melodies[0] else: # If no melody could be extracted, create an empty melody that starts 1 # step before the request start_step. This will result in 1 step of # silence when the melody is extended below. steps_per_bar = int( mm.steps_per_bar_in_quantized_sequence( quantized_primer_sequence)) melody = mm.Melody([], start_step=max(0, start_step - 1), steps_per_bar=steps_per_bar, steps_per_quarter=self.steps_per_quarter) extracted_chords, _ = mm.extract_chords(quantized_backing_sequence) chords = extracted_chords[0] # Make sure that chords and melody start on the same step. if chords.start_step < melody.start_step: chords.set_length( len(chords) - melody.start_step + chords.start_step) assert chords.end_step == end_step # Ensure that the melody extends up to the step we want to start generating. melody.set_length(start_step - melody.start_step) # Extract generation arguments from generator options. arg_types = { '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) generated_melody = self._model.generate_melody(melody, chords, **args) generated_lead_sheet = mm.LeadSheet(generated_melody, chords) generated_sequence = generated_lead_sheet.to_sequence(qpm=qpm) 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.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)) qpm = (input_sequence.tempos[0].qpm if input_sequence and input_sequence.tempos else mm.DEFAULT_QUARTERS_PER_MINUTE) steps_per_second = mm.steps_per_quarter_to_steps_per_second( self.steps_per_quarter, qpm) 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, steps_per_second) 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 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_sequence = mm.quantize_note_sequence(primer_sequence, self.steps_per_quarter) # Setting gap_bars to infinite ensures that the entire input will be used. extracted_melodies, _ = mm.extract_melodies( quantized_sequence, search_start_step=input_start_step, min_bars=0, min_unique_pitches=1, gap_bars=float('inf'), ignore_polyphonic_notes=True) assert len(extracted_melodies) <= 1 start_step = mm.quantize_to_step(generate_section.start_time, steps_per_second) end_step = mm.quantize_to_step(generate_section.end_time, steps_per_second) if extracted_melodies and extracted_melodies[0]: melody = extracted_melodies[0] else: # If no melody could be extracted, create an empty melody that starts 1 # step before the request start_step. This will result in 1 step of # silence when the melody is extended below. steps_per_bar = int( mm.steps_per_bar_in_quantized_sequence(quantized_sequence)) melody = mm.Melody([], start_step=max(0, start_step - 1), steps_per_bar=steps_per_bar, steps_per_quarter=self.steps_per_quarter) # Ensure that the melody extends up to the step we want to start generating. melody.set_length(start_step - melody.start_step) # Extract generation arguments from generator options. arg_types = { '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) generated_melody = self._model.generate_melody( end_step - melody.start_step, melody, **args) generated_sequence = generated_melody.to_sequence(qpm=qpm) 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.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)) qpm = (input_sequence.tempos[0].qpm if input_sequence and input_sequence.tempos else mm.DEFAULT_QUARTERS_PER_MINUTE) steps_per_second = mm.steps_per_quarter_to_steps_per_second( self.steps_per_quarter, qpm) 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, steps_per_second, quantize_cutoff=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 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_sequence = mm.quantize_note_sequence( primer_sequence, self.steps_per_quarter) # Setting gap_bars to infinite ensures that the entire input will be used. extracted_melodies, _ = mm.extract_melodies( quantized_sequence, search_start_step=input_start_step, min_bars=0, min_unique_pitches=1, gap_bars=float('inf'), ignore_polyphonic_notes=True) assert len(extracted_melodies) <= 1 start_step = mm.quantize_to_step( generate_section.start_time, steps_per_second, quantize_cutoff=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. end_step = mm.quantize_to_step( generate_section.end_time, steps_per_second, quantize_cutoff=1.0) if extracted_melodies and extracted_melodies[0]: melody = extracted_melodies[0] else: # If no melody could be extracted, create an empty melody that starts 1 # step before the request start_step. This will result in 1 step of # silence when the melody is extended below. steps_per_bar = int( mm.steps_per_bar_in_quantized_sequence(quantized_sequence)) melody = mm.Melody([], start_step=max(0, start_step - 1), steps_per_bar=steps_per_bar, steps_per_quarter=self.steps_per_quarter) # Ensure that the melody extends up to the step we want to start generating. melody.set_length(start_step - melody.start_step) # Extract generation arguments from generator options. arg_types = { '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) generated_melody = self._model.generate_melody( end_step - melody.start_step, melody, **args) generated_sequence = generated_melody.to_sequence(qpm=qpm) 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 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)) if input_sequence and input_sequence.tempos: qpm = input_sequence.tempos[0].qpm else: qpm = mm.DEFAULT_QUARTERS_PER_MINUTE steps_per_second = mm.steps_per_quarter_to_steps_per_second( self.steps_per_quarter, qpm) 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, 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 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_sequence = mm.quantize_note_sequence(primer_sequence, self.steps_per_quarter) # Setting gap_bars to infinite ensures that the entire input will be used. extracted_drum_tracks, _ = drum_pipelines.extract_drum_tracks( quantized_sequence, search_start_step=input_start_step, min_bars=0, gap_bars=float('inf'), ignore_is_drum=True) assert len(extracted_drum_tracks) <= 1 start_step = mm.quantize_to_step(generate_section.start_time, 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. end_step = mm.quantize_to_step(generate_section.end_time, steps_per_second, quantize_cutoff=1.0) if extracted_drum_tracks and extracted_drum_tracks[0]: drums = extracted_drum_tracks[0] else: # If no drum track could be extracted, create an empty drum track that # starts 1 step before the request start_step. This will result in 1 step # of silence when the drum track is extended below. steps_per_bar = int( mm.steps_per_bar_in_quantized_sequence(quantized_sequence)) drums = mm.DrumTrack([], start_step=max(0, start_step - 1), steps_per_bar=steps_per_bar, steps_per_quarter=self.steps_per_quarter) # Ensure that the drum track extends up to the step we want to start # generating. drums.set_length(start_step - drums.start_step) # Extract generation arguments from generator options. arg_types = { '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) generated_drums = self._model.generate_drum_track( end_step - drums.start_step, drums, **args) generated_sequence = generated_drums.to_sequence(qpm=qpm) 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)) if input_sequence and input_sequence.tempos: qpm = input_sequence.tempos[0].qpm else: qpm = mm.DEFAULT_QUARTERS_PER_MINUTE steps_per_second = mm.steps_per_quarter_to_steps_per_second( self.steps_per_quarter, qpm) generate_section = generator_options.generate_sections[0] if generator_options.input_sections: # Use primer melody from input section only. Take backing chords from # beginning of input section through end of generate section. input_section = generator_options.input_sections[0] primer_sequence = mm.trim_note_sequence( input_sequence, input_section.start_time, input_section.end_time) backing_sequence = mm.trim_note_sequence( input_sequence, input_section.start_time, generate_section.end_time) input_start_step = mm.quantize_to_step( input_section.start_time, steps_per_second, quantize_cutoff=0.0) else: # No input section. Take primer melody from the beginning of the sequence # up until the start of the generate section. primer_sequence = mm.trim_note_sequence( input_sequence, 0.0, generate_section.start_time) backing_sequence = mm.trim_note_sequence( input_sequence, 0.0, generate_section.end_time) 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 input section. This model can only extend melodies. ' 'Requested start time: %s, Final note end time: %s' % (generate_section.start_time, last_end_time)) # Quantize the priming and backing sequences. quantized_primer_sequence = mm.quantize_note_sequence( primer_sequence, self.steps_per_quarter) quantized_backing_sequence = mm.quantize_note_sequence( backing_sequence, self.steps_per_quarter) # Setting gap_bars to infinite ensures that the entire input will be used. extracted_melodies, _ = mm.extract_melodies( quantized_primer_sequence, search_start_step=input_start_step, min_bars=0, min_unique_pitches=1, gap_bars=float('inf'), ignore_polyphonic_notes=True) assert len(extracted_melodies) <= 1 start_step = mm.quantize_to_step( generate_section.start_time, 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. end_step = mm.quantize_to_step( generate_section.end_time, steps_per_second, quantize_cutoff=1.0) if extracted_melodies and extracted_melodies[0]: melody = extracted_melodies[0] else: # If no melody could be extracted, create an empty melody that starts 1 # step before the request start_step. This will result in 1 step of # silence when the melody is extended below. steps_per_bar = int( mm.steps_per_bar_in_quantized_sequence(quantized_primer_sequence)) melody = mm.Melody([], start_step=max(0, start_step - 1), steps_per_bar=steps_per_bar, steps_per_quarter=self.steps_per_quarter) extracted_chords, _ = mm.extract_chords(quantized_backing_sequence) chords = extracted_chords[0] # Make sure that chords and melody start on the same step. if chords.start_step < melody.start_step: chords.set_length(len(chords) - melody.start_step + chords.start_step) assert chords.end_step == end_step # Ensure that the melody extends up to the step we want to start generating. melody.set_length(start_step - melody.start_step) # Extract generation arguments from generator options. arg_types = { '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) generated_melody = self._model.generate_melody(melody, chords, **args) generated_lead_sheet = mm.LeadSheet(generated_melody, chords) generated_sequence = generated_lead_sheet.to_sequence(qpm=qpm) assert (generated_sequence.total_time - generate_section.end_time) <= 1e-5 return generated_sequence
def _to_tensors(self, note_sequence): """Converts NoteSequence to unique, one-hot tensor sequences.""" try: if self._steps_per_quarter: quantized_sequence = mm.quantize_note_sequence( note_sequence, self._steps_per_quarter) if (mm.steps_per_bar_in_quantized_sequence(quantized_sequence) != self._steps_per_bar): return ConverterTensors() else: quantized_sequence = mm.quantize_note_sequence_absolute( note_sequence, self._steps_per_second) except (mm.BadTimeSignatureException, mm.NonIntegerStepsPerBarException, mm.NegativeTimeException) as e: return ConverterTensors() if self._chord_encoding and not any( ta.annotation_type == CHORD_SYMBOL for ta in quantized_sequence.text_annotations): # We are conditioning on chords but sequence does not have chords. Try to # infer them. try: mm.infer_chords_for_sequence(quantized_sequence) except mm.ChordInferenceException: return ConverterTensors() event_lists, unused_stats = self._event_extractor_fn(quantized_sequence) if self._pad_to_total_time: for e in event_lists: e.set_length(len(e) + e.start_step, from_left=True) e.set_length(quantized_sequence.total_quantized_steps) if self._slice_steps: sliced_event_lists = [] for l in event_lists: for i in range(self._slice_steps, len(l) + 1, self._steps_per_bar): sliced_event_lists.append(l[i - self._slice_steps: i]) else: sliced_event_lists = event_lists if self._chord_encoding: try: sliced_chord_lists = chords_lib.event_list_chords( quantized_sequence, sliced_event_lists) except chords_lib.CoincidentChordsException: return ConverterTensors() sliced_event_lists = [zip(el, cl) for el, cl in zip(sliced_event_lists, sliced_chord_lists)] # TODO(adarob): Consider handling the fact that different event lists can # be mapped to identical tensors by the encoder_decoder (e.g., Drums). unique_event_tuples = list(set(tuple(l) for l in sliced_event_lists)) unique_event_tuples = self._maybe_sample_outputs(unique_event_tuples) if not unique_event_tuples: return ConverterTensors() control_seqs = [] if self._chord_encoding: unique_event_tuples, unique_chord_tuples = zip( *[zip(*t) for t in unique_event_tuples if t]) for t in unique_chord_tuples: try: chord_tokens = [self._chord_encoding.encode_event(e) for e in t] if self.end_token: # Repeat the last chord instead of using a special token; otherwise # the model may learn to rely on the special token to detect # endings. chord_tokens.append(chord_tokens[-1] if chord_tokens else self._chord_encoding.encode_event(mm.NO_CHORD)) except (mm.ChordSymbolException, mm.ChordEncodingException): return ConverterTensors() control_seqs.append( np_onehot(chord_tokens, self.control_depth, self.control_dtype)) seqs = [] for t in unique_event_tuples: seqs.append(np_onehot( [self._legacy_encoder_decoder.encode_event(e) for e in t] + ([] if self.end_token is None else [self.end_token]), self.output_depth, self.output_dtype)) return ConverterTensors(inputs=seqs, outputs=seqs, controls=control_seqs)
def _to_tensors(self, note_sequence): try: quantized_sequence = mm.quantize_note_sequence( note_sequence, self._steps_per_quarter) if (mm.steps_per_bar_in_quantized_sequence(quantized_sequence) != self._steps_per_bar): return [] except (mm.BadTimeSignatureException, mm.NonIntegerStepsPerBarException, mm.NegativeTimeException): return [] total_bars = int( np.ceil(quantized_sequence.total_quantized_steps / self._steps_per_bar)) total_bars = min(total_bars, self._max_bars) # Assign an instrument class for each instrument, and compute its coverage. # If an instrument has multiple classes, it is considered INVALID. instrument_type = np.zeros(MAX_INSTRUMENT_NUMBER + 1, np.uint8) coverage = np.zeros((total_bars, MAX_INSTRUMENT_NUMBER + 1), np.bool) for note in quantized_sequence.notes: i = note.instrument if i > MAX_INSTRUMENT_NUMBER: tf.logging.warning('Skipping invalid instrument number: %d', i) continue inferred_type = ( self.InstrumentType.DRUMS if note.is_drum else self._program_map.get(note.program, self.InstrumentType.INVALID)) if not instrument_type[i]: instrument_type[i] = inferred_type elif instrument_type[i] != inferred_type: instrument_type[i] = self.InstrumentType.INVALID start_bar = note.quantized_start_step // self._steps_per_bar end_bar = int(np.ceil(note.quantized_end_step / self._steps_per_bar)) if start_bar >= total_bars: continue coverage[start_bar:min(end_bar, total_bars), i] = True # Group instruments by type. instruments_by_type = collections.defaultdict(list) for i, type_ in enumerate(instrument_type): if type_ not in (self.InstrumentType.UNK, self.InstrumentType.INVALID): instruments_by_type[type_].append(i) if len(instruments_by_type) < 3: # This NoteSequence doesn't have all 3 types. return [], [] # Encode individual instruments. # Set total time so that instruments will be padded correctly. note_sequence.total_time = ( total_bars * self._steps_per_bar * 60 / note_sequence.tempos[0].qpm / self._steps_per_quarter) encoded_instruments = {} for i in (instruments_by_type[self.InstrumentType.MEL] + instruments_by_type[self.InstrumentType.BASS]): _, t = self._melody_converter.to_tensors( _extract_instrument(note_sequence, i)) if t: encoded_instruments[i] = t[0] else: coverage[:, i] = False for i in instruments_by_type[self.InstrumentType.DRUMS]: _, t = self._drums_converter.to_tensors( _extract_instrument(note_sequence, i)) if t: encoded_instruments[i] = t[0] else: coverage[:, i] = False # Fill in coverage gaps up to self._gap_bars. og_coverage = coverage.copy() for j in range(total_bars): coverage[j] = np.any( og_coverage[ max(0, j-self._gap_bars):min(total_bars, j+self._gap_bars) + 1], axis=0) # Take cross product of instruments from each class and compute combined # encodings where they overlap. seqs = [] for grp in itertools.product( instruments_by_type[self.InstrumentType.MEL], instruments_by_type[self.InstrumentType.BASS], instruments_by_type[self.InstrumentType.DRUMS]): # Consider an instrument covered within gap_bars from the end if any of # the other instruments are. This allows more leniency when re-encoding # slices. grp_coverage = np.all(coverage[:, grp], axis=1) grp_coverage[:self._gap_bars] = np.any(coverage[:self._gap_bars, grp]) grp_coverage[-self._gap_bars:] = np.any(coverage[-self._gap_bars:, grp]) for j in range(total_bars - self._slice_bars + 1): if np.all(grp_coverage[j:j + self._slice_bars]): start_step = j * self._steps_per_bar end_step = (j + self._slice_bars) * self._steps_per_bar seqs.append(np.concatenate( [encoded_instruments[i][start_step:end_step] for i in grp], axis=-1)) return seqs, seqs
def _to_tensors(self, note_sequence): try: quantized_sequence = mm.quantize_note_sequence( note_sequence, self._steps_per_quarter) if (mm.steps_per_bar_in_quantized_sequence(quantized_sequence) != self._steps_per_bar): return ConverterTensors() except (mm.BadTimeSignatureException, mm.NonIntegerStepsPerBarException, mm.NegativeTimeException): return ConverterTensors() if self._chord_encoding and not any( ta.annotation_type == CHORD_SYMBOL for ta in quantized_sequence.text_annotations): # We are conditioning on chords but sequence does not have chords. Try to # infer them. try: mm.infer_chords_for_sequence(quantized_sequence) except mm.ChordInferenceException: return ConverterTensors() # The trio parts get extracted from the original NoteSequence, so copy the # inferred chords back to that one. for qta in quantized_sequence.text_annotations: if qta.annotation_type == CHORD_SYMBOL: ta = note_sequence.text_annotations.add() ta.annotation_type = CHORD_SYMBOL ta.time = qta.time ta.text = qta.text total_bars = int( np.ceil(quantized_sequence.total_quantized_steps / self._steps_per_bar)) total_bars = min(total_bars, self._max_bars) # Assign an instrument class for each instrument, and compute its coverage. # If an instrument has multiple classes, it is considered INVALID. instrument_type = np.zeros(MAX_INSTRUMENT_NUMBER + 1, np.uint8) coverage = np.zeros((total_bars, MAX_INSTRUMENT_NUMBER + 1), np.bool) for note in quantized_sequence.notes: i = note.instrument if i > MAX_INSTRUMENT_NUMBER: tf.logging.warning('Skipping invalid instrument number: %d', i) continue inferred_type = ( self.InstrumentType.DRUMS if note.is_drum else self._program_map.get(note.program, self.InstrumentType.INVALID)) if not instrument_type[i]: instrument_type[i] = inferred_type elif instrument_type[i] != inferred_type: instrument_type[i] = self.InstrumentType.INVALID start_bar = note.quantized_start_step // self._steps_per_bar end_bar = int(np.ceil(note.quantized_end_step / self._steps_per_bar)) if start_bar >= total_bars: continue coverage[start_bar:min(end_bar, total_bars), i] = True # Group instruments by type. instruments_by_type = collections.defaultdict(list) for i, type_ in enumerate(instrument_type): if type_ not in (self.InstrumentType.UNK, self.InstrumentType.INVALID): instruments_by_type[type_].append(i) if len(instruments_by_type) < 3: # This NoteSequence doesn't have all 3 types. return ConverterTensors() # Encode individual instruments. # Set total time so that instruments will be padded correctly. note_sequence.total_time = ( total_bars * self._steps_per_bar * 60 / note_sequence.tempos[0].qpm / self._steps_per_quarter) encoded_instruments = {} encoded_chords = None for i in (instruments_by_type[self.InstrumentType.MEL] + instruments_by_type[self.InstrumentType.BASS]): tensors = self._melody_converter.to_tensors( _extract_instrument(note_sequence, i)) if tensors.outputs: encoded_instruments[i] = tensors.outputs[0] if encoded_chords is None: encoded_chords = tensors.controls[0] elif not np.array_equal(encoded_chords, tensors.controls[0]): tf.logging.warning('Trio chords disagreement between instruments.') else: coverage[:, i] = False for i in instruments_by_type[self.InstrumentType.DRUMS]: tensors = self._drums_converter.to_tensors( _extract_instrument(note_sequence, i)) if tensors.outputs: encoded_instruments[i] = tensors.outputs[0] else: coverage[:, i] = False # Fill in coverage gaps up to self._gap_bars. og_coverage = coverage.copy() for j in range(total_bars): coverage[j] = np.any( og_coverage[ max(0, j-self._gap_bars):min(total_bars, j+self._gap_bars) + 1], axis=0) # Take cross product of instruments from each class and compute combined # encodings where they overlap. seqs = [] control_seqs = [] for grp in itertools.product( instruments_by_type[self.InstrumentType.MEL], instruments_by_type[self.InstrumentType.BASS], instruments_by_type[self.InstrumentType.DRUMS]): # Consider an instrument covered within gap_bars from the end if any of # the other instruments are. This allows more leniency when re-encoding # slices. grp_coverage = np.all(coverage[:, grp], axis=1) grp_coverage[:self._gap_bars] = np.any(coverage[:self._gap_bars, grp]) grp_coverage[-self._gap_bars:] = np.any(coverage[-self._gap_bars:, grp]) for j in range(total_bars - self._slice_bars + 1): if (np.all(grp_coverage[j:j + self._slice_bars]) and all(i in encoded_instruments for i in grp)): start_step = j * self._steps_per_bar end_step = (j + self._slice_bars) * self._steps_per_bar seqs.append(np.concatenate( [encoded_instruments[i][start_step:end_step] for i in grp], axis=-1)) if encoded_chords is not None: control_seqs.append(encoded_chords[start_step:end_step]) return ConverterTensors(inputs=seqs, outputs=seqs, controls=control_seqs)
def _to_tensors(self, note_sequence): try: quantized_sequence = mm.quantize_note_sequence( note_sequence, self._steps_per_quarter) if (mm.steps_per_bar_in_quantized_sequence(quantized_sequence) != self._steps_per_bar): return [] except (mm.BadTimeSignatureException, mm.NonIntegerStepsPerBarException, mm.NegativeTimeException): return [] total_bars = int( np.ceil(quantized_sequence.total_quantized_steps / self._steps_per_bar)) total_bars = min(total_bars, self._max_bars) # Assign an instrument class for each instrument, and compute its coverage. # If an instrument has multiple classes, it is considered INVALID. instrument_type = np.zeros(MAX_INSTRUMENT_NUMBER + 1, np.uint8) coverage = np.zeros((total_bars, MAX_INSTRUMENT_NUMBER + 1), np.bool) for note in quantized_sequence.notes: i = note.instrument if i > MAX_INSTRUMENT_NUMBER: tf.logging.warning('Skipping invalid instrument number: %d', i) continue inferred_type = ( self.InstrumentType.DRUMS if note.is_drum else self._program_map.get(note.program, self.InstrumentType.INVALID)) if not instrument_type[i]: instrument_type[i] = inferred_type elif instrument_type[i] != inferred_type: instrument_type[i] = self.InstrumentType.INVALID start_bar = note.quantized_start_step // self._steps_per_bar end_bar = int(np.ceil(note.quantized_end_step / self._steps_per_bar)) if start_bar >= total_bars: continue coverage[start_bar:min(end_bar, total_bars), i] = True # Group instruments by type. instruments_by_type = collections.defaultdict(list) for i, type_ in enumerate(instrument_type): if type_ not in (self.InstrumentType.UNK, self.InstrumentType.INVALID): instruments_by_type[type_].append(i) if len(instruments_by_type) < 3: # This NoteSequence doesn't have all 3 types. return [], [] # Encode individual instruments. # Set total time so that instruments will be padded correctly. note_sequence.total_time = ( total_bars * self._steps_per_bar * 60 / note_sequence.tempos[0].qpm / self._steps_per_quarter) encoded_instruments = {} for i in (instruments_by_type[self.InstrumentType.MEL] + instruments_by_type[self.InstrumentType.BASS]): _, t = self._melody_converter.to_tensors( _extract_instrument(note_sequence, i)) if t: encoded_instruments[i] = t[0] else: coverage[:, i] = False for i in instruments_by_type[self.InstrumentType.DRUMS]: _, t = self._drums_converter.to_tensors( _extract_instrument(note_sequence, i)) if t: encoded_instruments[i] = t[0] else: coverage[:, i] = False # Fill in coverage gaps up to self._gap_bars. og_coverage = coverage.copy() for j in range(total_bars): coverage[j] = np.any( og_coverage[ max(0, j-self._gap_bars):min(total_bars, j+self._gap_bars) + 1], axis=0) # Take cross product of instruments from each class and compute combined # encodings where they overlap. seqs = [] for grp in itertools.product( instruments_by_type[self.InstrumentType.MEL], instruments_by_type[self.InstrumentType.BASS], instruments_by_type[self.InstrumentType.DRUMS]): # Consider an instrument covered within gap_bars from the end if any of # the other instruments are. This allows more leniency when re-encoding # slices. grp_coverage = np.all(coverage[:, grp], axis=1) grp_coverage[:self._gap_bars] = np.any(coverage[:self._gap_bars, grp]) grp_coverage[-self._gap_bars:] = np.any(coverage[-self._gap_bars:, grp]) for j in range(total_bars - self._slice_bars + 1): if np.all(grp_coverage[j:j + self._slice_bars]): start_step = j * self._steps_per_bar end_step = (j + self._slice_bars) * self._steps_per_bar seqs.append(np.concatenate( [encoded_instruments[i][start_step:end_step] for i in grp], axis=-1)) return seqs, seqs
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)) qpm = (input_sequence.tempos[0].qpm if input_sequence and input_sequence.tempos else mm.DEFAULT_QUARTERS_PER_MINUTE) 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 = self.seconds_to_steps(input_section.start_time, qpm) 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 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_sequence = mm.quantize_note_sequence( primer_sequence, self.steps_per_quarter) # Setting gap_bars to infinite ensures that the entire input will be used. extracted_drum_tracks, _ = mm.extract_drum_tracks( quantized_sequence, search_start_step=input_start_step, min_bars=0, gap_bars=float('inf')) assert len(extracted_drum_tracks) <= 1 start_step = self.seconds_to_steps( generate_section.start_time, qpm) end_step = self.seconds_to_steps(generate_section.end_time, qpm) if extracted_drum_tracks and extracted_drum_tracks[0]: drums = extracted_drum_tracks[0] else: # If no drum track could be extracted, create an empty drum track that # starts 1 step before the request start_step. This will result in 1 step # of silence when the drum track is extended below. steps_per_bar = int( mm.steps_per_bar_in_quantized_sequence(quantized_sequence)) drums = mm.DrumTrack([], start_step=max(0, start_step - 1), steps_per_bar=steps_per_bar, steps_per_quarter=self.steps_per_quarter) # Ensure that the drum track extends up to the step we want to start # generating. drums.set_length(start_step - drums.start_step) # Extract generation arguments from generator options. arg_types = { '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) generated_drums = self._model.generate_drum_track( end_step - drums.start_step, drums, **args) generated_sequence = generated_drums.to_sequence(qpm=qpm) assert (generated_sequence.total_time - generate_section.end_time) <= 1e-5 return generated_sequence
def _primer_melody_to_event_sequence(self, input_sequence, generator_options, config): qpm = (input_sequence.tempos[0].qpm if input_sequence and input_sequence.tempos else mm.DEFAULT_QUARTERS_PER_MINUTE) steps_per_second = mm.steps_per_quarter_to_steps_per_second( self.steps_per_quarter, qpm) 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, steps_per_second, quantize_cutoff=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 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_sequence = mm.quantize_note_sequence(primer_sequence, self.steps_per_quarter) # Setting gap_bars to infinite ensures that the entire input will be used. extracted_melodies, _ = mm.extract_melodies( quantized_sequence, search_start_step=input_start_step, min_bars=0, min_unique_pitches=1, gap_bars=float('inf'), ignore_polyphonic_notes=True) assert len(extracted_melodies) <= 1 start_step = mm.quantize_to_step(generate_section.start_time, steps_per_second, quantize_cutoff=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. end_step = mm.quantize_to_step(generate_section.end_time, steps_per_second, quantize_cutoff=1.0) if extracted_melodies and extracted_melodies[0]: melody = extracted_melodies[0] else: # If no melody could be extracted, create an empty melody that starts 1 # step before the request start_step. This will result in 1 step of # silence when the melody is extended below. steps_per_bar = int( mm.steps_per_bar_in_quantized_sequence(quantized_sequence)) melody = mm.Melody([], start_step=max(0, start_step - 1), steps_per_bar=steps_per_bar, steps_per_quarter=self.steps_per_quarter) # Ensure that the melody extends up to the step we want to start generating. melody.set_length(start_step - melody.start_step - 2) now_encoding = config.encoder_decoder._one_hot_encoding # Extract generation arguments from generator options. primer_events = self._model.primer_melody_to_events( end_step - melody.start_step, melody) for i, event in enumerate(primer_events): primer_events[i] = now_encoding.encode_event(event) return primer_events
def _to_tensors(self, note_sequence): # Performance sequences require sustain to be correctly interpreted. note_sequence = sequences_lib.apply_sustain_control_changes(note_sequence) if self._chord_encoding and not any( ta.annotation_type == CHORD_SYMBOL for ta in note_sequence.text_annotations): try: # Quantize just for the purpose of chord inference. # TODO(iansimon): Allow chord inference in unquantized sequences. quantized_sequence = mm.quantize_note_sequence( note_sequence, self._steps_per_quarter) if (mm.steps_per_bar_in_quantized_sequence(quantized_sequence) != self._steps_per_bar): return data.ConverterTensors() # Infer chords in quantized sequence. mm.infer_chords_for_sequence(quantized_sequence) except (mm.BadTimeSignatureError, mm.NonIntegerStepsPerBarError, mm.NegativeTimeError, mm.ChordInferenceError): return data.ConverterTensors() # Copy inferred chords back to original sequence. for qta in quantized_sequence.text_annotations: if qta.annotation_type == CHORD_SYMBOL: ta = note_sequence.text_annotations.add() ta.annotation_type = CHORD_SYMBOL ta.time = qta.time ta.text = qta.text if note_sequence.tempos: quarters_per_minute = note_sequence.tempos[0].qpm else: quarters_per_minute = mm.DEFAULT_QUARTERS_PER_MINUTE quarters_per_bar = self._steps_per_bar / self._steps_per_quarter hop_size_quarters = quarters_per_bar * self._hop_size_bars hop_size_seconds = 60.0 * hop_size_quarters / quarters_per_minute # Split note sequence by bar hop size (in seconds). subsequences = sequences_lib.split_note_sequence( note_sequence, hop_size_seconds) if self._first_subsequence_only and len(subsequences) > 1: return data.ConverterTensors() sequence_tensors = [] sequence_chord_tensors = [] for subsequence in subsequences: # Quantize this subsequence. try: quantized_subsequence = mm.quantize_note_sequence( subsequence, self._steps_per_quarter) if (mm.steps_per_bar_in_quantized_sequence(quantized_subsequence) != self._steps_per_bar): return data.ConverterTensors() except (mm.BadTimeSignatureError, mm.NonIntegerStepsPerBarError, mm.NegativeTimeError): return data.ConverterTensors() # Convert the quantized subsequence to tensors. tensors, chord_tensors = self._quantized_subsequence_to_tensors( quantized_subsequence) if tensors: sequence_tensors.append(tensors) if self._chord_encoding: sequence_chord_tensors.append(chord_tensors) return data.ConverterTensors( inputs=sequence_tensors, outputs=sequence_tensors, controls=sequence_chord_tensors)
def _to_tensors(self, note_sequence): try: quantized_sequence = mm.quantize_note_sequence( note_sequence, self._steps_per_quarter) if (mm.steps_per_bar_in_quantized_sequence(quantized_sequence) != self._steps_per_bar): return ConverterTensors() except (mm.BadTimeSignatureException, mm.NonIntegerStepsPerBarException, mm.NegativeTimeException): return ConverterTensors() if self._chord_encoding and not any( ta.annotation_type == CHORD_SYMBOL for ta in quantized_sequence.text_annotations): # We are conditioning on chords but sequence does not have chords. Try to # infer them. try: mm.infer_chords_for_sequence(quantized_sequence) except mm.ChordInferenceException: return ConverterTensors() # The trio parts get extracted from the original NoteSequence, so copy the # inferred chords back to that one. for qta in quantized_sequence.text_annotations: if qta.annotation_type == CHORD_SYMBOL: ta = note_sequence.text_annotations.add() ta.annotation_type = CHORD_SYMBOL ta.time = qta.time ta.text = qta.text total_bars = int( np.ceil(quantized_sequence.total_quantized_steps / self._steps_per_bar)) total_bars = min(total_bars, self._max_bars) # Assign an instrument class for each instrument, and compute its coverage. # If an instrument has multiple classes, it is considered INVALID. instrument_type = np.zeros(MAX_INSTRUMENT_NUMBER + 1, np.uint8) coverage = np.zeros((total_bars, MAX_INSTRUMENT_NUMBER + 1), np.bool) for note in quantized_sequence.notes: i = note.instrument if i > MAX_INSTRUMENT_NUMBER: tf.logging.warning('Skipping invalid instrument number: %d', i) continue inferred_type = ( self.InstrumentType.DRUMS if note.is_drum else self._program_map.get(note.program, self.InstrumentType.INVALID)) if not instrument_type[i]: instrument_type[i] = inferred_type elif instrument_type[i] != inferred_type: instrument_type[i] = self.InstrumentType.INVALID start_bar = note.quantized_start_step // self._steps_per_bar end_bar = int(np.ceil(note.quantized_end_step / self._steps_per_bar)) if start_bar >= total_bars: continue coverage[start_bar:min(end_bar, total_bars), i] = True # Group instruments by type. instruments_by_type = collections.defaultdict(list) for i, type_ in enumerate(instrument_type): if type_ not in (self.InstrumentType.UNK, self.InstrumentType.INVALID): instruments_by_type[type_].append(i) if len(instruments_by_type) < 3: # This NoteSequence doesn't have all 3 types. return ConverterTensors() # Encode individual instruments. # Set total time so that instruments will be padded correctly. note_sequence.total_time = ( total_bars * self._steps_per_bar * 60 / note_sequence.tempos[0].qpm / self._steps_per_quarter) encoded_instruments = {} encoded_chords = None for i in (instruments_by_type[self.InstrumentType.MEL] + instruments_by_type[self.InstrumentType.BASS]): tensors = self._melody_converter.to_tensors( _extract_instrument(note_sequence, i)) if tensors.outputs: encoded_instruments[i] = tensors.outputs[0] if encoded_chords is None: encoded_chords = tensors.controls[0] elif not np.array_equal(encoded_chords, tensors.controls[0]): tf.logging.warning('Trio chords disagreement between instruments.') else: coverage[:, i] = False for i in instruments_by_type[self.InstrumentType.DRUMS]: tensors = self._drums_converter.to_tensors( _extract_instrument(note_sequence, i)) if tensors.outputs: encoded_instruments[i] = tensors.outputs[0] else: coverage[:, i] = False # Fill in coverage gaps up to self._gap_bars. og_coverage = coverage.copy() for j in range(total_bars): coverage[j] = np.any( og_coverage[ max(0, j-self._gap_bars):min(total_bars, j+self._gap_bars) + 1], axis=0) # Take cross product of instruments from each class and compute combined # encodings where they overlap. seqs = [] control_seqs = [] for grp in itertools.product( instruments_by_type[self.InstrumentType.MEL], instruments_by_type[self.InstrumentType.BASS], instruments_by_type[self.InstrumentType.DRUMS]): # Consider an instrument covered within gap_bars from the end if any of # the other instruments are. This allows more leniency when re-encoding # slices. grp_coverage = np.all(coverage[:, grp], axis=1) grp_coverage[:self._gap_bars] = np.any(coverage[:self._gap_bars, grp]) grp_coverage[-self._gap_bars:] = np.any(coverage[-self._gap_bars:, grp]) for j in range(total_bars - self._slice_bars + 1): if (np.all(grp_coverage[j:j + self._slice_bars]) and all(i in encoded_instruments for i in grp)): start_step = j * self._steps_per_bar end_step = (j + self._slice_bars) * self._steps_per_bar seqs.append(np.concatenate( [encoded_instruments[i][start_step:end_step] for i in grp], axis=-1)) if encoded_chords is not None: control_seqs.append(encoded_chords[start_step:end_step]) return ConverterTensors(inputs=seqs, outputs=seqs, controls=control_seqs)
def _to_tensors(self, note_sequence): """Converts NoteSequence to unique, one-hot tensor sequences.""" try: if self._steps_per_quarter: quantized_sequence = mm.quantize_note_sequence( note_sequence, self._steps_per_quarter) if (mm.steps_per_bar_in_quantized_sequence(quantized_sequence) != self._steps_per_bar): return ConverterTensors() else: quantized_sequence = mm.quantize_note_sequence_absolute( note_sequence, self._steps_per_second) except (mm.BadTimeSignatureException, mm.NonIntegerStepsPerBarException, mm.NegativeTimeException) as e: return ConverterTensors() if self._chord_encoding and not any( ta.annotation_type == CHORD_SYMBOL for ta in quantized_sequence.text_annotations): # We are conditioning on chords but sequence does not have chords. Try to # infer them. try: mm.infer_chords_for_sequence(quantized_sequence) except mm.ChordInferenceException: return ConverterTensors() event_lists, unused_stats = self._event_extractor_fn(quantized_sequence) if self._pad_to_total_time: for e in event_lists: e.set_length(len(e) + e.start_step, from_left=True) e.set_length(quantized_sequence.total_quantized_steps) if self._slice_steps: sliced_event_lists = [] for l in event_lists: for i in range(self._slice_steps, len(l) + 1, self._steps_per_bar): sliced_event_lists.append(l[i - self._slice_steps: i]) else: sliced_event_lists = event_lists if self._chord_encoding: try: sliced_chord_lists = chords_lib.event_list_chords( quantized_sequence, sliced_event_lists) except chords_lib.CoincidentChordsException: return ConverterTensors() sliced_event_lists = [zip(el, cl) for el, cl in zip(sliced_event_lists, sliced_chord_lists)] # TODO(adarob): Consider handling the fact that different event lists can # be mapped to identical tensors by the encoder_decoder (e.g., Drums). unique_event_tuples = list(set(tuple(l) for l in sliced_event_lists)) unique_event_tuples = self._maybe_sample_outputs(unique_event_tuples) if not unique_event_tuples: return ConverterTensors() control_seqs = [] if self._chord_encoding: unique_event_tuples, unique_chord_tuples = zip( *[zip(*t) for t in unique_event_tuples if t]) for t in unique_chord_tuples: try: chord_tokens = [self._chord_encoding.encode_event(e) for e in t] if self.end_token: # Repeat the last chord instead of using a special token; otherwise # the model may learn to rely on the special token to detect # endings. chord_tokens.append(chord_tokens[-1] if chord_tokens else self._chord_encoding.encode_event(mm.NO_CHORD)) except (mm.ChordSymbolException, mm.ChordEncodingException): return ConverterTensors() control_seqs.append( np_onehot(chord_tokens, self.control_depth, self.control_dtype)) seqs = [] for t in unique_event_tuples: seqs.append(np_onehot( [self._legacy_encoder_decoder.encode_event(e) for e in t] + ([] if self.end_token is None else [self.end_token]), self.output_depth, self.output_dtype)) return ConverterTensors(inputs=seqs, outputs=seqs, controls=control_seqs)
def _to_tensors(self, note_sequence): # Performance sequences require sustain to be correctly interpreted. note_sequence = sequences_lib.apply_sustain_control_changes( note_sequence) if self._chord_encoding and not any( ta.annotation_type == CHORD_SYMBOL for ta in note_sequence.text_annotations): try: # Quantize just for the purpose of chord inference. # TODO(iansimon): Allow chord inference in unquantized sequences. quantized_sequence = mm.quantize_note_sequence( note_sequence, self._steps_per_quarter) if (mm.steps_per_bar_in_quantized_sequence(quantized_sequence) != self._steps_per_bar): return data.ConverterTensors() # Infer chords in quantized sequence. mm.infer_chords_for_sequence(quantized_sequence) except (mm.BadTimeSignatureException, mm.NonIntegerStepsPerBarException, mm.NegativeTimeException, mm.ChordInferenceException): return data.ConverterTensors() # Copy inferred chords back to original sequence. for qta in quantized_sequence.text_annotations: if qta.annotation_type == CHORD_SYMBOL: ta = note_sequence.text_annotations.add() ta.annotation_type = CHORD_SYMBOL ta.time = qta.time ta.text = qta.text quarters_per_minute = (note_sequence.tempos[0].qpm if note_sequence.tempos else mm.DEFAULT_QUARTERS_PER_MINUTE) quarters_per_bar = self._steps_per_bar / self._steps_per_quarter hop_size_quarters = quarters_per_bar * self._hop_size_bars hop_size_seconds = 60.0 * hop_size_quarters / quarters_per_minute # Split note sequence by bar hop size (in seconds). subsequences = sequences_lib.split_note_sequence( note_sequence, hop_size_seconds) if self._first_subsequence_only and len(subsequences) > 1: return data.ConverterTensors() sequence_tensors = [] sequence_chord_tensors = [] for subsequence in subsequences: # Quantize this subsequence. try: quantized_subsequence = mm.quantize_note_sequence( subsequence, self._steps_per_quarter) if (mm.steps_per_bar_in_quantized_sequence( quantized_subsequence) != self._steps_per_bar): return data.ConverterTensors() except (mm.BadTimeSignatureException, mm.NonIntegerStepsPerBarException, mm.NegativeTimeException): return data.ConverterTensors() # Convert the quantized subsequence to tensors. tensors, chord_tensors = self._quantized_subsequence_to_tensors( quantized_subsequence) if tensors: sequence_tensors.append(tensors) if self._chord_encoding: sequence_chord_tensors.append(chord_tensors) return data.ConverterTensors(inputs=sequence_tensors, outputs=sequence_tensors, controls=sequence_chord_tensors)
def _to_tensors(self, note_sequence): """Converts NoteSequence to unique sequences.""" try: quantized_sequence = mm.quantize_note_sequence( note_sequence, self._steps_per_quarter) if (mm.steps_per_bar_in_quantized_sequence(quantized_sequence) != self._steps_per_bar): return [], [] except (mm.BadTimeSignatureException, mm.NonIntegerStepsPerBarException, mm.NegativeTimeException) as e: return [], [] new_notes = [] for n in quantized_sequence.notes: if not n.is_drum: continue if n.pitch not in self._pitch_class_map: continue n.pitch = self._pitch_class_map[n.pitch] new_notes.append(n) del quantized_sequence.notes[:] quantized_sequence.notes.extend(new_notes) event_lists, unused_stats = self._drums_extractor_fn(quantized_sequence) if self._pad_to_total_time: for e in event_lists: e.set_length(len(e) + e.start_step, from_left=True) e.set_length(quantized_sequence.total_quantized_steps) if self._slice_steps: sliced_event_tuples = [] for l in event_lists: for i in range(self._slice_steps, len(l) + 1, self._steps_per_bar): sliced_event_tuples.append(tuple(l[i - self._slice_steps: i])) else: sliced_event_tuples = [tuple(l) for l in event_lists] unique_event_tuples = list(set(sliced_event_tuples)) unique_event_tuples = self._maybe_sample_outputs(unique_event_tuples) rolls = [] oh_vecs = [] for t in unique_event_tuples: if self._roll_input or self._roll_output: if self.end_token is not None: t_roll = list(t) + [(self._pr_encoder_decoder.input_size - 1,)] else: t_roll = t rolls.append(np.vstack([ self._pr_encoder_decoder.events_to_input(t_roll, i).astype(np.bool) for i in range(len(t_roll))])) if not (self._roll_input and self._roll_output): labels = [self._oh_encoder_decoder.encode_event(e) for e in t] if self.end_token is not None: labels += [self._oh_encoder_decoder.num_classes] oh_vecs.append(np_onehot( labels, self._oh_encoder_decoder.num_classes + (self.end_token is not None), np.bool)) if self._roll_input: input_seqs = [ np.append(roll, np.expand_dims(np.all(roll == 0, axis=1), axis=1), axis=1) for roll in rolls] else: input_seqs = oh_vecs output_seqs = rolls if self._roll_output else oh_vecs return input_seqs, output_seqs