def run(config_map):
"""Load model params, save config file and start trainer.
Args:
config_map: Dictionary mapping configuration name to Config object.
Raises:
ValueError: if required flags are missing or invalid.
"""
date_and_time = time.strftime('%Y-%m-%d_%H%M%S')
if FLAGS.run_dir is None == FLAGS.checkpoint_file is None:
raise ValueError(
'Exactly one of `--run_dir` or `--checkpoint_file` must be specified.')
if FLAGS.output_dir is None:
raise ValueError('`--output_dir` is required.')
tf.gfile.MakeDirs(FLAGS.output_dir)
if FLAGS.mode != 'sample' and FLAGS.mode != 'interpolate':
raise ValueError('Invalid value for `--mode`: %s' % FLAGS.mode)
if FLAGS.config not in config_map:
raise ValueError('Invalid config name: %s' % FLAGS.config)
config = config_map[FLAGS.config]
config.data_converter.max_tensors_per_item = None
if FLAGS.mode == 'interpolate':
if FLAGS.input_midi_1 is None or FLAGS.input_midi_2 is None:
raise ValueError(
'`--input_midi_1` and `--input_midi_2` must be specified in '
'`interpolate` mode.')
input_midi_1 = os.path.expanduser(FLAGS.input_midi_1)
input_midi_2 = os.path.expanduser(FLAGS.input_midi_2)
if not os.path.exists(input_midi_1):
raise ValueError('Input MIDI 1 not found: %s' % FLAGS.input_midi_1)
if not os.path.exists(input_midi_2):
raise ValueError('Input MIDI 2 not found: %s' % FLAGS.input_midi_2)
input_1 = note_seq.midi_file_to_note_sequence(input_midi_1)
input_2 = note_seq.midi_file_to_note_sequence(input_midi_2)
def _check_extract_examples(input_ns, path, input_number):
"""Make sure each input returns exactly one example from the converter."""
tensors = config.data_converter.to_tensors(input_ns).outputs
if not tensors:
print(
'MusicVAE configs have very specific input requirements. Could not '
'extract any valid inputs from `%s`. Try another MIDI file.' % path)
sys.exit()
elif len(tensors) > 1:
basename = os.path.join(
FLAGS.output_dir,
'%s_input%d-extractions_%s-*-of-%03d.mid' %
(FLAGS.config, input_number, date_and_time, len(tensors)))
for i, ns in enumerate(config.data_converter.from_tensors(tensors)):
note_seq.sequence_proto_to_midi_file(
ns, basename.replace('*', '%03d' % i))
print(
'%d valid inputs extracted from `%s`. Outputting these potential '
'inputs as `%s`. Call script again with one of these instead.' %
(len(tensors), path, basename))
sys.exit()
logging.info(
'Attempting to extract examples from input MIDIs using config `%s`...',
FLAGS.config)
_check_extract_examples(input_1, FLAGS.input_midi_1, 1)
_check_extract_examples(input_2, FLAGS.input_midi_2, 2)
logging.info('Loading model...')
if FLAGS.run_dir:
checkpoint_dir_or_path = os.path.expanduser(
os.path.join(FLAGS.run_dir, 'train'))
else:
checkpoint_dir_or_path = os.path.expanduser(FLAGS.checkpoint_file)
model = TrainedModel(
config, batch_size=min(FLAGS.max_batch_size, FLAGS.num_outputs),
checkpoint_dir_or_path=checkpoint_dir_or_path)
if FLAGS.mode == 'interpolate':
logging.info('Interpolating...')
_, mu, _ = model.encode([input_1, input_2])
z = np.array([
_slerp(mu[0], mu[1], t) for t in np.linspace(0, 1, FLAGS.num_outputs)])
results = model.decode(
length=config.hparams.max_seq_len,
z=z,
temperature=FLAGS.temperature)
elif FLAGS.mode == 'sample':
logging.info('Sampling...')
results = model.sample(
n=FLAGS.num_outputs,
length=config.hparams.max_seq_len,
temperature=FLAGS.temperature)
basename = os.path.join(
FLAGS.output_dir,
'%s_%s_%s-*-of-%03d.mid' %
(FLAGS.config, FLAGS.mode, date_and_time, FLAGS.num_outputs))
logging.info('Outputting %d files as `%s`...', FLAGS.num_outputs, basename)
for i, ns in enumerate(results):
note_seq.sequence_proto_to_midi_file(ns, basename.replace('*', '%03d' % i))
logging.info('Done.')
def run(config_map):
"""Load model params, save config file and start trainer.
Args:
config_map: Dictionary mapping configuration name to Config object.
Raises:
ValueError: if required flags are missing or invalid.
"""
date_and_time = time.strftime('%Y-%m-%d_%H%M%S')
if FLAGS.run_dir is None == FLAGS.checkpoint_file is None:
raise ValueError(
'Exactly one of `--run_dir` or `--checkpoint_file` must be specified.'
)
if FLAGS.output_dir is None:
raise ValueError('`--output_dir` is required.')
tf.gfile.MakeDirs(FLAGS.output_dir)
if FLAGS.mode != 'sample' and FLAGS.mode != 'interpolate' and FLAGS.mode != 'isample':
raise ValueError('Invalid value for `--mode`: %s' % FLAGS.mode)
if FLAGS.config not in config_map:
raise ValueError('Invalid config name: %s' % FLAGS.config)
config = config_map[FLAGS.config]
config.data_converter.max_tensors_per_item = None
if FLAGS.mode == 'interpolate':
if FLAGS.input_midi_1 is None or FLAGS.input_midi_2 is None:
raise ValueError(
'`--input_midi_1` and `--input_midi_2` must be specified in '
'`interpolate` mode.')
input_midi_1 = os.path.expanduser(FLAGS.input_midi_1)
input_midi_2 = os.path.expanduser(FLAGS.input_midi_2)
if not os.path.exists(input_midi_1):
raise ValueError('Input MIDI 1 not found: %s' % FLAGS.input_midi_1)
if not os.path.exists(input_midi_2):
raise ValueError('Input MIDI 2 not found: %s' % FLAGS.input_midi_2)
input_1 = mm.midi_file_to_note_sequence(input_midi_1)
input_2 = mm.midi_file_to_note_sequence(input_midi_2)
def _check_extract_examples(input_ns, path, input_number):
"""Make sure each input returns exactly one example from the converter."""
tensors = config.data_converter.to_tensors(input_ns).outputs
if not tensors:
print(
'MusicVAE configs have very specific input requirements. Could not '
'extract any valid inputs from `%s`. Try another MIDI file.'
% path)
sys.exit()
elif len(tensors) > 1:
basename = os.path.join(
FLAGS.output_dir,
'%s_input%d-extractions_%s-*-of-%03d.mid' %
(FLAGS.config, input_number, date_and_time, len(tensors)))
for i, ns in enumerate(
config.data_converter.to_notesequences(tensors)):
mm.sequence_proto_to_midi_file(
ns, basename.replace('*', '%03d' % i))
print(
'%d valid inputs extracted from `%s`. Outputting these potential '
'inputs as `%s`. Call script again with one of these instead.'
% (len(tensors), path, basename))
sys.exit()
logging.info(
'Attempting to extract examples from input MIDIs using config `%s`...',
FLAGS.config)
_check_extract_examples(input_1, FLAGS.input_midi_1, 1)
_check_extract_examples(input_2, FLAGS.input_midi_2, 2)
logging.info('Loading model...')
if FLAGS.run_dir:
checkpoint_dir_or_path = os.path.expanduser(
os.path.join(FLAGS.run_dir, 'train'))
else:
checkpoint_dir_or_path = os.path.expanduser(FLAGS.checkpoint_file)
model = TrainedModel(config,
batch_size=min(FLAGS.max_batch_size,
FLAGS.num_outputs),
checkpoint_dir_or_path=checkpoint_dir_or_path)
if FLAGS.mode == 'interpolate':
logging.info('Interpolating...')
_, mu, _ = model.encode([input_1, input_2])
z = np.array([
_slerp(mu[0], mu[1], t)
for t in np.linspace(0, 1, FLAGS.num_outputs)
])
results = model.decode(length=config.hparams.max_seq_len,
z=z,
temperature=FLAGS.temperature)
elif FLAGS.mode == 'sample':
logging.info('Sampling...')
results = model.sample(n=FLAGS.num_outputs,
length=config.hparams.max_seq_len,
temperature=FLAGS.temperature)
elif FLAGS.mode == 'isample':
assert FLAGS.input_image is not None, 'Provide an image to sample from'
assert FLAGS.input_midi_1 is not None, 'Provide a music to sample from'
logging.info('Sampling z from image vae...')
img = cv2.imread(FLAGS.input_image) / 255.
img = np.asarray(cv2.resize(img, (320, 240)))
img = np.expand_dims(img, axis=0)
print(img.shape)
latent = None
input_midi_1 = os.path.expanduser(FLAGS.input_midi_1)
input_1 = mm.midi_file_to_note_sequence(input_midi_1)
def _check_extract_examples(input_ns, path, input_number):
"""Make sure each input returns exactly one example from the converter."""
tensors = config.data_converter.to_tensors(input_ns).outputs
if not tensors:
print(
'MusicVAE configs have very specific input requirements. Could not '
'extract any valid inputs from `%s`. Try another MIDI file.'
% path)
sys.exit()
elif len(tensors) > 1:
basename = os.path.join(
FLAGS.output_dir,
'%s_input%d-extractions_%s-*-of-%03d.mid' %
(FLAGS.config, input_number, date_and_time, len(tensors)))
for i, ns in enumerate(
config.data_converter.to_notesequences(tensors)):
mm.sequence_proto_to_midi_file(
ns, basename.replace('*', '%03d' % i))
print(
'%d valid inputs extracted from `%s`. Outputting these potential '
'inputs as `%s`. Call script again with one of these instead.'
% (len(tensors), path, basename))
sys.exit()
logging.info(
'Attempting to extract examples from input MIDIs using config `%s`...',
FLAGS.config)
_check_extract_examples(input_1, FLAGS.input_midi_1, 1)
with model._sess as sess:
z_music, mu_music, sigma_music = model.encode([input_1])
dataset = tf.data.Dataset.from_tensors(img.astype(np.float32))
img = dataset.repeat().make_one_shot_iterator().get_next()
mu, sigma = model.vae.encode(img, config.hparams)
mu = mu.eval()
sigma = sigma.eval()
latent = ds.MultivariateNormalDiag(loc=mu + mu_music,
scale_diag=sigma +
sigma_music).sample().eval()
results = model.decode(length=config.hparams.max_seq_len,
z=latent,
temperature=FLAGS.temperature)
print(results)
basename = os.path.join(
FLAGS.output_dir, '%s_%s_%s-*-of-%03d.mid' %
(FLAGS.config, FLAGS.mode, date_and_time, FLAGS.num_outputs))
logging.info('Outputting %d files as `%s`...', FLAGS.num_outputs, basename)
for i, ns in enumerate(results):
mm.sequence_proto_to_midi_file(ns, basename.replace('*', '%03d' % i))
logging.info('Done.')
def run(config_map):
"""Load model params, save config file and start trainer.
Args:
config_map: Dictionary mapping configuration name to Config object.
Raises:
ValueError: if required flags are missing or invalid.
"""
date_and_time = time.strftime('%Y-%m-%d_%H%M%S')
if FLAGS.run_dir is None == FLAGS.checkpoint_file is None:
raise ValueError(
'Exactly one of `--run_dir` or `--checkpoint_file` must be specified.')
if FLAGS.output_dir is None:
raise ValueError('`--output_dir` is required.')
tf.gfile.MakeDirs(FLAGS.output_dir)
if FLAGS.mode != 'sample' and FLAGS.mode != 'interpolate':
raise ValueError('Invalid value for `--mode`: %s' % FLAGS.mode)
if FLAGS.config not in config_map:
raise ValueError('Invalid config name: %s' % FLAGS.config)
config = config_map[FLAGS.config]
config.data_converter.max_tensors_per_item = None
if FLAGS.mode == 'interpolate':
if FLAGS.input_midi_1 is None or FLAGS.input_midi_2 is None:
raise ValueError(
'`--input_midi_1` and `--input_midi_2` must be specified in '
'`interpolate` mode.')
input_midi_1 = os.path.expanduser(FLAGS.input_midi_1)
input_midi_2 = os.path.expanduser(FLAGS.input_midi_2)
if not os.path.exists(input_midi_1):
raise ValueError('Input MIDI 1 not found: %s' % FLAGS.input_midi_1)
if not os.path.exists(input_midi_2):
raise ValueError('Input MIDI 2 not found: %s' % FLAGS.input_midi_2)
input_1 = mm.midi_file_to_note_sequence(input_midi_1)
input_2 = mm.midi_file_to_note_sequence(input_midi_2)
def _check_extract_examples(input_ns, path, input_number):
"""Make sure each input returns exactly one example from the converter."""
tensors = config.data_converter.to_tensors(input_ns).outputs
if not tensors:
print(
'MusicVAE configs have very specific input requirements. Could not '
'extract any valid inputs from `%s`. Try another MIDI file.' % path)
sys.exit()
elif len(tensors) > 1:
basename = os.path.join(
FLAGS.output_dir,
'%s_input%d-extractions_%s-*-of-%03d.mid' %
(FLAGS.config, input_number, date_and_time, len(tensors)))
for i, ns in enumerate(config.data_converter.to_notesequences(tensors)):
mm.sequence_proto_to_midi_file(ns, basename.replace('*', '%03d' % i))
print(
'%d valid inputs extracted from `%s`. Outputting these potential '
'inputs as `%s`. Call script again with one of these instead.' %
(len(tensors), path, basename))
sys.exit()
logging.info(
'Attempting to extract examples from input MIDIs using config `%s`...',
FLAGS.config)
_check_extract_examples(input_1, FLAGS.input_midi_1, 1)
_check_extract_examples(input_2, FLAGS.input_midi_2, 2)
logging.info('Loading model...')
if FLAGS.run_dir:
checkpoint_dir_or_path = os.path.expanduser(
os.path.join(FLAGS.run_dir, 'train'))
else:
checkpoint_dir_or_path = os.path.expanduser(FLAGS.checkpoint_file)
model = TrainedModel(
config, batch_size=min(FLAGS.max_batch_size, FLAGS.num_outputs),
checkpoint_dir_or_path=checkpoint_dir_or_path)
if FLAGS.mode == 'interpolate':
logging.info('Interpolating...')
_, mu, _ = model.encode([input_1, input_2])
z = np.array([
_slerp(mu[0], mu[1], t) for t in np.linspace(0, 1, FLAGS.num_outputs)])
results = model.decode(
length=config.hparams.max_seq_len,
z=z,
temperature=FLAGS.temperature)
elif FLAGS.mode == 'sample':
logging.info('Sampling...')
results = model.sample(
n=FLAGS.num_outputs,
length=config.hparams.max_seq_len,
temperature=FLAGS.temperature)
basename = os.path.join(
FLAGS.output_dir,
'%s_%s_%s-*-of-%03d.mid' %
(FLAGS.config, FLAGS.mode, date_and_time, FLAGS.num_outputs))
logging.info('Outputting %d files as `%s`...', FLAGS.num_outputs, basename)
for i, ns in enumerate(results):
mm.sequence_proto_to_midi_file(ns, basename.replace('*', '%03d' % i))
logging.info('Done.')