def __init__(self,
graph,
scope,
checkpoint_dir,
checkpoint_file=None,
midi_primer=None,
training_file_list=None,
hparams=None,
note_rnn_type='default',
checkpoint_scope='rnn_model'):
"""Initialize by building the graph and loading a previous checkpoint.
Args:
graph: A tensorflow graph where the MelodyRNN's graph will be added.
scope: The tensorflow scope where this network will be saved.
checkpoint_dir: Path to the directory where the checkpoint file is saved.
checkpoint_file: Path to a checkpoint file to be used if none can be
found in the checkpoint_dir
midi_primer: Path to a single midi file that can be used to prime the
model.
training_file_list: List of paths to tfrecord files containing melody
training data.
hparams: A tf_lib.HParams object. Must match the hparams used to create
the checkpoint file.
note_rnn_type: If 'default', will use the basic LSTM described in the
research paper. If 'basic_rnn', will assume the checkpoint is from a
Magenta basic_rnn model.
checkpoint_scope: The scope in lstm which the model was originally defined
when it was first trained.
"""
self.graph = graph
self.session = None
self.saver = None
self.scope = scope
self.batch_size = 1
self.midi_primer = midi_primer
self.checkpoint_scope = checkpoint_scope
self.note_rnn_type = note_rnn_type
self.training_file_list = training_file_list
self.checkpoint_dir = checkpoint_dir
self.checkpoint_file = checkpoint_file
if hparams is not None:
tf.logging.info('Using custom hparams')
self.hparams = hparams
else:
tf.logging.info('Empty hparams string. Using defaults')
self.hparams = rl_tuner_ops.default_hparams()
self.build_graph()
self.state_value = self.get_zero_state()
if midi_primer is not None:
self.load_primer()
self.variable_names = rl_tuner_ops.get_variable_names(
self.graph, self.scope)
self.transpose_amount = 0
def main(_):
hparams = rl_tuner_ops.default_hparams()
if FLAGS.note_rnn_type == 'basic_rnn':
hparams = rl_tuner_ops.basic_rnn_hparams()
elif FLAGS.note_rnn_type == 'attention_rnn':
hparams = rl_tuner_ops.attention_rnn_hparams()
dqn_hparams = tf.contrib.training.HParams(random_action_probability=0.1,
store_every_nth=1,
train_every_nth=5,
minibatch_size=32,
discount_rate=0.5,
max_experience=100000,
target_network_update_rate=0.01)
output_dir = os.path.join(FLAGS.output_dir, FLAGS.algorithm)
output_ckpt = FLAGS.algorithm + '.ckpt'
backup_checkpoint_file = os.path.join(FLAGS.note_rnn_checkpoint_dir,
FLAGS.note_rnn_checkpoint_name)
rlt = rl_tuner.RLTuner(output_dir,
midi_primer=FLAGS.midi_primer,
dqn_hparams=dqn_hparams,
reward_scaler=FLAGS.reward_scaler,
save_name=output_ckpt,
output_every_nth=FLAGS.output_every_nth,
note_rnn_checkpoint_dir=FLAGS.note_rnn_checkpoint_dir,
note_rnn_checkpoint_file=backup_checkpoint_file,
note_rnn_type=FLAGS.note_rnn_type,
note_rnn_hparams=hparams,
num_notes_in_melody=FLAGS.num_notes_in_melody,
exploration_mode=FLAGS.exploration_mode,
algorithm=FLAGS.algorithm)
tf.logging.info('Saving images and melodies to: %s', rlt.output_dir)
tf.logging.info('Training...')
rlt.train(num_steps=FLAGS.training_steps,
exploration_period=FLAGS.exploration_steps)
tf.logging.info('Finished training. Saving output figures and composition.')
rlt.plot_rewards(image_name='Rewards-' + FLAGS.algorithm + '.eps')
rlt.generate_music_sequence(visualize_probs=True, title=FLAGS.algorithm,
prob_image_name=FLAGS.algorithm + '.png')
rlt.save_model_and_figs(FLAGS.algorithm)
tf.logging.info('Calculating music theory metric stats for 1000 '
'compositions.')
statistics = rlt.evaluate_music_theory_metrics(num_compositions=1000)
print("music theory evaluation statistic: {0}".format(statistics))
def __init__(self, graph, scope, checkpoint_dir, checkpoint_file=None,
midi_primer=None, training_file_list=None, hparams=None,
note_rnn_type='default', checkpoint_scope='rnn_model'):
"""Initialize by building the graph and loading a previous checkpoint.
Args:
graph: A tensorflow graph where the MelodyRNN's graph will be added.
scope: The tensorflow scope where this network will be saved.
checkpoint_dir: Path to the directory where the checkpoint file is saved.
checkpoint_file: Path to a checkpoint file to be used if none can be
found in the checkpoint_dir
midi_primer: Path to a single midi file that can be used to prime the
model.
training_file_list: List of paths to tfrecord files containing melody
training data.
hparams: A tf_lib.HParams object. Must match the hparams used to create
the checkpoint file.
note_rnn_type: If 'default', will use the basic LSTM described in the
research paper. If 'basic_rnn', will assume the checkpoint is from a
Magenta basic_rnn model.
checkpoint_scope: The scope in lstm which the model was originally defined
when it was first trained.
"""
self.graph = graph
self.session = None
self.scope = scope
self.batch_size = 1
self.midi_primer = midi_primer
self.checkpoint_scope = checkpoint_scope
self.note_rnn_type = note_rnn_type
self.training_file_list = training_file_list
self.checkpoint_dir = checkpoint_dir
self.checkpoint_file = checkpoint_file
if hparams is not None:
tf.logging.info('Using custom hparams')
self.hparams = hparams
else:
tf.logging.info('Empty hparams string. Using defaults')
self.hparams = rl_tuner_ops.default_hparams()
self.build_graph()
self.state_value = self.get_zero_state()
if midi_primer is not None:
self.load_primer()
self.variable_names = rl_tuner_ops.get_variable_names(self.graph,
self.scope)
self.transpose_amount = 0
def main(_):
if FLAGS.note_rnn_type == 'basic_rnn':
hparams = rl_tuner_ops.basic_rnn_hparams()
else:
hparams = rl_tuner_ops.default_hparams()
dqn_hparams = tf.contrib.training.HParams(random_action_probability=0.1,
store_every_nth=1,
train_every_nth=5,
minibatch_size=32,
discount_rate=0.5,
max_experience=100000,
target_network_update_rate=0.01)
output_dir = os.path.join(FLAGS.output_dir, FLAGS.algorithm)
output_ckpt = FLAGS.algorithm + '.ckpt'
backup_checkpoint_file = os.path.join(FLAGS.note_rnn_checkpoint_dir,
FLAGS.note_rnn_checkpoint_name)
rlt = rl_tuner.RLTuner(output_dir,
midi_primer=FLAGS.midi_primer,
dqn_hparams=dqn_hparams,
reward_scaler=FLAGS.reward_scaler,
save_name=output_ckpt,
output_every_nth=FLAGS.output_every_nth,
note_rnn_checkpoint_dir=FLAGS.note_rnn_checkpoint_dir,
note_rnn_checkpoint_file=backup_checkpoint_file,
note_rnn_type=FLAGS.note_rnn_type,
note_rnn_hparams=hparams,
num_notes_in_melody=FLAGS.num_notes_in_melody,
exploration_mode=FLAGS.exploration_mode,
algorithm=FLAGS.algorithm)
tf.logging.info('Saving images and melodies to: %s', rlt.output_dir)
tf.logging.info('Training...')
rlt.train(num_steps=FLAGS.training_steps,
exploration_period=FLAGS.exploration_steps)
tf.logging.info('Finished training. Saving output figures and composition.')
rlt.plot_rewards(image_name='Rewards-' + FLAGS.algorithm + '.eps')
rlt.generate_music_sequence(visualize_probs=True, title=FLAGS.algorithm,
prob_image_name=FLAGS.algorithm + '.png')
rlt.save_model_and_figs(FLAGS.algorithm)
tf.logging.info('Calculating music theory metric stats for 1000 '
'compositions.')
rlt.evaluate_music_theory_metrics(num_compositions=1000)
def main(_):
if FLAGS.note_rnn_type == 'basic_rnn':
hparams = rl_tuner_ops.basic_rnn_hparams()
else:
hparams = rl_tuner_ops.default_hparams()
dqn_hparams = tf.contrib.training.HParams(random_action_probability=0.1,
store_every_nth=1,
train_every_nth=5,
minibatch_size=32,
discount_rate=0.5,
max_experience=100000,
target_network_update_rate=0.01)
num_compositions = 100000
if (FLAGS.running_mode == 'comparison'):
tf.logging.info("Running for comparison")
# num_compositions = 1000
defaultRlt = run_algorithm(FLAGS.algorithm, dqn_hparams, hparams, num_compositions)
if (FLAGS.running_mode == 'comparison'):
pureRlt = run_algorithm('pure_rl', dqn_hparams, hparams, num_compositions)
plot_comparison(defaultRlt, pureRlt)
