def default_dqn_hparams():
"""Generates the default hparams for RLTuner DQN model."""
return tf_lib.HParams(random_action_probability=0.1,
store_every_nth=1,
train_every_nth=5,
minibatch_size=32,
discount_rate=0.95,
max_experience=100000,
target_network_update_rate=0.01)
def default_hparams():
return tf_lib.HParams(batch_size=128,
rnn_layer_sizes=[128, 128],
dropout_keep_prob=0.5,
skip_first_n_losses=0,
clip_norm=5,
initial_learning_rate=0.01,
decay_steps=1000,
decay_rate=0.85)
def setUp(self):
self.encoder_decoder = melodies_lib.OneHotEncoderDecoder(0, 12, 0)
self.hparams = tf_lib.HParams(batch_size=128,
rnn_layer_sizes=[128, 128],
dropout_keep_prob=0.5,
skip_first_n_losses=0,
clip_norm=5,
initial_learning_rate=0.01,
decay_steps=1000,
decay_rate=0.85)
def main(_):
hparams = (rl_tuner_ops.basic_rnn_hparams() if FLAGS.note_rnn_type
== 'basic_rnn' else rl_tuner_ops.default_hparams())
dqn_hparams = tf_lib.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 basic_rnn_hparams():
"""Generates the hparams used to train a basic_rnn.
These are the hparams used in the .mag file found at
https://github.com/tensorflow/magenta/tree/master/magenta/models/
melody_rnn#pre-trained
Returns:
Hyperparameters of the downloadable basic_rnn pre-trained model.
"""
# TODO(natashajaques): ability to restore basic_rnn from any .mag file.
return tf_lib.HParams(batch_size=128,
rnn_layer_sizes=[512, 512],
one_hot_length=NUM_CLASSES)
def default_hparams():
"""Generates the hparams used to train note rnn used in paper."""
return tf_lib.HParams(use_dynamic_rnn=True,
batch_size=BATCH_SIZE,
lr=0.0002,
l2_reg=2.5e-5,
clip_norm=5,
initial_learning_rate=0.5,
decay_steps=1000,
decay_rate=0.85,
rnn_layer_sizes=[100],
skip_first_n_losses=32,
one_hot_length=NUM_CLASSES,
exponentially_decay_learning_rate=True)
def basic_rnn_hparams():
"""Generates the hparams used to train a basic_rnn.
These are the hparams used in the .mag file found at
https://github.com/tensorflow/magenta/tree/master/magenta/models/
melody_rnn#pre-trained
Returns:
Hyperparameters of the downloadable basic_rnn pre-trained model.
"""
# TODO(natashajaques): ability to restore basic_rnn from any .mag
# file.
return tf_lib.HParams(batch_size=128,
dropout_keep_prob=0.5,
clip_norm=5,
initial_learning_rate=0.01,
decay_steps=1000,
decay_rate=0.85,
rnn_layer_sizes=[512, 512],
skip_first_n_losses=0,
one_hot_length=NUM_CLASSES,
exponentially_decay_learning_rate=True)
