def main(argv):
del argv # Unused
config = DotMap(getattr(config_maml, FLAGS.config))
print('MAML config: %s' % FLAGS.config)
tf.logging.info('MAML config: %s', FLAGS.config)
algo = maml_rl.MAMLReinforcementLearning(config)
sess_config = tf.ConfigProto(allow_soft_placement=True)
sess_config.gpu_options.allow_growth = True
with tf.Session(config=sess_config) as sess:
algo.init_logging(sess)
init = tf.global_variables_initializer()
sess.run(init)
done = False
while not done:
done, _ = algo.train(sess, 10)
algo.stop_logging()
def main(argv):
del argv # Unused
config = _load_config()
algo = maml_rl.MAMLReinforcementLearning(config,
logdir=FLAGS.model_dir,
save_config=False)
with tf.Session() as sess:
init = tf.global_variables_initializer()
sess.run(init)
algo.restore(sess, FLAGS.model_dir)
task = config.task_generator()
task_modifier = config.task_env_modifiers[FLAGS.test_task_index]
for attr in task_modifier:
task.__setattr__(attr, task_modifier[attr])
if FLAGS.eval_meta:
sum_reward = _rollout_and_save(
sess,
task,
algo.train_policies[0],
os.path.join(FLAGS.output_dir,
'task_{}'.format(FLAGS.test_task_index), 'meta'),
max_rollout_len=config.max_rollout_len)
print('Total reward for meta policy is: {}'.format(sum_reward))
if FLAGS.eval_finetune:
for step in range(FLAGS.num_finetune_steps):
tf.logging.info('Finetune step: {}'.format(step))
algo.finetune(sess, task_modifier)
sum_reward = _rollout_and_save(
sess,
task,
algo.train_policies[0],
os.path.join(FLAGS.output_dir,
'task_{}'.format(FLAGS.test_task_index),
'finetune_{}'.format(step)),
max_rollout_len=config.max_rollout_len)
print('Total reward for fine-tuned policy is: {}'.format(
sum_reward))