if args.teleop:
assert args.algo == "sac", "Teleoperation mode is not yet implemented for {}".format(
args.algo)
env = TeleopEnv(env, is_training=True)
model.set_env(env)
env.model = model
kwargs = {}
if args.log_interval > -1:
kwargs = {'log_interval': args.log_interval}
if args.algo == 'sac':
kwargs.update({
'callback':
create_callback(args.algo,
os.path.join(save_path, ENV_ID + "_best"),
verbose=1)
})
model.learn(n_timesteps, **kwargs)
if args.teleop:
env.wait()
env.exit()
time.sleep(0.5)
else:
# Close the connection properly
env.reset()
if isinstance(env, VecFrameStack):
env = env.venv
# HACK to bypass Monitor wrapper
exp_folder = args.trained_agent.split('.pkl')[0]
if normalize:
print("Loading saved running average")
env.load_running_average(exp_folder)
else:
# Train an agent from scratch
model = ALGOS[args.algo](env=env, tensorboard_log=tensorboard_log, verbose=1, **hyperparams)
kwargs = {}
if args.log_interval > -1:
kwargs.update({'log_interval': args.log_interval})
if args.algo == 'sac':
kwargs.update({'callback': create_callback(args.algo,
os.path.join(save_path, ENV_ID + "_best"),
verbose=1)})
kwargs.update({'save_path': save_path})
# Base policy
agent = None
if args.base_policy_path != '':
print("Loading Base Policy for JIRL ...")
agent = keras.models.load_model(args.base_policy_path)
kwargs.update({'base_policy': agent})
kwargs.update({'expert_guidance_steps': args.expert_guidance_steps})
kwargs.update({'joystick': js})
# Train agent using JIRL
model.learn_jirl(n_timesteps, **kwargs)