def run_forward_model(config, trial, env_name):
env = WrapperAtari(gym.make(env_name))
input_shape = env.observation_space.shape
action_dim = env.action_space.n
if config.n_env > 1:
env_list = []
print('Creating {0:d} environments'.format(config.n_env))
for i in range(config.n_env):
env_list.append(WrapperAtari(gym.make(env_name)))
print('Start training')
experiment = ExperimentNEnvPPO(env_name, env_list, config, input_shape, action_dim)
else:
experiment = ExperimentPPO(env_name, env, config)
experiment.add_preprocess(encode_state)
agent = PPOAtariForwardModelAgent(input_shape, action_dim, config, TYPE.discrete)
experiment.run_forward_model(agent, trial)
env.close()
if config.n_env > 1:
for i in range(config.n_env):
env_list[i].close()
def run_dop_ref_model(env_name, config, i):
env = create_env(env_name)
state_dim = env.observation_space.shape
action_dim = env.action_space.shape[0]
experiment = ExperimentPPO(env_name, env, config)
agent = PPOAerisDOPRefAgent(state_dim, action_dim, config, TYPE.continuous)
experiment.run_baseline(agent, i)
env.close()
def run_baseline(config, i):
env = gym.make('MountainCarContinuous-v0')
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
experiment = ExperimentPPO('MountainCarContinuous-v0', env, config)
agent = PPOSimpleAgent(state_dim, action_dim, config, TYPE.continuous)
experiment.run_baseline(agent, i)
env.close()
def run_baseline(config, i):
env = gym.make('Pendulum-v0')
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
experiment = ExperimentPPO('Pendulum-v0', env, config)
experiment.add_preprocess(encode)
agent = PPOAgent(state_dim, action_dim, config, TYPE.continuous)
experiment.run_baseline(agent, i)
env.close()
def test(config, path, env_name):
env = WrapperAtari(gym.make(env_name))
input_shape = env.observation_space.shape
action_dim = env.action_space.n
experiment = ExperimentPPO(env_name, env, config)
experiment.add_preprocess(encode_state)
agent = PPOAtariAgent(input_shape, action_dim, config, TYPE.discrete)
agent.load(path)
experiment.test(agent)
env.close()
def run_rnd_model(config, trial, env_name):
if config.n_env > 1:
print('Creating {0:d} environments'.format(config.n_env))
env = MultiEnvParallel([WrapperAtari(gym.make(env_name)) for _ in range(config.n_env)], config.n_env, config.num_threads)
else:
env = WrapperAtari(gym.make(env_name))
input_shape = env.observation_space.shape
action_dim = env.action_space.n
print('Start training')
if config.n_env > 1:
experiment = ExperimentNEnvPPO(env_name, env, config)
else:
experiment = ExperimentPPO(env_name, env, config)
experiment.add_preprocess(encode_state)
agent = PPOAtariRNDAgent(input_shape, action_dim, config, TYPE.discrete)
experiment.run_rnd_model(agent, trial)
env.close()
def run_baseline(env_name, config, trial):
if config.n_env > 1:
print('Creating {0:d} environments'.format(config.n_env))
env = MultiEnvParallel(
[create_env(env_name) for _ in range(config.n_env)], config.n_env,
config.num_threads)
else:
env = create_env(env_name)
input_shape = env.observation_space.shape
action_dim = env.action_space.shape[0]
print('Start training')
if config.n_env > 1:
experiment = ExperimentNEnvPPO(env_name, env, config)
else:
experiment = ExperimentPPO(env_name, env, config)
agent = PPOAerisAgent(input_shape, action_dim, config, TYPE.continuous)
experiment.run_baseline(agent, trial)
env.close()
def run_baseline(config, trial):
env = gym.make('CartPole-v0')
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.n
if config.n_env > 1:
env_list = []
print('Creating {0:d} environments'.format(config.n_env))
for i in range(config.n_env):
env_list.append(gym.make('CartPole-v0'))
print('Start training')
experiment = ExperimentNEnvPPO('Pitfall-v0', env_list, config)
else:
experiment = ExperimentPPO('Pitfall-v0', env, config)
agent = PPOSimpleAgent(state_dim, action_dim, config, TYPE.discrete)
experiment.run_baseline(agent, trial)
env.close()
if config.n_env > 1:
for i in range(config.n_env):
env_list[i].close()