def train(env_id, num_timesteps, seed, lam, sgd_steps, klcoeff, log):
"""
Train TRPO model for the mujoco environment, for testing purposes
:param env_id: (str) Environment ID
:param num_timesteps: (int) The total number of samples
:param seed: (int) The initial seed for training
"""
with tf_util.single_threaded_session():
rank = MPI.COMM_WORLD.Get_rank()
log_path = './experiments/' + str(
env_id) + './OURS-LOADED/noent_klcoeffanneal_samesgdsteps' + str(
sgd_steps) + '_longer_wgae0.95_exp1_2_' + str(seed)
#log_path = './experiments/'+str(env_id)+'./TRPO-3x/TRPOR-oldsampling/noent_klcoeff'+str(sgd_steps)+'_sgdstep_steps5_'+str(seed)
if not log:
if rank == 0:
logger.configure(log_path)
else:
logger.configure(log_path, format_strs=[])
logger.set_level(logger.DISABLED)
else:
if rank == 0:
logger.configure()
else:
logger.configure(format_strs=[])
logger.set_level(logger.DISABLED)
workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
#env = make_mujoco_env(env_id, workerseed)
def make_env():
env_out = gym.make(env_id)
env_out = bench.Monitor(env_out,
logger.get_dir(),
allow_early_resets=True)
env_out.seed(seed)
return env_out
env = DummyVecEnv([make_env])
env = VecNormalize(env) #, norm_reward=False, norm_obs=False)
#env = VecNormalize(env)
model = TRPO(MlpPolicy,
env,
timesteps_per_batch=2048,
max_kl=0.01,
cg_iters=10,
cg_damping=0.1,
entcoeff=0.0,
gamma=0.99,
lam=0.95,
vf_iters=5,
vf_stepsize=1e-3,
verbose=1,
seed=seed,
sgd_steps=sgd_steps,
klcoeff=klcoeff,
method="multistep-SGD")
model.learn(total_timesteps=10e6) #num_timesteps, seed=seed)
env.close()
def __exit__(self, exc_type, exc_val, exc_tb):
if self.verbose <= 1:
os.environ['TF_CPP_MIN_LOG_LEVEL'] = self.tf_level
if self.verbose <= 0:
logger.set_level(self.log_level)
gym.logger.set_level(self.gym_level)
def train(env_id, num_timesteps, seed):
"""
Train TRPO model for the mujoco environment, for testing purposes
:param env_id: (str) Environment ID
:param num_timesteps: (int) The total number of samples
:param seed: (int) The initial seed for training
"""
with tf_util.single_threaded_session():
rank = MPI.COMM_WORLD.Get_rank()
if rank == 0:
logger.configure()
else:
logger.configure(format_strs=[])
logger.set_level(logger.DISABLED)
workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
tblog = "/cvgl2/u/surajn/workspace/tb_logs/reacher/"
env = make_mujoco_env(env_id, workerseed)
model = TRPO(MlpPolicy,
env,
timesteps_per_batch=1024,
max_kl=0.01,
cg_iters=10,
cg_damping=0.1,
entcoeff=0.0,
gamma=0.99,
lam=0.98,
vf_iters=5,
vf_stepsize=1e-3,
tensorboard_log)
model.learn(total_timesteps=num_timesteps)
env.close()
def train(env_id, num_timesteps, run, kappa, vf_phi_update_interval, log):
"""
Train TRPO model for the mujoco environment, for testing purposes
:param env_id: (str) Environment ID
:param num_timesteps: (int) The total number of samples
:param seed: (int) The initial seed for training
"""
with tf_util.single_threaded_session():
rank = MPI.COMM_WORLD.Get_rank()
log_path = './experiments/'+str(env_id)+'./updated_nkappa_x7_ent_0.01_new/'+str(kappa)+'_'+str(vf_phi_update_interval)+'_'+str(run)
if not log:
if rank == 0:
logger.configure(log_path)
else:
logger.configure(log_path, format_strs=[])
logger.set_level(logger.DISABLED)
else:
if rank == 0:
logger.configure()
else:
logger.configure(format_strs=[])
logger.set_level(logger.DISABLED)
seed = run
workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
#set_global_seeds(run)
env = make_mujoco_env(env_id, workerseed)
test_env = None#make_mujoco_env(env_id, workerseed)
model = TRPO(MlpPolicy, env, test_env=test_env, timesteps_per_batch=1024, max_kl=0.01, cg_iters=10, cg_damping=0.1, entcoeff=0.01,
gamma=0.99, kappa=kappa, vf_iters=5, vf_stepsize=1e-3, verbose=1, vf_phi_update_interval=vf_phi_update_interval, seed=run)
model.learn(total_timesteps=int(2e6), seed=run)
#model.save("./"+str(env_id)+"./models/"+str(kappa)+"_"+str(run)+'_xnew_longer_slower'+str(vf_phi_update_interval)+'.pkl')
env.close()
def train(env_id, num_timesteps, seed, lam, sgd_steps, klcoeff, log):
"""
Train TRPO model for the mujoco environment, for testing purposes
:param env_id: (str) Environment ID
:param num_timesteps: (int) The total number of samples
:param seed: (int) The initial seed for training
"""
with tf_util.single_threaded_session():
rank = MPI.COMM_WORLD.Get_rank()
log_path = './experiments/' + str(
env_id) + './SAC-M/nips_test19/m' + str(sgd_steps) + '_c' + str(
0.5) + '_e' + str(klcoeff) + '_' + str(seed)
#log_path = './experiments/'+str(env_id)+'./TRPO-3x/TRPOR-oldsampling/noent_klcoeff'+str(sgd_steps)+'_sgdstep_steps5_'+str(seed)
if not log:
if rank == 0:
logger.configure(log_path)
else:
logger.configure(log_path, format_strs=[])
logger.set_level(logger.DISABLED)
else:
if rank == 0:
logger.configure()
else:
logger.configure(format_strs=[])
logger.set_level(logger.DISABLED)
workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
#env = make_mujoco_env(env_id, workerseed)
def make_env():
env_out = gym.make(env_id)
env_out = bench.Monitor(env_out,
logger.get_dir(),
allow_early_resets=True)
env_out.seed(seed)
return env_out
env = DummyVecEnv([make_env])
env = VecNormalize(env, norm_reward=False, norm_obs=False)
#env = VecNormalize(env)
model = MDPO(MlpPolicy,
env,
gamma=0.99,
verbose=1,
seed=seed,
buffer_size=1000000,
ent_coef=1.0,
gradient_steps=sgd_steps,
lam=klcoeff,
train_freq=1,
tsallis_q=1,
reparameterize=True,
klconst=0.5)
model.learn(
total_timesteps=int(num_timesteps)) #num_timesteps, seed=seed)
env.close()
def __enter__(self):
self.tf_level = os.environ.get('TF_CPP_MIN_LOG_LEVEL', '0')
self.log_level = logger.get_level()
self.gym_level = gym.logger.MIN_LEVEL
if self.verbose <= 1:
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
if self.verbose <= 0:
logger.set_level(logger.DISABLED)
gym.logger.set_level(gym.logger.DISABLED)
def run(env_id, seed, noise_type, layer_norm, evaluation, **kwargs):
"""
run the training of DDPG
:param env_id: (str) the environment ID
:param seed: (int) the initial random seed
:param noise_type: (str) the wanted noises ('adaptive-param', 'normal' or 'ou'), can use multiple noise type by
seperating them with commas
:param layer_norm: (bool) use layer normalization
:param evaluation: (bool) enable evaluation of DDPG training
:param kwargs: (dict) extra keywords for the training.train function
"""
# Configure things.
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
logger.set_level(logger.DISABLED)
# Create envs.
env = gym.make(env_id)
env = bench.Monitor(
env,
logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
if evaluation and rank == 0:
eval_env = gym.make(env_id)
eval_env = bench.Monitor(eval_env,
os.path.join(logger.get_dir(), 'gym_eval'))
env = bench.Monitor(env, None)
else:
eval_env = None
# Parse noise_type
action_noise = None
param_noise = None
nb_actions = env.action_space.shape[-1]
for current_noise_type in noise_type.split(','):
current_noise_type = current_noise_type.strip()
if current_noise_type == 'none':
pass
elif 'adaptive-param' in current_noise_type:
_, stddev = current_noise_type.split('_')
param_noise = AdaptiveParamNoiseSpec(
initial_stddev=float(stddev),
desired_action_stddev=float(stddev))
elif 'normal' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = NormalActionNoise(mean=np.zeros(nb_actions),
sigma=float(stddev) *
np.ones(nb_actions))
elif 'ou' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = OrnsteinUhlenbeckActionNoise(
mean=np.zeros(nb_actions),
sigma=float(stddev) * np.ones(nb_actions))
else:
raise RuntimeError(
'unknown noise type "{}"'.format(current_noise_type))
# Seed everything to make things reproducible.
seed = seed + 1000000 * rank
logger.info('rank {}: seed={}, logdir={}'.format(rank, seed,
logger.get_dir()))
tf.reset_default_graph()
set_global_seeds(seed)
env.seed(seed)
if eval_env is not None:
eval_env.seed(seed)
# Disable logging for rank != 0 to avoid noise.
start_time = 0
if rank == 0:
start_time = time.time()
model = DDPG(policy=MlpPolicy,
env=env,
memory_policy=Memory,
eval_env=eval_env,
param_noise=param_noise,
action_noise=action_noise,
memory_limit=int(1e6),
layer_norm=layer_norm,
verbose=2,
**kwargs)
model.learn(total_timesteps=10000)
env.close()
if eval_env is not None:
eval_env.close()
if rank == 0:
logger.info('total runtime: {}s'.format(time.time() - start_time))
def main(args):
logger.configure(SIMPLE.config.LOGDIR)
if args.debug:
logger.set_level(SIMPLE.config.DEBUG)
else:
logger.set_level(SIMPLE.config.INFO)
#make environment
env = get_environment(args.env_name)(verbose=args.verbose,
manual=args.manual)
env.seed(args.seed)
total_rewards = {}
if args.recommend:
ppo_model = load_model(env, 'best_model.zip')
ppo_agent = Agent('best_model', ppo_model)
else:
ppo_agent = None
agents = []
#load the agents
if len(args.agents) != env.n_players:
raise Exception(
f'{len(args.agents)} players specified but this is a {env.n_players} player game!'
)
for i, agent in enumerate(args.agents):
if agent == 'human':
agent_obj = Agent('human')
elif agent == 'rules':
agent_obj = Agent('rules')
elif agent == 'base':
base_model = load_model(env, 'base.zip')
agent_obj = Agent('base', base_model)
else:
ppo_model = load_model(env, f'{agent}.zip')
agent_obj = Agent(agent, ppo_model)
agents.append(agent_obj)
total_rewards[agent_obj.id] = 0
#play games
logger.info(f'\nPlaying {args.games} games...')
for game in range(args.games):
players = agents[:]
if args.randomise_players:
random.shuffle(players)
obs = env.reset()
done = False
for i, p in enumerate(players):
logger.debug(f'Player {i+1} = {p.name}')
while not done:
current_player = players[env.current_player_num]
env.render()
logger.debug(f'\nCurrent player name: {current_player.name}')
if args.recommend and current_player.name in ['human', 'rules']:
# show recommendation from last loaded model
logger.debug(f'\nRecommendation by {ppo_agent.name}:')
action = ppo_agent.choose_action(env,
choose_best_action=True,
mask_invalid_actions=True)
if current_player.name == 'human':
action = input('\nPlease choose an action: ')
try:
# for int actions
action = int(action)
except:
# for MulitDiscrete action input as list TODO
action = eval(action)
elif current_player.name == 'rules':
logger.debug(f'\n{current_player.name} model choices')
action = current_player.choose_action(
env, choose_best_action=False, mask_invalid_actions=True)
else:
logger.debug(f'\n{current_player.name} model choices')
action = current_player.choose_action(
env,
choose_best_action=args.best,
mask_invalid_actions=True)
obs, reward, done, _ = env.step(action)
for r, player in zip(reward, players):
total_rewards[player.id] += r
player.points += r
if args.cont:
input('Press any key to continue')
env.render()
logger.info(f"Played {game + 1} games: {total_rewards}")
if args.write_results:
write_results(players, game, args.games, env.turns_taken)
for p in players:
p.points = 0
env.close()
def main(args):
rank = MPI.COMM_WORLD.Get_rank()
model_dir = os.path.join(config.MODELDIR, args.env_name)
if rank == 0:
try:
os.makedirs(model_dir)
except:
pass
if args.reset:
reset_files(model_dir)
logger.configure(config.LOGDIR)
else:
logger.configure(format_strs=[])
if args.debug:
logger.set_level(config.DEBUG)
else:
time.sleep(5)
logger.set_level(config.INFO)
workerseed = args.seed + 10000 * MPI.COMM_WORLD.Get_rank()
set_global_seeds(workerseed)
logger.info('\nSetting up the selfplay training environment opponents...')
base_env = get_environment(args.env_name)
env = selfplay_wrapper(base_env)(opponent_type=args.opponent_type,
verbose=args.verbose)
env.seed(workerseed)
CustomPolicy = get_network_arch(args.env_name)
params = {
'gamma': args.gamma,
'timesteps_per_actorbatch': args.timesteps_per_actorbatch,
'clip_param': args.clip_param,
'entcoeff': args.entcoeff,
'optim_epochs': args.optim_epochs,
'optim_stepsize': args.optim_stepsize,
'optim_batchsize': args.optim_batchsize,
'lam': args.lam,
'adam_epsilon': args.adam_epsilon,
'schedule': 'linear',
'verbose': 1,
'tensorboard_log': config.LOGDIR
}
time.sleep(
5
) # allow time for the base model to be saved out when the environment is created
if args.reset or not os.path.exists(
os.path.join(model_dir, 'best_model.zip')):
logger.info('\nLoading the base PPO agent to train...')
model = PPO1.load(os.path.join(model_dir, 'base.zip'), env, **params)
else:
logger.info(
'\nLoading the best_model.zip PPO agent to continue training...')
model = PPO1.load(os.path.join(model_dir, 'best_model.zip'), env,
**params)
#Callbacks
logger.info(
'\nSetting up the selfplay evaluation environment opponents...')
callback_args = {
'eval_env':
selfplay_wrapper(base_env)(opponent_type=args.opponent_type,
verbose=args.verbose),
'best_model_save_path':
config.TMPMODELDIR,
'log_path':
config.LOGDIR,
'eval_freq':
args.eval_freq,
'n_eval_episodes':
args.n_eval_episodes,
'deterministic':
False,
'render':
True,
'verbose':
0
}
if args.rules:
logger.info(
'\nSetting up the evaluation environment against the rules-based agent...'
)
# Evaluate against a 'rules' agent as well
eval_actual_callback = EvalCallback(
eval_env=selfplay_wrapper(base_env)(opponent_type='rules',
verbose=args.verbose),
eval_freq=1,
n_eval_episodes=args.n_eval_episodes,
deterministic=args.best,
render=True,
verbose=0)
callback_args['callback_on_new_best'] = eval_actual_callback
# Evaluate the agent against previous versions
eval_callback = SelfPlayCallback(args.opponent_type, args.threshold,
args.env_name, **callback_args)
logger.info('\nSetup complete - commencing learning...\n')
model.learn(total_timesteps=int(1e9),
callback=[eval_callback],
reset_num_timesteps=False,
tb_log_name="tb")
env.close()
del env
def train(env_id, algo, num_timesteps, seed, sgd_steps, t_pi, t_c, lam, log,
expert_path, pretrain, pretrain_epochs, mdpo_update_steps,
num_trajectories, expert_model, exploration_bonus, bonus_coef,
random_action_len, is_action_features, dir_name, neural, lipschitz,
args):
"""
Train TRPO model for the mujoco environment, for testing purposes
:param env_id: (str) Environment ID
:param num_timesteps: (int) The total number of samples
:param seed: (int) The initial seed for training
"""
with tf_util.single_threaded_session():
# from mpi4py import MPI
# rank = MPI.COMM_WORLD.Get_rank()
rank = 0
env_name = env_id[:-3].lower()
log_dir = './experiments/' + env_name + '/' + str(algo).lower() + '/'\
+ 'tpi' + str(t_pi) + '_tc' + str(t_c) + '_lam' + str(lam)
log_dir += '_' + dir_name + '/'
log_name = str(algo) + '_updateSteps' + str(mdpo_update_steps)
# log_name += '_randLen' + str(random_action_len)
if exploration_bonus:
log_name += '_exploration' + str(bonus_coef)
if pretrain:
log_name += '_pretrain' + str(pretrain_epochs)
if not is_action_features:
log_name += "_states_only"
log_name += '_s' + str(seed)
log_path = log_dir + log_name
expert_path = './experts/' + expert_path
num_timesteps = int(num_timesteps)
args = args.__dict__
dir_path = os.getcwd() + log_dir[1:]
if not os.path.exists(dir_path):
os.makedirs(dir_path)
with open(os.getcwd() + log_dir[1:] + 'args.txt', 'w') as file:
file.write("Experiment Arguments:")
for key, val in args.items():
print(key, ": ", val, file=file)
if log:
if rank == 0:
logger.configure(log_path)
else:
logger.configure(log_path, format_strs=[])
logger.set_level(logger.DISABLED)
else:
if rank == 0:
logger.configure()
else:
logger.configure(format_strs=[])
logger.set_level(logger.DISABLED)
# workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
# env = make_mujoco_env(env_id, workerseed)
def make_env():
# env_out = gym.make(env_id, reset_noise_scale=1.0)
env_out = gym.make(env_id)
env_out = bench.Monitor(env_out,
logger.get_dir(),
allow_early_resets=True)
env_out.seed(seed)
env_out = wrap_mujoco(env_out, random_action_len=random_action_len)
return env_out
#
env = DummyVecEnv([make_env])
# env = VecNormalize(env)
if algo == 'Train':
train = True
else:
train = False
if algo == 'Evaluate':
eval = True
else:
eval = False
if train:
from stable_baselines import SAC
env = VecNormalize(env, norm_reward=False, norm_obs=False)
if num_timesteps > 0:
model = SAC('MlpPolicy',
env_id,
verbose=1,
buffer_size=1000000,
batch_size=256,
ent_coef='auto',
train_freq=1,
tau=0.01,
gradient_steps=1,
learning_starts=10000)
else:
model = SAC.load(expert_model, env)
generate_expert_traj(model,
expert_path,
n_timesteps=num_timesteps,
n_episodes=num_trajectories)
if num_timesteps > 0:
model.save('sac_' + env_name + '_' + str(num_timesteps))
elif eval:
from stable_baselines import SAC
env = VecNormalize(env, norm_reward=False, norm_obs=False)
model = SAC.load(expert_model, env)
generate_expert_traj(model,
expert_path,
n_timesteps=num_timesteps,
n_episodes=10,
evaluate=True)
else:
expert_path = expert_path + '.npz'
dataset = ExpertDataset(expert_path=expert_path,
traj_limitation=10,
verbose=1)
if algo == 'MDAL':
model = MDAL_MDPO_OFF('MlpPolicy',
env,
dataset,
verbose=1,
tensorboard_log="./experiments/" +
env_name + "/mdal/",
seed=seed,
buffer_size=1000000,
ent_coef=0.0,
learning_starts=10000,
batch_size=256,
tau=0.01,
gamma=0.99,
gradient_steps=sgd_steps,
mdpo_update_steps=mdpo_update_steps,
lam=0.0,
train_freq=1,
d_step=10,
tsallis_q=1,
reparameterize=True,
t_pi=t_pi,
t_c=t_c,
exploration_bonus=exploration_bonus,
bonus_coef=bonus_coef,
is_action_features=is_action_features,
neural=neural,
lipschitz=lipschitz)
elif algo == 'MDAL_ON_POLICY':
model = MDAL_MDPO_ON('MlpPolicy',
env,
dataset,
verbose=1,
timesteps_per_batch=2048,
tensorboard_log="./experiments/" +
env_name + "/mdal_mdpo_on/",
seed=seed,
max_kl=0.01,
cg_iters=10,
cg_damping=0.1,
entcoeff=0.0,
adversary_entcoeff=0.001,
gamma=0.99,
lam=0.95,
vf_iters=5,
vf_stepsize=1e-3,
sgd_steps=sgd_steps,
klcoeff=1.0,
method="multistep-SGD",
tsallis_q=1.0,
t_pi=t_pi,
t_c=t_c,
exploration_bonus=exploration_bonus,
bonus_coef=bonus_coef,
is_action_features=is_action_features,
neural=neural)
elif algo == 'MDAL_TRPO':
model = MDAL_TRPO('MlpPolicy',
env,
dataset,
verbose=1,
tensorboard_log="./experiments/" + env_name +
"/mdal_trpo/",
seed=seed,
gamma=0.99,
g_step=3,
d_step=5,
sgd_steps=1,
d_stepsize=9e-5,
entcoeff=0.0,
adversary_entcoeff=0.001,
max_kl=t_pi,
t_pi=t_pi,
t_c=t_c,
exploration_bonus=exploration_bonus,
bonus_coef=bonus_coef,
is_action_features=is_action_features,
neural=neural,
lam=0.98,
timesteps_per_batch=2000,
lipschitz=lipschitz)
elif algo == 'GAIL':
from mpi4py import MPI
from stable_baselines import GAIL
model = GAIL('MlpPolicy',
env,
dataset,
verbose=1,
tensorboard_log="./experiments/" + env_name +
"/gail/",
seed=seed,
entcoeff=0.0,
adversary_entcoeff=0.001,
lipschitz=lipschitz)
elif algo == 'GAIL_MDPO_OFF':
# from mpi4py import MPI
from stable_baselines import GAIL_MDPO_OFF
model = GAIL_MDPO_OFF('MlpPolicy',
env,
dataset,
verbose=1,
tensorboard_log="./experiments/" +
env_name + "/gail_mdpo_off/",
seed=seed,
ent_coef=0.0,
adversary_entcoeff=0.001,
buffer_size=1000000,
learning_starts=10000,
batch_size=256,
tau=0.01,
gamma=0.99,
gradient_steps=sgd_steps,
mdpo_update_steps=mdpo_update_steps,
lam=0.0,
train_freq=1,
tsallis_q=1,
reparameterize=True,
t_pi=t_pi,
t_c=t_c,
exploration_bonus=exploration_bonus,
bonus_coef=bonus_coef,
is_action_features=is_action_features,
lipschitz=lipschitz)
else:
raise ValueError("Not a valid algorithm.")
if pretrain:
model.pretrain(dataset, n_epochs=pretrain_epochs)
model.learn(total_timesteps=num_timesteps, tb_log_name=log_name)
env.close()
def train(params, model=None, path=None):
if model: # indicate in filename that this is a finetune
if params['name']:
params['name'] += '_Finetune'
else:
params['name'] = 'Finetune'
data_dir, tb_path = get_paths(params, path=path)
print("Training Parameters: ", params)
os.makedirs(data_dir, exist_ok=True)
# Save parameters immediatly
params.save(data_dir)
rank = mpi_rank_or_zero()
if rank != 0:
logger.set_level(logger.DISABLED)
def make_env(i):
env = get_env(params)
env = Monitor(env, data_dir + '/' + str(i), allow_early_resets=params['early_reset'])
return env
use_her = params['env_args']['use_her'] if 'use_her' in params['env_args'] else False
if use_her:
env = make_env(0)
goal_selection_strategy = 'future'
else:
env = DummyVecEnv([(lambda n: lambda: make_env(n))(i) for i in range(params['num_proc'])])
if model: # indicate in filename that this is a finetune
print("Model action space", model.action_space, model.action_space.low)
print("Env action space", env.action_space, env.action_space.low)
if params['normalize']:
env = VecNormalize(env)
if params['seed']:
seed = params['seed'] + 100000 * rank
set_global_seeds(seed)
params['alg_args']['seed'] = seed
if 'noise' in params and params['noise']:
from stable_baselines.ddpg import OrnsteinUhlenbeckActionNoise
n_actions = env.action_space.shape[-1]
params['alg_args']['action_noise'] = OrnsteinUhlenbeckActionNoise(mean=np.zeros(n_actions), sigma=float(params['noise'])*np.ones(n_actions))
if model is None:
alg = get_alg(params)
policy = get_policy(params)
if use_her:
from stable_baselines import HER
model = HER(policy, env, alg, n_sampled_goal=4, goal_selection_strategy=goal_selection_strategy, verbose=1,
tensorboard_log=tb_path, policy_kwargs=params['policy_args'], **params['alg_args'])
else:
model = alg(policy, env, verbose=1, tensorboard_log=tb_path, policy_kwargs=params['policy_args'], **params['alg_args'])
else:
model.set_env(env)
model.learn(total_timesteps=params['timesteps'], log_interval=params['log_interval'], callback=create_training_callback(data_dir,
freq=params['eval_freq'], checkpoint_freq=params['checkpoint_freq']))
print("######## SAVING MODEL TO", data_dir)
model.save(data_dir +'/final_model')
if params['normalize']:
env.save(data_dir + '/normalized_environment.env')
env.close()
def train(params, model=None, env=None):
print("Training Parameters: ", params)
data_dir, tb_path = get_paths(params)
os.makedirs(data_dir, exist_ok=True)
# Save parameters immediately
params.save(data_dir)
rank = mpi_rank_or_zero()
if rank != 0:
logger.set_level(logger.DISABLED)
# Create the environment if not given
if env is None:
def make_env(i):
env = get_env(params)
print("ENV IN UTIL" ,env)
# TODO: make monitor work for multiple agent.
env = Monitor(env, data_dir + '/' + str(i), allow_early_resets=params['early_reset'])
return env
# if 'PPO' in params['alg']:
# env = DummyVecEnv([(lambda n: lambda: make_env(n))(i) for i in range(params['num_proc'])])
# else:
# env = make_env(0)
env = make_env(0)
if params['normalize']:
env = VecNormalize(env)
# Set the seeds
if params['seed']:
seed = params['seed'] + 100000 * rank
set_global_seeds(seed)
params['alg_args']['seed'] = seed
if 'noise' in params and params['noise']:
from stable_baselines.ddpg import OrnsteinUhlenbeckActionNoise
n_actions = env.action_space.shape[-1]
params['alg_args']['action_noise'] = OrnsteinUhlenbeckActionNoise(mean=np.zeros(n_actions), sigma=float(params['noise'])*np.ones(n_actions))
print("ENV", env, env.action_space)
if model is None:
alg = get_alg(params)
policy = get_policy(params)
model = alg(policy, env, verbose=1, tensorboard_log=tb_path, policy_kwargs=params['policy_args'], **params['alg_args'])
else:
model.set_env(env)
print("\n===============================\n")
print("TENSORBOARD PATH:", tb_path)
print("\n===============================\n")
model.learn(total_timesteps=params['timesteps'], log_interval=params['log_interval'],
callback=create_training_callback(data_dir, params, env, freq=params['eval_freq'], checkpoint_freq=params['checkpoint_freq']))
print("Saving model to", data_dir)
model.save(data_dir +'/final_model')
if params['normalize']:
env.save(data_dir + '/environment.pkl')
env.close()
def train(env, seed, policy_fn, reward_giver, dataset, algo, g_step, d_step, policy_entcoeff, num_timesteps,
save_per_iter, checkpoint_dir, pretrained, bc_max_iter, task_name=None):
"""
train gail on mujoco
:param env: (Gym Environment) the environment
:param seed: (int) the initial random seed
:param policy_fn: (function (str, Gym Space, Gym Space, bool): MLPPolicy) policy generator
:param reward_giver: (TransitionClassifier) the reward predicter from obsevation and action
:param dataset: (MujocoDset) the dataset manager
:param algo: (str) the algorithm type (only 'trpo' is supported)
:param g_step: (int) number of steps to train policy in each epoch
:param d_step: (int) number of steps to train discriminator in each epoch
:param policy_entcoeff: (float) the weight of the entropy loss for the policy
:param num_timesteps: (int) the number of timesteps to run
:param save_per_iter: (int) the number of iterations before saving
:param checkpoint_dir: (str) the location for saving checkpoints
:param pretrained: (bool) use a pretrained behavior clone
:param bc_max_iter: (int) the maximum number of training iterations for the behavior clone
:param task_name: (str) the name of the task (can be None)
"""
pretrained_weight = None
#if pretrained and (bc_max_iter > 0):
# # Pretrain with behavior cloning
# pretrained_weight = behavior_clone.learn(env, policy_fn, dataset, max_iters=bc_max_iter)
if algo == 'trpo':
# Set up for MPI seed
# rank = 0 는 메인 쓰레드임을 의미합니다.
# rank != 0 인 쓰레드가 워커워커
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
logger.set_level(logger.DISABLED)
workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
set_global_seeds(workerseed)
env.seed(workerseed)
model = TRPO(policy_fn, env, timesteps_per_batch=1024, max_kl=0.01, cg_iters=10, gamma=0.995, lam=0.97,
entcoeff=policy_entcoeff, cg_damping=0.1, vf_stepsize=1e-3, vf_iters=5, _init_setup_model=False)
# GAIL param
# pretrained_weight = None
model.pretrained_weight = pretrained_weight
# Discriminator
model.reward_giver = reward_giver
model.expert_dataset = dataset
model.save_per_iter = save_per_iter
model.checkpoint_dir = checkpoint_dir
model.g_step = g_step
model.d_step = d_step
model.task_name = task_name
model.using_gail = True
# policy model setup!
model.setup_model()
# policy model update!
model.learn(total_timesteps=num_timesteps)
else:
raise NotImplementedError