def make_env(i, this_seed):
# Previously, we directly called `gym.make(env_name)`, but running
# `imitation.scripts.train_adversarial` within `imitation.scripts.parallel`
# created a weird interaction between Gym and Ray -- `gym.make` would fail
# inside this function for any of our custom environment unless those
# environments were also `gym.register()`ed inside `make_env`. Even
# registering the custom environment in the scope of `make_vec_env` didn't
# work. For more discussion and hypotheses on this issue see PR #160:
# https://github.com/HumanCompatibleAI/imitation/pull/160.
env = spec.make()
# Seed each environment with a different, non-sequential seed for diversity
# (even if caller is passing us sequentially-assigned base seeds). int() is
# necessary to work around gym bug where it chokes on numpy int64s.
env.seed(int(this_seed))
if max_episode_steps is not None:
env = TimeLimit(env, max_episode_steps)
elif spec.max_episode_steps is not None:
env = TimeLimit(env, max_episode_steps=spec.max_episode_steps)
# Use Monitor to record statistics needed for Baselines algorithms logging
# Optionally, save to disk
log_path = None
if log_dir is not None:
log_subdir = os.path.join(log_dir, "monitor")
os.makedirs(log_subdir, exist_ok=True)
log_path = os.path.join(log_subdir, f"mon{i:03d}")
env = bench.Monitor(env, log_path)
env = wrappers.RolloutInfoWrapper(env)
return env
def train():
"""
Train PPO1 model for slime volleyball, in MPI multiprocessing. Tested for 96 CPUs.
"""
rank = MPI.COMM_WORLD.Get_rank()
if rank == 0:
logger.configure(folder=LOGDIR)
else:
logger.configure(format_strs=[])
workerseed = SEED + 10000 * MPI.COMM_WORLD.Get_rank()
set_global_seeds(workerseed)
env = make_env(workerseed)
env = bench.Monitor(env, logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
env.seed(workerseed)
model = PPO1(MlpPolicy, env, timesteps_per_actorbatch=4096, clip_param=0.2, entcoeff=0.0, optim_epochs=10,
optim_stepsize=3e-4, optim_batchsize=64, gamma=0.99, lam=0.95, schedule='linear',
verbose=1)
eval_callback = EvalCallback(env, best_model_save_path=LOGDIR, log_path=LOGDIR, eval_freq=EVAL_FREQ, n_eval_episodes=EVAL_EPISODES)
model.learn(total_timesteps=NUM_TIMESTEPS, callback=eval_callback)
env.close()
del env
if rank == 0:
model.save(os.path.join(LOGDIR, "final_model")) # probably never get to this point.
def test_deepq():
"""
test DeepQ on atari
"""
logger.configure()
set_global_seeds(SEED)
env = make_atari(ENV_ID)
env = bench.Monitor(env, logger.get_dir())
env = wrap_atari_dqn(env)
model = DQN(env=env,
policy=CnnPolicy,
learning_rate=1e-4,
buffer_size=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
prioritized_replay=True,
prioritized_replay_alpha=0.6,
checkpoint_freq=10000)
model.learn(total_timesteps=NUM_TIMESTEPS)
env.close()
del model, env
def make_env(datapaths):
if len(datapaths) > 1:
env = EnsembleEnv(datapaths)
else:
env = Env(datapaths[0])
env = bench.Monitor(env, logger.get_dir())
return env
def test_deepq():
"""
test DeepQ on atari
"""
logger.configure()
set_global_seeds(SEED)
env = make_atari(ENV_ID)
env = bench.Monitor(env, logger.get_dir())
env = wrap_atari_dqn(env)
q_func = deepq_models.cnn_to_mlp(convs=[(32, 8, 4), (64, 4, 2),
(64, 3, 1)],
hiddens=[256],
dueling=True)
model = DeepQ(env=env,
policy=q_func,
learning_rate=1e-4,
buffer_size=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
prioritized_replay=True,
prioritized_replay_alpha=0.6,
checkpoint_freq=10000)
model.learn(total_timesteps=NUM_TIMESTEPS)
env.close()
del model, env
def train():
"""
Train PPO1 model for slime volleyball, in MPI multiprocessing. Tested for 96 CPUs.
"""
rank = MPI.COMM_WORLD.Get_rank()
if rank == 0:
logger.configure(folder=LOGDIR)
else:
logger.configure(format_strs=[])
workerseed = SEED + 10000 * MPI.COMM_WORLD.Get_rank()
set_global_seeds(workerseed)
env = make_env(workerseed)
env = bench.Monitor(
env,
logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
env.seed(workerseed)
model = PPO1.load(BEST_MODEL_PATH, env=env)
eval_callback = EvalCallback(env,
best_model_save_path=LOGDIR,
log_path=LOGDIR,
eval_freq=EVAL_FREQ,
n_eval_episodes=EVAL_EPISODES)
model.learn(total_timesteps=NUM_TIMESTEPS, callback=eval_callback)
env.close()
del env
if rank == 0:
model.save(os.path.join(
LOGDIR, "final_model")) # probably never get to this point.
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
def train(env_id, num_timesteps, seed):
"""
Train PPO1 model for Atari environments, 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
"""
rank = MPI.COMM_WORLD.Get_rank()
if rank == 0:
logger.configure()
else:
logger.configure(format_strs=[])
workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
set_global_seeds(workerseed)
env = make_atari(env_id)
env = bench.Monitor(env, logger.get_dir() and
os.path.join(logger.get_dir(), str(rank)))
env.seed(workerseed)
env = wrap_deepmind(env)
env.seed(workerseed)
model = PPO1(CnnPolicy, env, timesteps_per_actorbatch=256, clip_param=0.2, entcoeff=0.01, optim_epochs=4,
optim_stepsize=1e-3, optim_batchsize=64, gamma=0.99, lam=0.95, schedule='linear', verbose=2)
model.learn(total_timesteps=num_timesteps)
env.close()
del env
def make_env():
env_out = gym.make(args.env)
env_out = bench.Monitor(env_out,
logger.get_dir(),
allow_early_resets=True)
return env_out
def make_env():
env_out = gym.make(args.env)
env_out.env.disableViewer = True
env_out.env.visualize = False
env_out = bench.Monitor(env_out,
logger.get_dir(),
allow_early_resets=True)
return env_out
def _thunk():
local_env_kwargs = dict(env_kwargs) # copy this to avoid altering the others
local_env_kwargs["env_rank"] = rank
env = _make(env_id, env_kwargs=local_env_kwargs)
env.seed(seed + rank)
if log_dir is not None:
env = bench.Monitor(env, os.path.join(log_dir, str(rank)), allow_early_resets=allow_early_resets)
return env
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
def create_env(env_id, delay_step, env_str=str(0)):
# if env_type in ["mujoco", "Mujoco", "MuJoCo", "raw", "mujoco_raw", "raw_mujoco"]:
env = gym.make(env_id)
env = TimestepWrapper(env)
env = DelayedRewardWrapper(env, delay_step)
env = bench.Monitor(
env,
logger.get_dir() and os.path.join(logger.get_dir(), env_str))
return env
def main(args):
"""
start training the model
:param args: (ArgumentParser) the training argument
"""
with tf_util.make_session(num_cpu=1):
set_global_seeds(args.seed)
env = gym.make(args.env_id)
def policy_fn(name, ob_space, ac_space, reuse=False, placeholders=None, sess=None):
return mlp_policy.MlpPolicy(name=name, ob_space=ob_space, ac_space=ac_space, reuse=reuse, sess=sess,
hid_size=args.policy_hidden_size, num_hid_layers=2, placeholders=placeholders)
#========================================================================================================
env = bench.Monitor(env, logger.get_dir() and
os.path.join(logger.get_dir(), "monitor.json"))
env.seed(args.seed)
gym.logger.setLevel(logging.WARN)
# 길고 긴 task name을 받아옵니다. ===========================================================================
task_name = get_task_name(args)
args.checkpoint_dir = os.path.join(args.checkpoint_dir, task_name)
args.log_dir = os.path.join(args.log_dir, task_name)
# =======================================================================================================
if args.task == 'train':
dataset = MujocoDset(expert_path=args.expert_path, traj_limitation=args.traj_limitation)
#discriminator 네트워크 생성
reward_giver = TransitionClassifier(env, args.adversary_hidden_size, entcoeff=args.adversary_entcoeff)
#policy 네트워크 학습
#policy network 가 policy_fn 으로 선언되어있다는 것에 주의
train(env, args.seed, policy_fn, reward_giver,
dataset, args.algo, args.g_step, args.d_step,
args.policy_entcoeff, args.num_timesteps, args.save_per_iter, args.checkpoint_dir, args.pretrained, args.bc_max_iter, task_name)
# ======================================= 이것은 나중에 이해하보는 거시여=============================================
# 학습된 모델을 evaluate 할 때 사용합니다.
elif args.task == 'evaluate':
runner(env,
policy_fn,
args.load_model_path,
timesteps_per_batch=1024,
number_trajs=10,
stochastic_policy=args.stochastic_policy,
save=args.save_sample
)
else:
raise NotImplementedError
env.close()
def make_envs(env_id,
do_eval,
seed,
conf,
normalize_observations=False,
normalize_returns=False):
# Create envs.
env_params = conf.pop('env_params', {})
env = base_env = gym.make(env_id)
if hasattr(base_env, 'env'):
base_env = base_env.env
for attr in env_params:
setattr(base_env, attr, env_params[attr])
env = bench.Monitor(env, logger.get_dir(), allow_early_resets=True)
# Seed everything to make things reproducible.
logger.info('seed={}, logdir={}'.format(seed, logger.get_dir()))
tf.reset_default_graph()
set_global_seeds(seed)
env.seed(seed)
if normalize_observations or normalize_returns:
env = DummyVecEnv([lambda: env])
env = VecNormalize(env,
norm_obs=normalize_observations,
norm_reward=normalize_returns)
if do_eval:
eval_env = base_eval_env = gym.make(env_id)
if hasattr(base_eval_env, 'env'):
base_eval_env = base_eval_env.env
for attr in env_params:
setattr(base_eval_env, attr, env_params[attr])
eval_env = bench.Monitor(eval_env,
os.path.join(logger.get_dir(), 'gym_eval'),
allow_early_resets=True)
eval_env.seed(seed)
eval_env.base_env = base_eval_env
else:
base_eval_env = None
eval_env = None
env.base_env = base_env
return base_env, env, base_eval_env, eval_env
def make_env(i):
env = gym.make(env_id)
env.seed(seed + i) # seed each environment separately for diversity
# Use Monitor to record statistics needed for Baselines algorithms logging
# Optionally, save to disk
log_path = None
if log_dir is not None:
log_subdir = os.path.join(log_dir, 'monitor')
os.makedirs(log_subdir, exist_ok=True)
log_path = os.path.join(log_subdir, f'mon{i:03d}')
return bench.Monitor(env, log_path, allow_early_resets=True)
def train(env_id, num_timesteps, seed, num_options, app, saves, wsaves, epoch,
dc):
U.make_session(num_cpu=1).__enter__()
set_global_seeds(seed)
env = gym.make(env_id)
def policy_fn(name, ob_space, ac_space):
return mlp_policy.MlpPolicy(name=name,
ob_space=ob_space,
ac_space=ac_space,
hid_size=64,
num_hid_layers=2,
num_options=num_options,
dc=dc)
env = bench.Monitor(
env,
logger.get_dir() and osp.join(logger.get_dir(), "monitor.json"))
env.seed(seed)
gym.logger.setLevel(logging.WARN)
if num_options == 1:
optimsize = 64
elif num_options == 2:
optimsize = 32
else:
print("Only two options or primitive actions is currently supported.")
sys.exit()
pposgd_simple.learn(env,
policy_fn,
max_timesteps=num_timesteps,
timesteps_per_batch=2048,
clip_param=0.2,
entcoeff=0.0,
optim_epochs=10,
optim_stepsize=3e-4,
optim_batchsize=optimsize,
gamma=0.99,
lam=0.95,
schedule='constant',
num_options=num_options,
app=app,
saves=saves,
wsaves=wsaves,
epoch=epoch,
seed=seed,
dc=dc)
env.close()
def main(args):
"""
start training the model
:param args: (ArgumentParser) the training argument
"""
with tf_util.make_session(num_cpu=1):
set_global_seeds(args.seed)
env = gym.make(args.env_id)
def policy_fn(name, ob_space, ac_space, reuse=False, sess=None):
return mlp_policy.MlpPolicy(name=name,
ob_space=ob_space,
ac_space=ac_space,
sess=sess,
reuse=reuse,
hid_size=args.policy_hidden_size,
num_hid_layers=2)
env = bench.Monitor(
env,
logger.get_dir()
and os.path.join(logger.get_dir(), "monitor.json"))
env.seed(args.seed)
gym.logger.setLevel(logging.WARN)
task_name = get_task_name(args)
args.checkpoint_dir = os.path.join(args.checkpoint_dir, task_name)
args.log_dir = os.path.join(args.log_dir, task_name)
dataset = MujocoDset(expert_path=args.expert_path,
traj_limitation=args.traj_limitation)
savedir_fname = learn(env,
policy_fn,
dataset,
max_iters=args.BC_max_iter,
ckpt_dir=args.checkpoint_dir,
task_name=task_name,
verbose=True)
runner(env,
policy_fn,
savedir_fname,
timesteps_per_batch=1024,
number_trajs=10,
stochastic_policy=args.stochastic_policy,
save=args.save_sample,
reuse=True)
def main():
"""
run the atari test
"""
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--env', help='environment ID', default='BreakoutNoFrameskip-v4')
parser.add_argument('--seed', help='RNG seed', type=int, default=0)
parser.add_argument('--prioritized', type=int, default=1)
parser.add_argument('--prioritized-replay-alpha', type=float, default=0.6)
parser.add_argument('--dueling', type=int, default=1)
parser.add_argument('--num-timesteps', type=int, default=int(10e6))
parser.add_argument('--checkpoint-freq', type=int, default=10000)
parser.add_argument('--checkpoint-path', type=str, default=None)
args = parser.parse_args()
logger.configure()
set_global_seeds(args.seed)
env = make_atari(args.env)
env = bench.Monitor(env, logger.get_dir())
env = wrap_atari_dqn(env)
q_func = deepq_models.cnn_to_mlp(
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
hiddens=[256],
dueling=bool(args.dueling),
)
model = DeepQ(
env=env,
policy=q_func,
learning_rate=1e-4,
buffer_size=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
prioritized_replay=bool(args.prioritized),
prioritized_replay_alpha=args.prioritized_replay_alpha,
checkpoint_freq=args.checkpoint_freq,
checkpoint_path=args.checkpoint_path,
)
model.learn(total_timesteps=args.num_timesteps)
env.close()
def test_rollout_stats():
"""Applying `ObsRewIncrementWrapper` halves the reward mean.
`rollout_stats` should reflect this.
"""
env = gym.make("CartPole-v1")
env = bench.Monitor(env, None)
env = ObsRewHalveWrapper(env)
venv = vec_env.DummyVecEnv([lambda: env])
with serialize.load_policy("zero", "UNUSED", venv) as policy:
trajs = rollout.generate_trajectories(policy, venv, rollout.min_episodes(10))
s = rollout.rollout_stats(trajs)
np.testing.assert_allclose(s["return_mean"], s["monitor_return_mean"] / 2)
np.testing.assert_allclose(s["return_std"], s["monitor_return_std"] / 2)
np.testing.assert_allclose(s["return_min"], s["monitor_return_min"] / 2)
np.testing.assert_allclose(s["return_max"], s["monitor_return_max"] / 2)
def train(env_id, num_timesteps, seed):
"""
Train TRPO model for the atari 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
"""
rank = MPI.COMM_WORLD.Get_rank()
if rank == 0:
logger.configure()
else:
logger.configure(format_strs=[])
workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
set_global_seeds(workerseed)
env = make_atari(env_id)
# def policy_fn(name, ob_space, ac_space, sess=None, placeholders=None): # pylint: disable=W0613
# return CnnPolicy(name=name, ob_space=ob_space, ac_space=ac_space, sess=sess, placeholders=placeholders)
env = bench.Monitor(
env,
logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
env.seed(workerseed)
env = wrap_deepmind(env)
env.seed(workerseed)
model = TRPO(CnnPolicy,
env,
timesteps_per_batch=512,
max_kl=0.001,
cg_iters=10,
cg_damping=1e-3,
entcoeff=0.0,
gamma=0.98,
lam=1,
vf_iters=3,
vf_stepsize=1e-4)
model.learn(total_timesteps=int(num_timesteps * 1.1))
env.close()
def make_env():
env_out = GymWrapper(
suite.make(
"SawyerLift",
use_camera_obs=False, # do not use pixel observations
has_offscreen_renderer=
False, # not needed since not using pixel obs
has_renderer=True, # make sure we can render to the screen
reward_shaping=True, # use dense rewards
control_freq=
10, # control should happen fast enough so that simulation looks smooth
))
env_out.reward_range = None
env_out.metadata = None
env_out.spec = None
env_out = bench.Monitor(env_out,
logger.get_dir(),
allow_early_resets=True)
return env_out
def main():
"""
Run the atari test
"""
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--env',
help='environment ID',
default='BreakoutNoFrameskip-v4')
parser.add_argument('--seed', help='RNG seed', type=int, default=0)
parser.add_argument('--prioritized', type=int, default=1)
parser.add_argument('--dueling', type=int, default=1)
parser.add_argument('--prioritized-replay-alpha', type=float, default=0.6)
parser.add_argument('--num-timesteps', type=int, default=int(10e6))
args = parser.parse_args()
logger.configure()
set_global_seeds(args.seed)
env = make_atari(args.env)
env = bench.Monitor(env, logger.get_dir())
env = wrap_atari_dqn(env)
policy = partial(CnnPolicy, dueling=args.dueling == 1)
model = DQN(
env=env,
policy=policy,
learning_rate=1e-4,
buffer_size=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
prioritized_replay=bool(args.prioritized),
prioritized_replay_alpha=args.prioritized_replay_alpha,
)
model.learn(total_timesteps=args.num_timesteps)
env.close()
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 make_env(i):
env = CartPoleNoVelEnv()
env = bench.Monitor(env, None, allow_early_resets=True)
env.seed(i)
return env
def _init():
env = gym.make(env_id)
env.seed(seed + rank)
env = bench.Monitor(env, logger.get_dir(), allow_early_resets=True)
return env
def make_env(i):
env = CartPoleNoVelEnv()
env = TimeLimit(env, max_episode_steps=500)
env = bench.Monitor(env, None, allow_early_resets=True)
env.seed(i)
return env
def make_env(i):
env = env = gym.make("Breakout-v0")
env = bench.Monitor(env, None, allow_early_resets=True)
env.seed(i)
return env
def make_env():
env_out = StudentEnv()
env_out = bench.Monitor(env_out,
logger.get_dir(),
allow_early_resets=True)
return env_out
def make_env():
env_out = gym.make('CartPole-v0')
env_out = bench.Monitor(env_out,
logger.get_dir(),
allow_early_resets=True)
return env_out
