def main(args):
U.make_session(num_cpu=1).__enter__()
set_global_seeds(args.seed)
env = gym.make(args.env_id)
task_name = get_task_short_name(args)
logger.configure(dir='log_trpo_cartpole/%s' % task_name)
def policy_fn(name, ob_space, ac_space, reuse=False):
return build_policy(env, 'mlp', value_network='copy')
import logging
import os.path as osp
import bench
env = bench.Monitor(
env,
logger.get_dir() and osp.join(logger.get_dir(), "monitor.json"))
env.seed(args.seed)
gym.logger.setLevel(logging.WARN)
args.checkpoint_dir = osp.join(args.checkpoint_dir, task_name)
args.log_dir = osp.join(args.log_dir, task_name)
if args.task == 'train':
from utils.mujoco_dset import Dset_gym
expert_observations = np.genfromtxt(
'expert_data/cartpole/observations.csv')
expert_actions = np.genfromtxt('expert_data/cartpole/actions.csv',
dtype=np.int32)
expert_dataset = Dset_gym(inputs=expert_observations,
labels=expert_actions,
randomize=True)
# expert_dataset = (expert_observations, expert_actions)
reward_giver = Discriminator(env,
args.adversary_hidden_size,
entcoeff=args.adversary_entcoeff)
reward_guidance = Guidance(env,
args.policy_hidden_size,
expert_dataset=expert_dataset)
train(env, args.seed, policy_fn, reward_giver, reward_guidance,
expert_dataset, args.algo, args.g_step, args.d_step,
args.policy_entcoeff, args.num_timesteps, args.save_per_iter,
args.checkpoint_dir, args.log_dir, args.pretrained,
args.BC_max_iter, args.loss_percent, task_name)
elif args.task == 'evaluate':
avg_len, avg_ret = runner(env,
policy_fn,
args.load_model_path,
timesteps_per_batch=1024,
number_trajs=100,
stochastic_policy=args.stochastic_policy,
save=args.save_sample)
result = np.array([avg_ret, avg_len])
txt_name = args.load_model_path + 'result.txt'
np.savetxt(txt_name, result, fmt="%d", delimiter=" ")
print(args.load_model_path, avg_ret, avg_len)
print('保存成功')
else:
raise NotImplementedError
env.close()
def main(args):
U.make_session(num_cpu=1).__enter__()
set_global_seeds(args.seed)
env = gym.make(args.env_id)
task_name = get_task_short_name(args)
logger.configure(dir='log_trpo_mujoco/%s' % task_name)
def policy_fn(name, ob_space, ac_space, reuse=False):
return MlpPolicy(name=name,
ob_space=ob_space,
ac_space=ac_space,
reuse=reuse,
hid_size=args.policy_hidden_size,
num_hid_layers=2)
import logging
import os.path as osp
import bench
env = bench.Monitor(
env,
logger.get_dir() and osp.join(logger.get_dir(), "monitor.json"))
env.seed(args.seed)
gym.logger.setLevel(logging.WARN)
args.checkpoint_dir = osp.join(args.checkpoint_dir, task_name)
args.log_dir = osp.join(args.log_dir, task_name)
if args.task == 'train':
expert_dataset = Mujoco_Dset(expert_path=args.expert_path,
traj_limitation=args.traj_limitation)
reward_giver = Discriminator(env,
args.adversary_hidden_size,
entcoeff=args.adversary_entcoeff)
reward_guidance = Guidance(env,
args.policy_hidden_size,
expert_dataset=expert_dataset)
train(env, args.seed, policy_fn, reward_giver, reward_guidance,
expert_dataset, args.algo, args.g_step, args.d_step,
args.policy_entcoeff, args.num_timesteps, args.save_per_iter,
args.checkpoint_dir, args.log_dir, args.pretrained,
args.BC_max_iter, args.loss_percent, task_name)
elif args.task == 'evaluate':
avg_len, avg_ret = runner(env,
policy_fn,
args.load_model_path,
timesteps_per_batch=1024,
number_trajs=100,
stochastic_policy=args.stochastic_policy,
save=args.save_sample)
result = np.array([avg_ret, avg_len])
txt_name = args.load_model_path + 'result.txt'
np.savetxt(txt_name, result, fmt="%d", delimiter=" ")
print(args.load_model_path, avg_ret, avg_len)
print('保存成功')
else:
raise NotImplementedError
env.close()
def main(args):
U.make_session(num_cpu=1).__enter__()
set_global_seeds(args.seed)
from dp_env_v3 import DPEnv
env = DPEnv
task_name = get_task_short_name(args)
if rank == 0:
logger.configure(dir='log_gail/%s' % task_name)
if rank != 0:
logger.set_level(logger.DISABLED)
def policy_fn(name, ob_space, ac_space, reuse=False):
return MlpPolicy(name=name,
ob_space=ob_space,
ac_space=ac_space,
reuse=reuse,
hid_size=args.policy_hidden_size,
num_hid_layers=2)
import logging
import os.path as osp
import bench
env = bench.Monitor(
env,
logger.get_dir() and osp.join(logger.get_dir(), "monitor.json"))
env.seed(args.seed)
gym.logger.setLevel(logging.WARN)
task_name = get_task_name(args)
args.checkpoint_dir = osp.join(args.checkpoint_dir, task_name)
args.log_dir = osp.join(args.log_dir, task_name)
if args.task == 'train':
dataset = Mujoco_Dset(expert_path=args.expert_path,
traj_limitation=args.traj_limitation)
reward_giver = TransitionClassifier(env,
args.adversary_hidden_size,
entcoeff=args.adversary_entcoeff)
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.log_dir, args.pretrained, args.BC_max_iter, task_name)
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 main(args):
U.make_session(num_cpu=1).__enter__()
set_global_seeds(args.seed)
# from dp_env_v2 import DPEnv
from dp_env_v3 import DPEnv
# from dp_env_test import DPEnv
env = DPEnv()
# env = gym.make('Humanoid-v2')
task_name = get_task_short_name(args)
def policy_fn(name, ob_space, ac_space, reuse=False):
return MlpPolicy(name=name, ob_space=ob_space, ac_space=ac_space,
reuse=reuse, hid_size=args.policy_hidden_size, num_hid_layers=2)
if args.task == 'train':
import logging
import os.path as osp
import bench
if MPI is None or MPI.COMM_WORLD.Get_rank() == 0:
logger.configure(dir='log_tmp/%s'%task_name)
if MPI.COMM_WORLD.Get_rank() != 0:
logger.set_level(logger.DISABLED)
env = bench.Monitor(env, logger.get_dir() and
osp.join(logger.get_dir(), "monitor.json"))
env.seed(args.seed)
gym.logger.setLevel(logging.WARN)
task_name = get_task_short_name(args)
args.checkpoint_dir = osp.join(args.checkpoint_dir, task_name)
args.log_dir = osp.join(args.log_dir, task_name)
train(env,
args.seed,
policy_fn,
args.g_step,
args.policy_entcoeff,
args.pretrained_weight_path,
args.num_timesteps,
args.save_per_iter,
args.checkpoint_dir,
args.log_dir,
task_name)
elif args.task == 'evaluate':
runner(env,
policy_fn,
args.load_model_path,
timesteps_per_batch=1024,
number_trajs=100,
stochastic_policy=args.stochastic_policy,
save=args.save_sample)
else:
raise NotImplementedError
env.close()
def load_policy(model_path,
input_dim,
output_dim,
num_hidden,
num_layers,
init_logstd=1.,
discrete=False,
beta=1.0):
observation_space = Box(low=-np.inf, high=np.inf, shape=(input_dim, ))
if discrete:
action_space = Discrete(n=output_dim)
else:
action_space = Box(low=-np.inf, high=np.inf, shape=(output_dim, ))
tf.reset_default_graph()
config = tf.ConfigProto(allow_soft_placement=True,
inter_op_parallelism_threads=8,
intra_op_parallelism_threads=8,
device_count={'CPU': 8})
config.gpu_options.allow_growth = True
sess = U.make_session(make_default=True, config=config)
network = mlp(num_hidden=num_hidden, num_layers=num_layers)
# policy_train = build_policy(observation_space, action_space, network, trainable_variance=True,
# state_dependent_variance=True, beta=beta, init_logstd=init_logstd)()
ob_space = env.observation_space
ac_space = env.action_space
ob = observation_placeholder(ob_space)
policy_train = build_policy_trpo(
env, network, value_network='copy')(observ_placeholder=ob)
U.initialize()
if model_path != '':
policy_train.load(model_path)
return policy_train
