Python tf_util.single_threaded_session方法代码示例

# 需要导入模块: from baselines.common import tf_util [as 别名]
# 或者: from baselines.common.tf_util import single_threaded_session [as 别名]
def train(env_id, num_timesteps, seed):
    import baselines.common.tf_util as U
    sess = U.single_threaded_session()
    sess.__enter__()

    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()
    def policy_fn(name, ob_space, ac_space):
        return MlpPolicy(name=name, ob_space=ob_space, ac_space=ac_space,
            hid_size=32, num_hid_layers=2)
    env = make_mujoco_env(env_id, workerseed)
    trpo_mpi.learn(env, policy_fn, timesteps_per_batch=1024, max_kl=0.01, cg_iters=10, cg_damping=0.1,
        max_timesteps=num_timesteps, gamma=0.99, lam=0.98, vf_iters=5, vf_stepsize=1e-3)
    env.close() 

# 需要导入模块: from baselines.common import tf_util [as 别名]
# 或者: from baselines.common.tf_util import single_threaded_session [as 别名]
def train(env_id, num_timesteps, seed):
    from baselines.ppo1 import mlp_policy, pposgd_simple
    import baselines.common.tf_util as U
    rank = MPI.COMM_WORLD.Get_rank()
    sess = U.single_threaded_session()
    sess.__enter__()
    mujoco_py.ignore_mujoco_warnings().__enter__()
    workerseed = seed + 10000 * rank
    set_global_seeds(workerseed)
    env = make_robotics_env(env_id, workerseed, rank=rank)
    def policy_fn(name, ob_space, ac_space):
        return mlp_policy.MlpPolicy(name=name, ob_space=ob_space, ac_space=ac_space,
            hid_size=256, num_hid_layers=3)

    pposgd_simple.learn(env, policy_fn,
            max_timesteps=num_timesteps,
            timesteps_per_actorbatch=2048,
            clip_param=0.2, entcoeff=0.0,
            optim_epochs=5, optim_stepsize=3e-4, optim_batchsize=256,
            gamma=0.99, lam=0.95, schedule='linear',
        )
    env.close() 

# 需要导入模块: from baselines.common import tf_util [as 别名]
# 或者: from baselines.common.tf_util import single_threaded_session [as 别名]
def train(env_id, num_timesteps, seed):
    whoami  = mpi_fork(num_cpu)
    if whoami == "parent":
        return
    import baselines.common.tf_util as U
    logger.session().__enter__()
    sess = U.single_threaded_session()
    sess.__enter__()

    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 = gym.make(env_id)
    def policy_fn(name, ob_space, ac_space):
        return MlpPolicy(name=name, ob_space=env.observation_space, ac_space=env.action_space,
            hid_size=32, num_hid_layers=2)
    env = bench.Monitor(env, osp.join(logger.get_dir(), "%i.monitor.json" % rank))
    env.seed(workerseed)
    gym.logger.setLevel(logging.WARN)

    trpo_mpi.learn(env, policy_fn, timesteps_per_batch=1024, max_kl=0.01, cg_iters=10, cg_damping=0.1,
        max_timesteps=num_timesteps, gamma=0.99, lam=0.98, vf_iters=5, vf_stepsize=1e-3)
    env.close() 

# 需要导入模块: from baselines.common import tf_util [as 别名]
# 或者: from baselines.common.tf_util import single_threaded_session [as 别名]
def train(env_id, num_timesteps, seed):
    import baselines.common.tf_util as U
    sess = U.single_threaded_session()
    sess.__enter__()

    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 = gym.make(env_id)
    def policy_fn(name, ob_space, ac_space):
        return MlpPolicy(name=name, ob_space=env.observation_space, ac_space=env.action_space,
            hid_size=32, num_hid_layers=2)
    env = bench.Monitor(env, logger.get_dir() and
        osp.join(logger.get_dir(), str(rank)))
    env.seed(workerseed)
    gym.logger.setLevel(logging.WARN)

    trpo_mpi.learn(env, policy_fn, timesteps_per_batch=1024, max_kl=0.01, cg_iters=10, cg_damping=0.1,
        max_timesteps=num_timesteps, gamma=0.99, lam=0.98, vf_iters=5, vf_stepsize=1e-3)
    env.close() 

# 需要导入模块: from baselines.common import tf_util [as 别名]
# 或者: from baselines.common.tf_util import single_threaded_session [as 别名]
def train(env_id, num_timesteps, seed):
    from baselines.trpo_mpi.nosharing_cnn_policy import CnnPolicy
    from baselines.trpo_mpi import trpo_mpi
    import baselines.common.tf_util as U
    rank = MPI.COMM_WORLD.Get_rank()
    sess = U.single_threaded_session()
    sess.__enter__()
    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): #pylint: disable=W0613
        return CnnPolicy(name=name, ob_space=env.observation_space, ac_space=env.action_space)
    env = bench.Monitor(env, logger.get_dir() and osp.join(logger.get_dir(), str(rank)))
    env.seed(workerseed)

    env = wrap_deepmind(env)
    env.seed(workerseed)

    trpo_mpi.learn(env, policy_fn, timesteps_per_batch=512, max_kl=0.001, cg_iters=10, cg_damping=1e-3,
        max_timesteps=int(num_timesteps * 1.1), gamma=0.98, lam=1.0, vf_iters=3, vf_stepsize=1e-4, entcoeff=0.00)
    env.close() 

# 需要导入模块: from baselines.common import tf_util [as 别名]
# 或者: from baselines.common.tf_util import single_threaded_session [as 别名]
def train(env_id, num_timesteps, seed):
    from baselines.ppo1 import pposgd_simple, cnn_policy
    import baselines.common.tf_util as U
    rank = MPI.COMM_WORLD.Get_rank()
    sess = U.single_threaded_session()
    sess.__enter__()
    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): #pylint: disable=W0613
        return cnn_policy.CnnPolicy(name=name, ob_space=ob_space, ac_space=ac_space)
    env = bench.Monitor(env, logger.get_dir() and
        osp.join(logger.get_dir(), str(rank)))
    env.seed(workerseed)

    env = wrap_deepmind(env)
    env.seed(workerseed)

    pposgd_simple.learn(env, policy_fn,
        max_timesteps=int(num_timesteps * 1.1),
        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'
    )
    env.close() 

# 需要导入模块: from baselines.common import tf_util [as 别名]
# 或者: from baselines.common.tf_util import single_threaded_session [as 别名]
def train(env_id, num_timesteps, seed):
    from baselines.ppo1 import pposgd_simple, cnn_policy
    import baselines.common.tf_util as U
    rank = MPI.COMM_WORLD.Get_rank()
    sess = U.single_threaded_session()
    sess.__enter__()
    if rank == 0:
        logger.configure()
    else:
        logger.configure(format_strs=[])
    workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank() if seed is not None else None
    set_global_seeds(workerseed)
    env = make_atari(env_id)
    def policy_fn(name, ob_space, ac_space): #pylint: disable=W0613
        return cnn_policy.CnnPolicy(name=name, ob_space=ob_space, ac_space=ac_space)
    env = bench.Monitor(env, logger.get_dir() and
        osp.join(logger.get_dir(), str(rank)))
    env.seed(workerseed)

    env = wrap_deepmind(env)
    env.seed(workerseed)

    pposgd_simple.learn(env, policy_fn,
        max_timesteps=int(num_timesteps * 1.1),
        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'
    )
    env.close() 

# 需要导入模块: from baselines.common import tf_util [as 别名]
# 或者: from baselines.common.tf_util import single_threaded_session [as 别名]
def train(args,seed,writer=None):
    from baselines.ppo1 import pposgd_simple_gcn, gcn_policy
    import baselines.common.tf_util as U
    rank = MPI.COMM_WORLD.Get_rank()
    sess = U.single_threaded_session()
    sess.__enter__()
    if rank == 0:
        logger.configure()
    else:
        logger.configure(format_strs=[])
    workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
    set_global_seeds(workerseed)
    if args.env=='molecule':
        env = gym.make('molecule-v0')
        env.init(data_type=args.dataset,logp_ratio=args.logp_ratio,qed_ratio=args.qed_ratio,sa_ratio=args.sa_ratio,reward_step_total=args.reward_step_total,is_normalize=args.normalize_adj,reward_type=args.reward_type,reward_target=args.reward_target,has_feature=bool(args.has_feature),is_conditional=bool(args.is_conditional),conditional=args.conditional,max_action=args.max_action,min_action=args.min_action) # remember call this after gym.make!!
    elif args.env=='graph':
        env = GraphEnv()
        env.init(reward_step_total=args.reward_step_total,is_normalize=args.normalize_adj,dataset=args.dataset) # remember call this after gym.make!!
    print(env.observation_space)
    def policy_fn(name, ob_space, ac_space):
        return gcn_policy.GCNPolicy(name=name, ob_space=ob_space, ac_space=ac_space, atom_type_num=env.atom_type_num,args=args)
    env.seed(workerseed)

    pposgd_simple_gcn.learn(args,env, policy_fn,
        max_timesteps=args.num_steps,
        timesteps_per_actorbatch=256,
        clip_param=0.2, entcoeff=0.01,
        optim_epochs=8, optim_stepsize=args.lr, optim_batchsize=32,
        gamma=1, lam=0.95,
        schedule='linear', writer=writer
    )
    env.close() 

# 需要导入模块: from baselines.common import tf_util [as 别名]
# 或者: from baselines.common.tf_util import single_threaded_session [as 别名]
def train(env_id, num_timesteps, seed, num_cpu):
    from baselines.trpo_mpi.nosharing_cnn_policy import CnnPolicy
    from baselines.trpo_mpi import trpo_mpi
    import baselines.common.tf_util as U
    whoami  = mpi_fork(num_cpu)
    if whoami == "parent":
        return
    rank = MPI.COMM_WORLD.Get_rank()
    sess = U.single_threaded_session()
    sess.__enter__()
    logger.session().__enter__()
    if rank != 0:
        logger.set_level(logger.DISABLED)


    workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
    set_global_seeds(workerseed)
    env = gym.make(env_id)
    def policy_fn(name, ob_space, ac_space): #pylint: disable=W0613
        return CnnPolicy(name=name, ob_space=env.observation_space, ac_space=env.action_space)
    env = bench.Monitor(env, osp.join(logger.get_dir(), "%i.monitor.json"%rank))
    env.seed(workerseed)
    gym.logger.setLevel(logging.WARN)

    env = wrap_train(env)
    num_timesteps /= 4 # because we're wrapping the envs to do frame skip
    env.seed(workerseed)

    trpo_mpi.learn(env, policy_fn, timesteps_per_batch=512, max_kl=0.001, cg_iters=10, cg_damping=1e-3,
        max_timesteps=num_timesteps, gamma=0.98, lam=1.0, vf_iters=3, vf_stepsize=1e-4, entcoeff=0.00)
    env.close() 

# 需要导入模块: from baselines.common import tf_util [as 别名]
# 或者: from baselines.common.tf_util import single_threaded_session [as 别名]
def train(env_id, num_timesteps, seed, num_cpu):
    from baselines.pposgd import pposgd_simple, cnn_policy
    import baselines.common.tf_util as U
    whoami  = mpi_fork(num_cpu)
    if whoami == "parent": return
    rank = MPI.COMM_WORLD.Get_rank()
    sess = U.single_threaded_session()
    sess.__enter__()
    logger.session().__enter__()
    if rank != 0: logger.set_level(logger.DISABLED)
    workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
    set_global_seeds(workerseed)
    env = gym.make(env_id)
    def policy_fn(name, ob_space, ac_space): #pylint: disable=W0613
        return cnn_policy.CnnPolicy(name=name, ob_space=ob_space, ac_space=ac_space)
    env = bench.Monitor(env, osp.join(logger.get_dir(), "%i.monitor.json" % rank))
    env.seed(workerseed)
    gym.logger.setLevel(logging.WARN)

    env = wrap_train(env)
    num_timesteps /= 4 # because we're wrapping the envs to do frame skip
    env.seed(workerseed)

    pposgd_simple.learn(env, policy_fn,
        max_timesteps=num_timesteps,
        timesteps_per_batch=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'
    )
    env.close() 

# 需要导入模块: from baselines.common import tf_util [as 别名]
# 或者: from baselines.common.tf_util import single_threaded_session [as 别名]
def train(num_timesteps, seed):
    rank = MPI.COMM_WORLD.Get_rank()
    sess = U.single_threaded_session()
    sess.__enter__()
    if rank == 0:
        logger.configure()
    else:
        logger.configure(format_strs=[])
    workerseed = seed + 10000 * MPI.COMM_WORLD.Get_rank()
    set_global_seeds(workerseed)

    config_file = os.path.join(os.path.dirname(os.path.realpath(__file__)), '..', 'configs',
                               'ant_gibson_flagrun.yaml')
    print(config_file)

    env = AntGibsonFlagRunEnv(config = config_file)
    
    def mlp_policy_fn(name, sensor_space, ac_space):
        return mlp_policy.MlpPolicy(name=name, ob_space=sensor_space, ac_space=ac_space, hid_size=64, num_hid_layers=2)

    env.seed(workerseed)
    gym.logger.setLevel(logging.WARN)

    pposgd_sensor.learn(env, mlp_policy_fn,
        max_timesteps=int(num_timesteps * 1.1 * 5),
        timesteps_per_actorbatch=6000,
        clip_param=0.2, entcoeff=0.00,
        optim_epochs=4, optim_stepsize=1e-4, optim_batchsize=64,
        gamma=0.99, lam=0.95,
        schedule='linear',
        save_per_acts=500
    )
    env.close()