Python baselines.common.tf_util.single_threaded_session() Examples

The following are 30 code examples of baselines.common.tf_util.single_threaded_session(). You can vote up the ones you like or vote down the ones you don't like, and go to the original project or source file by following the links above each example. You may also want to check out all available functions/classes of the module baselines.common.tf_util , or try the search function .
Example #1
Source File: run_mujoco.py    From lirpg with MIT License 6 votes vote down vote up 
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()
Example #2
Source File: run_mujoco.py    From sonic_contest with MIT License 6 votes vote down vote up 
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()
Example #3
Source File: run_mujoco.py    From self-imitation-learning with MIT License 6 votes vote down vote up 
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()
Example #4
Source File: run_robotics.py    From baselines with MIT License 6 votes vote down vote up 
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()
Example #5
Source File: run_robotics.py    From ICML2019-TREX with MIT License 6 votes vote down vote up 
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()
Example #6
Source File: run_robotics.py    From ICML2019-TREX with MIT License 6 votes vote down vote up 
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()
Example #7
Source File: run_mujoco.py    From DRL_DeliveryDuel with MIT License 6 votes vote down vote up 
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()
Example #8
Source File: run_mujoco.py    From learning2run with MIT License 6 votes vote down vote up 
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()
Example #9
Source File: run_mujoco.py    From rl_graph_generation with BSD 3-Clause "New" or "Revised" License 6 votes vote down vote up 
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()
Example #10
Source File: run_robotics.py    From HardRLWithYoutube with MIT License 6 votes vote down vote up 
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()
Example #11
Source File: run_mujoco.py    From deeprl-baselines with MIT License 6 votes vote down vote up 
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()
Example #12
Source File: run_mujoco.py    From BackpropThroughTheVoidRL with MIT License 6 votes vote down vote up 
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()
Example #13
Source File: run_atari.py    From BackpropThroughTheVoidRL with MIT License 5 votes vote down vote up 
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)
    gym.logger.setLevel(logging.WARN)

    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()
Example #14
Source File: run_atari.py    From BackpropThroughTheVoidRL with MIT License 5 votes vote down vote up 
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)
    gym.logger.setLevel(logging.WARN)

    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()
Example #15
Source File: run_atari.py    From sonic_contest with MIT License 5 votes vote down vote up 
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()
Example #16
Source File: run_atari.py    From self-imitation-learning with MIT License 5 votes vote down vote up 
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()
Example #17
Source File: run_atari.py    From lirpg with MIT License 5 votes vote down vote up 
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()
Example #18
Source File: run_atari.py    From self-imitation-learning with MIT License 5 votes vote down vote up 
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()
Example #19
Source File: run_atari.py    From baselines with MIT License 5 votes vote down vote up 
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()
Example #20
Source File: train_ant_gibson_flagrun_ppo1.py    From midlevel-reps with MIT License 5 votes vote down vote up 
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()
Example #21
Source File: run_atari.py    From deeprl-baselines with MIT License 5 votes vote down vote up 
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)
    gym.logger.setLevel(logging.WARN)

    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()
Example #22
Source File: train_ant_navigate_ppo1.py    From midlevel-reps with MIT License 5 votes vote down vote up 
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_navigate.yaml')
    print(config_file)

    env = AntNavigateEnv(config = config_file)
    
    def policy_fn(name, ob_space, ac_space):
        #return mlp_policy.MlpPolicy(name=name, ob_space=sensor_space, ac_space=ac_space, hid_size=64, num_hid_layers=2)
        return cnn_policy.CnnPolicy(name=name, ob_space=ob_space, ac_space=ac_space, save_per_acts=10000, session=sess, kind='small')


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

    pposgd_simple.learn(env, 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()
Example #23
Source File: train_humanoid_navigate_ppo1.py    From midlevel-reps with MIT License 5 votes vote down vote up 
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',
                               'humanoid_navigate_nonviz.yaml')
    print(config_file)

    env = HumanoidNavigateEnv(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()
Example #24
Source File: run_energyplus.py    From rl-testbed-for-energyplus with MIT License 5 votes vote down vote up 
def train(env_id, num_timesteps, seed):
    import baselines.common.tf_util as U
    sess = U.single_threaded_session()
    sess.__enter__()
    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)

    # Create a new base directory like /tmp/openai-2018-05-21-12-27-22-552435
    log_dir = os.path.join(energyplus_logbase_dir(), datetime.datetime.now().strftime("openai-%Y-%m-%d-%H-%M-%S-%f"))
    if not os.path.exists(log_dir + '/output'):
        os.makedirs(log_dir + '/output')
    os.environ["ENERGYPLUS_LOG"] = log_dir
    model = os.getenv('ENERGYPLUS_MODEL')
    if model is None:
        print('Environment variable ENERGYPLUS_MODEL is not defined')
        os.exit()
    weather = os.getenv('ENERGYPLUS_WEATHER')
    if weather is None:
        print('Environment variable ENERGYPLUS_WEATHER is not defined')
        os.exit()

    rank = MPI.COMM_WORLD.Get_rank()
    if rank == 0:
        print('train: init logger with dir={}'.format(log_dir)) #XXX
        logger.configure(log_dir)
    else:
        logger.configure(format_strs=[])
        logger.set_level(logger.DISABLED)

    env = make_energyplus_env(env_id, workerseed)

    trpo_mpi.learn(env, policy_fn,
                   max_timesteps=num_timesteps,
                   #timesteps_per_batch=1*1024, max_kl=0.01, cg_iters=10, cg_damping=0.1,
                   timesteps_per_batch=16*1024, max_kl=0.01, cg_iters=10, cg_damping=0.1,
                   gamma=0.99, lam=0.98, vf_iters=5, vf_stepsize=1e-3)
    env.close()
Example #25
Source File: run_atari.py    From deeprl-baselines with MIT License 5 votes vote down vote up 
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)
    gym.logger.setLevel(logging.WARN)

    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()
Example #26
Source File: run_atari.py    From DRL_DeliveryDuel with MIT License 5 votes vote down vote up 
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()
Example #27
Source File: run_atari.py    From lirpg with MIT License 5 votes vote down vote up 
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()
Example #28
Source File: run_atari.py    From HardRLWithYoutube with MIT License 5 votes vote down vote up 
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()
Example #29
Source File: run_molecule.py    From rl_graph_generation with BSD 3-Clause "New" or "Revised" License 5 votes vote down vote up 
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()
Example #30
Source File: run_atari.py    From rl_graph_generation with BSD 3-Clause "New" or "Revised" License 5 votes vote down vote up 
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()