def train(env_id, num_timesteps, seed):
"""
Train PPO1 model for Robotics 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()
with mujoco_py.ignore_mujoco_warnings():
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, sess=None, placeholders=None):
return mlp_policy.MlpPolicy(name=name, ob_space=ob_space, ac_space=ac_space, hid_size=256, num_hid_layers=3,
sess=sess, placeholders=placeholders)
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()
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()
def train(env_id, num_timesteps, seed, hid_size=64, num_hid_layers=2):
from baselines.ppo1 import mlp_policy, pposgd_simple
assert env_id in (_MujocoEnvs + _RoboticsEnvs)
def policy_fn(name, ob_space, ac_space):
return mlp_policy.MlpPolicy(name=name, ob_space=ob_space, ac_space=ac_space,
hid_size=hid_size, num_hid_layers=num_hid_layers)
if env_id in _MujocoEnvs:
env = make_mujoco_env(env_id, seed)
elif env_id in _RoboticsEnvs:
env = make_robotics_env(env_id, seed)
else:
raise ValueError('Environment `{0}` is not supported.'.format(env_id))
# Not putting these params in config as we do not plan on changing them.
optim_epochs = 10 if env_id in _MujocoEnvs else 5
optim_batchsize = 64 if env_id in _MujocoEnvs else 256
pi = pposgd_simple.learn(env, policy_fn,
max_timesteps=num_timesteps,
timesteps_per_actorbatch=2048,
clip_param=0.2, entcoeff=0.0,
optim_epochs=optim_epochs, optim_stepsize=3e-4,
optim_batchsize=optim_batchsize,
gamma=0.99, lam=0.95, schedule='linear',
)
env.close()
return pi