def load_trained_policy_and_mdp(env_name, policy_state_str):
""" Creates the specialized MDP and policy objects needed to sample expert
trajectories for a given environment.
Returns:
mdp: An instance of `RLGymMDP`, similar to a real gym env except with
customized obs/action spaces and an internal `RLGyMSim` object.
policy: The agent's policy, encoded as either rl.GaussianPolicy for
continuous actions, or rl.GibbsPolicy for discrete actions.
train_args: A dictionary of arguments (like argparse dicts) based on the
trained policy's TRPO run.
"""
import gym
import policyopt
from policyopt import nn, rl
from environments import rlgymenv
# Load the saved state
policy_file, policy_key = util.split_h5_name(policy_state_str)
print 'Loading policy parameters from %s in %s' % (policy_key, policy_file)
with h5py.File(policy_file, 'r') as f:
train_args = json.loads(f.attrs['args'])
# Initialize the MDP
print 'Loading environment', env_name
mdp = rlgymenv.RLGymMDP(env_name)
print 'MDP observation space, action space sizes: %d, %d\n' % (mdp.obs_space.dim, mdp.action_space.storage_size)
# Initialize the policy
nn.reset_global_scope()
enable_obsnorm = bool(train_args['enable_obsnorm']) if 'enable_obsnorm' in train_args else train_args['obsnorm_mode'] != 'none'
if isinstance(mdp.action_space, policyopt.ContinuousSpace):
policy_cfg = rl.GaussianPolicyConfig(
hidden_spec=train_args['policy_hidden_spec'],
min_stdev=0.,
init_logstdev=0.,
enable_obsnorm=enable_obsnorm)
policy = rl.GaussianPolicy(policy_cfg, mdp.obs_space, mdp.action_space, 'GaussianPolicy')
else:
policy_cfg = rl.GibbsPolicyConfig(
hidden_spec=train_args['policy_hidden_spec'],
enable_obsnorm=enable_obsnorm)
policy = rl.GibbsPolicy(policy_cfg, mdp.obs_space, mdp.action_space, 'GibbsPolicy')
# Load the policy parameters
policy.load_h5(policy_file, policy_key)
return mdp, policy, train_args
def load_trained_policy_and_mdp(env_name, policy_state_str):
import gym
import policyopt
from policyopt import nn, rl
from environments import rlgymenv
# Load the saved state
policy_file, policy_key = util.split_h5_name(policy_state_str)
print('Loading policy parameters from %s in %s' %
(policy_key, policy_file))
with h5py.File(policy_file, 'r') as f:
train_args = json.loads(f.attrs['args'])
# Initialize the MDP
print('Loading environment', env_name)
mdp = rlgymenv.RLGymMDP(env_name)
print('MDP observation space, action space sizes: %d, %d\n' %
(mdp.obs_space.dim, mdp.action_space.storage_size))
# Initialize the policy
nn.reset_global_scope()
enable_obsnorm = bool(train_args['enable_obsnorm']
) if 'enable_obsnorm' in train_args else train_args[
'obsnorm_mode'] != 'none'
if isinstance(mdp.action_space, policyopt.ContinuousSpace):
policy_cfg = rl.GaussianPolicyConfig(
hidden_spec=train_args['policy_hidden_spec'],
min_stdev=0.,
init_logstdev=0.,
enable_obsnorm=enable_obsnorm)
policy = rl.GaussianPolicy(policy_cfg, mdp.obs_space, mdp.action_space,
'GaussianPolicy')
else:
policy_cfg = rl.GibbsPolicyConfig(
hidden_spec=train_args['policy_hidden_spec'],
enable_obsnorm=enable_obsnorm)
policy = rl.GibbsPolicy(policy_cfg, mdp.obs_space, mdp.action_space,
'GibbsPolicy')
# Load the policy parameters
policy.load_h5(policy_file, policy_key)
return mdp, policy, train_args
def load_trained_policy_and_mdp(env_name, policy_state_str):
import gym
import policyopt
from policyopt import nn, rl
from environments import rlgymenv
# Load the saved state
policy_file, policy_key = util.split_h5_name(policy_state_str)
print 'Loading policy parameters from %s in %s' % (policy_key, policy_file)
with h5py.File(policy_file, 'r') as f:
train_args = json.loads(f.attrs['args'])
# Initialize the MDP
print 'Loading environment', env_name
mdp = rlgymenv.RLGymMDP(env_name)
print 'MDP observation space, action space sizes: %d, %d\n' % (mdp.obs_space.dim, mdp.action_space.storage_size)
# Initialize the policy
nn.reset_global_scope()
enable_obsnorm = bool(train_args['enable_obsnorm']) if 'enable_obsnorm' in train_args else train_args['obsnorm_mode'] != 'none'
if isinstance(mdp.action_space, policyopt.ContinuousSpace):
policy_cfg = rl.GaussianPolicyConfig(
hidden_spec=train_args['policy_hidden_spec'],
min_stdev=0.,
init_logstdev=0.,
enable_obsnorm=enable_obsnorm)
policy = rl.GaussianPolicy(policy_cfg, mdp.obs_space, mdp.action_space, 'GaussianPolicy')
else:
policy_cfg = rl.GibbsPolicyConfig(
hidden_spec=train_args['policy_hidden_spec'],
enable_obsnorm=enable_obsnorm)
policy = rl.GibbsPolicy(policy_cfg, mdp.obs_space, mdp.action_space, 'GibbsPolicy')
# Load the policy parameters
policy.load_h5(policy_file, policy_key)
return mdp, policy, train_args