def get_domain(self, tag, args, path, debug=True):
if tag == 'Gridworld_SAS' or tag == 'SAS_Reco':
obj = utils.dynamic_load(path, tag, load_class=True)
env = obj(n_actions=args.n_actions,
debug=debug,
action_prob=args.action_prob)
return env, False, env.action_space.dtype == np.float32
elif tag == 'SF_Map':
obj = utils.dynamic_load(path, tag, load_class=True)
env = obj(debug=debug,
action_prob=args.action_prob,
bridge_prob=args.SF_bridge_prob)
return env, False, env.action_space.dtype == np.float32
else:
try:
import gym
from gym.spaces.box import Box
env = gym.make(tag)
return env, True, isinstance(env.action_space, Box)
except:
raise ValueError(
"Error! Environment neither available in Gym nor locally.")
def get_domain(self, tag, args, path, debug=True):
if tag == 'Gridworld_CL':
obj = utils.dynamic_load(path, tag, load_class=True)
env = obj(n_actions=args.n_actions, change_count=self.change_count, debug=debug, max_episodes=self.max_episodes)
return env, False, env.action_space.dtype == np.float32
else:
try:
import gym
from gym.spaces.box import Box
env = gym.make(tag)
return env, True, isinstance(env.action_space, Box)
except:
raise ValueError("Error! Environment neither available in Gym nor locally.")
def get_domain(self, tag, args, path, debug=True):
if tag == 'helpx' or tag == 'highbeam':
obj = utils.dynamic_load(path, 'RecommenderMDP', load_class=True)
env = obj(file=tag, ngram_n=2, debug=debug)
return env, False, False
elif tag == 'helpx_CL' or tag == 'highbeam_CL':
obj = utils.dynamic_load(path, 'RecommenderMDP_CL', load_class=True)
env = obj(file=tag[:-3], ngram_n=2, debug=debug, change_count=self.change_count,
dynamic_reward_action=self.dynamic_reward_action, max_episodes=self.max_episodes, change_remove=args.change_remove)
return env, False, False
elif tag == 'Gridworld_CL':
obj = utils.dynamic_load(path, tag, load_class=True)
env = obj(n_actions=args.n_actions, uniform_action_spread=self.uniform_action_spread,
change_count=self.change_count, debug=debug, max_episodes=self.max_episodes, change_remove=args.change_remove)
return env, False, env.action_space.dtype == np.float32
elif tag == 'Gridworld' or tag == 'Gridworld_vanilla':
obj = utils.dynamic_load(path, tag, load_class=True)
env = obj(n_actions=args.n_actions, debug=debug)
return env, False, env.action_space.dtype == np.float32
elif tag == 'six_room':
obj = utils.dynamic_load(path, tag, load_class=True)
env = obj(difficulty=args.difficulty, debug=debug)
return env, False, env.action_space.dtype == np.float32
elif tag == 'Gridworld_SAS' or tag == 'SAS_Bandit' or tag == 'SAS_Reco':
obj = utils.dynamic_load(path, tag, load_class=True)
env = obj(n_actions=args.n_actions, debug=debug, action_prob=args.action_prob)
return env, False, env.action_space.dtype == np.float32
elif tag == 'SF_Map':
obj = utils.dynamic_load(path, tag, load_class=True)
env = obj(debug=debug, action_prob=args.action_prob, bridge_prob=args.SF_bridge_prob)
return env, False, env.action_space.dtype == np.float32
else:
try:
import gym
from gym.spaces.box import Box
env = gym.make(tag)
return env, True, isinstance(env.action_space, Box)
except:
raise ValueError("Error! Environment neither available in Gym nor locally.")
def get_domain(self, tag, args, path, debug=True):
if tag == 'NS_Reco' or tag == 'NS_Reacher':
obj = utils.dynamic_load(path, tag, load_class=True)
env = obj(speed=args.speed, oracle=args.oracle, debug=debug)
return env, False, env.action_space.dtype == np.float32
else:
import gym
from gym.spaces.box import Box
# Register gym environment. By specifying kwargs,
# you are able to choose which patient to simulate.
# patient_name must be 'adolescent#001' to 'adolescent#010',
# or 'adult#001' to 'adult#010', or 'child#001' to 'child#010'
from gym.envs.registration import register
register(
id='NS_SimGlucose-v0',
entry_point='Environments.SimGlucose.simglucose.envs:NS_T1DSimEnv',
kwargs={'patient_name': 'adolescent#002', 'speed': self.speed, 'oracle': self.oracle, 'seed': self.seed}
)
env = gym.make(tag)
return env, True, isinstance(env.action_space, Box)
def __init__(self, args):
# SET UP PATHS
self.paths = OrderedDict()
self.paths['root'] = path.abspath(path.join(path.dirname(__file__), '..'))
# Do Hyper-parameter sweep, if needed
self.idx = args.base + args.inc
if self.idx >= 0 and args.hyper != 'default':
self.hyperparam_sweep = utils.dynamic_load(self.paths['root'], "random_search_" + str(args.hyper), load_class=False)
self.hyperparam_sweep.set(args, self.idx)
del self.hyperparam_sweep # *IMP: CanNOT deepcopy an object having reference to an imported library (DPG)
# Make results reproducible
seed = args.seed
self.seed = seed
np.random.seed(seed)
torch.manual_seed(seed)
# Copy all the variables from args to config
self.__dict__.update(vars(args))
# Frequency of saving results and models.
self.save_after = args.max_episodes // args.save_count if args.max_episodes > args.save_count else args.max_episodes
# add path to models
folder_suffix = args.experiment + args.folder_suffix
self.paths['Experiments'] = path.join(self.paths['root'], 'Experiments')
self.paths['experiment'] = path.join(self.paths['Experiments'], args.env_name, args.algo_name, folder_suffix)
if args.swarm:
# DO not create different folder for each seed in swarm
# Takes up a lot of space and makes data aggregation slow
self.paths['logs'] = self.paths['experiment'] + '/'
self.paths['ckpt'] = self.paths['experiment'] + '/'
self.paths['results'] = self.paths['experiment'] + '/'
else:
path_prefix = [self.paths['experiment'], str(args.seed)]
self.paths['logs'] = path.join(*path_prefix, 'Logs/')
self.paths['ckpt'] = path.join(*path_prefix, 'Checkpoints/')
self.paths['results'] = path.join(*path_prefix, 'Results/')
# Create directories
for (key, val) in self.paths.items():
if key not in ['root', 'datasets', 'data']:
utils.create_directory_tree(val)
# Save the all the configuration settings
dump(args.__dict__, open(path.join(self.paths['experiment'], 'args.yaml'), 'w'), default_flow_style=False,
explicit_start=True)
# Output logging
sys.stdout = utils.Logger(self.paths['logs'], args.restore, args.log_output)
# Get the domain and algorithm
self.env, self.gym_env, self.cont_actions = self.get_domain(args.env_name, args=args, debug=args.debug,
path=path.join(self.paths['root'], 'Environments'))
try:
self.env.seed(seed.item())
except:
self.env.seed(seed)
# Set Model
self.algo = utils.dynamic_load(path.join(self.paths['root'], 'Src', 'Algorithms'), args.algo_name, load_class=True)
self.feature_dim = [int(size) for size in args.NN_basis_dim.split(',')]
self.policy_basis_dim = [int(size) for size in args.Policy_basis_dim.split(',')]
# GPU
self.device = torch.device('cuda' if args.gpu else 'cpu')
# optimizer
if args.optim == 'adam':
self.optim = torch.optim.Adam
elif args.optim == 'rmsprop':
self.optim = torch.optim.RMSprop
elif args.optim == 'sgd':
self.optim = torch.optim.SGD
else:
raise ValueError('Undefined type of optmizer')
print("=====Configurations=====\n", args)
