def main():
args = parser.parse_args()
run_experiment(json_to_dict(args.experiment),
monitor_path=args.monitor_path,
only_last=args.only_last,
description=args.description,
seed=args.seed)
def run_experiment(spec, monitor_path=None, only_last=False):
"""Run an experiment using a specification dictionary."""
args = spec["agent"]["args"]
if monitor_path:
args["monitor_path"] = monitor_path
else:
monitor_path = args["monitor_path"]
if not os.path.exists(monitor_path):
os.makedirs(monitor_path)
envs_type = spec["environments"]["type"]
if envs_type == "single":
envs = [make_environment(spec["environments"]["source"])]
elif envs_type == "json":
envs = make_environments(json_to_dict(spec["environments"]["source"]))
args["envs"] = envs
if len(envs) == 1 or only_last:
args["env"] = envs[-1]
action_space_type = "discrete" if isinstance(envs[0].action_space,
Discrete) else "continuous"
state_dimensions = "single" if len(
envs[0].observation_space.shape) == 1 else "multi"
agent = make_agent(spec["agent"]["name"], state_dimensions,
action_space_type, **args)
save_config(monitor_path, agent.config, [env.to_dict() for env in envs])
agent.learn()
def main():
args = parser.parse_args()
spec = cluster_spec(args.n_tasks, 1)
cluster = tf.train.ClusterSpec(spec).as_cluster_def()
cls = load("agents.actorcritic.a3c_worker:" + args.cls)
config = json_to_dict(args.config)
task = cls(args.env_id, args.task_id, cluster, args.monitor_path, config, video=args.video)
task.learn()
def main():
comm = MPI.Comm.Get_parent()
task_id = comm.Get_rank()
args = parser.parse_args()
cls = load("agents.ppo.dppo_worker:" + args.cls)
config = json_to_dict(args.config)
task = cls(args.env_id, task_id, comm, args.monitor_path, config,
args.seed)
task.run()
def run_experiment(spec,
monitor_path=None,
only_last=False,
description=None,
seed=None):
"""Run an experiment using a specification dictionary."""
import os
if seed is not None:
set_seed(seed)
import datetime
import gym
gym.logger.set_level(gym.logger.ERROR)
from gym.spaces import Discrete
from environment.registration import make, make_environments
from agents.registration import make_agent
args = spec["agent"]["args"]
args["config_path"] = os.path.join(monitor_path, "config.json")
if monitor_path:
args["monitor_path"] = monitor_path
else:
monitor_path = args["monitor_path"]
if not os.path.exists(monitor_path):
os.makedirs(monitor_path)
envs_type = spec["environments"]["type"]
if envs_type == "single":
envs = [make(spec["environments"]["source"])]
elif envs_type == "json":
envs = make_environments(json_to_dict(spec["environments"]["source"]))
if seed is not None:
for env in envs:
env.seed(seed)
args["seed"] = seed
args["envs"] = envs
if len(envs) == 1 or only_last:
args["env"] = envs[-1]
action_space_type = "discrete" if isinstance(envs[0].action_space,
Discrete) else "continuous"
state_dimensions = "single" if len(
envs[0].observation_space.shape) == 1 else "multi"
agent = make_agent(spec["agent"]["name"], state_dimensions,
action_space_type, **args)
config = agent.config.copy()
if description is not None:
config["description"] = description
config["seed"] = str(seed)
config["start_time"] = datetime.datetime.now().astimezone().isoformat()
save_config(monitor_path, config,
[env.metadata["parameters"] for env in envs])
agent.learn()
def main():
args = parser.parse_args()
run_experiment(json_to_dict(args.experiment),
monitor_path=args.monitor_path,
only_last=args.only_last)
