def check_support(alg, config, stats):
for a_name, action_space in ACTION_SPACES_TO_TEST.items():
for o_name, obs_space in OBSERVATION_SPACES_TO_TEST.items():
print("=== Testing", alg, action_space, obs_space, "===")
stub_env = make_stub_env(action_space, obs_space)
register_env(
"stub_env", lambda c: stub_env())
stat = "ok"
a = None
try:
a = get_agent_class(alg)(config=config, env="stub_env")
a.train()
except UnsupportedSpaceException as e:
stat = "unsupported"
except Exception as e:
stat = "ERROR"
print(e)
print(traceback.format_exc())
finally:
if a:
try:
a.stop()
except Exception as e:
print("Ignoring error stopping agent", e)
pass
print(stat)
print()
stats[alg, a_name, o_name] = stat
def _register_all():
for key in ["PPO", "ES", "DQN", "APEX", "A3C", "BC", "PG", "__fake",
"__sigmoid_fake_data", "__parameter_tuning"]:
try:
from ray.rllib.agent import get_agent_class
register_trainable(key, get_agent_class(key))
except ImportError as e:
print("Warning: could not import {}: {}".format(key, e))
def test(use_object_store, alg_name):
cls = get_agent_class(alg_name)
alg1 = cls(config=CONFIGS[name], env="CartPole-v0")
alg2 = cls(config=CONFIGS[name], env="CartPole-v0")
for _ in range(3):
res = alg1.train()
print("current status: " + str(res))
# Sync the models
if use_object_store:
alg2.restore_from_object(alg1.save_to_object())
else:
alg2.restore(alg1.save())
for _ in range(10):
obs = np.random.uniform(size=4)
a1 = get_mean_action(alg1, obs)
a2 = get_mean_action(alg2, obs)
print("Checking computed actions", alg1, obs, a1, a2)
assert abs(a1 - a2) < .1, (a1, a2)
if not args.config:
# Load configuration from file
config_dir = os.path.dirname(args.checkpoint)
config_path = os.path.join(config_dir, "params.json")
with open(config_path) as f:
args.config = json.load(f)
if not args.env:
if not args.config.get("env"):
parser.error("the following arguments are required: --env")
args.env = args.config.get("env")
ray.init()
cls = get_agent_class(args.run)
agent = cls(env=args.env, config=args.config)
agent.restore(args.checkpoint)
num_steps = int(args.steps)
if args.run == "DQN":
env = gym.make(args.env)
env = wrap_dqn(get_registry(), env, args.config.get("model", {}))
else:
env = ModelCatalog.get_preprocessor_as_wrapper(get_registry(),
gym.make(args.env))
if args.out is not None:
rollouts = []
steps = 0
while steps < (num_steps or steps + 1):
if args.out is not None:
if not args.config:
# Load configuration from file
config_dir = os.path.dirname(args.checkpoint)
config_path = os.path.join(config_dir, "params.json")
with open(config_path) as f:
args.config = json.load(f)
if not args.env:
if not args.config.get("env"):
parser.error("the following arguments are required: --env")
args.env = args.config.get("env")
ray.init()
cls = get_agent_class(args.run)
agent = cls(env=args.env, config=args.config)
agent.restore(args.checkpoint)
num_steps = int(args.steps)
if args.run == "DQN":
env = gym.make(args.env)
env = wrap_dqn(get_registry(), env, args.config.get("model", {}))
else:
env = ModelCatalog.get_preprocessor_as_wrapper(get_registry(),
gym.make(args.env))
if args.out is not None:
rollouts = []
steps = 0
while steps < (num_steps or steps + 1):
if args.out is not None:
ray.init()
CONFIGS = {
"DQN": {},
"PPO": {
"num_sgd_iter": 5,
"timesteps_per_batch": 1000
},
"A3C": {
"use_lstm": False
},
}
# https://github.com/ray-project/ray/issues/1062 for enabling ES test as well
for name in ["DQN", "PPO", "A3C"]:
cls = get_agent_class(name)
alg1 = cls("CartPole-v0", CONFIGS[name])
alg2 = cls("CartPole-v0", CONFIGS[name])
for _ in range(3):
res = alg1.train()
print("current status: " + str(res))
# Sync the models
alg2.restore(alg1.save())
for _ in range(10):
obs = [
random.random(),
random.random(),
random.random(),
def _register_all():
for key in ["PPO", "ES", "DQN", "A3C", "__fake", "__sigmoid_fake_data"]:
try:
register_trainable(key, get_agent_class(key))
except ImportError as e:
print("Warning: could not import {}: {}".format(key, e))
def _register_all():
for key in ["PPO", "ES", "DQN", "APEX", "A3C", "BC", "PG", "DDPG",
"APEX_DDPG", "__fake", "__sigmoid_fake_data",
"__parameter_tuning"]:
from ray.rllib.agent import get_agent_class
register_trainable(key, get_agent_class(key))