def run(self, env: Env, n_steps=1000000):
env = self.wrap_env(env)
env.start()
try:
self._run(env, n_steps)
except KeyboardInterrupt:
env.stop()
def run(self, env: Env, expt, event_buffer=None, n_steps=1000000):
env = self.wrap_env(env)
env.start()
try:
self._run(env, expt, event_buffer, n_steps)
except KeyboardInterrupt:
env.stop()
self.on_finish()
def wrap_env(self, env: Env) -> Env:
render, env.render = env.render, False
envs = [env] + [copy.deepcopy(env) for _ in range(self.n_envs - 1)]
env.render = render
return MultiProcEnv(envs)
def run(self, env: Env, n_steps=1000000):
if self.args.test or (env.id
not in SUB_ENV_DICT) or (not self.args.HRL):
# either testing or training without HRL at all
# or the env selected does not have the subenvs
if not self.args.HRL or self.args.HRL != 'separate':
env = self.wrap_env(env)
env.start()
try:
self.on_start()
self._run(env, n_steps)
except KeyboardInterrupt:
env.stop()
self.on_finish()
# testing with HRL and separate subenvs
else:
env = self.wrap_env(env)
env.start()
subenvs = SUB_ENV_DICT[env.id]
print(
LOGGING_MSG_HEADER +
": Ensure that you are testing models trained using the <HRL_separate> approach."
)
print(
LOGGING_MSG_HEADER +
": Testing the {} with combined subpolicies trained seperately from subenvs-{}"
.format(env.id, subenvs))
try:
self.on_start()
self._run_subenvs(env, n_steps, subenvs=subenvs)
except KeyboardInterrupt:
env.stop()
self.on_finish()
else:
assert self.args.HRL in [
'human', 'systematic', 'random', 'sequential', 'separate'
]
subenvs = SUB_ENV_DICT[env.id]
print(LOGGING_MSG_HEADER + "： Subenvs are: ", subenvs)
subenv_steps = [n_steps // len(subenvs) for subenv in subenvs]
thresholds = [None for subenv in subenvs]
if self.args.HRL in ['human', 'sequential', 'separate']:
thresholds = HRL_thredhold(env.id)
print(LOGGING_MSG_HEADER + "： Reward thresholds are: ",
thresholds)
elif self.args.HRL == 'random':
import numpy as np
np.random.seed(1234)
indices = sorted(
np.random.choice(n_steps, len(subenvs) - 1, replace=False))
indices = [0] + sorted(
np.random.choice(n_steps, I - 1,
replace=False)) + [n_steps]
subenv_steps = np.ediff1d(indices)
elif self.args.HRL == 'systematic':
pass
# subenv_steps already defined and initializied
for i, (subenv, subenv_step, threshold) in enumerate(
zip(subenvs, subenv_steps, thresholds)):
env = SC2Env(subenv, env.render, max_ep_len=env.max_ep_len)
print(
LOGGING_MSG_HEADER +
": Creating and Running subenv : {} with maximum {} steps, and reward threshold is {}."
.format(env.id, subenv_step, threshold))
env = self.wrap_env(env)
env.start()
try:
self.on_start()
if i != 0:
self.reset()
self._run(env, subenv_step, threshold, subenv_id=i)
except KeyboardInterrupt:
env.stop()
break
self.on_finish()
"""