def testFakeMTRandomness(self):
mutex = threading.Lock()
@contextlib.contextmanager
def fake_with_rng(rrr):
from tartist.random import rng
with mutex:
backup = rng._rng
rng._rng = rrr
yield rrr
with mutex:
rng._rng = backup
q = queue.Queue()
def proc():
rng = tar.gen_rng()
with fake_with_rng(rng):
time.sleep(0.5)
state = tar.get_rng().get_state()
time.sleep(0.5)
q.put(state)
threads = [Thread(target=proc) for i in range(2)]
map_exec(Thread.start, threads)
map_exec(Thread.join, threads)
v1, v2 = q.get(), q.get()
self.assertFalse(not np.allclose(v1[1], v2[1]))
def main_inference_play_multithread(trainer):
def runner():
func = trainer.env.make_func()
func.compile({'theta': trainer.env.network.outputs['theta']})
player = make_player()
score = _evaluate(player, func)
mgr = trainer.runtime.get('summary_histories', None)
if mgr is not None:
mgr.put_async_scalar('inference/score', score)
nr_players = get_env('ppo.inference.nr_plays')
pool = [threading.Thread(target=runner) for _ in range(nr_players)]
map_exec(threading.Thread.start, pool)
map_exec(threading.Thread.join, pool)
def main_inference_play_multithread(trainer):
def runner():
func = trainer.env.make_func()
func.compile(trainer.env.network.outputs['q_argmax'])
player = make_player()
score = player.evaluate_one_episode(
lambda state: func(state=state[np.newaxis])[0])
mgr = trainer.runtime.get('summary_histories', None)
if mgr is not None:
mgr.put_async_scalar('inference/score', score)
nr_players = get_env('dqn.inference.nr_plays')
pool = [threading.Thread(target=runner) for _ in range(nr_players)]
map_exec(threading.Thread.start, pool)
map_exec(threading.Thread.join, pool)
def testMTRandomness(self):
q = queue.Queue()
def proc():
rng = tar.gen_rng()
with tar.with_rng(rng):
time.sleep(0.5)
state = tar.get_rng().get_state()
time.sleep(0.5)
q.put(state)
threads = [Thread(target=proc) for i in range(2)]
map_exec(Thread.start, threads)
map_exec(Thread.join, threads)
v1, v2 = q.get(), q.get()
self.assertFalse(np.allclose(v1[1], v2[1]))
def _compute_e_var(rs, ret_variance):
e_r = sum(rs) / len(rs)
if ret_variance:
rs_sqr = map_exec(lambda x: x**2, rs)
var_r = sum(rs_sqr) / len(rs_sqr) - e_r**2
return e_r, var_r
return e_r