def test_alf_metrics(self, num_envs, learn_queue_cap, unroll_length,
actor_queue_cap, num_actors, num_iterations):
episode_length = 5
env_f = lambda: TFPyEnvironment(
ValueUnittestEnv(batch_size=1, episode_length=episode_length))
envs = [env_f() for _ in range(num_envs)]
common.set_global_env(envs[0])
alg = _create_ac_algorithm()
driver = AsyncOffPolicyDriver(envs, alg, num_actors, unroll_length,
learn_queue_cap, actor_queue_cap)
driver.start()
total_num_steps_ = 0
for _ in range(num_iterations):
total_num_steps_ += driver.run_async()
driver.stop()
total_num_steps = int(driver.get_metrics()[1].result())
self.assertGreaterEqual(total_num_steps_, total_num_steps)
# An exp is only put in the log queue after it's put in the learning queue
# So when we stop the driver (which will force all queues to stop),
# some exps might be missing from the metric. Here we assert an arbitrary
# lower bound of 2/5. The upper bound is due to the fact that StepType.LAST
# is not recorded by the metric (episode_length==5).
self.assertLessEqual(total_num_steps, int(total_num_steps_ * 4 // 5))
self.assertGreaterEqual(total_num_steps,
int(total_num_steps_ * 2 // 5))
average_reward = int(driver.get_metrics()[2].result())
self.assertEqual(average_reward, episode_length - 1)
episode_length = int(driver.get_metrics()[3].result())
self.assertEqual(episode_length, episode_length)
def test_alf_metrics(self, num_envs, learn_queue_cap, unroll_length,
actor_queue_cap, num_actors, num_iterations):
episode_length = 5
env_f = lambda: TFPyEnvironment(
ValueUnittestEnv(batch_size=1, episode_length=episode_length))
alg = _create_ac_algorithm(env_f())
driver = AsyncOffPolicyDriver(env_f, alg, num_envs, num_actors,
unroll_length, learn_queue_cap,
actor_queue_cap)
driver.start()
total_num_steps_ = 0
for _ in range(num_iterations):
total_num_steps_ += driver.run_async()
driver.stop()
total_num_steps = int(driver.get_metrics()[1].result())
self.assertGreaterEqual(total_num_steps_, total_num_steps)
self.assertGreaterEqual(
total_num_steps, # multiply by 2/3 because 1/3 of steps are StepType.LAST
total_num_steps_ * 2 // 3)
average_reward = int(driver.get_metrics()[2].result())
self.assertEqual(average_reward, episode_length - 1)
episode_length = int(driver.get_metrics()[3].result())
self.assertEqual(episode_length, episode_length)