本文整理汇总了Python中ray.rllib.evaluation.policy_evaluator.PolicyEvaluator类的典型用法代码示例。如果您正苦于以下问题：Python PolicyEvaluator类的具体用法？Python PolicyEvaluator怎么用？Python PolicyEvaluator使用的例子？那么恭喜您, 这里精选的类代码示例或许可以为您提供帮助。
在下文中一共展示了PolicyEvaluator类的15个代码示例，这些例子默认根据受欢迎程度排序。您可以为喜欢或者感觉有用的代码点赞，您的评价将有助于系统推荐出更棒的Python代码示例。
示例1: testMultiAgentSampleRoundRobin
def testMultiAgentSampleRoundRobin(self):
act_space = gym.spaces.Discrete(2)
obs_space = gym.spaces.Discrete(10)
ev = PolicyEvaluator(
env_creator=lambda _: RoundRobinMultiAgent(5, increment_obs=True),
policy_graph={
"p0": (MockPolicyGraph, obs_space, act_space, {}),
},
policy_mapping_fn=lambda agent_id: "p0",
batch_steps=50)
batch = ev.sample()
self.assertEqual(batch.count, 50)
# since we round robin introduce agents into the env, some of the env
# steps don't count as proper transitions
self.assertEqual(batch.policy_batches["p0"].count, 42)
self.assertEqual(batch.policy_batches["p0"]["obs"].tolist()[:10], [
one_hot(0, 10),
one_hot(1, 10),
one_hot(2, 10),
one_hot(3, 10),
one_hot(4, 10),
] * 2)
self.assertEqual(batch.policy_batches["p0"]["new_obs"].tolist()[:10], [
one_hot(1, 10),
one_hot(2, 10),
one_hot(3, 10),
one_hot(4, 10),
one_hot(5, 10),
] * 2)
self.assertEqual(batch.policy_batches["p0"]["rewards"].tolist()[:10],
[100, 100, 100, 100, 0] * 2)
self.assertEqual(batch.policy_batches["p0"]["dones"].tolist()[:10],
[False, False, False, False, True] * 2)
self.assertEqual(batch.policy_batches["p0"]["t"].tolist()[:10],
[4, 9, 14, 19, 24, 5, 10, 15, 20, 25])示例2: testCustomRNNStateValues
def testCustomRNNStateValues(self):
h = {"some": {"arbitrary": "structure", "here": [1, 2, 3]}}
class StatefulPolicyGraph(PolicyGraph):
def compute_actions(self,
obs_batch,
state_batches,
prev_action_batch=None,
prev_reward_batch=None,
episodes=None,
**kwargs):
return [0] * len(obs_batch), [[h] * len(obs_batch)], {}
def get_initial_state(self):
return [{}] # empty dict
ev = PolicyEvaluator(
env_creator=lambda _: gym.make("CartPole-v0"),
policy_graph=StatefulPolicyGraph,
batch_steps=5)
batch = ev.sample()
self.assertEqual(batch.count, 5)
self.assertEqual(batch["state_in_0"][0], {})
self.assertEqual(batch["state_out_0"][0], h)
self.assertEqual(batch["state_in_0"][1], h)
self.assertEqual(batch["state_out_0"][1], h)示例3: testCompleteEpisodes
def testCompleteEpisodes(self):
ev = PolicyEvaluator(
env_creator=lambda _: MockEnv(10),
policy_graph=MockPolicyGraph,
batch_steps=5,
batch_mode="complete_episodes")
batch = ev.sample()
self.assertEqual(batch.count, 10)示例4: testExternalEnvHorizonNotSupported
def testExternalEnvHorizonNotSupported(self):
ev = PolicyEvaluator(
env_creator=lambda _: SimpleServing(MockEnv(25)),
policy_graph=MockPolicyGraph,
episode_horizon=20,
batch_steps=10,
batch_mode="complete_episodes")
self.assertRaises(ValueError, lambda: ev.sample())示例5: testExternalEnvBadActions
def testExternalEnvBadActions(self):
ev = PolicyEvaluator(
env_creator=lambda _: SimpleServing(MockEnv(25)),
policy_graph=BadPolicyGraph,
sample_async=True,
batch_steps=40,
batch_mode="truncate_episodes")
self.assertRaises(Exception, lambda: ev.sample())示例6: testAsync
def testAsync(self):
ev = PolicyEvaluator(
env_creator=lambda _: gym.make("CartPole-v0"),
sample_async=True,
policy_graph=MockPolicyGraph)
batch = ev.sample()
for key in ["obs", "actions", "rewards", "dones", "advantages"]:
self.assertIn(key, batch)
self.assertGreater(batch["advantages"][0], 1)示例7: testExternalEnvTruncateEpisodes
def testExternalEnvTruncateEpisodes(self):
ev = PolicyEvaluator(
env_creator=lambda _: SimpleServing(MockEnv(25)),
policy_graph=MockPolicyGraph,
batch_steps=40,
batch_mode="truncate_episodes")
for _ in range(3):
batch = ev.sample()
self.assertEqual(batch.count, 40)示例8: testCompleteEpisodesPacking
def testCompleteEpisodesPacking(self):
ev = PolicyEvaluator(
env_creator=lambda _: MockEnv(10),
policy_graph=MockPolicyGraph,
batch_steps=15,
batch_mode="complete_episodes")
batch = ev.sample()
self.assertEqual(batch.count, 20)
self.assertEqual(
batch["t"].tolist(),
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9])示例9: testExternalEnvOffPolicy
def testExternalEnvOffPolicy(self):
ev = PolicyEvaluator(
env_creator=lambda _: SimpleOffPolicyServing(MockEnv(25), 42),
policy_graph=MockPolicyGraph,
batch_steps=40,
batch_mode="complete_episodes")
for _ in range(3):
batch = ev.sample()
self.assertEqual(batch.count, 50)
self.assertEqual(batch["actions"][0], 42)
self.assertEqual(batch["actions"][-1], 42)示例10: testBaselinePerformance
def testBaselinePerformance(self):
ev = PolicyEvaluator(
env_creator=lambda _: gym.make("CartPole-v0"),
policy_graph=MockPolicyGraph,
batch_steps=100)
start = time.time()
count = 0
while time.time() - start < 1:
count += ev.sample().count
print()
print("Samples per second {}".format(count / (time.time() - start)))
print()示例11: testFilterSync
def testFilterSync(self):
ev = PolicyEvaluator(
env_creator=lambda _: gym.make("CartPole-v0"),
policy_graph=MockPolicyGraph,
sample_async=True,
observation_filter="ConcurrentMeanStdFilter")
time.sleep(2)
ev.sample()
filters = ev.get_filters(flush_after=True)
obs_f = filters["default"]
self.assertNotEqual(obs_f.rs.n, 0)
self.assertNotEqual(obs_f.buffer.n, 0)示例12: testBatchesLargerWhenVectorized
def testBatchesLargerWhenVectorized(self):
ev = PolicyEvaluator(
env_creator=lambda _: MockEnv(episode_length=8),
policy_graph=MockPolicyGraph,
batch_mode="truncate_episodes",
batch_steps=4,
num_envs=4)
batch = ev.sample()
self.assertEqual(batch.count, 16)
result = collect_metrics(ev, [])
self.assertEqual(result["episodes_this_iter"], 0)
batch = ev.sample()
result = collect_metrics(ev, [])
self.assertEqual(result["episodes_this_iter"], 4)示例13: testGetFilters
def testGetFilters(self):
ev = PolicyEvaluator(
env_creator=lambda _: gym.make("CartPole-v0"),
policy_graph=MockPolicyGraph,
sample_async=True,
observation_filter="ConcurrentMeanStdFilter")
self.sample_and_flush(ev)
filters = ev.get_filters(flush_after=False)
time.sleep(2)
filters2 = ev.get_filters(flush_after=False)
obs_f = filters["default"]
obs_f2 = filters2["default"]
self.assertGreaterEqual(obs_f2.rs.n, obs_f.rs.n)
self.assertGreaterEqual(obs_f2.buffer.n, obs_f.buffer.n)示例14: testMultiAgentSampleWithHorizon
def testMultiAgentSampleWithHorizon(self):
act_space = gym.spaces.Discrete(2)
obs_space = gym.spaces.Discrete(2)
ev = PolicyEvaluator(
env_creator=lambda _: BasicMultiAgent(5),
policy_graph={
"p0": (MockPolicyGraph, obs_space, act_space, {}),
"p1": (MockPolicyGraph, obs_space, act_space, {}),
},
policy_mapping_fn=lambda agent_id: "p{}".format(agent_id % 2),
episode_horizon=10, # test with episode horizon set
batch_steps=50)
batch = ev.sample()
self.assertEqual(batch.count, 50)示例15: testMetrics
def testMetrics(self):
ev = PolicyEvaluator(
env_creator=lambda _: MockEnv(episode_length=10),
policy_graph=MockPolicyGraph,
batch_mode="complete_episodes")
remote_ev = PolicyEvaluator.as_remote().remote(
env_creator=lambda _: MockEnv(episode_length=10),
policy_graph=MockPolicyGraph,
batch_mode="complete_episodes")
ev.sample()
ray.get(remote_ev.sample.remote())
result = collect_metrics(ev, [remote_ev])
self.assertEqual(result["episodes_this_iter"], 20)
self.assertEqual(result["episode_reward_mean"], 10)