def testServingEnvTruncateEpisodes(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)
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)
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)
def testBatchDivisibilityCheck(self):
self.assertRaises(
ValueError,
lambda: PolicyEvaluator(
env_creator=lambda _: MockEnv(episode_length=8),
policy_graph=MockPolicyGraph,
batch_mode="truncate_episodes",
batch_steps=15, num_envs=4))
def testServingEnvHorizonNotSupported(self):
ev = PolicyEvaluator(env_creator=lambda _: SimpleServing(MockEnv(25)),
policy_graph=MockPolicyGraph,
episode_horizon=20,
batch_steps=10,
batch_mode="complete_episodes")
ev.sample()
self.assertRaises(Exception, lambda: ev.sample())
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())
def _testWithOptimizer(self, optimizer_cls):
n = 3
env = gym.make("CartPole-v0")
act_space = env.action_space
obs_space = env.observation_space
dqn_config = {"gamma": 0.95, "n_step": 3}
if optimizer_cls == SyncReplayOptimizer:
# TODO: support replay with non-DQN graphs. Currently this can't
# happen since the replay buffer doesn't encode extra fields like
# "advantages" that PG uses.
policies = {
"p1": (DQNTFPolicy, obs_space, act_space, dqn_config),
"p2": (DQNTFPolicy, obs_space, act_space, dqn_config),
}
else:
policies = {
"p1": (PGTFPolicy, obs_space, act_space, {}),
"p2": (DQNTFPolicy, obs_space, act_space, dqn_config),
}
ev = PolicyEvaluator(
env_creator=lambda _: MultiCartpole(n),
policy=policies,
policy_mapping_fn=lambda agent_id: ["p1", "p2"][agent_id % 2],
batch_steps=50)
if optimizer_cls == AsyncGradientsOptimizer:
def policy_mapper(agent_id):
return ["p1", "p2"][agent_id % 2]
remote_evs = [
PolicyEvaluator.as_remote().remote(
env_creator=lambda _: MultiCartpole(n),
policy=policies,
policy_mapping_fn=policy_mapper,
batch_steps=50)
]
else:
remote_evs = []
optimizer = optimizer_cls(ev, remote_evs)
for i in range(200):
ev.foreach_policy(lambda p, _: p.set_epsilon(max(
0.02, 1 - i * .02)) if isinstance(p, DQNTFPolicy) else None)
optimizer.step()
result = collect_metrics(ev, remote_evs)
if i % 20 == 0:
def do_update(p):
if isinstance(p, DQNTFPolicy):
p.update_target()
ev.foreach_policy(lambda p, _: do_update(p))
print("Iter {}, rew {}".format(i,
result["policy_reward_mean"]))
print("Total reward", result["episode_reward_mean"])
if result["episode_reward_mean"] >= 25 * n:
return
print(result)
raise Exception("failed to improve reward")
def testBatchIds(self):
ev = PolicyEvaluator(env_creator=lambda _: gym.make("CartPole-v0"),
policy_graph=MockPolicyGraph)
batch1 = ev.sample()
batch2 = ev.sample()
self.assertEqual(len(set(batch1["unroll_id"])), 1)
self.assertEqual(len(set(batch2["unroll_id"])), 1)
self.assertEqual(
len(set(SampleBatch.concat(batch1, batch2)["unroll_id"])), 2)
def testAutoConcat(self):
ev = PolicyEvaluator(env_creator=lambda _: MockEnv(episode_length=40),
policy_graph=MockPolicyGraph,
sample_async=True,
batch_steps=10,
batch_mode="truncate_episodes",
observation_filter="ConcurrentMeanStdFilter")
time.sleep(2)
batch = ev.sample()
self.assertEqual(batch.count, 40) # auto-concat up to 5 episodes
def testRewardClipping(self):
# clipping on
ev = PolicyEvaluator(env_creator=lambda _: MockEnv2(episode_length=10),
policy_graph=MockPolicyGraph,
clip_rewards=True,
batch_mode="complete_episodes")
self.assertEqual(max(ev.sample()["rewards"]), 1)
result = collect_metrics(ev, [])
self.assertEqual(result["episode_reward_mean"], 1000)
# clipping off
ev2 = PolicyEvaluator(
env_creator=lambda _: MockEnv2(episode_length=10),
policy_graph=MockPolicyGraph,
clip_rewards=False,
batch_mode="complete_episodes")
self.assertEqual(max(ev2.sample()["rewards"]), 100)
result2 = collect_metrics(ev2, [])
self.assertEqual(result2["episode_reward_mean"], 1000)
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])
def make(cls,
env_creator,
policy_graph,
optimizer_batch_size=None,
num_workers=0,
num_envs_per_worker=None,
optimizer_config=None,
remote_num_cpus=None,
remote_num_gpus=None,
**eval_kwargs):
"""Creates an Optimizer with local and remote evaluators.
Args:
env_creator(func): Function that returns a gym.Env given an
EnvContext wrapped configuration.
policy_graph (class|dict): Either a class implementing
PolicyGraph, or a dictionary of policy id strings to
(PolicyGraph, obs_space, action_space, config) tuples.
See PolicyEvaluator documentation.
optimizer_batch_size (int): Batch size summed across all workers.
Will override worker `batch_steps`.
num_workers (int): Number of remote evaluators
num_envs_per_worker (int): (Optional) Sets the number
environments per evaluator for vectorization.
If set, overrides `num_envs` in kwargs
for PolicyEvaluator.__init__.
optimizer_config (dict): Config passed to the optimizer.
remote_num_cpus (int): CPU specification for remote evaluator.
remote_num_gpus (int): GPU specification for remote evaluator.
**eval_kwargs: PolicyEvaluator Class non-positional args.
Returns:
(Optimizer) Instance of `cls` with evaluators configured
accordingly.
"""
optimizer_config = optimizer_config or {}
if num_envs_per_worker:
assert num_envs_per_worker > 0, "Improper num_envs_per_worker!"
eval_kwargs["num_envs"] = int(num_envs_per_worker)
if optimizer_batch_size:
assert optimizer_batch_size > 0
if num_workers > 1:
eval_kwargs["batch_steps"] = \
optimizer_batch_size // num_workers
else:
eval_kwargs["batch_steps"] = optimizer_batch_size
evaluator = PolicyEvaluator(env_creator, policy_graph, **eval_kwargs)
remote_cls = PolicyEvaluator.as_remote(remote_num_cpus,
remote_num_gpus)
remote_evaluators = [
remote_cls.remote(env_creator, policy_graph, **eval_kwargs)
for i in range(num_workers)
]
return cls(evaluator, remote_evaluators, optimizer_config)
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()
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)
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)
def testExternalMultiAgentEnvTruncateEpisodes(self):
agents = 4
ev = PolicyEvaluator(
env_creator=lambda _: SimpleMultiServing(BasicMultiAgent(agents)),
policy_graph=MockPolicyGraph,
batch_steps=40,
batch_mode="truncate_episodes")
for _ in range(3):
batch = ev.sample()
self.assertEqual(batch.count, 160)
self.assertEqual(len(np.unique(batch["agent_index"])), agents)
def testSoftHorizon(self):
ev = PolicyEvaluator(env_creator=lambda _: MockEnv(episode_length=10),
policy_graph=MockPolicyGraph,
batch_mode="complete_episodes",
batch_steps=10,
episode_horizon=4,
soft_horizon=True)
samples = ev.sample()
# three logical episodes
self.assertEqual(len(set(samples["eps_id"])), 3)
# only 1 hard done value
self.assertEqual(sum(samples["dones"]), 1)
def testHardHorizon(self):
ev = PolicyEvaluator(env_creator=lambda _: MockEnv(episode_length=10),
policy=MockPolicy,
batch_mode="complete_episodes",
batch_steps=10,
episode_horizon=4,
soft_horizon=False)
samples = ev.sample()
# three logical episodes
self.assertEqual(len(set(samples["eps_id"])), 3)
# 3 done values
self.assertEqual(sum(samples["dones"]), 3)
def testBatchesLargerWhenVectorized(self):
ev = PolicyEvaluator(env_creator=lambda _: MockEnv(episode_length=8),
policy=MockPolicy,
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)
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)
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_total, 20)
self.assertEqual(result.episode_reward_mean, 10)
def testBatchesSmallerWhenVectorized(self):
ev = PolicyEvaluator(env_creator=lambda _: MockEnv(episode_length=8),
policy_graph=MockPolicyGraph,
batch_mode="truncate_episodes",
batch_steps=16,
num_envs=4)
batch = ev.sample()
self.assertEqual(batch.count, 16)
result = collect_metrics(ev, [])
self.assertEqual(result.episodes_total, 0)
batch = ev.sample()
result = collect_metrics(ev, [])
self.assertEqual(result.episodes_total, 4)
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)
def testExternalMultiAgentEnvSample(self):
agents = 2
act_space = gym.spaces.Discrete(2)
obs_space = gym.spaces.Discrete(2)
ev = PolicyEvaluator(
env_creator=lambda _: SimpleMultiServing(BasicMultiAgent(agents)),
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),
batch_steps=50)
batch = ev.sample()
self.assertEqual(batch.count, 50)
def _make_evs(self):
def make_sess():
return tf.Session(config=tf.ConfigProto(device_count={"CPU": 2}))
local = PolicyEvaluator(env_creator=lambda _: gym.make("CartPole-v0"),
policy_graph=PPOPolicyGraph,
tf_session_creator=make_sess)
remotes = [
PolicyEvaluator.as_remote().remote(
env_creator=lambda _: gym.make("CartPole-v0"),
policy_graph=PPOPolicyGraph,
tf_session_creator=make_sess)
]
return local, remotes
def testSampleFromEarlyDoneEnv(self):
act_space = gym.spaces.Discrete(2)
obs_space = gym.spaces.Discrete(2)
ev = PolicyEvaluator(
env_creator=lambda _: EarlyDoneMultiAgent(),
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),
batch_mode="complete_episodes",
batch_steps=1)
self.assertRaisesRegexp(ValueError,
".*don't have a last observation.*",
lambda: ev.sample())
def testMultiAgentSampleAsyncRemote(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),
batch_steps=50,
num_envs=4,
async_remote_worker_envs=True)
batch = ev.sample()
self.assertEqual(batch.count, 200)
def testVectorEnvSupport(self):
ev = PolicyEvaluator(
env_creator=lambda _: MockVectorEnv(episode_length=20, num_envs=8),
policy_graph=MockPolicyGraph,
batch_mode="truncate_episodes",
batch_steps=10)
for _ in range(8):
batch = ev.sample()
self.assertEqual(batch.count, 10)
result = collect_metrics(ev, [])
self.assertEqual(result.episodes_total, 0)
for _ in range(8):
batch = ev.sample()
self.assertEqual(batch.count, 10)
result = collect_metrics(ev, [])
self.assertEqual(result.episodes_total, 8)
def testReturningModelBasedRolloutsData(self):
class ModelBasedPolicyGraph(PGPolicyGraph):
def compute_actions(self,
obs_batch,
state_batches,
prev_action_batch=None,
prev_reward_batch=None,
is_training=False,
episodes=None):
# Pretend we did a model-based rollout and want to return
# the extra trajectory.
builder = episodes[0].new_batch_builder()
rollout_id = random.randint(0, 10000)
for t in range(5):
builder.add_values(
agent_id="extra_0",
policy_id="p1", # use p1 so we can easily check it
t=t,
eps_id=rollout_id, # new id for each rollout
obs=obs_batch[0],
actions=0,
rewards=0,
dones=t == 4,
infos={},
new_obs=obs_batch[0])
batch = builder.build_and_reset()
episodes[0].add_extra_batch(batch)
# Just return zeros for actions
return [0] * len(obs_batch), [], {}
single_env = gym.make("CartPole-v0")
obs_space = single_env.observation_space
act_space = single_env.action_space
ev = PolicyEvaluator(env_creator=lambda _: MultiCartpole(2),
policy_graph={
"p0": (ModelBasedPolicyGraph, obs_space,
act_space, {}),
"p1": (ModelBasedPolicyGraph, obs_space,
act_space, {}),
},
policy_mapping_fn=lambda agent_id: "p0",
batch_steps=5)
batch = ev.sample()
self.assertEqual(batch.count, 5)
self.assertEqual(batch.policy_batches["p0"].count, 10)
self.assertEqual(batch.policy_batches["p1"].count, 25)
def testMultiAgentSample(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),
batch_steps=50)
batch = ev.sample()
self.assertEqual(batch.count, 50)
self.assertEqual(batch.policy_batches["p0"].count, 150)
self.assertEqual(batch.policy_batches["p1"].count, 100)
self.assertEqual(batch.policy_batches["p0"]["t"].tolist(),
list(range(25)) * 6)
