def test_callbacks(self):
for fw in framework_iterator(frameworks=("torch", "tf")):
counts = Counter()
pg = PG(
env="CartPole-v0",
config={
"num_workers": 0,
"rollout_fragment_length": 50,
"train_batch_size": 50,
"callbacks": {
"on_episode_start":
lambda x: counts.update({"start": 1}),
"on_episode_step":
lambda x: counts.update({"step": 1}),
"on_episode_end": lambda x: counts.update({"end": 1}),
"on_sample_end":
lambda x: counts.update({"sample": 1}),
},
"framework": fw,
},
)
pg.train()
pg.train()
self.assertGreater(counts["sample"], 0)
self.assertGreater(counts["start"], 0)
self.assertGreater(counts["end"], 0)
self.assertGreater(counts["step"], 0)
pg.stop()
def test_nested_action_spaces(self):
config = DEFAULT_CONFIG.copy()
config["env"] = RandomEnv
# Write output to check, whether actions are written correctly.
tmp_dir = os.popen("mktemp -d").read()[:-1]
if not os.path.exists(tmp_dir):
# Last resort: Resolve via underlying tempdir (and cut tmp_.
tmp_dir = ray._private.utils.tempfile.gettempdir() + tmp_dir[4:]
assert os.path.exists(tmp_dir), f"'{tmp_dir}' not found!"
config["output"] = tmp_dir
# Switch off OPE as we don't write action-probs.
# TODO: We should probably always write those if `output` is given.
config["off_policy_estimation_methods"] = {}
# Pretend actions in offline files are already normalized.
config["actions_in_input_normalized"] = True
for _ in framework_iterator(config):
for name, action_space in SPACES.items():
config["env_config"] = {
"action_space": action_space,
}
for flatten in [True, False]:
print(f"A={action_space} flatten={flatten}")
shutil.rmtree(config["output"])
config["_disable_action_flattening"] = not flatten
trainer = PG(config)
trainer.train()
trainer.stop()
# Check actions in output file (whether properly flattened
# or not).
reader = JsonReader(
inputs=config["output"],
ioctx=trainer.workers.local_worker().io_context,
)
sample_batch = reader.next()
if flatten:
assert isinstance(sample_batch["actions"], np.ndarray)
assert len(sample_batch["actions"].shape) == 2
assert sample_batch["actions"].shape[0] == len(
sample_batch)
else:
tree.assert_same_structure(
trainer.get_policy().action_space_struct,
sample_batch["actions"],
)
# Test, whether offline data can be properly read by a
# BCTrainer, configured accordingly.
config["input"] = config["output"]
del config["output"]
bc_trainer = BC(config=config)
bc_trainer.train()
bc_trainer.stop()
config["output"] = tmp_dir
config["input"] = "sampler"
def test_local(self):
cf = DEFAULT_CONFIG.copy()
cf["model"]["fcnet_hiddens"] = [10]
cf["num_workers"] = 2
for _ in framework_iterator(cf):
agent = PG(cf, "CartPole-v0")
print(agent.train())
agent.stop()
def test_multi_agent(self):
register_env(
"multi_agent_cartpole", lambda _: MultiAgentCartPole({"num_agents": 10})
)
for fw in framework_iterator():
pg = PG(
env="multi_agent_cartpole",
config={
"num_workers": 0,
"output": self.test_dir,
"multiagent": {
"policies": {"policy_1", "policy_2"},
"policy_mapping_fn": (
lambda aid, **kwargs: random.choice(
["policy_1", "policy_2"]
)
),
},
"framework": fw,
},
)
pg.train()
self.assertEqual(len(os.listdir(self.test_dir)), 1)
pg.stop()
pg = PG(
env="multi_agent_cartpole",
config={
"num_workers": 0,
"input": self.test_dir,
"off_policy_estimation_methods": {
"simulation": {"type": "simulation"}
},
"train_batch_size": 2000,
"multiagent": {
"policies": {"policy_1", "policy_2"},
"policy_mapping_fn": (
lambda aid, **kwargs: random.choice(
["policy_1", "policy_2"]
)
),
},
"framework": fw,
},
)
for _ in range(50):
result = pg.train()
if not np.isnan(result["episode_reward_mean"]):
return # simulation ok
time.sleep(0.1)
assert False, "did not see any simulation results"
def test_multi_agent(self):
register_env("multi_agent_cartpole",
lambda _: MultiAgentCartPole({"num_agents": 10}))
for fw in framework_iterator():
pg = PG(
env="multi_agent_cartpole",
config={
"num_workers": 0,
"output": self.test_dir + fw,
"multiagent": {
"policies": {"policy_1", "policy_2"},
"policy_mapping_fn":
(lambda aid, **kwargs: random.choice(
["policy_1", "policy_2"])),
},
"framework": fw,
},
)
pg.train()
self.assertEqual(len(os.listdir(self.test_dir + fw)), 1)
pg.stop()
pg = PG(
env="multi_agent_cartpole",
config={
"num_workers": 0,
"input": self.test_dir + fw,
"train_batch_size": 2000,
"multiagent": {
"policies": {"policy_1", "policy_2"},
"policy_mapping_fn":
(lambda aid, **kwargs: random.choice(
["policy_1", "policy_2"])),
},
"framework": fw,
"evaluation_interval": 1,
"evaluation_config": {
"input": "sampler"
},
},
)
result = pg.train()
assert np.isnan(
result["episode_reward_mean"]
), "episode reward should not be computed for offline data"
assert not np.isnan(
result["evaluation"]["episode_reward_mean"]
), "Did not see simulation results during evaluation"
def test_query_evaluators(self):
register_env("test", lambda _: gym.make("CartPole-v0"))
for fw in framework_iterator(frameworks=("torch", "tf")):
pg = PG(
env="test",
config={
"num_workers": 2,
"rollout_fragment_length": 5,
"num_envs_per_worker": 2,
"framework": fw,
"create_env_on_driver": True,
},
)
results = pg.workers.foreach_worker(
lambda ev: ev.rollout_fragment_length)
results2 = pg.workers.foreach_worker_with_index(
lambda ev, i: (i, ev.rollout_fragment_length))
results3 = pg.workers.foreach_worker(
lambda ev: ev.foreach_env(lambda env: 1))
self.assertEqual(results, [10, 10, 10])
self.assertEqual(results2, [(0, 10), (1, 10), (2, 10)])
self.assertEqual(results3, [[1, 1], [1, 1], [1, 1]])
pg.stop()