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 write_outputs(self, output, fw, output_config=None):
agent = PG(
env="CartPole-v0",
config={
"output": output + (fw if output != "logdir" else ""),
"rollout_fragment_length": 250,
"framework": fw,
"output_config": output_config or {},
},
)
agent.train()
return agent
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_custom_input_procedure(self):
class CustomJsonReader(JsonReader):
def __init__(self, ioctx: IOContext):
super().__init__(ioctx.input_config["input_files"], ioctx)
def input_creator(ioctx: IOContext) -> InputReader:
return ShuffledInput(CustomJsonReader(ioctx))
register_input("custom_input", input_creator)
test_input_procedure = [
"custom_input",
input_creator,
"ray.rllib.examples.custom_input_api.CustomJsonReader",
]
for input_procedure in test_input_procedure:
for fw in framework_iterator(frameworks=("torch", "tf")):
self.write_outputs(self.test_dir, fw)
agent = PG(
env="CartPole-v0",
config={
"input": input_procedure,
"input_config": {
"input_files": self.test_dir + fw
},
"off_policy_estimation_methods": {},
"framework": fw,
},
)
result = agent.train()
self.assertEqual(result["timesteps_total"], 250)
self.assertTrue(np.isnan(result["episode_reward_mean"]))
def test_multiple_output_workers(self):
ray.shutdown()
ray.init(num_cpus=4, ignore_reinit_error=True)
for fw in framework_iterator(frameworks=["tf", "torch"]):
agent = PG(
env="CartPole-v0",
config={
"num_workers": 2,
"output": self.test_dir + fw,
"rollout_fragment_length": 250,
"framework": fw,
},
)
agent.train()
self.assertEqual(len(os.listdir(self.test_dir + fw)), 2)
reader = JsonReader(self.test_dir + fw + "/*.json")
reader.next()
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_with_flex_agents(self):
register_env("flex_agents_multi_agent_cartpole",
lambda _: FlexAgentsMultiAgent())
pg = PG(
env="flex_agents_multi_agent_cartpole",
config={
"num_workers": 0,
"framework": "tf",
},
)
for i in range(10):
result = pg.train()
print("Iteration {}, reward {}, timesteps {}".format(
i, result["episode_reward_mean"], result["timesteps_total"]))
def test_agent_input_dir(self):
for fw in framework_iterator(frameworks=("torch", "tf")):
self.write_outputs(self.test_dir, fw)
print("WROTE TO: ", self.test_dir)
agent = PG(
env="CartPole-v0",
config={
"input": self.test_dir + fw,
"off_policy_estimation_methods": {},
"framework": fw,
},
)
result = agent.train()
self.assertEqual(result["timesteps_total"], 250) # read from input
self.assertTrue(np.isnan(result["episode_reward_mean"]))
def test_agent_input_list(self):
for fw in framework_iterator(frameworks=("torch", "tf")):
self.write_outputs(self.test_dir, fw)
agent = PG(
env="CartPole-v0",
config={
"input": glob.glob(self.test_dir + fw + "/*.json"),
"off_policy_estimation_methods": {},
"rollout_fragment_length": 99,
"framework": fw,
},
)
result = agent.train()
self.assertEqual(result["timesteps_total"], 250) # read from input
self.assertTrue(np.isnan(result["episode_reward_mean"]))
def test_train_cartpole(self):
register_env("test", lambda _: SimpleServing(gym.make("CartPole-v0")))
config = {"num_workers": 0}
for _ in framework_iterator(config, frameworks=("tf", "torch")):
pg = PG(env="test", config=config)
reached = False
for i in range(80):
result = pg.train()
print("Iteration {}, reward {}, timesteps {}".format(
i, result["episode_reward_mean"],
result["timesteps_total"]))
if result["episode_reward_mean"] >= 80:
reached = True
break
if not reached:
raise Exception("failed to improve reward")
def test_agent_input_dict(self):
for fw in framework_iterator():
self.write_outputs(self.test_dir, fw)
agent = PG(
env="CartPole-v0",
config={
"input": {
self.test_dir + fw: 0.1,
"sampler": 0.9,
},
"train_batch_size": 2000,
"framework": fw,
},
)
result = agent.train()
self.assertTrue(not np.isnan(result["episode_reward_mean"]))
def test_train_multi_agent_cartpole_single_policy(self):
n = 10
register_env("multi_agent_cartpole",
lambda _: MultiAgentCartPole({"num_agents": n}))
pg = PG(
env="multi_agent_cartpole",
config={
"num_workers": 0,
"framework": "tf",
},
)
for i in range(50):
result = pg.train()
print("Iteration {}, reward {}, timesteps {}".format(
i, result["episode_reward_mean"], result["timesteps_total"]))
if result["episode_reward_mean"] >= 50 * n:
return
raise Exception("failed to improve reward")
def test_train_multi_agent_cartpole_multi_policy(self):
n = 10
register_env("multi_agent_cartpole",
lambda _: MultiAgentCartPole({"num_agents": n}))
def gen_policy():
config = {
"gamma": random.choice([0.5, 0.8, 0.9, 0.95, 0.99]),
"n_step": random.choice([1, 2, 3, 4, 5]),
}
return PolicySpec(config=config)
pg = PG(
env="multi_agent_cartpole",
config={
"num_workers": 0,
"multiagent": {
"policies": {
"policy_1": gen_policy(),
"policy_2": gen_policy(),
},
"policy_mapping_fn": lambda aid, **kwargs: "policy_1",
},
"framework": "tf",
},
)
# Just check that it runs without crashing
for i in range(10):
result = pg.train()
print("Iteration {}, reward {}, timesteps {}".format(
i, result["episode_reward_mean"], result["timesteps_total"]))
self.assertTrue(
pg.compute_single_action([0, 0, 0, 0], policy_id="policy_1") in
[0, 1])
self.assertTrue(
pg.compute_single_action([0, 0, 0, 0], policy_id="policy_2") in
[0, 1])
self.assertRaisesRegex(
KeyError,
"not found in PolicyMap",
lambda: pg.compute_single_action([0, 0, 0, 0],
policy_id="policy_3"),
)
def test_agent_input_eval_sim(self):
for fw in framework_iterator():
self.write_outputs(self.test_dir, fw)
agent = PG(
env="CartPole-v0",
config={
"input": self.test_dir + fw,
"off_policy_estimation_methods": {
"simulation": {"type": "simulation"}
},
"framework": fw,
},
)
for _ in range(50):
result = agent.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_agent_input_eval_sampler(self):
for fw in framework_iterator(frameworks=["tf", "torch"]):
self.write_outputs(self.test_dir, fw)
agent = PG(
env="CartPole-v0",
config={
"input": self.test_dir + fw,
"framework": fw,
"evaluation_interval": 1,
"evaluation_config": {
"input": "sampler"
},
},
)
result = agent.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_agent_input_postprocessing_enabled(self):
for fw in framework_iterator(frameworks=("tf", "torch")):
self.write_outputs(self.test_dir, fw)
# Rewrite the files to drop advantages and value_targets for
# testing
for path in glob.glob(self.test_dir + fw + "/*.json"):
out = []
with open(path) as f:
for line in f.readlines():
data = json.loads(line)
# Data won't contain rewards as these are not included
# in the write_outputs run (not needed in the
# SampleBatch). Flip out "rewards" for "advantages"
# just for testing.
data["rewards"] = data["advantages"]
del data["advantages"]
if "value_targets" in data:
del data["value_targets"]
out.append(data)
with open(path, "w") as f:
for data in out:
f.write(json.dumps(data))
agent = PG(
env="CartPole-v0",
config={
"input": self.test_dir + fw,
"off_policy_estimation_methods": {},
"postprocess_inputs": True, # adds back 'advantages'
"framework": fw,
},
)
result = agent.train()
self.assertEqual(result["timesteps_total"], 250) # read from input
self.assertTrue(np.isnan(result["episode_reward_mean"]))