def doTestNestedDict(self, make_env, test_lstm=False):
ModelCatalog.register_custom_model("composite", DictSpyModel)
register_env("nested", make_env)
pg = PGTrainer(
env="nested",
config={
"num_workers": 0,
"sample_batch_size": 5,
"train_batch_size": 5,
"model": {
"custom_model": "composite",
"use_lstm": test_lstm,
},
})
pg.train()
# Check that the model sees the correct reconstructed observations
for i in range(4):
seen = pickle.loads(
ray.experimental.internal_kv._internal_kv_get(
"d_spy_in_{}".format(i)))
pos_i = DICT_SAMPLES[i]["sensors"]["position"].tolist()
cam_i = DICT_SAMPLES[i]["sensors"]["front_cam"][0].tolist()
task_i = one_hot(
DICT_SAMPLES[i]["inner_state"]["job_status"]["task"], 5)
self.assertEqual(seen[0][0].tolist(), pos_i)
self.assertEqual(seen[1][0].tolist(), cam_i)
self.assertEqual(seen[2][0].tolist(), task_i)
def doTestNestedTuple(self, make_env):
ModelCatalog.register_custom_model("composite2", TupleSpyModel)
register_env("nested2", make_env)
pg = PGTrainer(
env="nested2",
config={
"num_workers": 0,
"sample_batch_size": 5,
"train_batch_size": 5,
"model": {
"custom_model": "composite2",
},
})
pg.train()
# Check that the model sees the correct reconstructed observations
for i in range(4):
seen = pickle.loads(
ray.experimental.internal_kv._internal_kv_get(
"t_spy_in_{}".format(i)))
pos_i = TUPLE_SAMPLES[i][0].tolist()
cam_i = TUPLE_SAMPLES[i][1][0].tolist()
task_i = one_hot(TUPLE_SAMPLES[i][2], 5)
self.assertEqual(seen[0][0].tolist(), pos_i)
self.assertEqual(seen[1][0].tolist(), cam_i)
self.assertEqual(seen[2][0].tolist(), task_i)
def do_test_nested_tuple(self, make_env):
ModelCatalog.register_custom_model("composite2", TupleSpyModel)
register_env("nested2", make_env)
pg = PGTrainer(env="nested2",
config={
"num_workers": 0,
"rollout_fragment_length": 5,
"train_batch_size": 5,
"model": {
"custom_model": "composite2",
},
"framework": "tf",
})
# Skip first passes as they came from the TorchPolicy loss
# initialization.
TupleSpyModel.capture_index = 0
pg.train()
# Check that the model sees the correct reconstructed observations
for i in range(4):
seen = pickle.loads(
ray.experimental.internal_kv._internal_kv_get(
"t_spy_in_{}".format(i)))
pos_i = TUPLE_SAMPLES[i][0].tolist()
cam_i = TUPLE_SAMPLES[i][1][0].tolist()
task_i = one_hot(TUPLE_SAMPLES[i][2], 5)
self.assertEqual(seen[0][0].tolist(), pos_i)
self.assertEqual(seen[1][0].tolist(), cam_i)
check(seen[2][0], task_i)
def do_test_nested_dict(self, make_env, test_lstm=False):
ModelCatalog.register_custom_model("composite", DictSpyModel)
register_env("nested", make_env)
pg = PGTrainer(env="nested",
config={
"num_workers": 0,
"rollout_fragment_length": 5,
"train_batch_size": 5,
"model": {
"custom_model": "composite",
"use_lstm": test_lstm,
},
"framework": "tf",
})
# Skip first passes as they came from the TorchPolicy loss
# initialization.
DictSpyModel.capture_index = 0
pg.train()
# Check that the model sees the correct reconstructed observations
for i in range(4):
seen = pickle.loads(
ray.experimental.internal_kv._internal_kv_get(
"d_spy_in_{}".format(i)))
pos_i = DICT_SAMPLES[i]["sensors"]["position"].tolist()
cam_i = DICT_SAMPLES[i]["sensors"]["front_cam"][0].tolist()
task_i = one_hot(
DICT_SAMPLES[i]["inner_state"]["job_status"]["task"], 5)
self.assertEqual(seen[0][0].tolist(), pos_i)
self.assertEqual(seen[1][0].tolist(), cam_i)
check(seen[2][0], task_i)
def test_callbacks(self):
for fw in framework_iterator(frameworks=("torch", "tf")):
counts = Counter()
pg = PGTrainer(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 testCallbacks(self):
counts = Counter()
pg = PGTrainer(env="CartPole-v0",
config={
"num_workers": 0,
"sample_batch_size": 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}),
},
})
pg.train()
pg.train()
pg.train()
pg.train()
self.assertEqual(counts["sample"], 4)
self.assertGreater(counts["start"], 0)
self.assertGreater(counts["end"], 0)
self.assertGreater(counts["step"], 200)
self.assertLess(counts["step"], 400)
def test_multi_agent_complex_spaces(self):
ModelCatalog.register_custom_model("dict_spy", DictSpyModel)
ModelCatalog.register_custom_model("tuple_spy", TupleSpyModel)
register_env("nested_ma", lambda _: NestedMultiAgentEnv())
act_space = spaces.Discrete(2)
pg = PGTrainer(env="nested_ma",
config={
"num_workers": 0,
"rollout_fragment_length": 5,
"train_batch_size": 5,
"multiagent": {
"policies": {
"tuple_policy":
(PGTFPolicy, TUPLE_SPACE, act_space, {
"model": {
"custom_model": "tuple_spy"
}
}),
"dict_policy":
(PGTFPolicy, DICT_SPACE, act_space, {
"model": {
"custom_model": "dict_spy"
}
}),
},
"policy_mapping_fn": lambda a: {
"tuple_agent": "tuple_policy",
"dict_agent": "dict_policy"
}[a],
},
"framework": "tf",
})
# Skip first passes as they came from the TorchPolicy loss
# initialization.
TupleSpyModel.capture_index = DictSpyModel.capture_index = 0
pg.train()
for i in range(4):
seen = pickle.loads(
ray.experimental.internal_kv._internal_kv_get(
"d_spy_in_{}".format(i)))
pos_i = DICT_SAMPLES[i]["sensors"]["position"].tolist()
cam_i = DICT_SAMPLES[i]["sensors"]["front_cam"][0].tolist()
task_i = one_hot(
DICT_SAMPLES[i]["inner_state"]["job_status"]["task"], 5)
self.assertEqual(seen[0][0].tolist(), pos_i)
self.assertEqual(seen[1][0].tolist(), cam_i)
check(seen[2][0], task_i)
for i in range(4):
seen = pickle.loads(
ray.experimental.internal_kv._internal_kv_get(
"t_spy_in_{}".format(i)))
pos_i = TUPLE_SAMPLES[i][0].tolist()
cam_i = TUPLE_SAMPLES[i][1][0].tolist()
task_i = one_hot(TUPLE_SAMPLES[i][2], 5)
self.assertEqual(seen[0][0].tolist(), pos_i)
self.assertEqual(seen[1][0].tolist(), cam_i)
check(seen[2][0], task_i)
def writeOutputs(self, output):
agent = PGTrainer(env="CartPole-v0",
config={
"output": output,
"sample_batch_size": 250,
})
agent.train()
return agent
def writeOutputs(self, output):
agent = PGTrainer(env="CartPole-v0",
config={
"output": output,
"rollout_fragment_length": 250,
})
agent.train()
return agent
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["input_evaluation"] = []
# 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 [False, True]:
print(f"A={action_space} flatten={flatten}")
shutil.rmtree(config["output"])
config["_disable_action_flattening"] = not flatten
trainer = PGTrainer(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 = BCTrainer(config=config)
bc_trainer.train()
bc_trainer.stop()
config["output"] = tmp_dir
config["input"] = "sampler"
def write_outputs(self, output, fw):
agent = PGTrainer(env="CartPole-v0",
config={
"output":
output + (fw if output != "logdir" else ""),
"rollout_fragment_length": 250,
"framework": fw,
})
agent.train()
return agent
def testMultiAgent(self):
register_env("multi_agent_cartpole",
lambda _: MultiAgentCartPole({"num_agents": 10}))
single_env = gym.make("CartPole-v0")
def gen_policy():
obs_space = single_env.observation_space
act_space = single_env.action_space
return (PGTFPolicy, obs_space, act_space, {})
for fw in framework_iterator():
pg = PGTrainer(
env="multi_agent_cartpole",
config={
"num_workers": 0,
"output": self.test_dir,
"multiagent": {
"policies": {
"policy_1": gen_policy(),
"policy_2": gen_policy(),
},
"policy_mapping_fn": (
lambda agent_id: random.choice(
["policy_1", "policy_2"])),
},
"framework": fw,
})
pg.train()
self.assertEqual(len(os.listdir(self.test_dir)), 1)
pg.stop()
pg = PGTrainer(
env="multi_agent_cartpole",
config={
"num_workers": 0,
"input": self.test_dir,
"input_evaluation": ["simulation"],
"train_batch_size": 2000,
"multiagent": {
"policies": {
"policy_1": gen_policy(),
"policy_2": gen_policy(),
},
"policy_mapping_fn": (
lambda agent_id: 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 testMultiAgentComplexSpaces(self):
ModelCatalog.register_custom_model("dict_spy", DictSpyModel)
ModelCatalog.register_custom_model("tuple_spy", TupleSpyModel)
register_env("nested_ma", lambda _: NestedMultiAgentEnv())
act_space = spaces.Discrete(2)
pg = PGTrainer(env="nested_ma",
config={
"num_workers": 0,
"sample_batch_size": 5,
"train_batch_size": 5,
"multiagent": {
"policies": {
"tuple_policy":
(PGTFPolicy, TUPLE_SPACE, act_space, {
"model": {
"custom_model": "tuple_spy"
}
}),
"dict_policy":
(PGTFPolicy, DICT_SPACE, act_space, {
"model": {
"custom_model": "dict_spy"
}
}),
},
"policy_mapping_fn": lambda a: {
"tuple_agent": "tuple_policy",
"dict_agent": "dict_policy"
}[a],
},
})
pg.train()
for i in range(4):
seen = pickle.loads(
ray.experimental.internal_kv._internal_kv_get(
"d_spy_in_{}".format(i)))
pos_i = DICT_SAMPLES[i]["sensors"]["position"].tolist()
cam_i = DICT_SAMPLES[i]["sensors"]["front_cam"][0].tolist()
task_i = one_hot(
DICT_SAMPLES[i]["inner_state"]["job_status"]["task"], 5)
self.assertEqual(seen[0][0].tolist(), pos_i)
self.assertEqual(seen[1][0].tolist(), cam_i)
self.assertEqual(seen[2][0].tolist(), task_i)
for i in range(4):
seen = pickle.loads(
ray.experimental.internal_kv._internal_kv_get(
"t_spy_in_{}".format(i)))
pos_i = TUPLE_SAMPLES[i][0].tolist()
cam_i = TUPLE_SAMPLES[i][1][0].tolist()
task_i = one_hot(TUPLE_SAMPLES[i][2], 5)
self.assertEqual(seen[0][0].tolist(), pos_i)
self.assertEqual(seen[1][0].tolist(), cam_i)
self.assertEqual(seen[2][0].tolist(), task_i)
def test_rollout_dict_space(self):
register_env("nested", lambda _: NestedDictEnv())
agent = PGTrainer(env="nested", config={"framework": "tf"})
agent.train()
path = agent.save()
agent.stop()
# Test train works on restore
agent2 = PGTrainer(env="nested", config={"framework": "tf"})
agent2.restore(path)
agent2.train()
# Test rollout works on restore
rollout(agent2, "nested", 100)
def testRolloutDictSpace(self):
register_env("nested", lambda _: NestedDictEnv())
agent = PGTrainer(env="nested")
agent.train()
path = agent.save()
agent.stop()
# Test train works on restore
agent2 = PGTrainer(env="nested")
agent2.restore(path)
agent2.train()
# Test rollout works on restore
rollout(agent2, "nested", 100)
def test_pg_exec_impl(ray_start_regular):
trainer = PGTrainer(env="CartPole-v0",
config={
"min_iter_time_s": 0,
"use_exec_api": True
})
assert isinstance(trainer.train(), dict)
def testAgentInputPostprocessingEnabled(self):
self.writeOutputs(self.test_dir)
# Rewrite the files to drop advantages and value_targets for testing
for path in glob.glob(self.test_dir + "/*.json"):
out = []
for line in open(path).readlines():
data = json.loads(line)
del data["advantages"]
del data["value_targets"]
out.append(data)
with open(path, "w") as f:
for data in out:
f.write(json.dumps(data))
agent = PGTrainer(
env="CartPole-v0",
config={
"input": self.test_dir,
"input_evaluation": [],
"postprocess_inputs": True, # adds back 'advantages'
})
result = agent.train()
self.assertEqual(result["timesteps_total"], 250) # read from input
self.assertTrue(np.isnan(result["episode_reward_mean"]))
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 = PGTrainer(env="CartPole-v0",
config={
"input": input_procedure,
"input_config": {
"input_files": self.test_dir + fw
},
"input_evaluation": [],
"framework": fw,
})
result = agent.train()
self.assertEqual(result["timesteps_total"], 250)
self.assertTrue(np.isnan(result["episode_reward_mean"]))
def test_local(self):
cf = DEFAULT_CONFIG.copy()
cf["model"]["fcnet_hiddens"] = [10]
for _ in framework_iterator(cf):
agent = PGTrainer(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 = PGTrainer(
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 = PGTrainer(
env="multi_agent_cartpole",
config={
"num_workers": 0,
"input": self.test_dir,
"input_evaluation": ["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_no_step_on_init(self):
register_env("fail", lambda _: FailOnStepEnv())
for fw in framework_iterator(frameworks=()):
pg = PGTrainer(
env="fail", config={
"num_workers": 1,
"framework": fw,
})
self.assertRaises(Exception, lambda: pg.train())
pg.stop()
def testTrainCartpole(self):
register_env("test", lambda _: SimpleServing(gym.make("CartPole-v0")))
pg = PGTrainer(env="test", config={"num_workers": 0})
for i in range(100):
result = pg.train()
print("Iteration {}, reward {}, timesteps {}".format(
i, result["episode_reward_mean"], result["timesteps_total"]))
if result["episode_reward_mean"] >= 100:
return
raise Exception("failed to improve reward")
def testAgentInputDir(self):
self.writeOutputs(self.test_dir)
agent = PGTrainer(env="CartPole-v0",
config={
"input": self.test_dir,
"input_evaluation": [],
})
result = agent.train()
self.assertEqual(result["timesteps_total"], 250) # read from input
self.assertTrue(np.isnan(result["episode_reward_mean"]))
def test_train_multi_cartpole_single_policy(self):
n = 10
register_env("multi_cartpole", lambda _: MultiCartpole(n))
pg = PGTrainer(env="multi_cartpole", config={"num_workers": 0})
for i in range(100):
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_no_step_on_init(self):
# Allow for Unittest run.
ray.init(num_cpus=5, ignore_reinit_error=True)
register_env("fail", lambda _: FailOnStepEnv())
for fw in framework_iterator(frameworks=()):
pg = PGTrainer(env="fail",
config={
"num_workers": 1,
"framework": fw,
})
self.assertRaises(Exception, lambda: pg.train())
def testAgentInputList(self):
self.writeOutputs(self.test_dir)
agent = PGTrainer(env="CartPole-v0",
config={
"input": glob.glob(self.test_dir + "/*.json"),
"input_evaluation": [],
"rollout_fragment_length": 99,
})
result = agent.train()
self.assertEqual(result["timesteps_total"], 250) # read from input
self.assertTrue(np.isnan(result["episode_reward_mean"]))
def test_agent_input_dir(self):
for fw in framework_iterator(frameworks=("torch", "tf")):
self.write_outputs(self.test_dir, fw)
agent = PGTrainer(env="CartPole-v0",
config={
"input": self.test_dir + fw,
"input_evaluation": [],
"framework": fw,
})
result = agent.train()
self.assertEqual(result["timesteps_total"], 250) # read from input
self.assertTrue(np.isnan(result["episode_reward_mean"]))
def test_multi_agent_with_flex_agents(self):
register_env("flex_agents_multi_agent_cartpole",
lambda _: FlexAgentsMultiAgent())
pg = PGTrainer(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 testAgentInputDict(self):
self.writeOutputs(self.test_dir)
agent = PGTrainer(env="CartPole-v0",
config={
"input": {
self.test_dir: 0.1,
"sampler": 0.9,
},
"train_batch_size": 2000,
"input_evaluation": [],
})
result = agent.train()
self.assertTrue(not np.isnan(result["episode_reward_mean"]))
def testAgentInputEvalSim(self):
self.writeOutputs(self.test_dir)
agent = PGTrainer(env="CartPole-v0",
config={
"input": self.test_dir,
"input_evaluation": ["simulation"],
})
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"
