def test_driver_lives_sequential(shutdown_only):
ray.init(num_cpus=1)
ray.worker._global_node.kill_raylet()
ray.worker._global_node.kill_plasma_store()
ray.worker._global_node.kill_log_monitor()
ray.worker._global_node.kill_monitor()
ray.worker._global_node.kill_raylet_monitor()
def testCachingReusables(self):
# Test that we can define reusable variables before the driver is connected.
def foo_initializer():
return 1
def bar_initializer():
return []
def bar_reinitializer(bar):
return []
ray.reusables.foo = ray.Reusable(foo_initializer)
ray.reusables.bar = ray.Reusable(bar_initializer, bar_reinitializer)
@ray.remote
def use_foo():
return ray.reusables.foo
@ray.remote
def use_bar():
ray.reusables.bar.append(1)
return ray.reusables.bar
ray.init(start_ray_local=True, num_workers=2)
self.assertEqual(ray.get(use_foo.remote()), 1)
self.assertEqual(ray.get(use_foo.remote()), 1)
self.assertEqual(ray.get(use_bar.remote()), [1])
self.assertEqual(ray.get(use_bar.remote()), [1])
ray.worker.cleanup()
def test_connect_with_disconnected_node(shutdown_only):
config = json.dumps({
"num_heartbeats_timeout": 50,
"heartbeat_timeout_milliseconds": 10,
})
cluster = Cluster()
cluster.add_node(num_cpus=0, _internal_config=config)
ray.init(redis_address=cluster.redis_address)
info = relevant_errors(ray_constants.REMOVED_NODE_ERROR)
assert len(info) == 0
# This node is killed by SIGKILL, ray_monitor will mark it to dead.
dead_node = cluster.add_node(num_cpus=0, _internal_config=config)
cluster.remove_node(dead_node, allow_graceful=False)
wait_for_errors(ray_constants.REMOVED_NODE_ERROR, 1, timeout=2)
# This node is killed by SIGKILL, ray_monitor will mark it to dead.
dead_node = cluster.add_node(num_cpus=0, _internal_config=config)
cluster.remove_node(dead_node, allow_graceful=False)
wait_for_errors(ray_constants.REMOVED_NODE_ERROR, 2, timeout=2)
# This node is killed by SIGTERM, ray_monitor will not mark it again.
removing_node = cluster.add_node(num_cpus=0, _internal_config=config)
cluster.remove_node(removing_node, allow_graceful=True)
with pytest.raises(Exception, match=('Timing out of wait.')):
wait_for_errors(ray_constants.REMOVED_NODE_ERROR, 3, timeout=2)
# There is no connection error to a dead node.
info = relevant_errors(ray_constants.RAYLET_CONNECTION_ERROR)
assert len(info) == 0
def ray_start_object_store_memory():
# Start the Ray processes.
store_size = 10**6
ray.init(num_cpus=1, object_store_memory=store_size)
yield None
# The code after the yield will run as teardown code.
ray.shutdown()
def testPutGet(self):
ray.init(start_ray_local=True, num_workers=0)
for i in range(100):
value_before = i * 10 ** 6
objectid = ray.put(value_before)
value_after = ray.get(objectid)
self.assertEqual(value_before, value_after)
for i in range(100):
value_before = i * 10 ** 6 * 1.0
objectid = ray.put(value_before)
value_after = ray.get(objectid)
self.assertEqual(value_before, value_after)
for i in range(100):
value_before = "h" * i
objectid = ray.put(value_before)
value_after = ray.get(objectid)
self.assertEqual(value_before, value_after)
for i in range(100):
value_before = [1] * i
objectid = ray.put(value_before)
value_after = ray.get(objectid)
self.assertEqual(value_before, value_after)
ray.worker.cleanup()
def testWait(self):
ray.init(start_ray_local=True, num_workers=1)
@ray.remote
def f(delay):
time.sleep(delay)
return 1
objectids = [f.remote(1.0), f.remote(0.5), f.remote(0.5), f.remote(0.5)]
ready_ids, remaining_ids = ray.wait(objectids)
self.assertTrue(len(ready_ids) == 1)
self.assertTrue(len(remaining_ids) == 3)
ready_ids, remaining_ids = ray.wait(objectids, num_returns=4)
self.assertEqual(ready_ids, objectids)
self.assertEqual(remaining_ids, [])
objectids = [f.remote(0.5), f.remote(0.5), f.remote(0.5), f.remote(0.5)]
start_time = time.time()
ready_ids, remaining_ids = ray.wait(objectids, timeout=1.75, num_returns=4)
self.assertTrue(time.time() - start_time < 2)
self.assertEqual(len(ready_ids), 3)
self.assertEqual(len(remaining_ids), 1)
ray.wait(objectids)
objectids = [f.remote(1.0), f.remote(0.5), f.remote(0.5), f.remote(0.5)]
start_time = time.time()
ready_ids, remaining_ids = ray.wait(objectids, timeout=5)
self.assertTrue(time.time() - start_time < 5)
self.assertEqual(len(ready_ids), 1)
self.assertEqual(len(remaining_ids), 3)
ray.worker.cleanup()
def create_cluster(num_nodes):
cluster = Cluster()
for i in range(num_nodes):
cluster.add_node(resources={str(i): 100}, object_store_memory=10**9)
ray.init(redis_address=cluster.redis_address)
return cluster
def testRecursiveObjects(self):
ray.init(start_ray_local=True, num_workers=0)
class ClassA(object):
pass
ray.register_class(ClassA)
# Make a list that contains itself.
l = []
l.append(l)
# Make an object that contains itself as a field.
a1 = ClassA()
a1.field = a1
# Make two objects that contain each other as fields.
a2 = ClassA()
a3 = ClassA()
a2.field = a3
a3.field = a2
# Make a dictionary that contains itself.
d1 = {}
d1["key"] = d1
# Create a list of recursive objects.
recursive_objects = [l, a1, a2, a3, d1]
# Check that exceptions are thrown when we serialize the recursive objects.
for obj in recursive_objects:
self.assertRaises(Exception, lambda : ray.put(obj))
ray.worker.cleanup()
def ray_start_driver_put_errors():
plasma_store_memory = 10**9
# Start the Ray processes.
ray.init(num_cpus=1, object_store_memory=plasma_store_memory)
yield plasma_store_memory
# The code after the yield will run as teardown code.
ray.shutdown()
def testFailImportingRemoteFunction(self):
ray.init(start_ray_local=True, num_workers=2, driver_mode=ray.SILENT_MODE)
# This example is somewhat contrived. It should be successfully pickled, and
# then it should throw an exception when it is unpickled. This may depend a
# bit on the specifics of our pickler.
def reducer(*args):
raise Exception("There is a problem here.")
class Foo(object):
def __init__(self):
self.__name__ = "Foo_object"
self.func_doc = ""
self.__globals__ = {}
def __reduce__(self):
return reducer, ()
def __call__(self):
return
ray.remote(Foo())
for _ in range(100): # Retry if we need to wait longer.
if len(ray.task_info()["failed_remote_function_imports"]) >= 1:
break
time.sleep(0.1)
self.assertTrue("There is a problem here." in ray.task_info()["failed_remote_function_imports"][0]["error_message"])
ray.worker.cleanup()
def init():
ray.init(num_cpus=4)
async_api.init()
asyncio.get_event_loop().set_debug(False)
yield
async_api.shutdown()
ray.shutdown()
def launch(self):
"""Actual entry point into the class instance where everything happens.
Lots of delegating to classes that are in subclass or can be over-ridden.
"""
self.register_env_creator()
# All worker nodes will block at this step during training
ray_cluster_config = self.ray_init_config()
if not self.is_master_node:
return
# Start the driver on master node
ray.init(**ray_cluster_config)
experiment_config = self.get_experiment_config()
experiment_config = self.customize_experiment_config(experiment_config)
print("Running experiment with config %s" % json.dumps(experiment_config, indent=2))
run_experiments(experiment_config)
all_wokers_host_names = self.get_all_host_names()[1:]
# If distributed job, send TERMINATION_SIGNAL to all workers.
if len(all_wokers_host_names) > 0:
self.sage_cluster_communicator.create_s3_signal(TERMINATION_SIGNAL)
algo = experiment_config["training"]["run"]
env_string = experiment_config["training"]["config"]["env"]
config = experiment_config["training"]["config"]
self.save_checkpoint_and_serving_model(algorithm=algo,
env_string=env_string,
config=config)
def testMethods(self):
for module in [ra.core, ra.random, ra.linalg, da.core, da.random,
da.linalg]:
reload(module)
ray.init()
# test eye
object_id = ra.eye.remote(3)
val = ray.get(object_id)
assert_almost_equal(val, np.eye(3))
# test zeros
object_id = ra.zeros.remote([3, 4, 5])
val = ray.get(object_id)
assert_equal(val, np.zeros([3, 4, 5]))
# test qr - pass by value
a_val = np.random.normal(size=[10, 11])
q_id, r_id = ra.linalg.qr.remote(a_val)
q_val = ray.get(q_id)
r_val = ray.get(r_id)
assert_almost_equal(np.dot(q_val, r_val), a_val)
# test qr - pass by objectid
a = ra.random.normal.remote([10, 13])
q_id, r_id = ra.linalg.qr.remote(a)
a_val = ray.get(a)
q_val = ray.get(q_id)
r_val = ray.get(r_id)
assert_almost_equal(np.dot(q_val, r_val), a_val)
def router():
# We need at least 5 workers so resource won't be oversubscribed
ray.init(num_cpus=5)
# The following two blobs are equivalent
#
# handle = DeadlineAwareRouter.remote("DefaultTestRouter")
# ray.experimental.register_actor("DefaultTestRouter", handle)
# handle.start.remote()
#
# handle = start_router(DeadlineAwareRouter, "DefaultRouter")
handle = start_router(DeadlineAwareRouter, "DefaultRouter")
handle.register_actor.remote(
"VAdder", VectorizedAdder,
init_kwargs={"scaler_increment": 1}) # init args
handle.register_actor.remote(
"SAdder", ScalerAdder, init_kwargs={"scaler_increment": 2})
handle.register_actor.remote(
"SleepFirst", SleepOnFirst, init_kwargs={"sleep_time": 1})
handle.register_actor.remote(
"SleepCounter", SleepCounter, max_batch_size=1)
yield handle
ray.shutdown()
def testFailedTask(self):
reload(test_functions)
ray.init(num_workers=3, driver_mode=ray.SILENT_MODE)
test_functions.throw_exception_fct1.remote()
test_functions.throw_exception_fct1.remote()
wait_for_errors(b"task", 2)
self.assertEqual(len(relevant_errors(b"task")), 2)
for task in relevant_errors(b"task"):
self.assertIn(b"Test function 1 intentionally failed.",
task.get(b"message"))
x = test_functions.throw_exception_fct2.remote()
try:
ray.get(x)
except Exception as e:
self.assertIn("Test function 2 intentionally failed.", str(e))
else:
# ray.get should throw an exception.
self.assertTrue(False)
x, y, z = test_functions.throw_exception_fct3.remote(1.0)
for ref in [x, y, z]:
try:
ray.get(ref)
except Exception as e:
self.assertIn("Test function 3 intentionally failed.", str(e))
else:
# ray.get should throw an exception.
self.assertTrue(False)
def testFailedActorInit(self):
ray.init(num_workers=0, driver_mode=ray.SILENT_MODE)
error_message1 = "actor constructor failed"
error_message2 = "actor method failed"
@ray.remote
class FailedActor(object):
def __init__(self):
raise Exception(error_message1)
def get_val(self):
return 1
def fail_method(self):
raise Exception(error_message2)
a = FailedActor.remote()
# Make sure that we get errors from a failed constructor.
wait_for_errors(b"task", 1)
self.assertEqual(len(ray.error_info()), 1)
self.assertIn(error_message1,
ray.error_info()[0][b"message"].decode("ascii"))
# Make sure that we get errors from a failed method.
a.fail_method.remote()
wait_for_errors(b"task", 2)
self.assertEqual(len(ray.error_info()), 2)
self.assertIn(error_message2,
ray.error_info()[1][b"message"].decode("ascii"))
def run(args, parser):
config = {}
# Load configuration from file
config_dir = os.path.dirname(args.checkpoint)
config_path = os.path.join(config_dir, "params.pkl")
if not os.path.exists(config_path):
config_path = os.path.join(config_dir, "../params.pkl")
if not os.path.exists(config_path):
if not args.config:
raise ValueError(
"Could not find params.pkl in either the checkpoint dir or "
"its parent directory.")
else:
with open(config_path, 'rb') as f:
config = pickle.load(f)
if "num_workers" in config:
config["num_workers"] = min(2, config["num_workers"])
config = merge_dicts(config, args.config)
if not args.env:
if not config.get("env"):
parser.error("the following arguments are required: --env")
args.env = config.get("env")
ray.init()
cls = get_agent_class(args.run)
agent = cls(env=args.env, config=config)
agent.restore(args.checkpoint)
num_steps = int(args.steps)
rollout(agent, args.env, num_steps, args.out, args.no_render)
def test_dying_worker_wait(shutdown_only):
ray.init(num_cpus=2)
@ray.remote
def sleep_forever():
time.sleep(10**6)
@ray.remote
def get_pid():
return os.getpid()
x_id = sleep_forever.remote()
# Get the PID of the worker that block_in_wait will run on (sleep a little
# to make sure that sleep_forever has already started).
time.sleep(0.1)
worker_pid = ray.get(get_pid.remote())
@ray.remote
def block_in_wait(object_id_in_list):
ray.wait(object_id_in_list)
# Have the worker wait in a wait call.
block_in_wait.remote([x_id])
time.sleep(0.1)
# Kill the worker.
os.kill(worker_pid, signal.SIGKILL)
time.sleep(0.1)
# Create the object.
ray.worker.global_worker.put_object(x_id, 1)
time.sleep(0.1)
# Make sure that nothing has died.
assert ray.services.all_processes_alive()
def testNetworksIndependent(self):
# Note we use only one worker to ensure that all of the remote
# functions run on the same worker.
ray.init(num_workers=1)
net1 = NetActor()
net2 = NetActor()
# Make sure the two networks have different weights. TODO(rkn): Note
# that equality comparisons of numpy arrays normally does not work.
# This only works because at the moment they have size 1.
weights1 = net1.get_weights()
weights2 = net2.get_weights()
self.assertNotEqual(weights1, weights2)
# Set the weights and get the weights, and make sure they are
# unchanged.
new_weights1 = net1.set_and_get_weights(weights1)
new_weights2 = net2.set_and_get_weights(weights2)
self.assertEqual(weights1, new_weights1)
self.assertEqual(weights2, new_weights2)
# Swap the weights.
new_weights1 = net2.set_and_get_weights(weights1)
new_weights2 = net1.set_and_get_weights(weights2)
self.assertEqual(weights1, new_weights1)
self.assertEqual(weights2, new_weights2)
def ray_start_reconstruction(request):
num_nodes = request.param
plasma_store_memory = int(0.5 * 10**9)
cluster = Cluster(
initialize_head=True,
head_node_args={
"num_cpus": 1,
"object_store_memory": plasma_store_memory // num_nodes,
"redis_max_memory": 10**7,
"_internal_config": json.dumps({
"initial_reconstruction_timeout_milliseconds": 200
})
})
for i in range(num_nodes - 1):
cluster.add_node(
num_cpus=1,
object_store_memory=plasma_store_memory // num_nodes,
_internal_config=json.dumps({
"initial_reconstruction_timeout_milliseconds": 200
}))
ray.init(redis_address=cluster.redis_address)
yield plasma_store_memory, num_nodes, cluster
# Clean up the Ray cluster.
ray.shutdown()
cluster.shutdown()
def driver_0(redis_address, driver_index):
"""The script for driver 0.
This driver should create five actors that each use one GPU and some actors
that use no GPUs. After a while, it should exit.
"""
ray.init(redis_address=redis_address)
# Wait for all the nodes to join the cluster.
_wait_for_nodes_to_join(total_num_nodes)
# Start some long running task. Driver 2 will make sure the worker running
# this task has been killed.
for i in range(num_long_running_tasks_per_driver):
long_running_task.remote(driver_index, i, redis_address)
# Create some actors that require one GPU.
actors_one_gpu = [Actor1.remote(driver_index, i, redis_address)
for i in range(5)]
# Create some actors that don't require any GPUs.
actors_no_gpus = [Actor0.remote(driver_index, 5 + i, redis_address)
for i in range(5)]
for _ in range(1000):
ray.get([actor.check_ids.remote() for actor in actors_one_gpu])
ray.get([actor.check_ids.remote() for actor in actors_no_gpus])
# Start a long-running method on one actor and make sure this doesn't
# affect anything.
actors_no_gpus[0].long_running_method.remote()
_broadcast_event("DRIVER_0_DONE", redis_address)
def testFailImportingActor(self):
ray.init(num_workers=2, driver_mode=ray.SILENT_MODE)
# Create the contents of a temporary Python file.
temporary_python_file = """
def temporary_helper_function():
return 1
"""
f = tempfile.NamedTemporaryFile(suffix=".py")
f.write(temporary_python_file.encode("ascii"))
f.flush()
directory = os.path.dirname(f.name)
# Get the module name and strip ".py" from the end.
module_name = os.path.basename(f.name)[:-3]
sys.path.append(directory)
module = __import__(module_name)
# Define an actor that closes over this temporary module. This should
# fail when it is unpickled.
@ray.remote
class Foo(object):
def __init__(self):
self.x = module.temporary_python_file()
def get_val(self):
return 1
# There should be no errors yet.
self.assertEqual(len(ray.error_info()), 0)
# Create an actor.
foo = Foo.remote()
# Wait for the error to arrive.
wait_for_errors(b"register_actor", 1)
self.assertIn(b"No module named", ray.error_info()[0][b"message"])
# Wait for the error from when the __init__ tries to run.
wait_for_errors(b"task", 1)
self.assertIn(
b"failed to be imported, and so cannot execute this method",
ray.error_info()[1][b"message"])
# Check that if we try to get the function it throws an exception and
# does not hang.
with self.assertRaises(Exception):
ray.get(foo.get_val.remote())
# Wait for the error from when the call to get_val.
wait_for_errors(b"task", 2)
self.assertIn(
b"failed to be imported, and so cannot execute this method",
ray.error_info()[2][b"message"])
f.close()
# Clean up the junk we added to sys.path.
sys.path.pop(-1)
def testCustomModel(self):
ray.init()
ModelCatalog.register_custom_model("foo", CustomModel)
p1 = ModelCatalog.get_model({
"obs": tf.constant([1, 2, 3])
}, Box(0, 1, shape=(3, ), dtype=np.float32), Discrete(5), 5,
{"custom_model": "foo"})
self.assertEqual(str(type(p1)), str(CustomModel))
def test_using_hostnames(ray_start_head_local):
ray.init(node_ip_address="localhost", redis_address="localhost:6379")
@ray.remote
def f():
return 1
assert ray.get(f.remote()) == 1
def testAdditionalVariablesWithLoss(self):
ray.init(num_workers=1)
net = LossActor()
self.assertEqual(len(net.values[0].variables.items()), 3)
self.assertEqual(len(net.values[0].placeholders.items()), 3)
net.values[0].set_weights(net.values[0].get_weights())
def initialize(self, env, policy, pool):
super(RemoteSampler, self).initialize(env, policy, pool)
ray.init()
env_pkl = pickle.dumps(env)
policy_pkl = pickle.dumps(policy)
self._remote_environment = _RemoteEnv.remote(env_pkl, policy_pkl)
def testVariableNameCollision(self):
ray.init(num_workers=2)
net1 = NetActor()
net2 = NetActor()
# This is checking that the variable names of the two nets are the
# same, i.e., that the names in the weight dictionaries are the same.
net1.values[0].set_weights(net2.values[0].get_weights())
def test_temp_plasma_store_socket():
ray.init(plasma_store_socket_name="/tmp/i_am_a_temp_socket")
assert os.path.exists(
"/tmp/i_am_a_temp_socket"), "Specified socket path not found."
ray.shutdown()
try:
os.remove("/tmp/i_am_a_temp_socket")
except Exception:
pass
def testVersionMismatch(self):
ray_version = ray.__version__
ray.__version__ = "fake ray version"
ray.init(num_workers=1, driver_mode=ray.SILENT_MODE)
wait_for_errors(b"version_mismatch", 1)
ray.__version__ = ray_version
def ray_start():
# Start ray instance
ray.init(num_cpus=1)
# Run test using this fixture
yield None
# Shutdown ray instance
ray.shutdown()
def test_capture_child_tasks(ray_start_cluster, connect_to_client):
cluster = ray_start_cluster
total_num_tasks = 4
for _ in range(2):
cluster.add_node(num_cpus=total_num_tasks, num_gpus=total_num_tasks)
ray.init(address=cluster.address)
with connect_to_client_or_not(connect_to_client):
pg = ray.util.placement_group(
[{
"CPU": 2,
"GPU": 2,
}, {
"CPU": 2,
"GPU": 2,
}],
strategy="STRICT_PACK")
ray.get(pg.ready())
# If get_current_placement_group is used when the current worker/driver
# doesn't belong to any of placement group, it should return None.
assert get_current_placement_group() is None
# Test if tasks capture child tasks.
@ray.remote
def task():
return get_current_placement_group()
@ray.remote
def create_nested_task(child_cpu, child_gpu, set_none=False):
assert get_current_placement_group() is not None
kwargs = {
"num_cpus": child_cpu,
"num_gpus": child_gpu,
}
if set_none:
kwargs["placement_group"] = None
return ray.get([task.options(**kwargs).remote() for _ in range(3)])
t = create_nested_task.options(
num_cpus=1,
num_gpus=0,
placement_group=pg,
placement_group_capture_child_tasks=True).remote(1, 0)
pgs = ray.get(t)
# Every task should have current placement group because they
# should be implicitly captured by default.
assert None not in pgs
t1 = create_nested_task.options(
num_cpus=1,
num_gpus=0,
placement_group=pg,
placement_group_capture_child_tasks=True).remote(1, 0, True)
pgs = ray.get(t1)
# Every task should have no placement group since it's set to None.
# should be implicitly captured by default.
assert set(pgs) == {None}
# Test if tasks don't capture child tasks when the option is off.
t2 = create_nested_task.options(
num_cpus=0, num_gpus=1, placement_group=pg).remote(0, 1)
pgs = ray.get(t2)
# All placement groups should be None since we don't capture child
# tasks.
assert not all(pgs)
def test_automatic_cleanup_job(ray_start_cluster):
# Make sure the placement groups created by a
# job, actor, and task are cleaned when the job is done.
cluster = ray_start_cluster
num_nodes = 3
num_cpu_per_node = 4
# Create 3 nodes cluster.
for _ in range(num_nodes):
cluster.add_node(num_cpus=num_cpu_per_node)
cluster.wait_for_nodes()
info = ray.init(address=cluster.address)
available_cpus = ray.available_resources()["CPU"]
assert available_cpus == num_nodes * num_cpu_per_node
driver_code = f"""
import ray
ray.init(address="{info["redis_address"]}")
def create_pg():
pg = ray.util.placement_group(
[{{"CPU": 1}} for _ in range(3)],
strategy="STRICT_SPREAD")
ray.get(pg.ready())
return pg
@ray.remote(num_cpus=0)
def f():
create_pg()
@ray.remote(num_cpus=0)
class A:
def create_pg(self):
create_pg()
ray.get(f.remote())
a = A.remote()
ray.get(a.create_pg.remote())
# Create 2 pgs to make sure multiple placement groups that belong
# to a single job will be properly cleaned.
create_pg()
create_pg()
ray.shutdown()
"""
run_string_as_driver(driver_code)
# Wait until the driver is reported as dead by GCS.
def is_job_done():
jobs = ray.state.jobs()
for job in jobs:
if job["IsDead"]:
return True
return False
def assert_num_cpus(expected_num_cpus):
if expected_num_cpus == 0:
return "CPU" not in ray.available_resources()
return ray.available_resources()["CPU"] == expected_num_cpus
wait_for_condition(is_job_done)
available_cpus = ray.available_resources()["CPU"]
wait_for_condition(lambda: assert_num_cpus(num_nodes * num_cpu_per_node))
def test_int_dataframe():
ray.init()
pandas_df = pd.DataFrame({'col1': [0, 1, 2, 3],
'col2': [4, 5, 6, 7],
'col3': [8, 9, 10, 11],
'col4': [12, 13, 14, 15],
'col5': [0, 0, 0, 0]})
ray_df = rdf.from_pandas(pandas_df, 2)
testfuncs = [lambda x: x + 1,
lambda x: str(x),
lambda x: x * x,
lambda x: x,
lambda x: False]
keys = ['col1',
'col2',
'col3',
'col4']
test_roundtrip(ray_df, pandas_df)
test_index(ray_df, pandas_df)
test_size(ray_df, pandas_df)
test_ndim(ray_df, pandas_df)
test_ftypes(ray_df, pandas_df)
test_values(ray_df, pandas_df)
test_axes(ray_df, pandas_df)
test_shape(ray_df, pandas_df)
test_add_prefix(ray_df, pandas_df)
test_add_suffix(ray_df, pandas_df)
for testfunc in testfuncs:
test_applymap(ray_df, pandas_df, testfunc)
test_copy(ray_df)
test_sum(ray_df, pandas_df)
test_abs(ray_df, pandas_df)
test_keys(ray_df, pandas_df)
test_transpose(ray_df, pandas_df)
test_round(ray_df, pandas_df)
test_all(ray_df, pandas_df)
test_any(ray_df, pandas_df)
test___getitem__(ray_df, pandas_df)
test___delitem__(ray_df, pandas_df)
test___copy__(ray_df, pandas_df)
test___deepcopy__(ray_df, pandas_df)
test_bool(ray_df, pandas_df)
test_count(ray_df, pandas_df)
test_head(ray_df, pandas_df)
test_tail(ray_df, pandas_df)
test_idxmax(ray_df, pandas_df)
test_idxmin(ray_df, pandas_df)
test_pop(ray_df, pandas_df)
for key in keys:
test_get(ray_df, pandas_df, key)
test_get_dtype_counts(ray_df, pandas_df)
test_get_ftype_counts(ray_df, pandas_df)
test_max(ray_df, pandas_df)
test_min(ray_df, pandas_df)
test_notna(ray_df, pandas_df)
test_notnull(ray_df, pandas_df)
def ray_start_single_node():
# Please start this fixture in a cluster with 2 cpus.
address_info = ray.init(num_cpus=8)
yield address_info
ray.shutdown()
def test_mini_integration(ray_start_cluster, connect_to_client):
# Create bundles as many as number of gpus in the cluster.
# Do some random work and make sure all resources are properly recovered.
cluster = ray_start_cluster
num_nodes = 5
per_bundle_gpus = 2
gpu_per_node = 4
total_gpus = num_nodes * per_bundle_gpus * gpu_per_node
per_node_gpus = per_bundle_gpus * gpu_per_node
bundles_per_pg = 2
total_num_pg = total_gpus // (bundles_per_pg * per_bundle_gpus)
[
cluster.add_node(num_cpus=2, num_gpus=per_bundle_gpus * gpu_per_node)
for _ in range(num_nodes)
]
cluster.wait_for_nodes()
ray.init(address=cluster.address)
with connect_to_client_or_not(connect_to_client):
@ray.remote(num_cpus=0, num_gpus=1)
def random_tasks():
import time
import random
sleep_time = random.uniform(0.1, 0.2)
time.sleep(sleep_time)
return True
pgs = []
pg_tasks = []
# total bundle gpu usage = bundles_per_pg*total_num_pg*per_bundle_gpus
# Note this is half of total
for index in range(total_num_pg):
pgs.append(
ray.util.placement_group(
name=f"name{index}",
strategy="PACK",
bundles=[{
"GPU": per_bundle_gpus
} for _ in range(bundles_per_pg)]))
# Schedule tasks.
for i in range(total_num_pg):
pg = pgs[i]
pg_tasks.append([
random_tasks.options(
placement_group=pg,
placement_group_bundle_index=bundle_index).remote()
for bundle_index in range(bundles_per_pg)
])
# Make sure tasks are done and we remove placement groups.
num_removed_pg = 0
pg_indexes = [2, 3, 1, 7, 8, 9, 0, 6, 4, 5]
while num_removed_pg < total_num_pg:
index = pg_indexes[num_removed_pg]
pg = pgs[index]
assert all(ray.get(pg_tasks[index]))
ray.util.remove_placement_group(pg)
num_removed_pg += 1
@ray.remote(num_cpus=2, num_gpus=per_node_gpus)
class A:
def ping(self):
return True
# Make sure all resources are properly returned by scheduling
# actors that take up all existing resources.
actors = [A.remote() for _ in range(num_nodes)]
assert all(ray.get([a.ping.remote() for a in actors]))
self.register_variables(self.base_model.variables)
@override(ModelV2)
def forward(self, input_dict, state, seq_lens):
out, self._value_out = self.base_model(
[input_dict["obs"], input_dict["is_training"]])
return out, []
@override(ModelV2)
def value_function(self):
return tf.reshape(self._value_out, [-1])
if __name__ == "__main__":
args = parser.parse_args()
ray.init()
ModelCatalog.register_custom_model("bn_model", BatchNormModel)
config = {
"env": "Pendulum-v0" if args.run == "DDPG" else "CartPole-v0",
"model": {
"custom_model": "bn_model",
},
"num_workers": 0,
}
from ray.rllib.agents.ppo import PPOTrainer
trainer = PPOTrainer(config=config)
trainer.train()
import unittest
import ray
from ray.rllib.evaluation.rollout_worker import RolloutWorker
from ray.rllib.tests.test_rollout_worker import MockPolicy
class TestPerf(unittest.TestCase):
# Tested on Intel(R) Core(TM) i7-4600U CPU @ 2.10GHz
# 11/23/18: Samples per second 8501.125113727468
# 03/01/19: Samples per second 8610.164353268685
def testBaselinePerformance(self):
for _ in range(20):
ev = RolloutWorker(
env_creator=lambda _: gym.make("CartPole-v0"),
policy=MockPolicy,
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()
if __name__ == "__main__":
ray.init(num_cpus=5)
unittest.main(verbosity=2)
def __init__(self, game_name, config=None, split_resources_in=1):
# Load the game and the config from the module with the game name
try:
game_module = importlib.import_module("games." + game_name)
self.Game = game_module.Game
self.config = game_module.MuZeroConfig()
except ModuleNotFoundError as err:
print(
f'{game_name} is not a supported game name, try "cartpole" or refer to the documentation for adding a new game.'
)
raise err
# Overwrite the config
if config:
if type(config) is dict:
for param, value in config.items():
setattr(self.config, param, value)
else:
self.config = config
# Fix random generator seed
numpy.random.seed(self.config.seed)
torch.manual_seed(self.config.seed)
# Manage GPUs
if self.config.max_num_gpus == 0 and (self.config.selfplay_on_gpu
or self.config.train_on_gpu
or self.config.reanalyse_on_gpu):
raise ValueError(
"Inconsistent MuZeroConfig: max_num_gpus = 0 but GPU requested by selfplay_on_gpu or train_on_gpu or reanalyse_on_gpu."
)
if (self.config.selfplay_on_gpu or self.config.train_on_gpu
or self.config.reanalyse_on_gpu):
total_gpus = (self.config.max_num_gpus if self.config.max_num_gpus
is not None else torch.cuda.device_count())
else:
total_gpus = 0
self.num_gpus = total_gpus / split_resources_in
if 1 < self.num_gpus:
self.num_gpus = math.floor(self.num_gpus)
ray.init(num_gpus=total_gpus, ignore_reinit_error=True)
# Checkpoint and replay buffer used to initialize workers
self.checkpoint = {
"weights": None,
"optimizer_state": None,
"total_reward": 0,
"muzero_reward": 0,
"opponent_reward": 0,
"episode_length": 0,
"mean_value": 0,
"training_step": 0,
"lr": 0,
"total_loss": 0,
"value_loss": 0,
"reward_loss": 0,
"policy_loss": 0,
"num_played_games": 0,
"num_played_steps": 0,
"num_reanalysed_games": 0,
"terminate": False,
}
self.replay_buffer = {}
cpu_actor = CPUActor.remote()
cpu_weights = cpu_actor.get_initial_weights.remote(self.config)
self.checkpoint["weights"], self.summary = copy.deepcopy(
ray.get(cpu_weights))
# Workers
self.self_play_workers = None
self.test_worker = None
self.training_worker = None
self.reanalyse_worker = None
self.replay_buffer_worker = None
self.shared_storage_worker = None
def ray_start_single_node_2_gpus():
# Please start this fixture in a cluster with 2 GPUs.
address_info = ray.init(num_gpus=2)
yield address_info
ray.shutdown()
def setUpClass(cls) -> None:
ray.init()
def main(num_actors,
env_name="CartPole-v0",
batch_size=16,
update_iter=16,
gamma=0.97,
eta=0.9,
alpha=0.9,
burnin_length=4,
unroll_length=4):
s = time.time()
ray.init(local_mode=False)
logdir = Path(__file__).parent / "log"
if logdir.exists():
shutil.rmtree(logdir)
summary_writer = tf.summary.create_file_writer(str(logdir))
history = []
epsilons = np.linspace(0.1, 0.8, num_actors) if num_actors > 1 else [0.3]
actors = [
Actor.remote(pid=i,
env_name=env_name,
epsilon=epsilons[i],
gamma=gamma,
eta=eta,
alpha=alpha,
burnin_length=burnin_length,
unroll_length=unroll_length) for i in range(num_actors)
]
replay = SegmentReplayBuffer(buffer_size=2**12)
learner = Learner.remote(env_name=env_name,
gamma=gamma,
eta=eta,
alpha=alpha,
burnin_length=burnin_length,
unroll_length=unroll_length)
current_weights = ray.get(learner.define_network.remote())
current_weights = ray.put(current_weights)
tester = Tester.remote(env_name=env_name)
wip_actors = [
actor.sync_weights_and_rollout.remote(current_weights)
for actor in actors
]
for _ in range(10):
finished, wip_actors = ray.wait(wip_actors, num_returns=1)
priorities, segments, pid = ray.get(finished[0])
replay.add(priorities, segments)
wip_actors.extend(
[actors[pid].sync_weights_and_rollout.remote(current_weights)])
# minibatchs: (indices, weights, segments)
minibatchs = [
replay.sample_minibatch(batch_size=batch_size)
for _ in range(update_iter)
]
wip_learner = learner.update_network.remote(minibatchs)
wip_tester = tester.test_play.remote(current_weights, epsilon=0.05)
minibatchs = [
replay.sample_minibatch(batch_size=batch_size)
for _ in range(update_iter)
]
learner_cycles = 1
actor_cycles = 0
n_segment_added = 0
print("Start learning!!")
while learner_cycles <= 50:
actor_cycles += 1
finished, wip_actors = ray.wait(wip_actors, num_returns=1)
priorities, segments, pid = ray.get(finished[0])
replay.add(priorities, segments)
wip_actors.extend(
[actors[pid].sync_weights_and_rollout.remote(current_weights)])
n_segment_added += len(segments)
finished_learner, _ = ray.wait([wip_learner], timeout=0)
if finished_learner:
current_weights, indices, priorities, loss = ray.get(
finished_learner[0])
wip_learner = learner.update_network.remote(minibatchs)
current_weights = ray.put(current_weights)
#: 優先度の更新とminibatchの作成はlearnerよりも十分に速いという前提
replay.update_priority(indices, priorities)
minibatchs = [
replay.sample_minibatch(batch_size=batch_size)
for _ in range(update_iter)
]
print("Actor cycle:", actor_cycles, "Added:", n_segment_added)
learner_cycles += 1
actor_cycles = 0
n_segment_added = 0
with summary_writer.as_default():
tf.summary.scalar("learner_loss", loss, step=learner_cycles)
if learner_cycles % 5 == 0:
test_rewards = ray.get(wip_tester)
history.append((learner_cycles - 5, test_rewards))
wip_tester = tester.test_play.remote(current_weights,
epsilon=0.05)
print("Cycle:", learner_cycles, "Score:", test_rewards)
with summary_writer.as_default():
tf.summary.scalar("test_rewards",
test_rewards,
step=learner_cycles)
tf.summary.scalar("buffer_size",
len(replay),
step=learner_cycles)
wallclocktime = round(time.time() - s, 2)
cycles, scores = zip(*history)
plt.plot(cycles, scores)
plt.title(f"total time: {wallclocktime} sec")
plt.ylabel("test_score(epsilon=0.1)")
plt.savefig("log/history.png")
arr = np.ones(OBJECT_SIZE, dtype=np.uint8)
ref = ray.put(arr)
for actor in tqdm(actors, desc="Ensure all actors have started."):
ray.get(actor.foo.remote())
result_refs = []
for actor in tqdm(actors, desc="Broadcasting objects"):
result_refs.append(actor.sum.remote(ref))
results = ray.get(result_refs)
for result in results:
assert result == OBJECT_SIZE
ray.init(address="auto")
start = perf_counter()
test_object_broadcast()
end = perf_counter()
print(f"Broadcast time: {end - start} ({OBJECT_SIZE} B x {NUM_NODES} nodes)")
if "TEST_OUTPUT_JSON" in os.environ:
out_file = open(os.environ["TEST_OUTPUT_JSON"], "w")
results = {
"broadcast_time": end - start,
"object_size": OBJECT_SIZE,
"num_nodes": NUM_NODES,
"success": "1"
}
json.dump(results, out_file)
def test_capture_child_actors(ray_start_cluster, connect_to_client):
cluster = ray_start_cluster
total_num_actors = 4
for _ in range(2):
cluster.add_node(num_cpus=total_num_actors)
ray.init(address=cluster.address)
with connect_to_client_or_not(connect_to_client):
pg = ray.util.placement_group(
[{
"CPU": 2
}, {
"CPU": 2
}], strategy="STRICT_PACK")
ray.get(pg.ready())
# If get_current_placement_group is used when the current worker/driver
# doesn't belong to any of placement group, it should return None.
assert get_current_placement_group() is None
# Test actors first.
@ray.remote(num_cpus=1)
class NestedActor:
def ready(self):
return True
@ray.remote(num_cpus=1)
class Actor:
def __init__(self):
self.actors = []
def ready(self):
return True
def schedule_nested_actor(self):
# Make sure we can capture the current placement group.
assert get_current_placement_group() is not None
# Actors should be implicitly captured.
actor = NestedActor.remote()
ray.get(actor.ready.remote())
self.actors.append(actor)
def schedule_nested_actor_outside_pg(self):
# Don't use placement group.
actor = NestedActor.options(placement_group=None).remote()
ray.get(actor.ready.remote())
self.actors.append(actor)
a = Actor.options(
placement_group=pg,
placement_group_capture_child_tasks=True).remote()
ray.get(a.ready.remote())
# 1 top level actor + 3 children.
for _ in range(total_num_actors - 1):
ray.get(a.schedule_nested_actor.remote())
# Make sure all the actors are scheduled on the same node.
# (why? The placement group has STRICT_PACK strategy).
node_id_set = set()
for actor_info in ray.state.actors().values():
if actor_info["State"] == convert_actor_state(
gcs_utils.ActorTableData.ALIVE):
node_id = actor_info["Address"]["NodeID"]
node_id_set.add(node_id)
# Since all node id should be identical, set should be equal to 1.
assert len(node_id_set) == 1
# Kill an actor and wait until it is killed.
kill_actor_and_wait_for_failure(a)
with pytest.raises(ray.exceptions.RayActorError):
ray.get(a.ready.remote())
# Now create an actor, but do not capture the current tasks
a = Actor.options(placement_group=pg).remote()
ray.get(a.ready.remote())
# 1 top level actor + 3 children.
for _ in range(total_num_actors - 1):
ray.get(a.schedule_nested_actor.remote())
# Make sure all the actors are not scheduled on the same node.
# It is because the child tasks are not scheduled on the same
# placement group.
node_id_set = set()
for actor_info in ray.state.actors().values():
if actor_info["State"] == convert_actor_state(
gcs_utils.ActorTableData.ALIVE):
node_id = actor_info["Address"]["NodeID"]
node_id_set.add(node_id)
assert len(node_id_set) == 2
# Kill an actor and wait until it is killed.
kill_actor_and_wait_for_failure(a)
with pytest.raises(ray.exceptions.RayActorError):
ray.get(a.ready.remote())
# Lastly, make sure when None is specified, actors are not scheduled
# on the same placement group.
a = Actor.options(placement_group=pg).remote()
ray.get(a.ready.remote())
# 1 top level actor + 3 children.
for _ in range(total_num_actors - 1):
ray.get(a.schedule_nested_actor_outside_pg.remote())
# Make sure all the actors are not scheduled on the same node.
# It is because the child tasks are not scheduled on the same
# placement group.
node_id_set = set()
for actor_info in ray.state.actors().values():
if actor_info["State"] == convert_actor_state(
gcs_utils.ActorTableData.ALIVE):
node_id = actor_info["Address"]["NodeID"]
node_id_set.add(node_id)
assert len(node_id_set) == 2
assert (num_nodes * object_store_memory + num_redis_shards * redis_max_memory <
ray.utils.get_system_memory() / 2)
# Simulate a cluster on one machine.
cluster = Cluster()
for i in range(num_nodes):
cluster.add_node(
redis_port=6379 if i == 0 else None,
num_redis_shards=num_redis_shards if i == 0 else None,
num_cpus=4,
num_gpus=0,
resources={str(i): 5},
object_store_memory=object_store_memory,
redis_max_memory=redis_max_memory)
ray.init(address=cluster.address)
# Run the workload.
# Define a driver script that runs a few tasks and actors on each node in the
# cluster.
driver_script = """
import ray
ray.init(address="{}")
num_nodes = {}
@ray.remote
def f():
parse.add_argument("--batch_size", default=256, type=int, help="train batch size")
parse.add_argument("--lr", default=0.625e-4, type=float, help="learning rate")
parse.add_argument("--suffix", default="default", help="suffix for saving folders")
parse.add_argument("--speed", default=8, type=int, help="training data consum speed. default 8x data generate speed")
args = parse.parse_args()
n_worker = args.num_agents
n_loader = args.num_loaders
env_name = "{}NoFrameskip-v4".format(args.env_name)
buffer = "multi_workers_buffer"
batch_size = args.batch_size
lr = args.lr
suffix = args.suffix
speed = args.speed
n_iter = 40*32//batch_size
ray.init(num_cpus=1+2*n_worker+n_loader, object_store_memory=1*1024**3, memory=6*1024**3)
buffer = lmdb_op.init(buffer)
workers = [ray.remote(DQN_Worker).options(num_gpus=0.1).remote(env_name=env_name, db=buffer, db_write=lmdb_op.write) for _ in range(n_worker)]
test_worker = ray.remote(DQN_Worker).options(num_gpus=0.1).remote(env_name=env_name, phase="test", suffix=suffix)
worker_id = {worker: "worker_{}".format(i) for i, worker in enumerate(workers)}
dataloader = Dataloader(buffer, lmdb_op, worker_num=n_loader, batch_size=batch_size, batch_num=n_iter)
opt = ray.remote(Optimizer).options(num_gpus=0.3).remote(dataloader, env_name, suffix=suffix, iter_steps=n_iter, update_period=10000, lr=lr)
sche = Sched()
eps = 1
opt_start = False
glog = SummaryWriter("./logdir/{}/{}/{}.lr{}.batch{}".format(env_name, suffix, Optimizer.__name__, lr, batch_size))
model_save_period = 100000
train_step = 0
model_idx = 0
register_env(gym_name, create_env)
return alg_run, gym_name, config
if __name__ == "__main__":
parser = ArgumentParser()
parser.add_argument("-s",
"--seeds_file",
dest="seeds_file",
help="pickle file containing seeds",
default=None)
args = parser.parse_args()
alg_run, gym_name, config = setup_exps(args.seeds_file)
ray.init(num_cpus=N_CPUS + 1)
trials = run_experiments(
{
flow_params["exp_tag"]: {
"run": alg_run,
"env": gym_name,
"config": {
**config
},
"restore":
'/home/flow/ray_results/central_i696_IDMJunction_horizon2000_Ignore/PPO_MergePOEnv_Ignore-v0_0_2020-04-25_13-49-44m7fk71uh/checkpoint_230',
#"resume":True,
"checkpoint_freq": 10, #20,
"checkpoint_at_end": True,
"max_failures": 999,
"stop": {
def test_atomic_creation(ray_start_cluster, connect_to_client):
# Setup cluster.
cluster = ray_start_cluster
bundle_cpu_size = 2
bundle_per_node = 2
num_nodes = 2
[
cluster.add_node(num_cpus=bundle_cpu_size * bundle_per_node)
for _ in range(num_nodes)
]
ray.init(address=cluster.address)
@ray.remote(num_cpus=1)
class NormalActor:
def ping(self):
pass
@ray.remote(num_cpus=3)
def bothering_task():
time.sleep(6)
return True
with connect_to_client_or_not(connect_to_client):
# Schedule tasks to fail initial placement group creation.
tasks = [bothering_task.remote() for _ in range(2)]
# Make sure the two common task has scheduled.
def tasks_scheduled():
return ray.available_resources()["CPU"] == 2.0
wait_for_condition(tasks_scheduled)
# Create an actor that will fail bundle scheduling.
# It is important to use pack strategy to make test less flaky.
pg = ray.util.placement_group(
name="name",
strategy="SPREAD",
bundles=[{
"CPU": bundle_cpu_size
} for _ in range(num_nodes * bundle_per_node)])
# Create a placement group actor.
# This shouldn't be scheduled because atomic
# placement group creation should've failed.
pg_actor = NormalActor.options(
placement_group=pg,
placement_group_bundle_index=num_nodes * bundle_per_node -
1).remote()
# Wait on the placement group now. It should be unready
# because normal actor takes resources that are required
# for one of bundle creation.
ready, unready = ray.wait([pg.ready()], timeout=0.5)
assert len(ready) == 0
assert len(unready) == 1
# Wait until all tasks are done.
assert all(ray.get(tasks))
# Wait on the placement group creation. Since resources are now
# available, it should be ready soon.
ready, unready = ray.wait([pg.ready()])
assert len(ready) == 1
assert len(unready) == 0
# Confirm that the placement group actor is created. It will
# raise an exception if actor was scheduled before placement
# group was created thus it checks atomicity.
ray.get(pg_actor.ping.remote(), timeout=3.0)
ray.kill(pg_actor)
# Make sure atomic creation failure didn't impact resources.
@ray.remote(num_cpus=bundle_cpu_size)
def resource_check():
return True
# This should hang because every resources
# are claimed by placement group.
check_without_pg = [
resource_check.remote() for _ in range(bundle_per_node * num_nodes)
]
# This all should scheduled on each bundle.
check_with_pg = [
resource_check.options(
placement_group=pg, placement_group_bundle_index=i).remote()
for i in range(bundle_per_node * num_nodes)
]
# Make sure these are hanging.
ready, unready = ray.wait(check_without_pg, timeout=0)
assert len(ready) == 0
assert len(unready) == bundle_per_node * num_nodes
# Make sure these are all scheduled.
assert all(ray.get(check_with_pg))
ray.util.remove_placement_group(pg)
def pg_removed():
return ray.util.placement_group_table(pg)["state"] == "REMOVED"
wait_for_condition(pg_removed)
# Make sure check without pgs are all
# scheduled properly because resources are cleaned up.
assert all(ray.get(check_without_pg))
def test_client_serialize_addon(call_ray_stop_only):
import ray
import pydantic
ray.init(num_cpus=0)
class User(pydantic.BaseModel):
name: str
with ray_start_client_server() as ray:
assert ray.get(ray.put(User(name="ray"))).name == "ray"
object_ref_cleanup_script = """
import ray
ray.init("ray://localhost:50051")
@ray.remote
def f():
return 42
@ray.remote
class SomeClass:
pass
obj_ref = f.remote()
actor_ref = SomeClass.remote()
max_episode_steps=200,
kwargs={})
def create_env(env_config):
pass_params_to_gym(env_name)
env = gym.envs.make(env_name)
return env
if __name__ == '__main__':
register_env(env_name, lambda env_config: create_env(env_config))
config = ppo.DEFAULT_CONFIG.copy()
horizon = 10
num_cpus = 4
ray.init(num_cpus=num_cpus, redirect_output=True)
config["num_workers"] = num_cpus
config["train_batch_size"] = 1000
config["num_sgd_iter"] = 10
config["gamma"] = 0.999
config["horizon"] = horizon
config["use_gae"] = False
config["model"].update({"fcnet_hiddens": [256, 256]})
options = {
"multiagent_obs_shapes": [2, 2],
"multiagent_act_shapes": [1, 1],
"multiagent_shared_model": False,
"multiagent_fcnet_hiddens": [[32, 32]] * 2
}
config["model"].update({"custom_options": options})
alg = ppo.PPOAgent(env=env_name, config=config)
def test_automatic_cleanup_detached_actors(ray_start_cluster):
# Make sure the placement groups created by a
# detached actors are cleaned properly.
cluster = ray_start_cluster
num_nodes = 3
num_cpu_per_node = 2
# Create 3 nodes cluster.
for _ in range(num_nodes):
cluster.add_node(num_cpus=num_cpu_per_node)
cluster.wait_for_nodes()
info = ray.init(
address=cluster.address, namespace="default_test_namespace")
available_cpus = ray.available_resources()["CPU"]
assert available_cpus == num_nodes * num_cpu_per_node
driver_code = f"""
import ray
ray.init(address="{info["redis_address"]}", namespace="default_test_namespace")
def create_pg():
pg = ray.util.placement_group(
[{{"CPU": 1}} for _ in range(3)],
strategy="STRICT_SPREAD")
ray.get(pg.ready())
return pg
# TODO(sang): Placement groups created by tasks launched by detached actor
# is not cleaned with the current protocol.
# @ray.remote(num_cpus=0)
# def f():
# create_pg()
@ray.remote(num_cpus=0, max_restarts=1)
class A:
def create_pg(self):
create_pg()
def create_child_pg(self):
self.a = A.options(name="B").remote()
ray.get(self.a.create_pg.remote())
def kill_child_actor(self):
ray.kill(self.a)
try:
ray.get(self.a.create_pg.remote())
except Exception:
pass
a = A.options(lifetime="detached", name="A").remote()
ray.get(a.create_pg.remote())
# TODO(sang): Currently, child tasks are cleaned when a detached actor
# is dead. We cannot test this scenario until it is fixed.
# ray.get(a.create_child_pg.remote())
ray.shutdown()
"""
run_string_as_driver(driver_code)
# Wait until the driver is reported as dead by GCS.
def is_job_done():
jobs = ray.state.jobs()
for job in jobs:
if job["IsDead"]:
return True
return False
def assert_num_cpus(expected_num_cpus):
if expected_num_cpus == 0:
return "CPU" not in ray.available_resources()
return ray.available_resources()["CPU"] == expected_num_cpus
wait_for_condition(is_job_done)
wait_for_condition(lambda: assert_num_cpus(num_nodes))
# Make sure when a child actor spawned by a detached actor
# is killed, the placement group is removed.
a = ray.get_actor("A")
# TODO(sang): child of detached actors
# seem to be killed when jobs are done. We should fix this before
# testing this scenario.
# ray.get(a.kill_child_actor.remote())
# assert assert_num_cpus(num_nodes)
# Make sure placement groups are cleaned when detached actors are killed.
ray.kill(a, no_restart=False)
wait_for_condition(lambda: assert_num_cpus(num_nodes * num_cpu_per_node))
# The detached actor a should've been restarted.
# Recreate a placement group.
ray.get(a.create_pg.remote())
wait_for_condition(lambda: assert_num_cpus(num_nodes))
# Kill it again and make sure the placement group
# that is created is deleted again.
ray.kill(a, no_restart=False)
wait_for_condition(lambda: assert_num_cpus(num_nodes * num_cpu_per_node))
import time
import ray
@ray.remote
def f():
time.sleep(0.01)
return ray.services.get_node_ip_address()
ray.init(redis_address='master:6379')
# Get a list of the IP addresses of the nodes that have joined the cluster.
ids = set(ray.get([f.remote() for _ in range(1000)]))
print(ids)
N_ROWS = args.num_rows # number of row of bidirectional lanes
N_COLUMNS = args.num_cols # number of columns of bidirectional lanes
flow_params = make_flow_params(N_ROWS, N_COLUMNS, EDGE_INFLOW)
upload_dir = args.upload_dir
RUN_MODE = args.run_mode
ALGO = args.algo
if ALGO == 'PPO':
alg_run, env_name, config = setup_exps_PPO(flow_params)
else:
raise NotImplementedError
if RUN_MODE == 'local':
ray.init(num_cpus=N_CPUS + 1)
N_ITER = 1
elif RUN_MODE == 'cluster':
ray.init(redis_address="localhost:6379")
N_ITER = 2000
exp_tag = {
'run': alg_run,
'env': env_name,
'checkpoint_freq': 25,
"max_failures": 10,
'stop': {
'training_iteration': N_ITER
},
'config': config,
"num_samples": 1,
def ray_start_2_cpus():
address_info = ray.init(num_cpus=2)
yield address_info
# The code after the yield will run as teardown code.
ray.shutdown()
import io
import base64
from IPython.display import HTML
from dqn_model import DQNModel
from dqn_model import _DQNModel
from memory_remote import ReplayBuffer_remote
import matplotlib.pyplot as plt
from custom_cartpole import CartPoleEnv
# matplotlib inline
ray.shutdown()
ray.init(include_webui=False, ignore_reinit_error=True, redis_max_memory=500000000, object_store_memory=5000000000, temp_dir="../../tmp/")
FloatTensor = torch.FloatTensor
# # Set the Env name and action space for CartPole
# ENV_NAME = 'CartPole-v0'
# # Move left, Move right
# ACTION_DICT = {"LEFT": 0, "RIGHT": 1}
# # Register the environment
# env_CartPole = gym.make(ENV_NAME)
#
# # Set result saveing floder
# result_floder = ENV_NAME
# result_file = ENV_NAME + "/results.txt"
def main(args):
# ====================================
# init env config
# ====================================
if args.no_debug:
ray.init(webui_host="127.0.0.1")
else:
ray.init(local_mode=True, webui_host="127.0.0.1")
# use ray cluster for training
# ray.init(
# address="auto" if args.address is None else args.address,
# redis_password="******",
# )
#
# print(
# "--------------- Ray startup ------------\n{}".format(
# ray.state.cluster_resources()
# )
# )
agent_specs = {"AGENT-007": agent_spec}
env_config = {
"seed": 42,
"scenarios": [scenario_paths],
"headless": args.headless,
"agent_specs": agent_specs,
}
# ====================================
# init tune config
# ====================================
class MultiEnv(RLlibHiWayEnv):
def __init__(self, env_config):
env_config["scenarios"] = [
scenario_paths[(env_config.worker_index - 1) %
len(scenario_paths)]
]
super(MultiEnv, self).__init__(config=env_config)
tune_config = {
"env": MultiEnv,
"env_config": env_config,
"multiagent": {
"policies": {
"default_policy": (
None,
OBSERVATION_SPACE,
ACTION_SPACE,
{},
)
},
"policy_mapping_fn": lambda agent_id: "default_policy",
},
"framework": "torch",
"callbacks": {
"on_episode_start": on_episode_start,
"on_episode_step": on_episode_step,
"on_episode_end": on_episode_end,
},
"lr": 1e-4,
"log_level": "WARN",
"num_workers": args.num_workers,
"horizon": args.horizon,
"learning_starts": 10000,
"buffer_size": 50000,
"rollout_fragment_length": 4,
"train_batch_size": 32,
# "exploration_config": {
# "epsilon_timesteps": 200000,
# "final_epsilon": .01
# },
"timesteps_per_iteration": 10000,
"dueling": False,
"prioritized_replay": False,
# "model": {
# "fcnet_hiddens": [256, 256],
# "fcnet_activation": "relu",
# },
"exploration_config": {
# Exploration sub-class by name or full path to module+class
# (e.g. “ray.rllib.utils.exploration.epsilon_greedy.EpsilonGreedy”)
# "type": "zhr_train_rllib.multiobj_episilongreedy.EpsilonGreedy",
# Parameters for the Exploration class' constructor:
"initial_epsilon": 1.0,
"final_epsilon": 0.02,
"epsilon_timesteps":
200000, # Timesteps over which to anneal epsilon.
},
# "observation_filter": "MeanStdFilter",
# "model":{
# "use_lstm": True,
# },
}
tune_config.update({
"gamma": 0.995,
"decompose_num": 2,
# "seed_global": tune.grid_search([10, 20, 30, 40])
})
num_samples = 4
if not args.no_debug:
tune_config.update({
"learning_starts": 1000,
"buffer_size": 5000,
"timesteps_per_iteration": 1000,
})
num_samples = 1
# ====================================
# init log and checkpoint dir_info
# ====================================
experiment_name = EXPERIMENT_NAME.format(
scenario=args.exper,
algorithm="DQN",
n_agent=1,
)
log_dir = Path(args.log_dir).expanduser().absolute() / RUN_NAME
log_dir.mkdir(parents=True, exist_ok=True)
print(f"Checkpointing at {log_dir}")
if args.restore:
restore_path = Path(args.restore).expanduser()
print(f"Loading model from {restore_path}")
else:
restore_path = None
# run experiments
analysis = tune.run(
DQN,
# "DQN",
name=experiment_name,
stop={"timesteps_total": 10000000 / 2},
checkpoint_freq=10,
checkpoint_at_end=True,
local_dir=str(log_dir),
resume=args.resume,
restore=restore_path,
max_failures=1000,
export_formats=["model", "checkpoint"],
config=tune_config,
num_samples=num_samples,
)
print(analysis.dataframe().head())
def tune_random(name, exp, num_samples=2, seed_value=None, **digit_kwargs):
"""Tune hyperparameters of any bandit experiment."""
# ------------------------------------------------------------------------
# Init seed
prng = np.random.RandomState(seed_value)
# Init ray
if not ray.is_initialized():
if "debug" in digit_kwargs:
ray.init(local_mode=digit_kwargs["debug"])
else:
ray.init()
# ------------------------------------------------------------------------
# Create the train function. We do it scope to control if it
# gets remote'ed to GPUs or not.
try:
if digit_kwargs["use_gpu"]:
@ray.remote(num_gpus=0.25)
def train(name=None, exp_func=None, config=None):
trial = exp_func(**config)
trial.update({"config": config, "name": name})
return trial
else:
@ray.remote
def train(name=None, exp_func=None, config=None):
trial = exp_func(**config)
trial.update({"config": config, "name": name})
return trial
except KeyError:
@ray.remote
def train(name=None, exp_func=None, config=None):
trial = exp_func(**config)
trial.update({"config": config, "name": name})
return trial
# ------------------------------------------------------------------------
# Init:
# Separate name from path
path, name = os.path.split(name)
# Look up the bandit run function were using in this tuning.
exp_func = getattr(glia_digits, exp)
# ------------------------------------------------------------------------
# Run!
# Setup the parallel workers
runs = []
for i in range(num_samples):
# Make a new HP sample
params = {}
for k, v in digit_kwargs.items():
if isinstance(v, str):
params[k] = v
elif isinstance(v, bool):
params[k] = v
elif isinstance(v, int):
params[k] = v
elif isinstance(v, float):
params[k] = v
else:
low, high = v
params[k] = prng.uniform(low=low, high=high)
runs.append(train.remote(i, exp_func, params))
trials = [ray.get(r) for r in runs]
# ------------------------------------------------------------------------
# Save configs and correct (full model data is dropped):
best = get_best_trial(trials, 'correct')
# Best trial config
best_config = best["config"]
best_config.update(get_best_result(trials, 'correct'))
save_checkpoint(best_config,
filename=os.path.join(path, name + "_best.pkl"))
# Sort and save the configs of all trials
sorted_configs = {}
for i, trial in enumerate(get_sorted_trials(trials, 'correct')):
sorted_configs[i] = trial["config"]
sorted_configs[i].update({"correct": trial["correct"]})
save_checkpoint(sorted_configs,
filename=os.path.join(path, name + "_sorted.pkl"))
# kill ray
ray.shutdown()
# -
return best, trials
def ray_init_with_task_retry_delay():
address = ray.init(_system_config={"task_retry_delay_ms": 100})
yield address
ray.shutdown()
import torch
from scipy.signal import savgol_filter
from tqdm import tqdm
import ffmpeg
import ray
from data_segments.get_data_segments import (get_flame_params_for_file,
get_segments_v2)
from misc.read_n_write import flame2glow
from misc.shared import DATA_DIR
from misc.utils import frames2ms, get_participant
WIN_LEN = 9
ONLY_ODD = True
ray.init(num_cpus=10)
segments = get_segments_v2()
sessions_dir = DATA_DIR / "Sessions_50fps_pytorch_data/sessions"
def to_iterator(obj_ids):
while obj_ids:
done, obj_ids = ray.wait(obj_ids)
yield ray.get(done[0])
def get_openface(file_):
tar_file = TarFile(file_)
filename="smoketest.parquet",
num_rows=args.num_workers * 500,
num_features=4,
num_classes=2,
num_partitions=args.num_workers * 10)
use_gpu = False
else:
path = args.file
if not os.path.exists(path):
raise ValueError(
f"Benchmarking data not found: {path}."
f"\nFIX THIS by running `python create_test_data.py` first.")
init_start = time.time()
if args.smoke_test:
ray.init(num_cpus=num_workers)
else:
ray.init(address="auto")
init_taken = time.time() - init_start
full_start = time.time()
bst, train_taken = train_ray(path=path,
num_workers=num_workers,
num_boost_rounds=num_boost_rounds,
num_files=num_files,
regression=args.regression,
use_gpu=use_gpu,
smoke_test=args.smoke_test)
full_taken = time.time() - full_start
print(f"TOTAL TIME TAKEN: {full_taken:.2f} seconds "
f"({init_taken:.2f} for init)")
def setUpClass(cls):
ray.init()
