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 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 testObjStore(self):
node_ip_address = "127.0.0.1"
scheduler_address = ray.services.start_ray_local(num_objstores=2, num_workers=0, worker_path=None)
ray.connect(node_ip_address, scheduler_address, mode=ray.SCRIPT_MODE)
objstore_addresses = [objstore_info["address"] for objstore_info in ray.scheduler_info()["objstores"]]
w1 = ray.worker.Worker()
w2 = ray.worker.Worker()
ray.reusables._cached_reusables = [] # This is a hack to make the test run.
ray.connect(node_ip_address, scheduler_address, objstore_address=objstore_addresses[0], mode=ray.SCRIPT_MODE, worker=w1)
ray.reusables._cached_reusables = [] # This is a hack to make the test run.
ray.connect(node_ip_address, scheduler_address, objstore_address=objstore_addresses[1], mode=ray.SCRIPT_MODE, worker=w2)
for cls in [Foo, Bar, Baz, Qux, SubQux, Exception, CustomError, Point, NamedTupleExample]:
ray.register_class(cls)
# putting and getting an object shouldn't change it
for data in RAY_TEST_OBJECTS:
objectid = ray.put(data, w1)
result = ray.get(objectid, w1)
assert_equal(result, data)
# putting an object, shipping it to another worker, and getting it shouldn't change it
for data in RAY_TEST_OBJECTS:
objectid = ray.put(data, w1)
result = ray.get(objectid, w2)
assert_equal(result, data)
# putting an object, shipping it to another worker, and getting it shouldn't change it
for data in RAY_TEST_OBJECTS:
objectid = ray.put(data, w2)
result = ray.get(objectid, w1)
assert_equal(result, data)
# This test fails. See https://github.com/ray-project/ray/issues/159.
# getting multiple times shouldn't matter
# for data in [np.zeros([10, 20]), np.random.normal(size=[45, 25]), np.zeros([10, 20], dtype=np.dtype("float64")), np.zeros([10, 20], dtype=np.dtype("float32")), np.zeros([10, 20], dtype=np.dtype("int64")), np.zeros([10, 20], dtype=np.dtype("int32"))]:
# objectid = worker.put(data, w1)
# result = worker.get(objectid, w2)
# result = worker.get(objectid, w2)
# result = worker.get(objectid, w2)
# assert_equal(result, data)
# Getting a buffer after modifying it before it finishes should return updated buffer
objectid = ray.libraylib.get_objectid(w1.handle)
buf = ray.libraylib.allocate_buffer(w1.handle, objectid, 100)
buf[0][0] = 1
ray.libraylib.finish_buffer(w1.handle, objectid, buf[1], 0)
completedbuffer = ray.libraylib.get_buffer(w1.handle, objectid)
self.assertEqual(completedbuffer[0][0], 1)
# We started multiple drivers manually, so we will disconnect them manually.
ray.disconnect(worker=w1)
ray.disconnect(worker=w2)
ray.worker.cleanup()
def testObjStore(self):
node_ip_address = "127.0.0.1"
scheduler_address = ray.services.start_ray_local(num_objstores=2, num_workers=0, worker_path=None)
ray.connect(node_ip_address, scheduler_address, mode=ray.SCRIPT_MODE)
objstore_addresses = [objstore_info["address"] for objstore_info in ray.scheduler_info()["objstores"]]
w1 = ray.worker.Worker()
w2 = ray.worker.Worker()
ray.reusables._cached_reusables = [] # This is a hack to make the test run.
ray.connect(node_ip_address, scheduler_address, objstore_address=objstore_addresses[0], mode=ray.SCRIPT_MODE, worker=w1)
ray.reusables._cached_reusables = [] # This is a hack to make the test run.
ray.connect(node_ip_address, scheduler_address, objstore_address=objstore_addresses[1], mode=ray.SCRIPT_MODE, worker=w2)
for cls in [Foo, Bar, Baz, Qux, SubQux, Exception, CustomError, Point, NamedTupleExample]:
ray.register_class(cls)
# putting and getting an object shouldn't change it
for data in RAY_TEST_OBJECTS:
objectid = ray.put(data, w1)
result = ray.get(objectid, w1)
assert_equal(result, data)
# putting an object, shipping it to another worker, and getting it shouldn't change it
for data in RAY_TEST_OBJECTS:
objectid = ray.put(data, w1)
result = ray.get(objectid, w2)
assert_equal(result, data)
# putting an object, shipping it to another worker, and getting it shouldn't change it
for data in RAY_TEST_OBJECTS:
objectid = ray.put(data, w2)
result = ray.get(objectid, w1)
assert_equal(result, data)
# This test fails. See https://github.com/ray-project/ray/issues/159.
# getting multiple times shouldn't matter
# for data in [np.zeros([10, 20]), np.random.normal(size=[45, 25]), np.zeros([10, 20], dtype=np.dtype("float64")), np.zeros([10, 20], dtype=np.dtype("float32")), np.zeros([10, 20], dtype=np.dtype("int64")), np.zeros([10, 20], dtype=np.dtype("int32"))]:
# objectid = worker.put(data, w1)
# result = worker.get(objectid, w2)
# result = worker.get(objectid, w2)
# result = worker.get(objectid, w2)
# assert_equal(result, data)
# Getting a buffer after modifying it before it finishes should return updated buffer
objectid = ray.libraylib.get_objectid(w1.handle)
buf = ray.libraylib.allocate_buffer(w1.handle, objectid, 100)
buf[0][0] = 1
ray.libraylib.finish_buffer(w1.handle, objectid, buf[1], 0)
completedbuffer = ray.libraylib.get_buffer(w1.handle, objectid)
self.assertEqual(completedbuffer[0][0], 1)
# We started multiple drivers manually, so we will disconnect them manually.
ray.disconnect(worker=w1)
ray.disconnect(worker=w2)
ray.worker.cleanup()
def testDeallocation(self):
reload(test_functions)
for module in [
ra.core, ra.random, ra.linalg, da.core, da.random, da.linalg
]:
reload(module)
ray.init(start_ray_local=True, num_workers=1)
ray.register_class(da.DistArray)
def check_not_deallocated(object_ids):
reference_counts = ray.scheduler_info()["reference_counts"]
for object_id in object_ids:
self.assertGreater(reference_counts[object_id.id], 0)
def check_everything_deallocated():
reference_counts = ray.scheduler_info()["reference_counts"]
self.assertEqual(reference_counts, len(reference_counts) * [-1])
z = da.zeros.remote([da.BLOCK_SIZE, 2 * da.BLOCK_SIZE])
time.sleep(0.1)
objectid_val = z.id
time.sleep(0.1)
check_not_deallocated([z])
del z
time.sleep(0.1)
check_everything_deallocated()
x = ra.zeros.remote([10, 10])
y = ra.zeros.remote([10, 10])
z = ra.dot.remote(x, y)
objectid_val = x.id
time.sleep(0.1)
check_not_deallocated([x, y, z])
del x
time.sleep(0.1)
check_not_deallocated([y, z])
del y
time.sleep(0.1)
check_not_deallocated([z])
del z
time.sleep(0.1)
check_everything_deallocated()
z = da.zeros.remote([4 * da.BLOCK_SIZE])
time.sleep(0.1)
check_not_deallocated(ray.get(z).objectids.tolist())
del z
time.sleep(0.1)
check_everything_deallocated()
ray.worker.cleanup()
def testRegisterClass(self):
ray.init(start_ray_local=True, num_workers=0)
# Check that putting an object of a class that has not been registered
# throws an exception.
class TempClass(object):
pass
self.assertRaises(Exception, lambda : ray.put(Foo))
# Check that registering a class that Ray cannot serialize efficiently
# raises an exception.
self.assertRaises(Exception, lambda : ray.register_class(type(True)))
# Check that registering the same class with pickle works.
ray.register_class(type(float), pickle=True)
self.assertEqual(ray.get(ray.put(float)), float)
ray.worker.cleanup()
def testRegisterClass(self):
ray.init(start_ray_local=True, num_workers=0)
# Check that putting an object of a class that has not been registered
# throws an exception.
class TempClass(object):
pass
self.assertRaises(Exception, lambda : ray.put(Foo))
# Check that registering a class that Ray cannot serialize efficiently
# raises an exception.
self.assertRaises(Exception, lambda : ray.register_class(type(True)))
# Check that registering the same class with pickle works.
ray.register_class(type(float), pickle=True)
self.assertEqual(ray.get(ray.put(float)), float)
ray.worker.cleanup()
def mat_mul(size, num_workers, block_size):
# Start Ray.
ray.init(start_ray_local=True, num_workers=num_workers)
ray.register_class(da.DistArray)
# Allocate two size x size arrays of all one's, and multiply them. Block on
# the result.
print "Matrix multiply with size {}, block size {}".format(
size, block_size)
start = time.time()
a = da.ones.remote([size, size], block_size=block_size)
b = da.ones.remote([size, size], block_size=block_size)
c = da.dot.remote(a, b)
c = da.assemble.remote(c)
print c.id
c_get = ray.get(c)
end = time.time()
print c_get
print "{}".format(end - start)
return ray
def testAssemble(self):
for module in [
ra.core, ra.random, ra.linalg, da.core, da.random, da.linalg
]:
reload(module)
ray.init(start_ray_local=True, num_workers=1)
ray.register_class(da.DistArray)
a = ra.ones.remote([da.BLOCK_SIZE, da.BLOCK_SIZE])
b = ra.zeros.remote([da.BLOCK_SIZE, da.BLOCK_SIZE])
x = da.DistArray([2 * da.BLOCK_SIZE, da.BLOCK_SIZE],
np.array([[a], [b]]))
assert_equal(
x.assemble(),
np.vstack([
np.ones([da.BLOCK_SIZE, da.BLOCK_SIZE]),
np.zeros([da.BLOCK_SIZE, da.BLOCK_SIZE])
]))
ray.worker.cleanup()
def testGet(self):
ray.init(start_ray_local=True, num_workers=3)
for cls in [Foo, Bar, Baz, Qux, SubQux, Exception, CustomError, Point, NamedTupleExample]:
ray.register_class(cls)
# Remote objects should be deallocated when the corresponding ObjectID goes
# out of scope, and all results of ray.get called on the ID go out of scope.
for val in RAY_TEST_OBJECTS:
x = ray.put(val)
objectid = x.id
xval = ray.get(x)
del x, xval
self.assertEqual(ray.scheduler_info()["reference_counts"][objectid], -1)
# Remote objects that do not contain numpy arrays should be deallocated when
# the corresponding ObjectID goes out of scope, even if ray.get has been
# called on the ObjectID.
for val in [True, False, None, 1, 1.0, 1L, "hi", u"hi", [1, 2, 3], (1, 2, 3), [(), {(): ()}]]:
x = ray.put(val)
objectid = x.id
xval = ray.get(x)
del x
self.assertEqual(ray.scheduler_info()["reference_counts"][objectid], -1)
def testGet(self):
ray.init(start_ray_local=True, num_workers=3)
for cls in [Foo, Bar, Baz, Qux, SubQux, Exception, CustomError, Point, NamedTupleExample]:
ray.register_class(cls)
# Remote objects should be deallocated when the corresponding ObjectID goes
# out of scope, and all results of ray.get called on the ID go out of scope.
for val in RAY_TEST_OBJECTS:
x = ray.put(val)
objectid = x.id
xval = ray.get(x)
del x, xval
self.assertEqual(ray.scheduler_info()["reference_counts"][objectid], -1)
# Remote objects that do not contain numpy arrays should be deallocated when
# the corresponding ObjectID goes out of scope, even if ray.get has been
# called on the ObjectID.
for val in [True, False, None, 1, 1.0, 1L, "hi", u"hi", [1, 2, 3], (1, 2, 3), [(), {(): ()}]]:
x = ray.put(val)
objectid = x.id
xval = ray.get(x)
del x
self.assertEqual(ray.scheduler_info()["reference_counts"][objectid], -1)
def testPassingArgumentsByValue(self):
ray.init(num_workers=1)
@ray.remote
def f(x):
return x
ray.register_class(Exception)
ray.register_class(CustomError)
ray.register_class(Point)
ray.register_class(Foo)
ray.register_class(Bar)
ray.register_class(Baz)
ray.register_class(NamedTupleExample)
# Check that we can pass arguments by value to remote functions and that
# they are uncorrupted.
for obj in RAY_TEST_OBJECTS:
assert_equal(obj, ray.get(f.remote(obj)))
ray.worker.cleanup()
# find most common word
most_popular_word = None
most_popular_ct = 0
for word, ct in ray.get(res).items():
if ct > most_popular_ct:
most_popular_word = word
most_popular_ct = ct
print "most popular word is '{}' with count {}".format(
most_popular_word, most_popular_ct)
def chunks(l, n):
chunk_size = (len(l) - 1) / n + 1
for i in xrange(0, len(l), chunk_size):
yield l[i:i + chunk_size]
if __name__ == '__main__':
bench_env = raybench.Env()
bench_env.ray_init()
ray.register_class(type(dict_merge.remote), pickle=True)
num_splits = bench_env.num_workers
input_splits = init_wc(num_splits, 100000)
print "number of splits", len(input_splits)
do_wc(input_splits)
ray.flush_log()
first_block = ray.get(self.objectids[(0,) * self.ndim])
dtype = first_block.dtype
result = np.zeros(self.shape, dtype=dtype)
for index in np.ndindex(*self.num_blocks):
lower = DistArray.compute_block_lower(index, self.shape)
upper = DistArray.compute_block_upper(index, self.shape)
result[[slice(l, u) for (l, u) in zip(lower, upper)]] = ray.get(self.objectids[index])
return result
def __getitem__(self, sliced):
# TODO(rkn): fix this, this is just a placeholder that should work but is inefficient
a = self.assemble()
return a[sliced]
# Register the DistArray class with Ray so that it knows how to serialize it.
ray.register_class(DistArray)
@ray.remote
def assemble(a):
return a.assemble()
# TODO(rkn): what should we call this method
@ray.remote
def numpy_to_dist(a):
result = DistArray(a.shape)
for index in np.ndindex(*result.num_blocks):
lower = DistArray.compute_block_lower(index, a.shape)
upper = DistArray.compute_block_upper(index, a.shape)
result.objectids[index] = ray.put(a[[slice(l, u) for (l, u) in zip(lower, upper)]])
return result
for index in np.ndindex(*self.num_blocks):
lower = DistArray.compute_block_lower(index, self.shape)
upper = DistArray.compute_block_upper(index, self.shape)
result[[slice(l, u) for (l, u) in zip(lower, upper)]] = ray.get(
self.objectids[index])
return result
def __getitem__(self, sliced):
# TODO(rkn): Fix this, this is just a placeholder that should work but is
# inefficient.
a = self.assemble()
return a[sliced]
# Register the DistArray class with Ray so that it knows how to serialize it.
ray.register_class(DistArray)
@ray.remote
def assemble(a):
return a.assemble()
# TODO(rkn): What should we call this method?
@ray.remote
def numpy_to_dist(a):
result = DistArray(a.shape)
for index in np.ndindex(*result.num_blocks):
lower = DistArray.compute_block_lower(index, a.shape)
upper = DistArray.compute_block_upper(index, a.shape)
result.objectids[index] = ray.put(a[[slice(l, u) for (l, u)
def testMethods(self):
for module in [
ra.core, ra.random, ra.linalg, da.core, da.random, da.linalg
]:
reload(module)
ray.init(start_ray_local=True, num_objstores=2, num_workers=10)
ray.register_class(da.DistArray)
x = da.zeros.remote([9, 25, 51], "float")
assert_equal(ray.get(da.assemble.remote(x)), np.zeros([9, 25, 51]))
x = da.ones.remote([11, 25, 49], dtype_name="float")
assert_equal(ray.get(da.assemble.remote(x)), np.ones([11, 25, 49]))
x = da.random.normal.remote([11, 25, 49])
y = da.copy.remote(x)
assert_equal(ray.get(da.assemble.remote(x)),
ray.get(da.assemble.remote(y)))
x = da.eye.remote(25, dtype_name="float")
assert_equal(ray.get(da.assemble.remote(x)), np.eye(25))
x = da.random.normal.remote([25, 49])
y = da.triu.remote(x)
assert_equal(ray.get(da.assemble.remote(y)),
np.triu(ray.get(da.assemble.remote(x))))
x = da.random.normal.remote([25, 49])
y = da.tril.remote(x)
assert_equal(ray.get(da.assemble.remote(y)),
np.tril(ray.get(da.assemble.remote(x))))
x = da.random.normal.remote([25, 49])
y = da.random.normal.remote([49, 18])
z = da.dot.remote(x, y)
w = da.assemble.remote(z)
u = da.assemble.remote(x)
v = da.assemble.remote(y)
assert_almost_equal(ray.get(w), np.dot(ray.get(u), ray.get(v)))
assert_almost_equal(ray.get(w), np.dot(ray.get(u), ray.get(v)))
# test add
x = da.random.normal.remote([23, 42])
y = da.random.normal.remote([23, 42])
z = da.add.remote(x, y)
assert_almost_equal(
ray.get(da.assemble.remote(z)),
ray.get(da.assemble.remote(x)) + ray.get(da.assemble.remote(y)))
# test subtract
x = da.random.normal.remote([33, 40])
y = da.random.normal.remote([33, 40])
z = da.subtract.remote(x, y)
assert_almost_equal(
ray.get(da.assemble.remote(z)),
ray.get(da.assemble.remote(x)) - ray.get(da.assemble.remote(y)))
# test transpose
x = da.random.normal.remote([234, 432])
y = da.transpose.remote(x)
assert_equal(
ray.get(da.assemble.remote(x)).T, ray.get(da.assemble.remote(y)))
# test numpy_to_dist
x = da.random.normal.remote([23, 45])
y = da.assemble.remote(x)
z = da.numpy_to_dist.remote(y)
w = da.assemble.remote(z)
assert_equal(ray.get(da.assemble.remote(x)),
ray.get(da.assemble.remote(z)))
assert_equal(ray.get(y), ray.get(w))
# test da.tsqr
for shape in [[123, da.BLOCK_SIZE], [7, da.BLOCK_SIZE],
[da.BLOCK_SIZE, da.BLOCK_SIZE], [da.BLOCK_SIZE, 7],
[10 * da.BLOCK_SIZE, da.BLOCK_SIZE]]:
x = da.random.normal.remote(shape)
K = min(shape)
q, r = da.linalg.tsqr.remote(x)
x_val = ray.get(da.assemble.remote(x))
q_val = ray.get(da.assemble.remote(q))
r_val = ray.get(r)
self.assertTrue(r_val.shape == (K, shape[1]))
assert_equal(r_val, np.triu(r_val))
assert_almost_equal(x_val, np.dot(q_val, r_val))
assert_almost_equal(np.dot(q_val.T, q_val), np.eye(K))
# test da.linalg.modified_lu
def test_modified_lu(d1, d2):
print "testing dist_modified_lu with d1 = " + str(
d1) + ", d2 = " + str(d2)
assert d1 >= d2
k = min(d1, d2)
m = ra.random.normal.remote([d1, d2])
q, r = ra.linalg.qr.remote(m)
l, u, s = da.linalg.modified_lu.remote(da.numpy_to_dist.remote(q))
q_val = ray.get(q)
r_val = ray.get(r)
l_val = ray.get(da.assemble.remote(l))
u_val = ray.get(u)
s_val = ray.get(s)
s_mat = np.zeros((d1, d2))
for i in range(len(s_val)):
s_mat[i, i] = s_val[i]
assert_almost_equal(q_val - s_mat,
np.dot(l_val,
u_val)) # check that q - s = l * u
assert_equal(np.triu(u_val),
u_val) # check that u is upper triangular
assert_equal(np.tril(l_val),
l_val) # check that l is lower triangular
for d1, d2 in [(100, 100), (99, 98), (7, 5), (7, 7), (20, 7),
(20, 10)]:
test_modified_lu(d1, d2)
# test dist_tsqr_hr
def test_dist_tsqr_hr(d1, d2):
print "testing dist_tsqr_hr with d1 = " + str(
d1) + ", d2 = " + str(d2)
a = da.random.normal.remote([d1, d2])
y, t, y_top, r = da.linalg.tsqr_hr.remote(a)
a_val = ray.get(da.assemble.remote(a))
y_val = ray.get(da.assemble.remote(y))
t_val = ray.get(t)
y_top_val = ray.get(y_top)
r_val = ray.get(r)
tall_eye = np.zeros((d1, min(d1, d2)))
np.fill_diagonal(tall_eye, 1)
q = tall_eye - np.dot(y_val, np.dot(t_val, y_top_val.T))
assert_almost_equal(np.dot(q.T, q),
np.eye(min(d1, d2))) # check that q.T * q = I
assert_almost_equal(np.dot(
q, r_val), a_val) # check that a = (I - y * t * y_top.T) * r
for d1, d2 in [(123, da.BLOCK_SIZE), (7, da.BLOCK_SIZE),
(da.BLOCK_SIZE, da.BLOCK_SIZE), (da.BLOCK_SIZE, 7),
(10 * da.BLOCK_SIZE, da.BLOCK_SIZE)]:
test_dist_tsqr_hr(d1, d2)
def test_dist_qr(d1, d2):
print "testing qr with d1 = {}, and d2 = {}.".format(d1, d2)
a = da.random.normal.remote([d1, d2])
K = min(d1, d2)
q, r = da.linalg.qr.remote(a)
a_val = ray.get(da.assemble.remote(a))
q_val = ray.get(da.assemble.remote(q))
r_val = ray.get(da.assemble.remote(r))
self.assertEqual(q_val.shape, (d1, K))
self.assertEqual(r_val.shape, (K, d2))
assert_almost_equal(np.dot(q_val.T, q_val), np.eye(K))
assert_equal(r_val, np.triu(r_val))
assert_almost_equal(a_val, np.dot(q_val, r_val))
for d1, d2 in [(123, da.BLOCK_SIZE), (7, da.BLOCK_SIZE),
(da.BLOCK_SIZE, da.BLOCK_SIZE), (da.BLOCK_SIZE, 7),
(13, 21), (34, 35), (8, 7)]:
test_dist_qr(d1, d2)
test_dist_qr(d2, d1)
for _ in range(20):
d1 = np.random.randint(1, 35)
d2 = np.random.randint(1, 35)
test_dist_qr(d1, d2)
ray.worker.cleanup()
class BlockMatrix(object):
def __init__(self, shape, block_size=10):
assert len(shape) == 2
self.shape = shape
self.block_size = block_size
self.num_blocks = [
int(np.ceil(a / self.block_size)) for a in self.shape
]
# The field "block_ids" will be a numpy array of object IDs, where each
# object ID refers to a numpy array.
self.block_ids = np.empty(self.num_blocks, dtype=object)
# Note that we need to call ray.register_class so that we can pass BlockMatrix
# objects to remote functions and return them from remote functions.
ray.register_class(BlockMatrix)
# This is a helper function which creates an array of zeros.
@ray.remote
def zeros_helper(shape):
return np.zeros(shape)
# EXERCISE: Define a remote function which returns a BlockMatrix of zeros. You
# can assume that len(shape) == 2. You can do this as follows.
# - First call the BlockMatrix constructor with the appropriate shape.
# - Then call "zeros_helper" to create an object ID for each entry of
# the "block_ids" field of the BlockMatrix.
#
# NOTE: You could call "zeros_helper" once for every block in the BlockMatrix,
parser = argparse.ArgumentParser(
description="Run the policy gradient algorithm.")
parser.add_argument("--environment",
default="Pong-v0",
type=str,
help="The gym environment to use.")
parser.add_argument("--redis-address",
default=None,
type=str,
help="The Redis address of the cluster.")
args = parser.parse_args()
ray.init(redis_address=args.redis_address)
ray.register_class(AtariRamPreprocessor)
ray.register_class(AtariPixelPreprocessor)
ray.register_class(NoPreprocessor)
mdp_name = args.environment
if args.environment == "Pong-v0":
preprocessor = AtariPixelPreprocessor()
elif mdp_name == "Pong-ram-v3":
preprocessor = AtariRamPreprocessor()
elif mdp_name == "CartPole-v0":
preprocessor = NoPreprocessor()
else:
print("No environment was chosen, so defaulting to Pong-v0.")
mdp_name = "Pong-v0"
preprocessor = AtariPixelPreprocessor()