def testMirroredStratParaAsync(self):
"""Tests RNG/MirrorStrategy interaction #3.
The user can create n independent RNGs outside strategy.scope(), where n
is the number of replicas, and give one to each replica. The replicas can
thus get different random-number streams.
"""
shape = [3, 4]
dtype = dtypes.int32
gens = random.get_global_generator().split(count=2)
devices = ["/cpu:0", test_util.gpu_device_name()]
strat = MirroredStrategy(devices=devices)
# Use `PerReplica` to specify which `gen` is sent to which replica
gens = dist_values.PerReplica(
device_map=dist_values.ReplicaDeviceMap(devices),
values=[[g] for g in gens])
with strat.scope():
def f(gen):
t1 = gen.uniform_full_int(shape=shape, dtype=dtype)
t2 = gen.uniform_full_int(shape=shape, dtype=dtype)
t = array_ops.stack([t1, t2])
return t
results = strat.extended.call_for_each_replica(
fn=f, args=gens)
values = results.values
self.assertAllEqual(2, len(values))
self.assertAllDifferent(values)
def testSameAsOldRandomOps(self):
"""Tests that the generated numbers are the same as the old random_ops.py .
"""
seed1, seed2 = 50, 60
# note how the two seeds for the old op correspond to the seed for the new
# op
random.get_global_generator().reset([0, seed2, seed1])
shape = constant_op.constant([2, 3])
dtype = dtypes.float32
# create a graph for the old op in order to call it many times
@def_function.function
def old():
return gen_random_ops.random_standard_normal(
shape, dtype=dtype, seed=seed1, seed2=seed2)
def new():
return random.get_global_generator().standard_normal(shape, dtype=dtype)
for _ in range(100):
self.assertAllEqual(old(), new())
def compare(fst_includes_print, snd_includes_print): random.get_global_generator().reset(50) fst = f(fst_includes_print) random.get_global_generator().reset(50) snd = f(snd_includes_print) self.assertAllEqual(fst, snd) # Now do the above again using accelerated (defunned) 'f'. # Running 'f' with two different Boolean arguments should cause # two different graphs to be generated, hence demonstrating the # insensitivity to graph changes. f_acc = def_function.function(f) random.get_global_generator().reset(50) fst = f_acc(fst_includes_print) random.get_global_generator().reset(50) snd = f_acc(snd_includes_print) self.assertAllEqual(fst, snd)
def testOpSeedSelectionAfterSetSeed(self):
"""Tests that op-seed selection is reset after reseting global generator.
Fixing GitHub issue 9171:
https://github.com/tensorflow/tensorflow/issues/9171
"""
shape = (3,)
random.get_global_generator().reset(1)
a = random.get_global_generator().normal(shape)
random.get_global_generator().reset(1)
b = random.get_global_generator().normal(shape)
self.assertAllEqual(a, b)
# Now do the above again using accelerated ('defun'ed) computation
@def_function.function
def f():
return random.get_global_generator().normal(shape)
random.get_global_generator().reset(1)
c = f()
random.get_global_generator().reset(1)
d = f()
self.assertAllEqual(c, d)
self.assertAllEqual(a, c)
def verify(counter1, counter2, key1, key2, expect1, expect2):
counter = uint32s_to_uint64(counter1, counter2)
key = uint32s_to_uint64(key1, key2)
random.get_global_generator().reset([counter, key])
got = random.get_global_generator().uniform_full_int(
shape=(2,), dtype=dtypes.uint32)
expect = [expect1, expect2]
self.assertAllEqual(expect, got)
random.get_global_generator().reset([counter, key])
got = random.get_global_generator().uniform_full_int(
shape=(), dtype=dtypes.uint64)
self.assertAllEqual(uint32s_to_uint64(*expect), got)
def f(): return random.get_global_generator().normal(shape)
def f(include_print):
shape = constant_op.constant([5])
if include_print:
shape = logging_ops.Print(shape, [shape])
return random.get_global_generator().normal(shape)
def testEagerAndDefun(self): """A simple test to make sure the op works in eager and defunned mode.""" random.get_global_generator().normal((3,))
def f(include_print):
shape = constant_op.constant([5])
if include_print:
shape = logging_ops.Print(shape, [shape])
return random.get_global_generator().normal(shape)
def f():
return random.get_global_generator().normal(shape)
def testEagerAndDefun(self):
"""A simple test to make sure the op works in eager and defunned mode."""
random.get_global_generator().normal((3, ))
def new(): return random.get_global_generator().standard_normal(shape, dtype=dtype)
def make_seed(): generator = random.get_global_generator() state = array_ops.identity(generator.state, name="state") return generator.uniform_full_int((2,), dtypes.int32, name="seed"), state
def new():
with ops.device(test_util.gpu_device_name()):
return random.get_global_generator().standard_normal(
shape, dtype=dtype)
def new():
with ops.device("/device:CPU:0"):
return random.get_global_generator().standard_normal(
shape, dtype=dtype)
def new():
return random.get_global_generator().standard_normal(shape,
dtype=dtype)
