def test_validating_dataset_input_tensors_with_dtype_mismatch(self):
with self.cached_session():
strategy = mirrored_strategy.MirroredStrategy(
['/device:GPU:0', '/device:CPU:0'])
a = constant_op.constant([1, 2], shape=(1, 2), dtype=dtypes.int32)
b = constant_op.constant([1, 2],
shape=(1, 2),
dtype=dtypes.float64)
x = values.DistributedValues({
'/device:CPU:0': a,
'/device:GPU:0': b
})
y = values.DistributedValues({
'/device:CPU:0': a,
'/device:GPU:0': a
})
with strategy.scope():
# Removed device and input tensor dtype details from the error message
# since the order of the device and the corresponding input tensor dtype
# is not deterministic over different runs.
with self.assertRaisesRegexp(
ValueError, 'Input tensor dtypes do not match for '
'distributed tensor inputs '
'DistributedValues:.+'):
distributed_training_utils.validate_distributed_dataset_inputs(
strategy, x, y)
def testCanonicalization(self):
canonical_cpu = ["/job:localhost/replica:0/task:0/device:CPU:0"]
v = values.DistributedValues({"": 42})
self.assertEqual(canonical_cpu, list(v._index.keys()))
v = values.DistributedValues({"/device:CPU:0": 42})
self.assertEqual(canonical_cpu, list(v._index.keys()))
v = values.DistributedValues({"/cpu:0": 42})
self.assertEqual(canonical_cpu, list(v._index.keys()))
v = values.DistributedValues({"/CPU:0": 42})
self.assertEqual(canonical_cpu, list(v._index.keys()))
with self.assertRaises(AssertionError):
v = values.DistributedValues({"/device:cpu:0": 42})
def testGetEager(self):
with ops.device("/device:CPU:0"):
one = constant_op.constant(1)
two = constant_op.constant(2)
v = values.DistributedValues({"/device:CPU:0": one, "/device:GPU:0": two})
self.assertEqual(two, v.get("/device:GPU:0"))
self.assertEqual(one, v.get())
with self.assertRaises(ValueError):
self.assertIsNone(v.get("/device:GPU:2"))
def testGetGraph(self):
with context.graph_mode(), \
ops.Graph().as_default(), \
ops.device("/device:CPU:0"):
one = constant_op.constant(1)
two = constant_op.constant(2)
v = values.DistributedValues({"/device:CPU:0": one, "/device:GPU:0": two})
self.assertEqual(two, v.get("/device:GPU:0"))
self.assertEqual(one, v.get())
with self.assertRaises(ValueError):
self.assertIsNone(v.get("/device:GPU:2"))
def testIsTensorLikeWithAConstant(self):
with context.graph_mode(), \
ops.Graph().as_default(), \
ops.device("/device:CPU:0"):
one = constant_op.constant(1)
two = 2.0
v = values.DistributedValues({"/device:CPU:0": one, "/device:GPU:0": two})
self.assertEqual(two, v.get("/device:GPU:0"))
self.assertEqual(one, v.get())
self.assertFalse(v.is_tensor_like)
self.assertFalse(tensor_util.is_tensor(v))
def testNonMatchingVariableCreation(self):
self._skip_eager_if_gpus_less_than(1)
def model_fn(name):
v = variable_scope.variable(1.0, name=name)
distribute_lib.get_tower_context().merge_call(lambda _: _)
return v
dist = mirrored_strategy.MirroredStrategy(
["/device:GPU:0", "/device:CPU:0"])
with dist.scope():
names = values.DistributedValues({
"/device:CPU:0": "foo",
"/device:GPU:0": "bar"
})
with self.assertRaises(RuntimeError):
_ = dist.call_for_each_tower(model_fn, names, run_concurrently=False)
