def _singleMetaGraphSavedModel(self):
export_graph = ops.Graph()
with export_graph.as_default():
start = array_ops.placeholder(shape=[1, 1],
dtype=dtypes.float32,
name="start")
distractor = variables.RefVariable(-1., name="distractor")
v = variables.RefVariable(3., name="v")
local_variable = variables.VariableV1(
1.,
collections=[ops.GraphKeys.LOCAL_VARIABLES],
trainable=False,
use_resource=True)
output = array_ops.identity(start * v * local_variable,
name="output")
with session_lib.Session() as session:
session.run([
v.initializer, distractor.initializer,
local_variable.initializer
])
path = os.path.join(self.get_temp_dir(), "saved_model",
str(ops.uid()))
simple_save.simple_save(
session,
path,
inputs={"start": start},
outputs={"output": output},
legacy_init_op=local_variable.initializer)
return path
def _v1_single_metagraph_saved_model(self, use_resource):
export_graph = ops.Graph()
with export_graph.as_default():
start = array_ops.placeholder(
shape=None, dtype=dtypes.float32, name="start")
if use_resource:
distractor = variables.RefVariable(-1., name="distractor")
v = resource_variable_ops.ResourceVariable(3., name="v")
else:
# "distractor" gets saved in the checkpoint and so used in the restore
# function, but not in the pruned function for the signature. This tests
# node naming: it needs to be consistent (and ideally always the same as
# the node in the original GraphDef) for the resource manager to find
# the right variable.
distractor = variables.RefVariable(-1., name="distractor")
v = variables.RefVariable(3., name="v")
local_variable = variables.VariableV1(
1.,
collections=[ops.GraphKeys.LOCAL_VARIABLES],
trainable=False,
use_resource=True)
output = array_ops.identity(start * v * local_variable, name="output")
with session_lib.Session() as session:
session.run([v.initializer, distractor.initializer,
local_variable.initializer])
path = os.path.join(self.get_temp_dir(), "saved_model", str(ops.uid()))
simple_save.simple_save(
session,
path,
inputs={"start": start},
outputs={"output": output},
legacy_init_op=local_variable.initializer)
return path
def doTestSparse(self, use_resource=False):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.cached_session():
# Initialize variables for numpy implementation.
m0, v0, m1, v1 = 0.0, 0.0, 0.0, 0.0
var0_np = np.array([1.0, 2.0], dtype=dtype.as_numpy_dtype)
grads0_np = np.array([0.1, 0.1], dtype=dtype.as_numpy_dtype)
var1_np = np.array([3.0, 4.0], dtype=dtype.as_numpy_dtype)
grads1_np = np.array([0.01, 0.01], dtype=dtype.as_numpy_dtype)
if use_resource:
var0 = resource_variable_ops.ResourceVariable(var0_np)
var1 = resource_variable_ops.ResourceVariable(var1_np)
else:
var0 = variables.RefVariable(var0_np)
var1 = variables.RefVariable(var1_np)
grads0_np_indices = np.array([0, 1], dtype=np.int32)
grads0 = ops.IndexedSlices(
constant_op.constant(grads0_np),
constant_op.constant(grads0_np_indices),
constant_op.constant([2]))
grads1_np_indices = np.array([0, 1], dtype=np.int32)
grads1 = ops.IndexedSlices(
constant_op.constant(grads1_np),
constant_op.constant(grads1_np_indices),
constant_op.constant([2]))
opt = adam.AdamOptimizer()
update = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]))
variables.global_variables_initializer().run()
# Fetch params to validate initial values
self.assertAllClose([1.0, 2.0], self.evaluate(var0))
self.assertAllClose([3.0, 4.0], self.evaluate(var1))
beta1_power, beta2_power = opt._get_beta_accumulators()
# Run 3 steps of Adam
for t in range(1, 4):
self.assertAllCloseAccordingToType(
0.9**t, self.evaluate(beta1_power))
self.assertAllCloseAccordingToType(
0.999**t, self.evaluate(beta2_power))
update.run()
var0_np, m0, v0 = adam_update_numpy(
var0_np, grads0_np, t, m0, v0)
var1_np, m1, v1 = adam_update_numpy(
var1_np, grads1_np, t, m1, v1)
# Validate updated params
self.assertAllCloseAccordingToType(var0_np,
self.evaluate(var0))
self.assertAllCloseAccordingToType(var1_np,
self.evaluate(var1))
def Rtt_default_variable_creator(next_creator=None, **kwargs):
"""Default variable creator."""
assert next_creator is None
initial_value = kwargs.get("initial_value", None)
trainable = kwargs.get("trainable", None)
collections = kwargs.get("collections", None)
validate_shape = kwargs.get("validate_shape", True)
caching_device = kwargs.get("caching_device", None)
name = kwargs.get("name", None)
variable_def = kwargs.get("variable_def", None)
dtype = kwargs.get("dtype", None)
expected_shape = kwargs.get("expected_shape", None)
import_scope = kwargs.get("import_scope", None)
constraint = kwargs.get("constraint", None)
use_resource = kwargs.get("use_resource", None)
synchronization = kwargs.get("synchronization", None)
aggregation = kwargs.get("aggregation", None)
shape = kwargs.get("shape", None)
initial_value = convert_init_value_to_string(initial_value, dtype)
if use_resource is None:
use_resource = variable_scope.get_variable_scope().use_resource
if use_resource is None:
use_resource = variable_scope._DEFAULT_USE_RESOURCE
use_resource = use_resource or context.executing_eagerly()
if use_resource:
distribute_strategy = kwargs.get("distribute_strategy", None)
return rtt_ts.convert_to_rtttensor(
resource_variable_ops.ResourceVariable(
initial_value=initial_value,
trainable=trainable,
collections=collections,
validate_shape=validate_shape,
caching_device=caching_device,
name=name,
dtype=dtype,
constraint=constraint,
variable_def=variable_def,
import_scope=import_scope,
distribute_strategy=distribute_strategy,
synchronization=synchronization,
aggregation=aggregation,
shape=shape))
else:
return rtt_ts.convert_to_rtttensor(
variables.RefVariable(initial_value=initial_value,
trainable=trainable,
collections=collections,
validate_shape=validate_shape,
caching_device=caching_device,
name=name,
dtype=dtype,
constraint=constraint,
variable_def=variable_def,
expected_shape=expected_shape,
import_scope=import_scope,
synchronization=synchronization,
aggregation=aggregation,
shape=shape))
def create_host_embedding(name,
shape,
dtype,
partition_strategy="TOKEN",
optimizer_spec=None,
initializer=None):
""" Create a HostEmbedding.
Args:
name: The name which uniquely identifies the embedding.
shape: The shape for the tensor which will hold the embedding.
dtype: The dtype for the tensor which will hold the embedding.
partition_strategy: When
`enable_experimental_remote_buffer_embedding` is `True` and using
replication, the embedding must be distributed across the replicas.
This option decides on which axis the embedding will be split. Options
are "TOKEN" or "ENCODING".
optimizer_spec: A description of how the embedding will be optimized.
When `None`, the embedding is assumed to not be trainable.
initializer: The initializer to use when creating the embedding tensor.
Returns:
A `HostEmbedding` object that wraps the created embedding tensor.
"""
if initializer is None:
initializer = array_ops.zeros(shape, dtype)
with ops.device('cpu'):
embedding_tensor = variables.RefVariable(initial_value=initializer,
name=name)
return HostEmbedding(name,
embedding_tensor,
partition_strategy=partition_strategy,
optimizer_spec=optimizer_spec)
def testGatherNdRefVariable(self):
with self.cached_session():
v = variables.RefVariable(constant_op.constant([[1, 2], [3, 4], [5, 6]]))
self.evaluate(variables.global_variables_initializer())
gather = array_ops.gather_nd(v, [[0, 1], [2, 0]])
if not context.executing_eagerly(): # .op doesn't make sense in Eager
self.assertEqual("GatherNd", gather.op.name)
self.assertAllEqual([2, 5], gather)
def doTestBasic(self, use_resource=False, use_callable_params=False):
if context.executing_eagerly() and not use_resource:
self.skipTest(
"Skipping test with use_resource=False and executing eagerly.")
for i, dtype in enumerate(
[dtypes.half, dtypes.float32, dtypes.float64]):
with self.session(graph=ops.Graph()):
# Initialize variables for numpy implementation.
m0, v0, m1, v1 = 0.0, 0.0, 0.0, 0.0
var0_np = np.array([1.0, 2.0], dtype=dtype.as_numpy_dtype)
grads0_np = np.array([0.1, 0.1], dtype=dtype.as_numpy_dtype)
var1_np = np.array([3.0, 4.0], dtype=dtype.as_numpy_dtype)
grads1_np = np.array([0.01, 0.01], dtype=dtype.as_numpy_dtype)
if use_resource:
var0 = resource_variable_ops.ResourceVariable(
var0_np, name="var0_%d" % i)
var1 = resource_variable_ops.ResourceVariable(
var1_np, name="var1_%d" % i)
else:
var0 = variables.RefVariable(var0_np)
var1 = variables.RefVariable(var1_np)
grads0 = constant_op.constant(grads0_np)
grads1 = constant_op.constant(grads1_np)
learning_rate = lambda: 0.001
beta1 = lambda: 0.9
beta2 = lambda: 0.999
epsilon = lambda: 1e-8
if not use_callable_params:
learning_rate = learning_rate()
beta1 = beta1()
beta2 = beta2()
epsilon = epsilon()
opt = adam.AdamOptimizer(learning_rate=learning_rate)
update = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]))
opt_variables = opt.variables()
beta1_power, beta2_power = opt._get_beta_accumulators()
self.assertTrue(beta1_power is not None)
self.assertTrue(beta2_power is not None)
self.assertIn(beta1_power, opt_variables)
self.assertIn(beta2_power, opt_variables)
# Ensure that non-slot variables are the same type as the requested
# variables.
self.assertEqual(
use_resource,
resource_variable_ops.is_resource_variable(beta1_power))
self.assertEqual(
use_resource,
resource_variable_ops.is_resource_variable(beta2_power))
if not context.executing_eagerly():
with ops.Graph().as_default():
# Shouldn't return non-slot variables from other graphs.
self.assertEqual(0, len(opt.variables()))
self.evaluate(variables.global_variables_initializer())
# Fetch params to validate initial values
self.assertAllClose([1.0, 2.0], self.evaluate(var0))
self.assertAllClose([3.0, 4.0], self.evaluate(var1))
beta1_power, beta2_power = opt._get_beta_accumulators()
# Run 3 steps of Adam
for t in range(1, 4):
if not context.executing_eagerly():
self.evaluate(update)
elif t > 1:
opt.apply_gradients(zip([grads0, grads1],
[var0, var1]))
self.assertAllCloseAccordingToType(
0.9**(t + 1), self.evaluate(beta1_power))
self.assertAllCloseAccordingToType(
0.999**(t + 1), self.evaluate(beta2_power))
var0_np, m0, v0 = adam_update_numpy(
var0_np, grads0_np, t, m0, v0)
var1_np, m1, v1 = adam_update_numpy(
var1_np, grads1_np, t, m1, v1)
# Validate updated params
self.assertAllCloseAccordingToType(var0_np,
self.evaluate(var0))
self.assertAllCloseAccordingToType(var1_np,
self.evaluate(var1))
if use_resource:
self.assertEqual("var0_%d/Adam:0" % (i, ),
opt.get_slot(var=var0, name="m").name)
