def testManyRestoresGraph(self):
"""Restores after the first should not modify the graph."""
with context.graph_mode():
graph = ops.Graph()
with graph.as_default(), self.session(graph):
checkpoint_directory = self.get_temp_dir()
checkpoint_prefix = os.path.join(checkpoint_directory, "ckpt")
obj = util.Checkpoint()
obj.var = variable_scope.get_variable(name="v", initializer=0.)
obj.opt = adam.AdamOptimizer(0.1)
obj.opt.minimize(obj.var.read_value())
self.evaluate(util.gather_initializers(obj))
save_path = obj.save(checkpoint_prefix)
obj.restore(save_path)
before_ops = graph.get_operations()
obj.restore(save_path)
self.assertEqual(before_ops, graph.get_operations())
def testSharing(self):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.test_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)
var0 = variables.Variable(var0_np)
var1 = variables.Variable(var1_np)
grads0 = constant_op.constant(grads0_np)
grads1 = constant_op.constant(grads1_np)
opt = adam.AdamOptimizer()
update1 = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]))
update2 = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]))
variables.global_variables_initializer().run()
beta1_power, beta2_power = opt._get_beta_accumulators()
# Fetch params to validate initial values
self.assertAllClose([1.0, 2.0], var0.eval())
self.assertAllClose([3.0, 4.0], var1.eval())
# Run 3 steps of intertwined Adam1 and Adam2.
for t in range(1, 4):
self.assertAllCloseAccordingToType(0.9**t,
beta1_power.eval())
self.assertAllCloseAccordingToType(0.999**t,
beta2_power.eval())
if t % 2 == 0:
update1.run()
else:
update2.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, var0.eval())
self.assertAllCloseAccordingToType(var1_np, var1.eval())
def testTwoSessions(self):
optimizer = adam.AdamOptimizer()
g = ops.Graph()
with g.as_default():
with session.Session():
var0 = variables.Variable(np.array([1.0, 2.0]), name="v0")
grads0 = constant_op.constant(np.array([0.1, 0.1]))
optimizer.apply_gradients([(grads0, var0)])
gg = ops.Graph()
with gg.as_default():
with session.Session():
var0 = variables.Variable(np.array([1.0, 2.0]), name="v0")
grads0 = constant_op.constant(np.array([0.1, 0.1]))
# If the optimizer saves any state not keyed by graph the following line
# fails.
optimizer.apply_gradients([(grads0, var0)])
def testWithDefun(self):
num_training_steps = 2
checkpoint_directory = self.get_temp_dir()
checkpoint_prefix = os.path.join(checkpoint_directory, "ckpt")
for training_continuation in range(3):
with ops.Graph().as_default(), self.test_session(
graph=ops.get_default_graph()), test_util.device(use_gpu=True):
model = MyModel()
# Don't actually train so we can test variable values
optimizer = adam.AdamOptimizer(0.)
root = util.Checkpoint(
optimizer=optimizer, model=model,
global_step=training_util.get_or_create_global_step())
checkpoint_path = checkpoint_management.latest_checkpoint(
checkpoint_directory)
status = root.restore(save_path=checkpoint_path)
def train_fn():
@function.defun
def _call_model(x):
return model(x)
with backprop.GradientTape() as tape:
loss = _call_model(constant_op.constant([[3.]]))
gradients = tape.gradient(loss, model.variables)
return optimizer.apply_gradients(zip(gradients, model.variables),
global_step=root.global_step)
if not context.executing_eagerly():
train_fn = functools.partial(
self.evaluate, train_fn())
status.initialize_or_restore()
for _ in range(num_training_steps):
train_fn()
if training_continuation > 0:
status.assert_consumed()
self.assertAllClose([[42.]], self.evaluate(model.variables[0]))
else:
self.evaluate(model.variables[0].assign([[42.]]))
root.save(file_prefix=checkpoint_prefix)
self.assertEqual((training_continuation + 1) * num_training_steps,
self.evaluate(root.global_step))
self.assertEqual(training_continuation + 1,
self.evaluate(root.save_counter))
def _initialized_model(self):
input_value = constant_op.constant([[3.]])
model = MyModel()
optimizer = adam.AdamOptimizer(0.001)
optimizer_step = training_util.get_or_create_global_step()
root_trackable = util.Checkpoint(
optimizer=optimizer, model=model, optimizer_step=optimizer_step)
train_op = optimizer.minimize(
functools.partial(model, input_value),
global_step=optimizer_step)
self.evaluate(util.gather_initializers(
root_trackable))
self.evaluate(train_op)
# A regular variable, a slot variable, and a non-slot Optimizer variable
# with known values to check when loading.
self.evaluate(model._named_dense.bias.assign([1.]))
self.evaluate(optimizer.get_slot(
var=model._named_dense.bias, name="m").assign([2.]))
beta_1_power, _ = optimizer._get_beta_accumulators()
self.evaluate(beta_1_power.assign(3.))
return root_trackable
def testDeferredRestorationUsageEager(self):
"""An idiomatic eager execution example."""
num_training_steps = 10
checkpoint_directory = self.get_temp_dir()
checkpoint_prefix = os.path.join(checkpoint_directory, "ckpt")
for training_continuation in range(3):
model = MyModel()
optimizer = adam.AdamOptimizer(0.001)
root = checkpointable_utils.Checkpoint(
optimizer=optimizer, model=model,
optimizer_step=training_util.get_or_create_global_step())
root.restore(core_saver.latest_checkpoint(checkpoint_directory))
for _ in range(num_training_steps):
# TODO(allenl): Use a Dataset and serialize/checkpoint it.
input_value = constant_op.constant([[3.]])
optimizer.minimize(
lambda: model(input_value), # pylint: disable=cell-var-from-loop
global_step=root.optimizer_step)
root.save(file_prefix=checkpoint_prefix)
self.assertEqual((training_continuation + 1) * num_training_steps,
root.optimizer_step.numpy())
def testUsageGraph(self):
"""Expected usage when graph building."""
with context.graph_mode():
num_training_steps = 10
checkpoint_directory = self.get_temp_dir()
checkpoint_prefix = os.path.join(checkpoint_directory, "ckpt")
for training_continuation in range(3):
with ops.Graph().as_default():
model = MyModel()
optimizer = adam.AdamOptimizer(0.001)
root = checkpointable_utils.Checkpoint(
optimizer=optimizer,
model=model,
global_step=training_util.get_or_create_global_step())
input_value = constant_op.constant([[3.]])
train_op = optimizer.minimize(model(input_value),
global_step=root.global_step)
checkpoint_path = core_saver.latest_checkpoint(
checkpoint_directory)
with self.test_session(
graph=ops.get_default_graph()) as session:
status = root.restore(save_path=checkpoint_path)
status.initialize_or_restore(session=session)
if checkpoint_path is None:
self.assertEqual(0, training_continuation)
with self.assertRaises(AssertionError):
status.assert_consumed()
else:
status.assert_consumed()
for _ in range(num_training_steps):
session.run(train_op)
root.save(file_prefix=checkpoint_prefix,
session=session)
self.assertEqual(
(training_continuation + 1) * num_training_steps,
session.run(root.global_step))
self.assertEqual(training_continuation + 1,
session.run(root.save_counter))
def testNamingWithOptimizer(self):
input_value = constant_op.constant([[3.]])
model = MyModel()
# A nuisance Model using the same optimizer. Its slot variables should not
# go in the checkpoint, since it is never depended on.
other_model = MyModel()
optimizer = adam.AdamOptimizer(0.001)
optimizer_step = training_util.get_or_create_global_step()
root_trackable = util.Checkpoint(
optimizer=optimizer, model=model, optimizer_step=optimizer_step)
if context.executing_eagerly():
optimizer.minimize(
lambda: model(input_value),
global_step=optimizer_step)
optimizer.minimize(
lambda: other_model(input_value),
global_step=optimizer_step)
else:
train_op = optimizer.minimize(
model(input_value), global_step=optimizer_step)
optimizer.minimize(
other_model(input_value),
global_step=optimizer_step)
self.evaluate(util.gather_initializers(
root_trackable))
self.evaluate(train_op)
named_variables, serialized_graph, _ = graph_view.ObjectGraphView(
root_trackable).serialize_object_graph()
expected_checkpoint_names = (
# Created in the root node, so no prefix.
"optimizer_step",
"model/_second/kernel",
"model/_named_dense/kernel",
"model/_named_dense/bias",
# non-Layer dependency of the model
"model/_non_layer/a_variable",
# The optimizer creates two non-slot variables
"optimizer/beta1_power",
"optimizer/beta2_power",
# Slot variables
"model/_second/kernel/.OPTIMIZER_SLOT/optimizer/m",
"model/_second/kernel/.OPTIMIZER_SLOT/optimizer/v",
"model/_named_dense/kernel/.OPTIMIZER_SLOT/optimizer/m",
"model/_named_dense/kernel/.OPTIMIZER_SLOT/optimizer/v",
"model/_named_dense/bias/.OPTIMIZER_SLOT/optimizer/m",
"model/_named_dense/bias/.OPTIMIZER_SLOT/optimizer/v",
)
suffix = "/.ATTRIBUTES/VARIABLE_VALUE"
expected_checkpoint_names = [
name + suffix for name in expected_checkpoint_names]
# The optimizer and Dense layers also save get_config() JSON
expected_checkpoint_names.extend([
"model/_second/.ATTRIBUTES/OBJECT_CONFIG_JSON",
"model/_named_dense/.ATTRIBUTES/OBJECT_CONFIG_JSON"
])
named_variables = {v.name: v for v in named_variables}
six.assertCountEqual(self, expected_checkpoint_names,
named_variables.keys())
# Check that we've mapped to the right variable objects (not exhaustive)
self.assertEqual(
"global_step",
named_variables["optimizer_step" + suffix].full_name)
self.assertEqual(
"my_model/dense_1/kernel",
named_variables["model/_second/kernel" + suffix].full_name)
self.assertEqual(
"my_model/dense/kernel",
named_variables["model/_named_dense/kernel" + suffix].full_name)
self.assertEqual(
"beta1_power",
named_variables["optimizer/beta1_power" + suffix].full_name)
self.assertEqual(
"beta2_power",
named_variables["optimizer/beta2_power" + suffix].full_name)
# Spot check the generated protocol buffers.
self.assertEqual("optimizer",
serialized_graph.nodes[0].children[1].local_name)
optimizer_node = serialized_graph.nodes[serialized_graph.nodes[0].children[
1].node_id]
self.assertEqual("beta1_power", optimizer_node.children[0].local_name)
self.assertEqual(
"beta1_power", serialized_graph.nodes[optimizer_node.children[0]
.node_id].attributes[0].full_name)
self.assertEqual(
"my_model/dense/kernel",
serialized_graph.nodes[optimizer_node.slot_variables[0]
.original_variable_node_id]
.attributes[0].full_name)
# We strip off the :0 suffix, as variable.name-based saving does.
self.assertEqual(
"my_model/dense/kernel/Adam",
serialized_graph.nodes[optimizer_node.slot_variables[0]
.slot_variable_node_id]
.attributes[0].full_name)
self.assertEqual(
"my_model/dense/kernel/Adam:0",
optimizer.get_slot(
var=model._named_dense.kernel,
name="m").name)
self.assertEqual(
"model/_named_dense/kernel" + suffix,
serialized_graph.nodes[
optimizer_node.slot_variables[0]
.original_variable_node_id].attributes[0].checkpoint_key)
self.assertEqual("m", optimizer_node.slot_variables[0].slot_name)
self.assertEqual(
"model/_named_dense/kernel/.OPTIMIZER_SLOT/optimizer/m" + suffix,
serialized_graph.nodes[
optimizer_node.slot_variables[0]
.slot_variable_node_id].attributes[0].checkpoint_key)
def testSaveRestore(self):
model = MyModel()
optimizer = adam.AdamOptimizer(0.001)
root_trackable = util.Checkpoint(
optimizer=optimizer, model=model)
input_value = constant_op.constant([[3.]])
if context.executing_eagerly():
optimizer.minimize(
lambda: model(input_value))
else:
train_op = optimizer.minimize(model(input_value))
# TODO(allenl): Make initialization more pleasant when graph building.
root_trackable.save_counter # pylint: disable=pointless-statement
self.evaluate(util.gather_initializers(
root_trackable))
self.evaluate(train_op)
prefix = os.path.join(self.get_temp_dir(), "ckpt")
self.evaluate(state_ops.assign(model._named_dense.variables[1], [42.]))
m_bias_slot = optimizer.get_slot(model._named_dense.variables[1], "m")
self.evaluate(state_ops.assign(m_bias_slot, [1.5]))
save_path = root_trackable.save(file_prefix=prefix)
self.evaluate(state_ops.assign(model._named_dense.variables[1], [43.]))
self.evaluate(state_ops.assign(root_trackable.save_counter, 3))
optimizer_variables = self.evaluate(optimizer.variables())
self.evaluate(state_ops.assign(m_bias_slot, [-2.]))
# Immediate restoration
status = root_trackable.restore(save_path=save_path).assert_consumed()
status.run_restore_ops()
self.assertAllEqual([42.], self.evaluate(model._named_dense.variables[1]))
self.assertAllEqual(1, self.evaluate(root_trackable.save_counter))
self.assertAllEqual([1.5], self.evaluate(m_bias_slot))
if not context.executing_eagerly():
return # Restore-on-create is only supported when executing eagerly
on_create_model = MyModel()
on_create_optimizer = adam.AdamOptimizer(
0.001,
# Preserve beta_1_power and beta_2_power when appying gradients
# so we can test that they've been restored correctly.
beta1=1.0,
beta2=1.0)
on_create_root = util.Checkpoint(
optimizer=on_create_optimizer, model=on_create_model)
# Deferred restoration
status = on_create_root.restore(save_path=save_path)
on_create_model(constant_op.constant([[3.]])) # create variables
self.assertAllEqual(1, self.evaluate(on_create_root.save_counter))
self.assertAllEqual([42.],
self.evaluate(
on_create_model._named_dense.variables[1]))
on_create_m_bias_slot = on_create_optimizer.get_slot(
on_create_model._named_dense.variables[1], "m")
# Optimizer slot variables are created when the original variable is
# restored.
self.assertAllEqual([1.5], self.evaluate(on_create_m_bias_slot))
self.assertAllEqual(optimizer_variables[2:],
self.evaluate(on_create_optimizer.variables()))
dummy_var = resource_variable_ops.ResourceVariable([1.])
on_create_optimizer.minimize(loss=dummy_var.read_value)
status.assert_consumed()
beta_1_power, beta_2_power = on_create_optimizer._get_beta_accumulators()
self.assertAllEqual(optimizer_variables[0], self.evaluate(beta_1_power))
self.assertAllEqual(optimizer_variables[1], self.evaluate(beta_2_power))
def doTestBasic(self, use_resource=False):
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.Variable(var0_np)
var1 = variables.Variable(var1_np)
grads0 = constant_op.constant(grads0_np)
grads1 = constant_op.constant(grads1_np)
opt = adam.AdamOptimizer()
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)
with ops.Graph().as_default():
# Shouldn't return non-slot variables from other graphs.
self.assertEqual(0, len(opt.variables()))
if not context.executing_eagerly():
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)
def testMultipleGraphsNonSlotVariables(self):
with context.graph_mode():
checkpoint_directory = self.get_temp_dir()
checkpoint_prefix = os.path.join(checkpoint_directory, "ckpt")
optimizer = adam.AdamOptimizer(0.001)
# Construct a model in one graph
first_graph = ops.Graph()
first_session = session_lib.Session(graph=first_graph)
with first_graph.as_default(), first_session.as_default():
first_variable = resource_variable_ops.ResourceVariable([1.])
first_root_checkpointable = checkpointable_utils.Checkpoint(
optimizer=optimizer, variable=first_variable)
train_op = optimizer.minimize(first_variable.read_value)
self.evaluate(
checkpointable_utils.gather_initializers(
first_root_checkpointable))
self.evaluate(train_op)
self.evaluate(first_variable.assign([1.]))
self.evaluate(
optimizer.get_slot(var=first_variable,
name="m").assign([2.]))
beta1_power, _ = optimizer._get_beta_accumulators()
self.evaluate(beta1_power.assign(3.))
# Save and load in a second graph
second_graph = ops.Graph()
with second_graph.as_default(), session_lib.Session(
graph=second_graph):
second_variable = resource_variable_ops.ResourceVariable([1.])
second_root_checkpointable = checkpointable_utils.Checkpoint(
optimizer=optimizer, variable=second_variable)
train_op = optimizer.minimize(second_variable.read_value)
second_root_checkpointable.restore(
None).initialize_or_restore()
self.evaluate(train_op)
self.evaluate(second_variable.assign([4.]))
self.evaluate(
optimizer.get_slot(var=second_variable,
name="m").assign([5.]))
beta1_power, _ = optimizer._get_beta_accumulators()
self.evaluate(beta1_power.assign(6.))
save_path = second_root_checkpointable.save(checkpoint_prefix)
self.evaluate(second_variable.assign([7.]))
self.evaluate(
optimizer.get_slot(var=second_variable,
name="m").assign([8.]))
beta1_power, _ = optimizer._get_beta_accumulators()
self.assertAllEqual(6., self.evaluate(beta1_power))
status = second_root_checkpointable.restore(save_path)
status.assert_consumed().run_restore_ops()
self.assertAllEqual([4.], self.evaluate(second_variable))
self.assertAllEqual([5.],
self.evaluate(
optimizer.get_slot(var=second_variable,
name="m")))
beta1_power, _ = optimizer._get_beta_accumulators()
self.assertAllEqual(6., self.evaluate(beta1_power))
# Check that the first graph is unmolested
with first_graph.as_default(), first_session.as_default():
self.assertAllEqual([1.], self.evaluate(first_variable))
self.assertAllEqual([2.],
self.evaluate(
optimizer.get_slot(var=first_variable,
name="m")))
beta1_power, _ = optimizer._get_beta_accumulators()
self.assertAllEqual(3., self.evaluate(beta1_power))
def testDeferredSlotRestoration(self):
checkpoint_directory = self.get_temp_dir()
root = checkpointable.Checkpointable()
root.var = checkpointable_utils.add_variable(root,
name="var",
initializer=0.)
optimizer = adam.AdamOptimizer(0.1)
if context.executing_eagerly():
optimizer.minimize(root.var.read_value)
else:
train_op = optimizer.minimize(root.var)
# Note that `optimizer` has not been added as a dependency of
# `root`. Create a one-off grouping so that slot variables for `root.var`
# get initialized too.
self.evaluate(
checkpointable_utils.gather_initializers(
checkpointable_utils.Checkpoint(root=root,
optimizer=optimizer)))
self.evaluate(train_op)
self.evaluate(state_ops.assign(root.var, 12.))
no_slots_path = checkpointable_utils.CheckpointableSaver(root).save(
os.path.join(checkpoint_directory, "no_slots"))
root.optimizer = optimizer
self.evaluate(state_ops.assign(root.var, 13.))
self.evaluate(
state_ops.assign(optimizer.get_slot(name="m", var=root.var), 14.))
slots_path = checkpointable_utils.CheckpointableSaver(root).save(
os.path.join(checkpoint_directory, "with_slots"))
new_root = checkpointable.Checkpointable()
# Load the slot-containing checkpoint (deferred), then immediately overwrite
# the non-slot variable (also deferred).
slot_status = checkpointable_utils.CheckpointableSaver(
new_root).restore(slots_path)
no_slot_status = checkpointable_utils.CheckpointableSaver(
new_root).restore(no_slots_path)
with self.assertRaises(AssertionError):
no_slot_status.assert_consumed()
new_root.var = checkpointable_utils.add_variable(new_root,
name="var",
shape=[])
no_slot_status.assert_consumed()
no_slot_status.run_restore_ops()
self.assertEqual(12., self.evaluate(new_root.var))
new_root.optimizer = adam.AdamOptimizer(0.1)
with self.assertRaisesRegexp(AssertionError, "beta1_power"):
slot_status.assert_consumed()
self.assertEqual(12., self.evaluate(new_root.var))
if context.executing_eagerly():
# Slot variables are only created with restoring initializers when
# executing eagerly.
self.assertEqual(
14.,
self.evaluate(
new_root.optimizer.get_slot(name="m", var=new_root.var)))
else:
self.assertIs(
new_root.optimizer.get_slot(name="m", var=new_root.var), None)
if context.executing_eagerly():
new_root.optimizer.minimize(new_root.var.read_value)
else:
train_op = new_root.optimizer.minimize(new_root.var)
# The slot variable now exists; restore() didn't create it, but we should
# now have a restore op for it.
slot_status.run_restore_ops()
self.assertEqual(
14.,
self.evaluate(
new_root.optimizer.get_slot(name="m", var=new_root.var)))
self.evaluate(train_op)
slot_status.assert_consumed()
adagrad_optimizer_v1_fn = NamedObject(
"AdagradV1", lambda: adagrad.AdagradOptimizer(0.001))
adam_optimizer_v1_fn = NamedObject("AdamV1",
lambda: adam.AdamOptimizer(0.001, epsilon=1))
rmsprop_optimizer_v1_fn = NamedObject(
"RmsPropV1", lambda: rmsprop.RMSPropOptimizer(0.001))
optimizers_v1 = [gradient_descent_optimizer_v1_fn, adagrad_optimizer_v1_fn]
gradient_descent_optimizer_v2_fn = NamedObject(
"GradientDescentV2",
lambda: gradient_descent_v2.GradientDescentOptimizer(0.2))
adagrad_optimizer_v2_fn = NamedObject(
"AdagradV2", lambda: adagrad_v2.AdagradOptimizer(0.001))
adam_optimizer_v2_fn = NamedObject(
"AdamV2", lambda: adam_v2.AdamOptimizer(0.001, epsilon=1))
optimizers_v2 = [gradient_descent_optimizer_v2_fn, adagrad_optimizer_v2_fn]
graph_and_eager_modes = ["graph", "eager"]
def distributions_and_v1_optimizers():
"""A common set of combination with DistributionStrategies and Optimizers."""
return combine(
distribution=[
one_device_strategy,
mirrored_strategy_with_gpu_and_cpu,
mirrored_strategy_with_two_gpus,
core_mirrored_strategy_with_gpu_and_cpu,
core_mirrored_strategy_with_two_gpus,