def testObjectsCombined(self):
# Currently fine to load two checkpoint objects into one Python object
checkpoint_directory = self.get_temp_dir()
save_root = checkpointable.Checkpointable()
save_root.dep_one = checkpointable.Checkpointable()
save_root.dep_two = checkpointable.Checkpointable()
checkpointable_utils.add_variable(
save_root.dep_one, name="var1", initializer=32., dtype=dtypes.float64)
checkpointable_utils.add_variable(
save_root.dep_two, name="var2", initializer=64., dtype=dtypes.float64)
self.evaluate(checkpointable_utils.gather_initializers(save_root))
save_path = checkpointable_utils.CheckpointableSaver(save_root).save(
os.path.join(checkpoint_directory, "ckpt"))
load_root = checkpointable.Checkpointable()
load_root.dep_one = checkpointable.Checkpointable()
load_root.dep_two = load_root.dep_one
v1 = checkpointable_utils.add_variable(
load_root.dep_one, name="var1", shape=[], dtype=dtypes.float64)
v2 = checkpointable_utils.add_variable(
load_root.dep_one, name="var2", shape=[], dtype=dtypes.float64)
status = checkpointable_utils.CheckpointableSaver(load_root).restore(
save_path).assert_consumed()
status.run_restore_ops()
self.assertEqual(32., self.evaluate(v1))
self.assertEqual(64., self.evaluate(v2))
def testAddVariable(self):
obj = NonLayerCheckpointable()
with self.assertRaisesRegexp(ValueError, "do not specify shape"):
checkpointable_utils.add_variable(
obj, name="shape_specified_twice", shape=[], initializer=1)
constant_initializer = checkpointable_utils.add_variable(
obj, name="constant_initializer", initializer=1)
with variable_scope.variable_scope("some_variable_scope"):
ones_initializer = checkpointable_utils.add_variable(
obj,
name="ones_initializer",
shape=[2],
initializer=init_ops.ones_initializer(dtype=dtypes.float32))
bare_initializer = checkpointable_utils.add_variable(
obj,
name="bare_initializer",
shape=[2, 2],
dtype=dtypes.float64,
initializer=init_ops.zeros_initializer)
# Even in graph mode, there are no naming conflicts between objects, only
# naming conflicts within an object.
other_duplicate = resource_variable_ops.ResourceVariable(
name="duplicate", initial_value=1.)
duplicate = checkpointable_utils.add_variable(
obj, name="duplicate", shape=[])
with self.assertRaisesRegexp(ValueError, "'duplicate' already exists"):
checkpointable_utils.add_variable(obj, name="duplicate", shape=[])
self.evaluate(checkpointable_utils.gather_initializers(obj))
self.assertEqual("constant_initializer:0", constant_initializer.name)
self.assertEqual(1, self.evaluate(constant_initializer))
self.assertEqual("some_variable_scope/ones_initializer:0",
ones_initializer.name)
self.assertAllEqual([1, 1], self.evaluate(ones_initializer))
self.assertAllEqual([[0., 0.],
[0., 0.]], self.evaluate(bare_initializer))
self.assertEqual("a_variable:0", obj.a_variable.name)
self.assertEqual("duplicate:0", other_duplicate.name)
if context.in_graph_mode():
# The .name attribute may be globally influenced, but the checkpoint name
# won't be (tested below).
self.assertEqual("duplicate_1:0", duplicate.name)
else:
# When executing eagerly, there's no uniquification of variable names. The
# checkpoint name will be the same.
self.assertEqual("duplicate:0", duplicate.name)
named_variables, _ = checkpointable_utils._serialize_object_graph(obj)
expected_checkpoint_names = (
"a_variable/.ATTRIBUTES/VARIABLE_VALUE",
"bare_initializer/.ATTRIBUTES/VARIABLE_VALUE",
"constant_initializer/.ATTRIBUTES/VARIABLE_VALUE",
"duplicate/.ATTRIBUTES/VARIABLE_VALUE",
"ones_initializer/.ATTRIBUTES/VARIABLE_VALUE",
)
six.assertCountEqual(
self, expected_checkpoint_names, named_variables.keys())
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 testSaveRestore(self):
network = MyNetwork()
optimizer = adam.AdamOptimizer(0.001)
root_checkpointable = checkpointable_utils.Checkpoint(
optimizer=optimizer, network=network)
input_value = constant_op.constant([[3.]])
if context.in_eager_mode():
optimizer.minimize(
lambda: network(input_value))
else:
train_op = optimizer.minimize(network(input_value))
# TODO(allenl): Make initialization more pleasant when graph building.
root_checkpointable.save_counter # pylint: disable=pointless-statement
self.evaluate(checkpointable_utils.gather_initializers(
root_checkpointable))
self.evaluate(train_op)
prefix = os.path.join(self.get_temp_dir(), "ckpt")
self.evaluate(state_ops.assign(network._named_dense.variables[1], [42.]))
m_bias_slot = optimizer.get_slot(network._named_dense.variables[1], "m")
self.evaluate(state_ops.assign(m_bias_slot, [1.5]))
save_path = root_checkpointable.save(file_prefix=prefix)
self.evaluate(state_ops.assign(network._named_dense.variables[1], [43.]))
self.evaluate(state_ops.assign(root_checkpointable.save_counter, 3))
optimizer_variables = self.evaluate(optimizer.variables())
self.evaluate(state_ops.assign(m_bias_slot, [-2.]))
# Immediate restoration
status = root_checkpointable.restore(save_path=save_path).assert_consumed()
status.run_restore_ops()
self.assertAllEqual([42.], self.evaluate(network._named_dense.variables[1]))
self.assertAllEqual(1, self.evaluate(root_checkpointable.save_counter))
self.assertAllEqual([1.5], self.evaluate(m_bias_slot))
if context.in_graph_mode():
return # Restore-on-create is only supported when executing eagerly
on_create_network = MyNetwork()
on_create_optimizer = adam.AdamOptimizer(0.001)
on_create_root = checkpointable_utils.Checkpoint(
optimizer=on_create_optimizer, network=on_create_network)
# Deferred restoration
status = on_create_root.restore(save_path=save_path)
on_create_network(constant_op.constant([[3.]])) # create variables
self.assertAllEqual(1, self.evaluate(on_create_root.save_counter))
self.assertAllEqual([42.],
self.evaluate(
on_create_network._named_dense.variables[1]))
on_create_m_bias_slot = on_create_optimizer.get_slot(
on_create_network._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()))
on_create_optimizer._create_slots(
[resource_variable_ops.ResourceVariable([1.])])
status.assert_consumed()
beta1_power, beta2_power = on_create_optimizer._get_beta_accumulators()
self.assertAllEqual(optimizer_variables[0], self.evaluate(beta1_power))
self.assertAllEqual(optimizer_variables[1], self.evaluate(beta2_power))
def testManySavesGraph(self):
"""Saves after the first should not modify the graph."""
with context.graph_mode():
graph = ops.Graph()
with graph.as_default(), self.test_session(graph):
checkpoint_directory = self.get_temp_dir()
checkpoint_prefix = os.path.join(checkpoint_directory, "ckpt")
obj = checkpointable.Checkpointable()
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(checkpointable_utils.gather_initializers(obj))
saver = checkpointable_utils.CheckpointableSaver(obj)
saver.save(checkpoint_prefix)
before_ops = graph.get_operations()
saver.save(checkpoint_prefix)
self.assertEqual(before_ops, graph.get_operations())
def testDependencyLoop(self):
# Note: this test creates garbage during eager execution because it
# purposefully creates a reference cycle.
first = checkpointable.Checkpointable()
second = checkpointable.Checkpointable()
first.second = second
second.first = first
first.v = checkpointable_utils.add_variable(
first, "v1", initializer=[3., 1., 4.])
second.v = checkpointable_utils.add_variable(
second, "v2", initializer=[1., 1., 2., 3.])
self.evaluate(checkpointable_utils.gather_initializers(first))
checkpoint_directory = self.get_temp_dir()
save_path = checkpointable_utils.CheckpointableSaver(first).save(
os.path.join(checkpoint_directory, "ckpt"))
# Test deferred loading
first_load = checkpointable.Checkpointable()
status = checkpointable_utils.CheckpointableSaver(
first_load).restore(save_path)
second_load = checkpointable.Checkpointable()
first_load.second = second_load
second_load.first = first_load
with self.assertRaises(AssertionError):
status.assert_consumed()
first_load.v = checkpointable_utils.add_variable(
first_load, "v1", shape=[3])
second_load.v = checkpointable_utils.add_variable(
second_load, "v2", shape=[4])
status.assert_consumed()
status.run_restore_ops()
self.assertAllEqual([3., 1., 4.], self.evaluate(first_load.v))
self.assertAllEqual([1., 1., 2., 3.], self.evaluate(second_load.v))
# Test loading when variables have already been created
self.evaluate(first_load.v.assign([2., 7., 1.]))
self.assertAllEqual([2., 7., 1.], self.evaluate(first_load.v))
self.evaluate(second_load.v.assign([2., 7., 1., 8.]))
self.assertAllEqual([2., 7., 1., 8.], self.evaluate(second_load.v))
status = checkpointable_utils.CheckpointableSaver(first_load).restore(
save_path).assert_consumed()
status.run_restore_ops()
self.assertAllEqual([3., 1., 4.], self.evaluate(first_load.v))
self.assertAllEqual([1., 1., 2., 3.], self.evaluate(second_load.v))
def _initialized_model(self):
input_value = constant_op.constant([[3.]])
network = MyNetwork()
optimizer = adam.AdamOptimizer(0.001)
optimizer_step = training_util.get_or_create_global_step()
root_checkpointable = checkpointable_utils.Checkpoint(
optimizer=optimizer, network=network, optimizer_step=optimizer_step)
train_op = optimizer.minimize(
functools.partial(network, input_value),
global_step=optimizer_step)
self.evaluate(checkpointable_utils.gather_initializers(
root_checkpointable))
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(network._named_dense.bias.assign([1.]))
self.evaluate(optimizer.get_slot(
var=network._named_dense.bias, name="m").assign([2.]))
beta1_power, _ = optimizer._get_beta_accumulators()
self.evaluate(beta1_power.assign(3.))
return root_checkpointable
def testAmbiguousLoad(self):
# Not OK to split one checkpoint object into two
checkpoint_directory = self.get_temp_dir()
save_root = checkpointable.Checkpointable()
save_root.dep_one = checkpointable.Checkpointable()
save_root.dep_two = checkpointable.Checkpointable()
dep_three = checkpointable.Checkpointable()
save_root.dep_one.dep_three = dep_three
save_root.dep_two.dep_three = dep_three
checkpointable_utils.add_variable(dep_three, name="var", initializer=0.)
self.evaluate(checkpointable_utils.gather_initializers(save_root))
save_path = checkpointable_utils.CheckpointableSaver(save_root).save(
os.path.join(checkpoint_directory, "ckpt"))
load_root = checkpointable.Checkpointable()
checkpointable_utils.CheckpointableSaver(load_root).restore(save_path)
load_root.dep_one = checkpointable.Checkpointable()
load_root.dep_two = checkpointable.Checkpointable()
load_root.dep_one.dep_three = checkpointable.Checkpointable()
with self.assertRaisesRegexp(AssertionError,
"resolved to different objects"):
load_root.dep_two.dep_three = checkpointable.Checkpointable()
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.in_graph_mode():
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)
else:
optimizer.minimize(root.var.read_value)
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.in_eager_mode():
# 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.in_graph_mode():
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)
else:
new_root.optimizer.minimize(new_root.var.read_value)
slot_status.assert_consumed()
def testNamingWithOptimizer(self):
input_value = constant_op.constant([[3.]])
network = MyNetwork()
# A nuisance Network using the same optimizer. Its slot variables should not
# go in the checkpoint, since it is never depended on.
other_network = MyNetwork()
optimizer = adam.AdamOptimizer(0.001)
optimizer_step = training_util.get_or_create_global_step()
root_checkpointable = checkpointable_utils.Checkpoint(
optimizer=optimizer, network=network, optimizer_step=optimizer_step)
if context.in_eager_mode():
optimizer.minimize(
lambda: network(input_value),
global_step=optimizer_step)
optimizer.minimize(
lambda: other_network(input_value),
global_step=optimizer_step)
else:
train_op = optimizer.minimize(
network(input_value), global_step=optimizer_step)
optimizer.minimize(
other_network(input_value),
global_step=optimizer_step)
self.evaluate(checkpointable_utils.gather_initializers(
root_checkpointable))
self.evaluate(train_op)
named_variables, serialized_graph = (
checkpointable_utils._serialize_object_graph(root_checkpointable))
expected_checkpoint_names = (
# Created in the root node, so no prefix.
"optimizer_step",
# No name provided to track_checkpointable(), so the position is used
# instead (one-based).
"network/via_track_layer/kernel",
# track_checkpointable() with a name provided, so that's used
"network/_named_dense/kernel",
"network/_named_dense/bias",
# non-Layer dependency of the network
"network/_non_layer/a_variable",
# The optimizer creates two non-slot variables
"optimizer/beta1_power",
"optimizer/beta2_power",
# Slot variables
"network/via_track_layer/kernel/.OPTIMIZER_SLOT/optimizer/m",
"network/via_track_layer/kernel/.OPTIMIZER_SLOT/optimizer/v",
"network/_named_dense/kernel/.OPTIMIZER_SLOT/optimizer/m",
"network/_named_dense/kernel/.OPTIMIZER_SLOT/optimizer/v",
"network/_named_dense/bias/.OPTIMIZER_SLOT/optimizer/m",
"network/_named_dense/bias/.OPTIMIZER_SLOT/optimizer/v",
)
suffix = "/.ATTRIBUTES/VARIABLE_VALUE"
expected_checkpoint_names = [
name + suffix for name in expected_checkpoint_names]
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:0",
named_variables["optimizer_step" + suffix].name)
self.assertEqual(
"my_network/dense_1/kernel:0",
named_variables["network/via_track_layer/kernel" + suffix].name)
self.assertEqual(
"my_network/dense/kernel:0",
named_variables["network/_named_dense/kernel" + suffix].name)
self.assertEqual(
"beta1_power:0",
named_variables["optimizer/beta1_power" + suffix].name)
self.assertEqual(
"beta2_power:0",
named_variables["optimizer/beta2_power" + suffix].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_network/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_network/dense/kernel/Adam",
serialized_graph.nodes[optimizer_node.slot_variables[0]
.slot_variable_node_id]
.attributes[0].full_name)
self.assertEqual(
"my_network/dense/kernel/Adam:0",
optimizer.get_slot(
var=named_variables["network/_named_dense/kernel" + suffix],
name="m").name)
self.assertEqual(
"network/_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(
"network/_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_checkpointable = checkpointable_utils.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_checkpointable.save_counter # pylint: disable=pointless-statement
self.evaluate(
checkpointable_utils.gather_initializers(root_checkpointable))
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_checkpointable.save(file_prefix=prefix)
self.evaluate(state_ops.assign(model._named_dense.variables[1], [43.]))
self.evaluate(state_ops.assign(root_checkpointable.save_counter, 3))
optimizer_variables = self.evaluate(optimizer.variables())
self.evaluate(state_ops.assign(m_bias_slot, [-2.]))
# Immediate restoration
status = root_checkpointable.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_checkpointable.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 beta1_power and beta2_power when appying gradients so we can
# test that they've been restored correctly.
beta1=1.0,
beta2=1.0)
on_create_root = checkpointable_utils.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()
beta1_power, beta2_power = on_create_optimizer._get_beta_accumulators()
self.assertAllEqual(optimizer_variables[0], self.evaluate(beta1_power))
self.assertAllEqual(optimizer_variables[1], self.evaluate(beta2_power))
def testSaveRestore(self):
model = MyModel()
optimizer = adam.AdamOptimizer(0.001)
root_checkpointable = checkpointable_utils.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_checkpointable.save_counter # pylint: disable=pointless-statement
self.evaluate(checkpointable_utils.gather_initializers(
root_checkpointable))
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_checkpointable.save(file_prefix=prefix)
self.evaluate(state_ops.assign(model._named_dense.variables[1], [43.]))
self.evaluate(state_ops.assign(root_checkpointable.save_counter, 3))
optimizer_variables = self.evaluate(optimizer.variables())
self.evaluate(state_ops.assign(m_bias_slot, [-2.]))
# Immediate restoration
status = root_checkpointable.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_checkpointable.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 beta1_power and beta2_power when appying gradients so we can
# test that they've been restored correctly.
beta1=1.0, beta2=1.0)
on_create_root = checkpointable_utils.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()
beta1_power, beta2_power = on_create_optimizer._get_beta_accumulators()
self.assertAllEqual(optimizer_variables[0], self.evaluate(beta1_power))
self.assertAllEqual(optimizer_variables[1], self.evaluate(beta2_power))
