def testConstructAdadeltaWithEpsilonValues(self):
opt = adadelta.Adadelta(epsilon=None)
config = opt.get_config()
self.assertEqual(config["epsilon"], 1e-7)
opt = adadelta.Adadelta(epsilon=1e-8)
config = opt.get_config()
self.assertEqual(config["epsilon"], 1e-8)
def testConstructAdadeltaWithLR(self):
opt = adadelta.Adadelta(lr=1.0, rho=0.9, epsilon=1.)
self.assertEqual(opt.lr, 1.0)
opt_2 = adadelta.Adadelta(learning_rate=0.1,
rho=0.9,
epsilon=1.,
lr=1.0)
self.assertEqual(opt_2.lr, 1.0)
opt_3 = adadelta.Adadelta(learning_rate=0.1, rho=0.9, epsilon=1.)
self.assertEqual(opt_3.lr, 0.1)
def testConstructAdadeltaWithLR(self):
opt = adadelta.Adadelta(lr=1.0, rho=0.9, epsilon=1.)
opt_2 = adadelta.Adadelta(learning_rate=0.1, rho=0.9, epsilon=1., lr=1.0)
opt_3 = adadelta.Adadelta(learning_rate=0.1, rho=0.9, epsilon=1.)
self.assertIsInstance(opt.lr, variables.Variable)
self.assertIsInstance(opt_2.lr, variables.Variable)
self.assertIsInstance(opt_3.lr, variables.Variable)
self.evaluate(variables.global_variables_initializer())
self.assertAllClose(self.evaluate(opt.lr), (1.0))
self.assertAllClose(self.evaluate(opt_2.lr), (1.0))
self.assertAllClose(self.evaluate(opt_3.lr), (0.1))
def testConfig(self):
def rho():
return ops.convert_to_tensor(1.0)
epsilon = ops.convert_to_tensor(1.0)
opt = adadelta.Adadelta(learning_rate=1.0, rho=rho, epsilon=epsilon)
config = opt.get_config()
opt2 = adadelta.Adadelta.from_config(config)
self.assertEqual(opt._hyper["learning_rate"][1],
opt2._hyper["learning_rate"][1])
self.assertEqual(opt._hyper["rho"][1].__name__,
opt2._hyper["rho"][1].__name__)
self.assertEqual(opt._hyper["epsilon"][1], opt2._hyper["epsilon"][1])
def testMinimizeSparseResourceVariable(self):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.cached_session():
var0 = resource_variable_ops.ResourceVariable([[1.0, 2.0]], dtype=dtype)
x = constant_op.constant([[4.0], [5.0]], dtype=dtype)
pred = math_ops.matmul(embedding_ops.embedding_lookup([var0], [0]), x)
loss = pred * pred
sgd_op = adadelta.Adadelta(1.0, 1.0, 1.0).minimize(loss)
variables.global_variables_initializer().run()
# Fetch params to validate initial values
self.assertAllCloseAccordingToType([[1.0, 2.0]], var0.eval())
# Run 1 step of sgd
sgd_op.run()
# Validate updated params
self.assertAllCloseAccordingToType([[-111, -138]], self.evaluate(var0))
def testMinimizeSparseResourceVariable(self):
for dtype in _DATA_TYPES:
var0 = resource_variable_ops.ResourceVariable([[1.0, 2.0]], dtype=dtype)
x = constant_op.constant([[4.0], [5.0]], dtype=dtype)
def loss():
pred = math_ops.matmul(embedding_ops.embedding_lookup([var0], [0]), x) # pylint: disable=cell-var-from-loop
return pred * pred
sgd_op = adadelta.Adadelta(1.0, 1.0, 1.0).minimize(loss, var_list=[var0])
self.evaluate(variables.global_variables_initializer())
# Fetch params to validate initial values
self.assertAllCloseAccordingToType([[1.0, 2.0]], self.evaluate(var0))
# Run 1 step of sgd
self.evaluate(sgd_op)
# Validate updated params
self.assertAllCloseAccordingToType([[-111, -138]], self.evaluate(var0))
gradient_descent_optimizer_v1_fn = combinations.NamedObject(
"GradientDescentV1",
lambda: gradient_descent.GradientDescentOptimizer(0.2))
adagrad_optimizer_v1_fn = combinations.NamedObject(
"AdagradV1", lambda: adagrad.AdagradOptimizer(0.001))
adam_optimizer_v1_fn = combinations.NamedObject(
"AdamV1", lambda: adam.AdamOptimizer(0.001, epsilon=1))
rmsprop_optimizer_v1_fn = combinations.NamedObject(
"RmsPropV1", lambda: rmsprop.RMSPropOptimizer(0.001))
# TODO(shiningsun): consider adding the other v1 optimizers
optimizers_v1 = [gradient_descent_optimizer_v1_fn, adagrad_optimizer_v1_fn]
adadelta_optimizer_keras_v2_fn = combinations.NamedObject(
"AdadeltaKerasV2", lambda: adadelta_keras_v2.Adadelta(0.001))
adagrad_optimizer_keras_v2_fn = combinations.NamedObject(
"AdagradKerasV2", lambda: adagrad_keras_v2.Adagrad(0.001))
adam_optimizer_keras_v2_fn = combinations.NamedObject(
"AdamKerasV2", lambda: adam_keras_v2.Adam(0.001, epsilon=1.0))
adamax_optimizer_keras_v2_fn = combinations.NamedObject(
"AdamaxKerasV2", lambda: adamax_keras_v2.Adamax(0.001, epsilon=1.0))
nadam_optimizer_keras_v2_fn = combinations.NamedObject(
"NadamKerasV2", lambda: nadam_keras_v2.Nadam(0.001, epsilon=1.0))
ftrl_optimizer_keras_v2_fn = combinations.NamedObject(
"FtrlKerasV2", lambda: ftrl_keras_v2.Ftrl(0.001))
gradient_descent_optimizer_keras_v2_fn = combinations.NamedObject(
"GradientDescentKerasV2", lambda: gradient_descent_keras_v2.SGD(0.2))
rmsprop_optimizer_keras_v2_fn = combinations.NamedObject(
"RmsPropKerasV2", lambda: rmsprop_keras_v2.RMSprop(0.001))
def testAdadeltaCompatibility(self):
opt_v1 = optimizers.Adadelta(lr=0.01)
opt_v2 = adadelta.Adadelta(learning_rate=0.01)
self._testOptimizersCompatibility(opt_v1, opt_v2)
def doTestBasic(self, use_resource=False, use_callable_params=False):
num_updates = 4 # number of ADADELTA steps to perform
for dtype in [dtypes.half, dtypes.float32]:
for grad in [0.2, 0.1, 0.01]:
for lr in [1.0, 0.5, 0.1]:
var0_init = [1.0, 2.0]
var1_init = [3.0, 4.0]
if use_resource:
var0 = resource_variable_ops.ResourceVariable(
var0_init, dtype=dtype)
var1 = resource_variable_ops.ResourceVariable(
var1_init, dtype=dtype)
else:
var0 = variables.Variable(var0_init, dtype=dtype)
var1 = variables.Variable(var1_init, dtype=dtype)
grads = constant_op.constant([grad, grad], dtype=dtype)
accum = 0.0
accum_update = 0.0
# ADADELTA gradient optimizer
rho = 0.95
epsilon = 1e-8
if use_callable_params:
adadelta_opt = adadelta.Adadelta(
learning_rate=lambda: lr, # pylint: disable=cell-var-from-loop
rho=lambda: rho, # pylint: disable=cell-var-from-loop
epsilon=lambda: epsilon) # pylint: disable=cell-var-from-loop
else:
adadelta_opt = adadelta.Adadelta(learning_rate=lr,
rho=rho,
epsilon=epsilon)
if not context.executing_eagerly():
adadelta_update = adadelta_opt.apply_gradients(
zip([grads, grads], [var0, var1]))
self.evaluate(variables.global_variables_initializer())
# Assign slots
slot = [None] * 2
slot_update = [None] * 2
slot[0] = adadelta_opt.get_slot(var0, "accum_grad")
self.assertEqual(slot[0].get_shape(), var0.get_shape())
slot_update[0] = adadelta_opt.get_slot(
var0, "accum_var")
self.assertEqual(slot_update[0].get_shape(),
var0.get_shape())
slot[1] = adadelta_opt.get_slot(var1, "accum_grad")
self.assertEqual(slot[1].get_shape(), var1.get_shape())
slot_update[1] = adadelta_opt.get_slot(
var1, "accum_var")
self.assertEqual(slot_update[1].get_shape(),
var1.get_shape())
# Fetch params to validate initial values
self.assertAllClose(var0_init, self.evaluate(var0))
self.assertAllClose(var1_init, self.evaluate(var1))
update = [None] * num_updates
tot_update = 0
for step in range(num_updates):
# Run adadelta update for comparison
if not context.executing_eagerly():
self.evaluate(adadelta_update)
else:
adadelta_opt.apply_gradients(
zip([grads, grads], [var0, var1]))
# Perform initial update without previous accum values
accum = accum * rho + (grad**2) * (1 - rho)
update[step] = (np.sqrt(accum_update + epsilon) *
(1. / np.sqrt(accum + epsilon)) * grad)
accum_update = (accum_update * rho +
(update[step]**2) * (1.0 - rho))
tot_update += update[step] * lr
if not context.executing_eagerly():
# Check that the accumulators have been updated
# TODO(lxuechen): This is hard to test in eager mode
for slot_idx in range(2):
self.assertAllCloseAccordingToType(
np.array([accum, accum],
dtype=dtype.as_numpy_dtype()),
self.evaluate(slot[slot_idx]),
rtol=1e-5)
self.assertAllCloseAccordingToType(
np.array([accum_update, accum_update],
dtype=dtype.as_numpy_dtype()),
self.evaluate(slot_update[slot_idx]),
rtol=1e-5)
# Check that the parameters have been updated
self.assertAllCloseAccordingToType(
np.array([
var0_init[0] - tot_update,
var0_init[1] - tot_update
],
dtype=dtype.as_numpy_dtype()),
self.evaluate(var0),
rtol=1e-5)
self.assertAllCloseAccordingToType(
np.array([
var1_init[0] - tot_update,
var1_init[1] - tot_update
],
dtype=dtype.as_numpy_dtype()),
self.evaluate(var1),
rtol=1e-5)
def testConstructAdadeltaWithEpsilonValues(self):
opt = adadelta.Adadelta(epsilon=None)
self.assertEqual(opt.epsilon, 1e-7)
opt = adadelta.Adadelta(epsilon=1e-8)
self.assertEqual(opt.epsilon, 1e-8)
def doTestBasic(self, use_resource=False):
num_updates = 4 # number of ADADELTA steps to perform
for dtype in [dtypes.half, dtypes.float32]:
for grad in [0.2, 0.1, 0.01]:
for lr in [1.0, 0.5, 0.1]:
with self.cached_session():
var0_init = [1.0, 2.0]
var1_init = [3.0, 4.0]
if use_resource:
var0 = resource_variable_ops.ResourceVariable(
var0_init, dtype=dtype)
var1 = resource_variable_ops.ResourceVariable(
var1_init, dtype=dtype)
else:
var0 = variables.Variable(var0_init, dtype=dtype)
var1 = variables.Variable(var1_init, dtype=dtype)
grads = constant_op.constant([grad, grad], dtype=dtype)
accum = 0.0
accum_update = 0.0
# ADADELTA gradient optimizer
rho = 0.95
epsilon = 1e-8
adadelta_opt = adadelta.Adadelta(lr, rho, epsilon)
adadelta_update = adadelta_opt.apply_gradients(
zip([grads, grads], [var0, var1]))
opt_vars = adadelta_opt.variables()
self.assertStartsWith(opt_vars[0].name, var0._shared_name)
self.assertStartsWith(opt_vars[1].name, var0._shared_name)
self.assertStartsWith(opt_vars[2].name, var1._shared_name)
self.assertStartsWith(opt_vars[3].name, var1._shared_name)
self.assertEqual(4, len(opt_vars))
variables.global_variables_initializer().run()
# Assign slots
slot = [None] * 2
slot_update = [None] * 2
slot[0] = adadelta_opt.get_slot(var0, "accum_grad")
self.assertEqual(slot[0].shape, var0.shape)
slot_update[0] = adadelta_opt.get_slot(var0, "accum_var")
self.assertEqual(slot_update[0].shape, var0.shape)
slot[1] = adadelta_opt.get_slot(var1, "accum_grad")
self.assertEqual(slot[1].shape, var1.shape)
slot_update[1] = adadelta_opt.get_slot(var1, "accum_var")
self.assertEqual(slot_update[1].shape, var1.shape)
# Fetch params to validate initial values
self.assertAllClose(var0_init, self.evaluate(var0))
self.assertAllClose(var1_init, self.evaluate(var1))
# Check that the accumulators have been updated
for slot_idx in range(2):
self.assertAllCloseAccordingToType(
np.array([accum, accum], dtype=dtype.as_numpy_dtype()),
slot[slot_idx].eval(),
rtol=1e-5)
self.assertAllCloseAccordingToType(
np.array(
[accum_update, accum_update],
dtype=dtype.as_numpy_dtype()),
slot_update[slot_idx].eval(),
rtol=1e-5)
# Check that the parameters have been updated
self.assertAllCloseAccordingToType(
np.array(
[var0_init[0] - tot_update, var0_init[1] - tot_update],
dtype=dtype.as_numpy_dtype()),
var0.eval(),
rtol=1e-5)
self.assertAllCloseAccordingToType(
np.array(
[var1_init[0] - tot_update, var1_init[1] - tot_update],
dtype=dtype.as_numpy_dtype()),
var1.eval(),
rtol=1e-5)
def doTestBasic(self, use_resource=False):
num_updates = 4 # number of ADADELTA steps to perform
for dtype in [dtypes.half, dtypes.float32]:
for grad in [0.2, 0.1, 0.01]:
for lr in [1.0, 0.5, 0.1]:
with self.cached_session():
var0_init = [1.0, 2.0]
var1_init = [3.0, 4.0]
if use_resource:
var0 = resource_variable_ops.ResourceVariable(
var0_init, dtype=dtype)
var1 = resource_variable_ops.ResourceVariable(
var1_init, dtype=dtype)
else:
var0 = variables.Variable(var0_init, dtype=dtype)
var1 = variables.Variable(var1_init, dtype=dtype)
grads = constant_op.constant([grad, grad], dtype=dtype)
accum = 0.0
accum_update = 0.0
# ADADELTA gradient optimizer
rho = 0.95
epsilon = 1e-8
adadelta_opt = adadelta.Adadelta(lr, rho, epsilon)
adadelta_update = adadelta_opt.apply_gradients(
zip([grads, grads], [var0, var1]))
opt_vars = adadelta_opt.variables()
self.assertStartsWith(opt_vars[0].name,
var0._shared_name)
self.assertStartsWith(opt_vars[1].name,
var0._shared_name)
self.assertStartsWith(opt_vars[2].name,
var1._shared_name)
self.assertStartsWith(opt_vars[3].name,
var1._shared_name)
self.assertEqual(4, len(opt_vars))
variables.global_variables_initializer().run()
# Assign slots
slot = [None] * 2
slot_update = [None] * 2
self.assertEqual(["accum", "accum_update"],
adadelta_opt.get_slot_names())
slot[0] = adadelta_opt.get_slot(var0, "accum")
self.assertEquals(slot[0].get_shape(),
var0.get_shape())
self.assertFalse(
slot[0] in variables.trainable_variables())
slot_update[0] = adadelta_opt.get_slot(
var0, "accum_update")
self.assertEquals(slot_update[0].get_shape(),
var0.get_shape())
self.assertFalse(
slot_update[0] in variables.trainable_variables())
slot[1] = adadelta_opt.get_slot(var1, "accum")
self.assertEquals(slot[1].get_shape(),
var1.get_shape())
self.assertFalse(
slot[1] in variables.trainable_variables())
slot_update[1] = adadelta_opt.get_slot(
var1, "accum_update")
self.assertEquals(slot_update[1].get_shape(),
var1.get_shape())
self.assertFalse(
slot_update[1] in variables.trainable_variables())
# Fetch params to validate initial values
self.assertAllClose(var0_init, var0.eval())
self.assertAllClose(var1_init, var1.eval())
update = [None] * num_updates
tot_update = 0
for step in range(num_updates):
# Run adadelta update for comparison
adadelta_update.run()
# Perform initial update without previous accum values
accum = accum * rho + (grad**2) * (1 - rho)
update[step] = (np.sqrt(accum_update + epsilon) *
(1. / np.sqrt(accum + epsilon)) *
grad)
accum_update = (accum_update * rho +
(update[step]**2) * (1.0 - rho))
tot_update += update[step] * lr
# Check that the accumulators have been updated
for slot_idx in range(2):
self.assertAllCloseAccordingToType(
np.array([accum, accum],
dtype=dtype.as_numpy_dtype()),
slot[slot_idx].eval(),
rtol=1e-5)
self.assertAllCloseAccordingToType(
np.array([accum_update, accum_update],
dtype=dtype.as_numpy_dtype()),
slot_update[slot_idx].eval(),
rtol=1e-5)
# Check that the parameters have been updated
self.assertAllCloseAccordingToType(np.array(
[
var0_init[0] - tot_update,
var0_init[1] - tot_update
],
dtype=dtype.as_numpy_dtype()),
var0.eval(),
rtol=1e-5)
self.assertAllCloseAccordingToType(np.array(
[
var1_init[0] - tot_update,
var1_init[1] - tot_update
],
dtype=dtype.as_numpy_dtype()),
var1.eval(),
rtol=1e-5)
