def testFtrlWithL1_L2(self):
for dtype in [dtypes.half, dtypes.float32]:
with self.cached_session() as sess:
var0 = variables.Variable([1.0, 2.0], dtype=dtype)
var1 = variables.Variable([4.0, 3.0], dtype=dtype)
grads0 = constant_op.constant([0.1, 0.2], dtype=dtype)
grads1 = constant_op.constant([0.01, 0.02], dtype=dtype)
opt = ftrl.Ftrl(3.0,
initial_accumulator_value=0.1,
l1_regularization_strength=0.001,
l2_regularization_strength=2.0)
update = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]))
variables.global_variables_initializer().run()
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllCloseAccordingToType([1.0, 2.0], v0_val)
self.assertAllCloseAccordingToType([4.0, 3.0], v1_val)
# Run 10 steps FTRL
for _ in range(10):
update.run()
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllCloseAccordingToType(
np.array([-0.24059935, -0.46829352]), v0_val)
self.assertAllCloseAccordingToType(
np.array([-0.02406147, -0.04830509]), v1_val)
def testFtrlWithL1_L2_L2ShrinkageSparse(self):
"""Tests the new FTRL op with support for l2 shrinkage on sparse grads."""
for dtype in [dtypes.half, dtypes.float32]:
with self.cached_session() as sess:
var0 = variables.Variable([[1.0], [2.0]], dtype=dtype)
var1 = variables.Variable([[4.0], [3.0]], dtype=dtype)
grads0 = ops.IndexedSlices(
constant_op.constant([0.1], shape=[1, 1], dtype=dtype),
constant_op.constant([0]), constant_op.constant([2, 1]))
grads1 = ops.IndexedSlices(
constant_op.constant([0.02], shape=[1, 1], dtype=dtype),
constant_op.constant([1]), constant_op.constant([2, 1]))
opt = ftrl.Ftrl(3.0,
initial_accumulator_value=0.1,
l1_regularization_strength=0.001,
l2_regularization_strength=2.0,
l2_shrinkage_regularization_strength=0.1)
update = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]))
variables.global_variables_initializer().run()
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllCloseAccordingToType([[1.0], [2.0]], v0_val)
self.assertAllCloseAccordingToType([[4.0], [3.0]], v1_val)
# Run 10 steps FTRL
for _ in range(10):
update.run()
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllCloseAccordingToType([[-0.22578995], [2.]],
v0_val)
self.assertAllCloseAccordingToType([[4.], [-0.13229476]],
v1_val)
def testFtrlwithoutRegularization2(self):
for dtype in [dtypes.half, dtypes.float32]:
with self.cached_session() as sess:
var0 = variables.Variable([1.0, 2.0], dtype=dtype)
var1 = variables.Variable([4.0, 3.0], dtype=dtype)
grads0 = constant_op.constant([0.1, 0.2], dtype=dtype)
grads1 = constant_op.constant([0.01, 0.02], dtype=dtype)
opt = ftrl.Ftrl(3.0,
initial_accumulator_value=0.1,
l1_regularization_strength=0.0,
l2_regularization_strength=0.0)
update = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]))
variables.global_variables_initializer().run()
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllCloseAccordingToType([1.0, 2.0], v0_val)
self.assertAllCloseAccordingToType([4.0, 3.0], v1_val)
# Run 3 steps FTRL
for _ in range(3):
update.run()
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllCloseAccordingToType(
np.array([-2.55607247, -3.98729396]), v0_val)
self.assertAllCloseAccordingToType(
np.array([-0.28232238, -0.56096673]), v1_val)
def testFtrlWithL2_Beta(self):
# TODO(tanzheny, omalleyt): Fix test in eager mode.
for dtype in [dtypes.half, dtypes.float32]:
with ops.Graph().as_default(), self.cached_session():
var0 = variables.Variable([1.0, 2.0], dtype=dtype)
var1 = variables.Variable([4.0, 3.0], dtype=dtype)
grads0 = constant_op.constant([0.1, 0.2], dtype=dtype)
grads1 = constant_op.constant([0.01, 0.02], dtype=dtype)
opt = ftrl.Ftrl(
3.0,
initial_accumulator_value=0.1,
l1_regularization_strength=0.0,
l2_regularization_strength=0.1,
beta=0.1)
update = opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
self.evaluate(variables.global_variables_initializer())
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllCloseAccordingToType([1.0, 2.0], v0_val)
self.assertAllCloseAccordingToType([4.0, 3.0], v1_val)
# Run 10 steps FTRL
for _ in range(10):
update.run()
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllCloseAccordingToType(
np.array([-2.735487, -4.704625]), v0_val)
self.assertAllCloseAccordingToType(
np.array([-0.294335, -0.586556]), v1_val)
def testFtrlWithL1(self):
# TODO(tanzheny, omalleyt): Fix test in eager mode.
for dtype in [dtypes.half, dtypes.float32]:
with ops.Graph().as_default(), self.cached_session(use_gpu=True):
var0 = variables.Variable([1.0, 2.0], dtype=dtype)
var1 = variables.Variable([4.0, 3.0], dtype=dtype)
grads0 = constant_op.constant([0.1, 0.2], dtype=dtype)
grads1 = constant_op.constant([0.01, 0.02], dtype=dtype)
opt = ftrl.Ftrl(3.0,
initial_accumulator_value=0.1,
l1_regularization_strength=0.001,
l2_regularization_strength=0.0)
update = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]))
variables.global_variables_initializer().run()
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllCloseAccordingToType([1.0, 2.0], v0_val)
self.assertAllCloseAccordingToType([4.0, 3.0], v1_val)
# Run 10 steps FTRL
for _ in range(10):
update.run()
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllCloseAccordingToType(
np.array([-7.66718769, -10.91273689]), v0_val)
self.assertAllCloseAccordingToType(
np.array([-0.93460727, -1.86147261]), v1_val)
def testFtrlWithBeta(self):
# TODO(tanzheny, omalleyt): Fix test in eager mode.
for dtype in [dtypes.half, dtypes.float32]:
with ops.Graph().as_default(), self.cached_session(use_gpu=True):
var0 = variables.Variable([1.0, 2.0], dtype=dtype)
var1 = variables.Variable([4.0, 3.0], dtype=dtype)
grads0 = constant_op.constant([0.1, 0.2], dtype=dtype)
grads1 = constant_op.constant([0.01, 0.02], dtype=dtype)
opt = ftrl.Ftrl(3.0, initial_accumulator_value=0.1, beta=0.1)
update = opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
self.evaluate(variables.global_variables_initializer())
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllCloseAccordingToType([1.0, 2.0], v0_val)
self.assertAllCloseAccordingToType([4.0, 3.0], v1_val)
# Run 10 steps FTRL
for _ in range(10):
update.run()
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllCloseAccordingToType(
np.array([-6.096838, -9.162214]), v0_val)
self.assertAllCloseAccordingToType(
np.array([-0.717741, -1.425132]), v1_val)
def doTestFtrlwithoutRegularization(self, use_resource=False):
for dtype in [dtypes.half, dtypes.float32]:
with self.cached_session() as sess:
if use_resource:
var0 = resource_variable_ops.ResourceVariable([0.0, 0.0],
dtype=dtype)
var1 = resource_variable_ops.ResourceVariable([0.0, 0.0],
dtype=dtype)
else:
var0 = variables.Variable([0.0, 0.0], dtype=dtype)
var1 = variables.Variable([0.0, 0.0], dtype=dtype)
grads0 = constant_op.constant([0.1, 0.2], dtype=dtype)
grads1 = constant_op.constant([0.01, 0.02], dtype=dtype)
opt = ftrl.Ftrl(3.0,
initial_accumulator_value=0.1,
l1_regularization_strength=0.0,
l2_regularization_strength=0.0)
update = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]))
variables.global_variables_initializer().run()
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllClose([0.0, 0.0], v0_val)
self.assertAllClose([0.0, 0.0], v1_val)
# Run 3 steps FTRL
for _ in range(3):
update.run()
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllCloseAccordingToType(
np.array([-2.60260963, -4.29698515]), v0_val)
self.assertAllCloseAccordingToType(
np.array([-0.28432083, -0.56694895]), v1_val)
def testFtrlWithL2ShrinkageDoesNotChangeLrSchedule(self):
"""Verifies that l2 shrinkage in FTRL does not change lr schedule."""
# TODO(tanzheny, omalleyt): Fix test in eager mode.
for dtype in [dtypes.half, dtypes.float32]:
with ops.Graph().as_default(), self.cached_session() as sess:
var0 = variables.Variable([1.0, 2.0], dtype=dtype)
var1 = variables.Variable([1.0, 2.0], dtype=dtype)
grads0 = constant_op.constant([0.1, 0.2], dtype=dtype)
grads1 = constant_op.constant([0.1, 0.2], dtype=dtype)
opt0 = ftrl.Ftrl(
3.0,
initial_accumulator_value=0.1,
l1_regularization_strength=0.001,
l2_regularization_strength=2.0,
l2_shrinkage_regularization_strength=0.1)
opt1 = ftrl.Ftrl(
3.0,
initial_accumulator_value=0.1,
l1_regularization_strength=0.001,
l2_regularization_strength=2.0)
update0 = opt0.apply_gradients([(grads0, var0)])
update1 = opt1.apply_gradients([(grads1, var1)])
self.evaluate(variables.global_variables_initializer())
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllCloseAccordingToType([1.0, 2.0], v0_val)
self.assertAllCloseAccordingToType([1.0, 2.0], v1_val)
# Run 10 steps FTRL
for _ in range(10):
update0.run()
update1.run()
v0_val, v1_val = self.evaluate([var0, var1])
# var0 is experiencing L2 shrinkage so it should be smaller than var1
# in magnitude.
self.assertTrue((v0_val**2 < v1_val**2).all())
accum0 = sess.run(opt0.get_slot(var0, "accumulator"))
accum1 = sess.run(opt1.get_slot(var1, "accumulator"))
# L2 shrinkage should not change how we update grad accumulator.
self.assertAllCloseAccordingToType(accum0, accum1)
def test_numpy_input_fn_with_optimizer_instance(self):
"""Tests complete flow with optimizer_v2 instance."""
label_dimension = 2
batch_size = 10
train_input_fn, eval_input_fn, predict_input_fn = self._create_input_fn(
label_dimension, batch_size)
self._test_complete_flow(
train_input_fn=train_input_fn,
eval_input_fn=eval_input_fn,
predict_input_fn=predict_input_fn,
input_dimension=label_dimension,
label_dimension=label_dimension,
batch_size=batch_size,
optimizer=ftrl_v2.Ftrl(0.01)) # Test with optimizer_v2 instance
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 = ftrl.Ftrl(1.0).minimize(loss, var_list=[var0])
variables.global_variables_initializer().run()
# Fetch params to validate initial values
self.assertAllCloseAccordingToType([[1.0, 2.0]], self.evaluate(var0))
# Run 1 step of sgd
sgd_op.run()
# Validate updated params
self.assertAllCloseAccordingToType([[0, 1]],
self.evaluate(var0),
atol=0.01)
def testEquivGradientDescentwithoutRegularization(self):
for dtype in [dtypes.half, dtypes.float32]:
with self.cached_session():
val0, val1 = self.applyOptimizer(
ftrl.Ftrl(
3.0,
# Fixed learning rate
learning_rate_power=-0.0,
initial_accumulator_value=0.1,
l1_regularization_strength=0.0,
l2_regularization_strength=0.0),
dtype)
with self.cached_session():
val2, val3 = self.applyOptimizer(
gradient_descent.GradientDescentOptimizer(3.0), dtype)
self.assertAllCloseAccordingToType(val0, val2)
self.assertAllCloseAccordingToType(val1, val3)
def testEquivGradientDescentwithoutRegularization(self):
# TODO(tanzheny, omalleyt): Fix test in eager mode.
for dtype in [dtypes.half, dtypes.float32]:
with ops.Graph().as_default(), self.cached_session(use_gpu=True):
val0, val1 = self.applyOptimizer(
ftrl.Ftrl(
3.0,
# Fixed learning rate
learning_rate_power=-0.0,
initial_accumulator_value=0.1,
l1_regularization_strength=0.0,
l2_regularization_strength=0.0),
dtype)
with ops.Graph().as_default(), self.cached_session(use_gpu=True):
val2, val3 = self.applyOptimizer(
gradient_descent.GradientDescentOptimizer(3.0), dtype)
self.assertAllCloseAccordingToType(val0, val2)
self.assertAllCloseAccordingToType(val1, val3)
def testMinimizeSparseResourceVariable(self):
# TODO(tanzheny, omalleyt): Fix test in eager mode.
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with ops.Graph().as_default(), self.cached_session():
var0 = variables.Variable([[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 = ftrl.Ftrl(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
sgd_op.run()
# Validate updated params
self.assertAllCloseAccordingToType([[0, 1]],
self.evaluate(var0),
atol=0.01)
def testEquivAdagradwithoutRegularization(self):
for dtype in [dtypes.half, dtypes.float32]:
with self.cached_session(use_gpu=True):
val0, val1 = self.applyOptimizer(
ftrl.Ftrl(
3.0,
# Adagrad learning rate
learning_rate_power=-0.5,
initial_accumulator_value=0.1,
l1_regularization_strength=0.0,
l2_regularization_strength=0.0),
dtype)
with self.cached_session(use_gpu=True):
val2, val3 = self.applyOptimizer(
adagrad.AdagradOptimizer(3.0,
initial_accumulator_value=0.1),
dtype)
self.assertAllCloseAccordingToType(val0, val2)
self.assertAllCloseAccordingToType(val1, val3)
def testFtrlWithL1_L2_L2Shrinkage(self):
"""Test the new FTRL op with support for l2 shrinkage.
The addition of this parameter which places a constant pressure on weights
towards the origin causes the gradient descent trajectory to differ. The
weights will tend to have smaller magnitudes with this parameter set.
"""
# TODO(tanzheny, omalleyt): Fix test in eager mode.
for dtype in [dtypes.half, dtypes.float32]:
with ops.Graph().as_default(), self.cached_session(use_gpu=True):
var0 = variables.Variable([1.0, 2.0], dtype=dtype)
var1 = variables.Variable([4.0, 3.0], dtype=dtype)
grads0 = constant_op.constant([0.1, 0.2], dtype=dtype)
grads1 = constant_op.constant([0.01, 0.02], dtype=dtype)
opt = ftrl.Ftrl(3.0,
initial_accumulator_value=0.1,
l1_regularization_strength=0.001,
l2_regularization_strength=2.0,
l2_shrinkage_regularization_strength=0.1)
update = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]))
variables.global_variables_initializer().run()
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllCloseAccordingToType([1.0, 2.0], v0_val)
self.assertAllCloseAccordingToType([4.0, 3.0], v1_val)
# Run 10 steps FTRL
for _ in range(10):
update.run()
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllCloseAccordingToType(
np.array([-0.22578995, -0.44345796]), v0_val)
self.assertAllCloseAccordingToType(
np.array([-0.14378493, -0.13229476]), v1_val)
# 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))
# TODO(shiningsun): consider adding the other v2 optimizers
optimizers_v2 = [
gradient_descent_optimizer_keras_v2_fn, adagrad_optimizer_keras_v2_fn
]
optimizers_v1_and_v2 = optimizers_v1 + optimizers_v2
graph_and_eager_modes = ["graph", "eager"]
HP_IN_GRAPH = {
'Adam': ['decay', 'learning_rate'],
'Ftrl': [
'decay', 'l1_regularization_strength', 'l2_regularization_strength',
'learning_rate', 'learning_rate_power'
],
'RMSProp': ['decay', 'learning_rate', 'momentum', 'rho'],
'Adagrad': ['decay', 'learning_rate'],
'SGD': ['decay', 'learning_rate', 'momentum'],
}
# optimizer v2 instance.
OPT_V2_INSTANCE = {
'Adagrad': adagrad.Adagrad(),
'Adam': adam.Adam(),
'Ftrl': ftrl.Ftrl(),
'RMSProp': rmsprop.RMSprop(),
'SGD': gradient_descent.SGD(),
}
def _add_new_variable(initial_value, var_name_v2, var_name_v1, var_map,
var_names_map):
"""Creates a new variable and add it to the variable maps."""
var = tf.Variable(initial_value, name=var_name_v2)
var_map[var_name_v2] = var
var_names_map[var_name_v2] = var_name_v1
def _add_opt_variable(opt_name_v2, var_name_v1, idx, suffix_v2, reader,
var_map, var_names_map):
