def testProximalAdagradwithoutRegularization2(self):
with self.session(), self.test_scope():
var0 = resource_variable_ops.ResourceVariable([1.0, 2.0])
var1 = resource_variable_ops.ResourceVariable([4.0, 3.0])
grads0 = constant_op.constant([0.1, 0.2])
grads1 = constant_op.constant([0.01, 0.02])
opt = proximal_adagrad.ProximalAdagradOptimizer(
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]))
self.evaluate(variables.global_variables_initializer())
self.assertAllClose([1.0, 2.0], self.evaluate(var0))
self.assertAllClose([4.0, 3.0], self.evaluate(var1))
# Run 3 steps Proximal Adagrad.
for _ in range(3):
update.run()
self.assertAllClose(np.array([-1.60261, -2.296985]),
self.evaluate(var0))
self.assertAllClose(np.array([3.715679, 2.433051]),
self.evaluate(var1))
def testProximalAdagradWithL1_L2(self):
with self.test_session() as sess:
var0 = variables.Variable([1.0, 2.0])
var1 = variables.Variable([4.0, 3.0])
grads0 = constant_op.constant([0.1, 0.2])
grads1 = constant_op.constant([0.01, 0.02])
opt = proximal_adagrad.ProximalAdagradOptimizer(
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 = sess.run([var0, var1])
self.assertAllClose([1.0, 2.0], v0_val)
self.assertAllClose([4.0, 3.0], v1_val)
# Run 10 steps Proximal Adagrad.
for _ in range(10):
update.run()
v0_val, v1_val = sess.run([var0, var1])
self.assertAllClose(np.array([-0.0495, -0.0995]), v0_val)
self.assertAllClose(np.array([-0.0045, -0.0095]), v1_val)
def testResourceProximalAdagradwithoutRegularization(self):
with self.session(), self.test_scope():
var0 = resource_variable_ops.ResourceVariable([0.0, 0.0])
var1 = resource_variable_ops.ResourceVariable([0.0, 0.0])
grads0 = constant_op.constant([0.1, 0.2])
grads1 = constant_op.constant([0.01, 0.02])
opt = proximal_adagrad.ProximalAdagradOptimizer(
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]))
self.evaluate(variables.global_variables_initializer())
self.assertAllClose([0.0, 0.0], self.evaluate(var0))
self.assertAllClose([0.0, 0.0], self.evaluate(var1))
# Run 3 steps Proximal Adagrad.
for _ in range(3):
update.run()
self.assertAllClose(np.array([-2.60260963, -4.29698515]),
self.evaluate(var0))
self.assertAllClose(np.array([-0.28432083, -0.56694895]),
self.evaluate(var1))
opt_vars = opt.variables()
self.assertStartsWith(opt_vars[0].name, var0._shared_name)
self.assertStartsWith(opt_vars[1].name, var1._shared_name)
self.assertEqual(2, len(opt_vars))
def testProximalAdagradWithL1(self):
# ProximalAdagradOptimizer is supported only in V1.
with ops.Graph().as_default(), self.cached_session():
var0 = variables.Variable([1.0, 2.0])
var1 = variables.Variable([4.0, 3.0])
grads0 = constant_op.constant([0.1, 0.2])
grads1 = constant_op.constant([0.01, 0.02])
opt = proximal_adagrad.ProximalAdagradOptimizer(
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]))
self.evaluate(variables.global_variables_initializer())
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllClose([1.0, 2.0], v0_val)
self.assertAllClose([4.0, 3.0], v1_val)
# Run 10 steps Proximal Adagrad
for _ in range(10):
update.run()
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllClose(np.array([-6.663634, -9.190331]), v0_val)
self.assertAllClose(np.array([2.959304, 1.029232]), v1_val)
def testProximalAdagradwithoutRegularization(self):
with self.test_session() as sess:
var0 = variables.Variable([0.0, 0.0])
var1 = variables.Variable([0.0, 0.0])
grads0 = constant_op.constant([0.1, 0.2])
grads1 = constant_op.constant([0.01, 0.02])
opt = proximal_adagrad.ProximalAdagradOptimizer(
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 = sess.run([var0, var1])
self.assertAllClose([0.0, 0.0], v0_val)
self.assertAllClose([0.0, 0.0], v1_val)
# Run 3 steps Proximal Adagrad.
for _ in range(3):
update.run()
v0_val, v1_val = sess.run([var0, var1])
self.assertAllClose(np.array([-2.60260963, -4.29698515]), v0_val)
self.assertAllClose(np.array([-0.28432083, -0.56694895]), v1_val)
def doTestProximalAdagradwithoutRegularization(self, use_resource=False):
# ProximalAdagradOptimizer is supported only in V1.
with ops.Graph().as_default(), self.cached_session():
var0 = variables.Variable([0.0, 0.0])
var1 = variables.Variable([0.0, 0.0])
grads0 = constant_op.constant([0.1, 0.2])
grads1 = constant_op.constant([0.01, 0.02])
opt = proximal_adagrad.ProximalAdagradOptimizer(
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]))
self.evaluate(variables.global_variables_initializer())
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 Proximal Adagrad.
for _ in range(3):
update.run()
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllClose(np.array([-2.60260963, -4.29698515]), v0_val)
self.assertAllClose(np.array([-0.28432083, -0.56694895]), v1_val)
opt_vars = opt.variables()
self.assertStartsWith(opt_vars[0].name, var0._shared_name)
self.assertStartsWith(opt_vars[1].name, var1._shared_name)
self.assertEqual(2, len(opt_vars))
def testProximalAdagradWithL1(self):
with self.session(), self.test_scope():
var0 = resource_variable_ops.ResourceVariable([1.0, 2.0])
var1 = resource_variable_ops.ResourceVariable([4.0, 3.0])
grads0 = constant_op.constant([0.1, 0.2])
grads1 = constant_op.constant([0.01, 0.02])
opt = proximal_adagrad.ProximalAdagradOptimizer(
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]))
self.evaluate(variables.global_variables_initializer())
self.assertAllClose([1.0, 2.0], self.evaluate(var0))
self.assertAllClose([4.0, 3.0], self.evaluate(var1))
# Run 10 steps Proximal Adagrad
for _ in range(10):
update.run()
self.assertAllClose(np.array([-6.663634, -9.190331]),
self.evaluate(var0))
self.assertAllClose(np.array([2.959304, 1.029232]),
self.evaluate(var1))
def testProximalAdagradwithoutRegularization2(self):
# ProximalAdagradOptimizer is supported only in V1.
with ops.Graph().as_default(), self.cached_session():
var0 = variables.Variable([1.0, 2.0])
var1 = variables.Variable([4.0, 3.0])
grads0 = constant_op.constant([0.1, 0.2])
grads1 = constant_op.constant([0.01, 0.02])
opt = proximal_adagrad.ProximalAdagradOptimizer(
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]))
self.evaluate(variables.global_variables_initializer())
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllClose([1.0, 2.0], v0_val)
self.assertAllClose([4.0, 3.0], v1_val)
# Run 3 steps Proximal Adagrad.
for _ in range(3):
update.run()
v0_val, v1_val = self.evaluate([var0, var1])
self.assertAllClose(np.array([-1.60261, -2.296985]), v0_val)
self.assertAllClose(np.array([3.715679, 2.433051]), v1_val)
def test_ops_with_var_and_proximal_adagrad(self):
var_list = [
deo.get_variable('sp_var', initializer=0.0, dim=2),
]
opt_list = [
proximal_adagrad.ProximalAdagradOptimizer(0.1),
]
self.common_run_context(var_list,
opt_list,
name='proximal_adagrad_test')
def testEquivAdagradwithoutRegularization(self):
with self.session(), self.test_scope():
val0, val1 = self.applyOptimizer(
proximal_adagrad.ProximalAdagradOptimizer(
3.0,
initial_accumulator_value=0.1,
l1_regularization_strength=0.0,
l2_regularization_strength=0.0))
with self.session(), self.test_scope():
val2, val3 = self.applyOptimizer(
adagrad.AdagradOptimizer(3.0, initial_accumulator_value=0.1))
self.assertAllClose(val0, val2)
self.assertAllClose(val1, val3)
def testEquivAdagradwithoutRegularization(self):
# ProximalAdagradOptimizer is supported only in V1.
with ops.Graph().as_default(), self.cached_session():
val0, val1 = self.applyOptimizer(
proximal_adagrad.ProximalAdagradOptimizer(
3.0,
initial_accumulator_value=0.1,
l1_regularization_strength=0.0,
l2_regularization_strength=0.0))
with ops.Graph().as_default(), self.cached_session():
val2, val3 = self.applyOptimizer(
adagrad.AdagradOptimizer(3.0, initial_accumulator_value=0.1))
self.assertAllClose(val0, val2)
self.assertAllClose(val1, val3)
def testEquivSparseAdagradwithoutRegularization(self):
with self.cached_session():
val0, val1 = self.applyOptimizer(
proximal_adagrad.ProximalAdagradOptimizer(
3.0,
initial_accumulator_value=0.1,
l1_regularization_strength=0.0,
l2_regularization_strength=0.0),
is_sparse=True)
with self.cached_session():
val2, val3 = self.applyOptimizer(adagrad.AdagradOptimizer(
3.0, initial_accumulator_value=0.1),
is_sparse=True)
self.assertAllClose(val0, val2)
self.assertAllClose(val1, val3)
def testMinimizeSparseResourceVariable(self):
for dtype in [dtypes.float32, dtypes.float64]:
with self.test_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 = proximal_adagrad.ProximalAdagradOptimizer(
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([[0, 1]],
var0.eval(),
atol=0.01)
def test_proximal_adagrad_minimize_trainable(self):
base_opt = proximal_adagrad.ProximalAdagradOptimizer(1.0)
test_opt = proximal_adagrad.ProximalAdagradOptimizer(1.0)
self.common_minimize_trainable(base_opt,
test_opt,
name="proximal_adagrad")
def test_proximal_adagrad_apply_restriction(self):
opt = proximal_adagrad.ProximalAdagradOptimizer(0.1)
self.commonly_apply_restriction_verify(opt)
def test_proximal_adagrad_restrict_on_policy(self):
opt = proximal_adagrad.ProximalAdagradOptimizer(0.1)
self.common_single_step_restrict_verification(opt)
def test_proximal_adagrad_restrictor_update(self):
opt = proximal_adagrad.ProximalAdagradOptimizer(0.1)
self.common_single_step_update_verification(opt)
