def testSparseRepeatedIndices(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():
repeated_index_update_var = variables.Variable([[1.0], [2.0]],
dtype=dtype)
aggregated_update_var = variables.Variable([[1.0], [2.0]],
dtype=dtype)
grad_repeated_index = indexed_slices.IndexedSlices(
constant_op.constant([0.1, 0.1], shape=[2, 1],
dtype=dtype),
constant_op.constant([1, 1]), constant_op.constant([2, 1]))
grad_aggregated = indexed_slices.IndexedSlices(
constant_op.constant([0.2], shape=[1, 1], dtype=dtype),
constant_op.constant([1]), constant_op.constant([2, 1]))
repeated_update = adamax.Adamax().apply_gradients([
(grad_repeated_index, repeated_index_update_var)
])
aggregated_update = adamax.Adamax().apply_gradients([
(grad_aggregated, aggregated_update_var)
])
self.evaluate(variables.global_variables_initializer())
self.assertAllClose(aggregated_update_var,
repeated_index_update_var.eval())
for _ in range(3):
repeated_update.run()
aggregated_update.run()
self.assertAllClose(aggregated_update_var,
repeated_index_update_var.eval())
def testSparseRepeatedIndices(self):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.cached_session():
repeated_index_update_var = variables.Variable(
[[1.0], [2.0]], dtype=dtype)
aggregated_update_var = variables.Variable(
[[1.0], [2.0]], dtype=dtype)
grad_repeated_index = ops.IndexedSlices(
constant_op.constant(
[0.1, 0.1], shape=[2, 1], dtype=dtype),
constant_op.constant([1, 1]),
constant_op.constant([2, 1]))
grad_aggregated = ops.IndexedSlices(
constant_op.constant(
[0.2], shape=[1, 1], dtype=dtype),
constant_op.constant([1]),
constant_op.constant([2, 1]))
repeated_update = adamax.Adamax().apply_gradients(
[(grad_repeated_index, repeated_index_update_var)])
aggregated_update = adamax.Adamax().apply_gradients(
[(grad_aggregated, aggregated_update_var)])
variables.global_variables_initializer().run()
self.assertAllClose(aggregated_update_var.eval(),
repeated_index_update_var.eval())
for _ in range(3):
repeated_update.run()
aggregated_update.run()
self.assertAllClose(aggregated_update_var.eval(),
repeated_index_update_var.eval())
def testConstructAdamaxWithLR(self): opt = adamax.Adamax(lr=1.0) self.assertEqual(opt.lr, 1.0) opt_2 = adamax.Adamax(learning_rate=0.1, lr=1.0) self.assertEqual(opt_2.lr, 1.0) opt_3 = adamax.Adamax(learning_rate=0.1) self.assertEqual(opt_3.lr, 0.1)
def testConstructAdamaxWithEpsilonValues(self):
opt = adamax.Adamax(epsilon=None)
config = opt.get_config()
self.assertEqual(config["epsilon"], 1e-7)
opt = adamax.Adamax(epsilon=1e-8)
config = opt.get_config()
self.assertEqual(config["epsilon"], 1e-8)
def testConstructAdamaxWithLR(self):
opt = adamax.Adamax(lr=1.0)
opt_2 = adamax.Adamax(learning_rate=0.1, lr=1.0)
opt_3 = adamax.Adamax(learning_rate=0.1)
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 testTensorLearningRate(self):
for dtype in [dtypes.half, dtypes.float32, dtypes.float64]:
with self.cached_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 = adamax.Adamax(constant_op.constant(0.001))
update = opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
variables.global_variables_initializer().run()
# Fetch params to validate initial values
self.assertAllClose([1.0, 2.0], var0.eval())
self.assertAllClose([3.0, 4.0], var1.eval())
beta1_power = get_beta_accumulators(opt, dtype)
# Run 3 steps of Adamax
for t in range(3):
self.assertAllCloseAccordingToType(0.9**(t + 1), beta1_power.eval())
update.run()
var0_np, m0, v0 = adamax_update_numpy(var0_np, grads0_np, t, m0, v0)
var1_np, m1, v1 = adamax_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 testSlotsUniqueEager(self):
v1 = variables.Variable(1.)
v2 = variables.Variable(1.)
opt = adamax.Adamax(1.)
opt.minimize(lambda: v1 + v2, var_list=[v1, v2])
# There should be iteration, and two unique slot variables for v1 and v2.
self.assertLen({id(v) for v in opt.variables()}, 5)
def testSlotsUniqueEager(self):
with context.eager_mode():
v1 = resource_variable_ops.ResourceVariable(1.)
v2 = resource_variable_ops.ResourceVariable(1.)
opt = adamax.Adamax(1.)
opt.minimize(lambda: v1 + v2, var_list=[v1, v2])
# There should be iteration, and two unique slot variables for v1 and v2.
self.assertEqual(5, len(set(opt.variables())))
def testBasic(self):
for i, dtype in enumerate(
[dtypes.half, dtypes.float32, dtypes.float64]):
with self.session(graph=ops.Graph(), use_gpu=True):
# Initialize variables for numpy implementation.
m0 = np.array([0.0, 0.0])
v0 = np.array([0.0, 0.0])
m1 = np.array([0.0, 0.0])
v1 = np.array([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 = resource_variable_ops.ResourceVariable(var0_np,
name="var0_%d" %
i)
var1 = resource_variable_ops.ResourceVariable(var1_np,
name="var1_%d" %
i)
grads0 = constant_op.constant(grads0_np)
grads1 = constant_op.constant(grads1_np)
opt = adamax.Adamax()
if not context.executing_eagerly():
update = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]))
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))
# Run 3 steps of Adamax
for t in range(3):
beta_1_power = get_beta_accumulators(opt, dtype)
self.assertAllCloseAccordingToType(
0.9**(t + 1), self.evaluate(beta_1_power))
if not context.executing_eagerly():
self.evaluate(update)
else:
opt.apply_gradients(zip([grads0, grads1],
[var0, var1]))
var0_np, m0, v0 = adamax_update_numpy(
var0_np, grads0_np, t, m0, v0)
var1_np, m1, v1 = adamax_update_numpy(
var1_np, grads1_np, t, m1, v1)
# Validate updated params
self.assertAllCloseAccordingToType(var0_np,
self.evaluate(var0),
rtol=1e-2)
self.assertAllCloseAccordingToType(var1_np,
self.evaluate(var1),
rtol=1e-2)
def testSparseDevicePlacement(self):
for index_dtype in [dtypes.int32, dtypes.int64]:
with self.cached_session(force_gpu=test.is_gpu_available()):
# If a GPU is available, tests that all optimizer ops can be placed on
# it (i.e. they have GPU kernels).
var = variables.Variable([[1.0], [2.0]])
indices = constant_op.constant([0, 1], dtype=index_dtype)
g_sum = lambda: math_ops.reduce_sum(array_ops.gather(var, indices)) # pylint: disable=cell-var-from-loop
optimizer = adamax.Adamax(3.0)
minimize_op = optimizer.minimize(g_sum, var_list=[var])
variables.global_variables_initializer().run()
minimize_op.run()
def testResourceSparse(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(use_gpu=True):
# Initialize variables for numpy implementation.
zero_slots = lambda: np.zeros((3), dtype=dtype.as_numpy_dtype) # pylint: disable=cell-var-from-loop
m0, v0, m1, v1 = zero_slots(), zero_slots(), zero_slots(
), zero_slots()
var0_np = np.array([1.0, 2.0, 3.0], dtype=dtype.as_numpy_dtype)
grads0_np = np.array([0.1, 0.1], dtype=dtype.as_numpy_dtype)
var1_np = np.array([4.0, 5.0, 6.0], dtype=dtype.as_numpy_dtype)
grads1_np = np.array([0.01, 0.01], dtype=dtype.as_numpy_dtype)
var0 = resource_variable_ops.ResourceVariable(var0_np)
var1 = resource_variable_ops.ResourceVariable(var1_np)
grads0_np_indices = np.array([0, 1], dtype=np.int32)
grads0 = ops.IndexedSlices(
constant_op.constant(grads0_np),
constant_op.constant(grads0_np_indices),
constant_op.constant([3]))
grads1_np_indices = np.array([2, 1], dtype=np.int32)
grads1 = ops.IndexedSlices(
constant_op.constant(grads1_np),
constant_op.constant(grads1_np_indices),
constant_op.constant([3]))
opt = adamax.Adamax()
update = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]))
variables.global_variables_initializer().run()
# Fetch params to validate initial values
self.assertAllClose([1.0, 2.0, 3.0], var0.eval())
self.assertAllClose([4.0, 5.0, 6.0], var1.eval())
beta1_power = get_beta_accumulators(opt, dtype)
# Run 3 steps of Adamax
for t in range(3):
self.assertAllCloseAccordingToType(0.9**(t + 1),
beta1_power.eval())
update.run()
var0_np, m0, v0 = adamax_sparse_update_numpy(
var0_np, grads0_np_indices, grads0_np, t, m0, v0)
var1_np, m1, v1 = adamax_sparse_update_numpy(
var1_np, grads1_np_indices, grads1_np, t, m1, v1)
# Validate updated params
self.assertAllCloseAccordingToType(var0_np, var0.eval())
self.assertAllCloseAccordingToType(var1_np, var1.eval())
def testSharing(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(use_gpu=True):
# 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 = adamax.Adamax()
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 = get_beta_accumulators(opt, dtype)
# 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 Adamax1 and Adamax2.
for t in range(3):
self.assertAllCloseAccordingToType(0.9**(t + 1),
beta1_power.eval())
if t % 2 == 0:
update1.run()
else:
update2.run()
var0_np, m0, v0 = adamax_update_numpy(
var0_np, grads0_np, t, m0, v0)
var1_np, m1, v1 = adamax_update_numpy(
var1_np, grads1_np, t, m1, v1)
# Validate updated params
self.assertAllCloseAccordingToType(var0_np, var0.eval())
self.assertAllCloseAccordingToType(var1_np, var1.eval())
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))
# 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
def testAdamaxCompatibility(self):
opt_v1 = optimizers.Adamax(lr=0.01)
opt_v2 = adamax.Adamax(learning_rate=0.01)
self._testOptimizersCompatibility(opt_v1, opt_v2)