def test_sparse_repeated_indices(dtype):
# todo: remove the with tf.device once the placement on cpu is enforced.
with tf.device("CPU:0"):
repeated_index_update_var = tf.Variable([[1], [2]], dtype=dtype)
aggregated_update_var = tf.Variable([[1], [2]], dtype=dtype)
grad_repeated_index = tf.IndexedSlices(
tf.constant([0.1, 0.1], shape=[2, 1], dtype=dtype),
tf.constant([1, 1]),
tf.constant([2, 1]),
)
grad_aggregated = tf.IndexedSlices(
tf.constant([0.2], shape=[1, 1], dtype=dtype),
tf.constant([1]),
tf.constant([2, 1]),
)
repeated_update_opt = lazy_adam.LazyAdam()
aggregated_update_opt = lazy_adam.LazyAdam()
for _ in range(3):
repeated_update_opt.apply_gradients(
[(grad_repeated_index, repeated_index_update_var)]
)
aggregated_update_opt.apply_gradients(
[(grad_aggregated, aggregated_update_var)]
)
np.testing.assert_allclose(
aggregated_update_var.numpy(), repeated_index_update_var.numpy()
)
def testSparseRepeatedIndices(self):
for dtype in [tf.dtypes.half, tf.dtypes.float32, tf.dtypes.float64]:
with self.cached_session():
repeated_index_update_var = tf.Variable([[1.0], [2.0]],
dtype=dtype)
aggregated_update_var = tf.Variable([[1.0], [2.0]],
dtype=dtype)
grad_repeated_index = tf.IndexedSlices(
tf.constant([0.1, 0.1], shape=[2, 1], dtype=dtype),
tf.constant([1, 1]),
tf.constant([2, 1]),
)
grad_aggregated = tf.IndexedSlices(
tf.constant([0.2], shape=[1, 1], dtype=dtype),
tf.constant([1]),
tf.constant([2, 1]),
)
repeated_update_opt = lazy_adam.LazyAdam()
repeated_update = repeated_update_opt.apply_gradients([
(grad_repeated_index, repeated_index_update_var)
])
aggregated_update_opt = lazy_adam.LazyAdam()
aggregated_update = aggregated_update_opt.apply_gradients([
(grad_aggregated, aggregated_update_var)
])
self.evaluate(tf.compat.v1.global_variables_initializer())
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 test_sharing(dtype):
# Initialize tf 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 = tf.Variable(var0_np)
var1 = tf.Variable(var1_np)
grads0 = tf.constant(grads0_np)
grads1 = tf.constant(grads1_np)
opt = lazy_adam.LazyAdam()
# Fetch params to validate initial values
np.testing.assert_allclose([1.0, 2.0], var0.numpy())
np.testing.assert_allclose([3.0, 4.0], var1.numpy())
# Run 3 steps of intertwined Adam1 and Adam2.
for t in range(3):
beta_1_power, beta_2_power = get_beta_accumulators(opt, dtype)
test_utils.assert_allclose_according_to_type(0.9 ** (t + 1), beta_1_power)
test_utils.assert_allclose_according_to_type(0.999 ** (t + 1), beta_2_power)
opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
var0_np, m0, v0 = adam_update_numpy(var0_np, grads0_np, t, m0, v0)
var1_np, m1, v1 = adam_update_numpy(var1_np, grads1_np, t, m1, v1)
# Validate updated params
test_utils.assert_allclose_according_to_type(var0_np, var0.numpy())
test_utils.assert_allclose_according_to_type(var1_np, var1.numpy())
def test_basic(use_callable_params, dtype):
# Initialize tf 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 = tf.Variable(var0_np)
var1 = tf.Variable(var1_np)
grads0 = tf.constant(grads0_np)
grads1 = tf.constant(grads1_np)
def learning_rate():
return 0.001
if not use_callable_params:
learning_rate = learning_rate()
opt = lazy_adam.LazyAdam(learning_rate=learning_rate)
# Run 3 steps of Adam
for t in range(3):
beta_1_power, beta_2_power = get_beta_accumulators(opt, dtype)
test_utils.assert_allclose_according_to_type(0.9 ** (t + 1), beta_1_power)
test_utils.assert_allclose_according_to_type(0.999 ** (t + 1), beta_2_power)
opt.apply_gradients(zip([grads0, grads1], [var0, var1]))
var0_np, m0, v0 = adam_update_numpy(var0_np, grads0_np, t, m0, v0)
var1_np, m1, v1 = adam_update_numpy(var1_np, grads1_np, t, m1, v1)
# Validate updated params
test_utils.assert_allclose_according_to_type(var0_np, var0.numpy())
test_utils.assert_allclose_according_to_type(var1_np, var1.numpy())
def test_slots_unique_eager():
v1 = tf.Variable(1.0)
v2 = tf.Variable(1.0)
opt = lazy_adam.LazyAdam(1.0)
opt.minimize(lambda: v1 + v2, var_list=[v1, v2])
# There should be iteration, and two unique slot variables for v1 and v2.
assert 5 == len(opt.variables())
assert opt.variables()[0] == opt.iterations
def testSlotsUniqueEager(self):
v1 = tf.Variable(1.0)
v2 = tf.Variable(1.0)
opt = lazy_adam.LazyAdam(1.0)
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(opt.variables()))
self.assertEqual(self.evaluate(opt.variables()[0]),
self.evaluate(opt.iterations))
def doTestBasic(self, use_callable_params=False):
for i, dtype in enumerate(
[tf.dtypes.half, tf.dtypes.float32, tf.dtypes.float64]):
with self.session(graph=tf.Graph()):
# Initialize tf 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 = tf.Variable(var0_np, name="var0_%d" % i)
var1 = tf.Variable(var1_np, name="var1_%d" % i)
grads0 = tf.constant(grads0_np)
grads1 = tf.constant(grads1_np)
def learning_rate():
return 0.001
if not use_callable_params:
learning_rate = learning_rate()
opt = lazy_adam.LazyAdam(learning_rate=learning_rate)
if not tf.executing_eagerly():
update = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]))
self.evaluate(tf.compat.v1.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 Adam
for t in range(3):
beta_1_power, beta_2_power = get_beta_accumulators(
opt, dtype)
self.assertAllCloseAccordingToType(
0.9**(t + 1), self.evaluate(beta_1_power))
self.assertAllCloseAccordingToType(
0.999**(t + 1), self.evaluate(beta_2_power))
if not tf.executing_eagerly():
self.evaluate(update)
else:
opt.apply_gradients(zip([grads0, grads1],
[var0, var1]))
var0_np, m0, v0 = adam_update_numpy(
var0_np, grads0_np, t, m0, v0)
var1_np, m1, v1 = adam_update_numpy(
var1_np, grads1_np, t, m1, v1)
# Validate updated params
self.assertAllCloseAccordingToType(var0_np,
self.evaluate(var0))
self.assertAllCloseAccordingToType(var1_np,
self.evaluate(var1))
self.assertEqual("LazyAdam/var0_%d/m:0" % (i, ),
opt.get_slot(var0, "m").name)
def testSparse(self):
for dtype in [tf.dtypes.half, tf.dtypes.float32, tf.dtypes.float64]:
with self.cached_session():
# Initialize tf for numpy implementation.
m0, v0, m1, v1 = 0.0, 0.0, 0.0, 0.0
var0_np = np.array([1.0, 1.0, 2.0], dtype=dtype.as_numpy_dtype)
grads0_np = np.array([0.1, 0.0, 0.1],
dtype=dtype.as_numpy_dtype)
var1_np = np.array([3.0, 3.0, 4.0], dtype=dtype.as_numpy_dtype)
grads1_np = np.array([0.01, 0.0, 0.01],
dtype=dtype.as_numpy_dtype)
var0 = tf.Variable(var0_np)
var1 = tf.Variable(var1_np)
grads0_np_indices = np.array([0, 2], dtype=np.int32)
grads0 = tf.IndexedSlices(
tf.constant(grads0_np[grads0_np_indices]),
tf.constant(grads0_np_indices),
tf.constant([3]),
)
grads1_np_indices = np.array([0, 2], dtype=np.int32)
grads1 = tf.IndexedSlices(
tf.constant(grads1_np[grads1_np_indices]),
tf.constant(grads1_np_indices),
tf.constant([3]),
)
opt = lazy_adam.LazyAdam()
update = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]))
self.evaluate(tf.compat.v1.global_variables_initializer())
# Fetch params to validate initial values
self.assertAllClose([1.0, 1.0, 2.0], self.evaluate(var0))
self.assertAllClose([3.0, 3.0, 4.0], self.evaluate(var1))
beta_1_power, beta_2_power = get_beta_accumulators(opt, dtype)
# Run 3 steps of Adam
for t in range(3):
self.assertAllCloseAccordingToType(
0.9**(t + 1), self.evaluate(beta_1_power))
self.assertAllCloseAccordingToType(
0.999**(t + 1), self.evaluate(beta_2_power))
self.evaluate(update)
var0_np, m0, v0 = adam_update_numpy(
var0_np, grads0_np, t, m0, v0)
var1_np, m1, v1 = adam_update_numpy(
var1_np, grads1_np, t, m1, v1)
# Validate updated params
self.assertAllCloseAccordingToType(var0_np,
self.evaluate(var0))
self.assertAllCloseAccordingToType(var1_np,
self.evaluate(var1))
def testSparseDevicePlacement(self):
for index_dtype in [tf.dtypes.int32, tf.dtypes.int64]:
with self.cached_session(force_gpu=tf.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 = tf.Variable([[1.0], [2.0]])
indices = tf.constant([0, 1], dtype=index_dtype)
g_sum = lambda: tf.math.reduce_sum(tf.gather(var, indices)) # pylint: disable=cell-var-from-loop
optimizer = lazy_adam.LazyAdam(3.0)
minimize_op = optimizer.minimize(g_sum, var_list=[var])
self.evaluate(tf.compat.v1.global_variables_initializer())
self.evaluate(minimize_op)
def test_sparse_device_placement(dtype):
# If a GPU is available, tests that all optimizer ops can be placed on
# it (i.e. they have GPU kernels).
var = tf.Variable([[1.0], [2.0]])
indices = tf.constant([0, 1], dtype=dtype)
def g_sum():
return tf.math.reduce_sum(tf.gather(var, indices))
optimizer = lazy_adam.LazyAdam(3.0)
optimizer.minimize(g_sum, var_list=[var])
def testSharing(self):
for dtype in [tf.dtypes.half, tf.dtypes.float32, tf.dtypes.float64]:
with self.cached_session():
# Initialize tf 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 = tf.Variable(var0_np)
var1 = tf.Variable(var1_np)
grads0 = tf.constant(grads0_np)
grads1 = tf.constant(grads1_np)
opt = lazy_adam.LazyAdam()
update1 = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]))
update2 = opt.apply_gradients(
zip([grads0, grads1], [var0, var1]))
self.evaluate(variables.global_variables_initializer())
beta_1_power, beta_2_power = get_beta_accumulators(opt, dtype)
# 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 intertwined Adam1 and Adam2.
for t in range(3):
self.assertAllCloseAccordingToType(
0.9**(t + 1), self.evaluate(beta_1_power))
self.assertAllCloseAccordingToType(
0.999**(t + 1), self.evaluate(beta_2_power))
if t % 2 == 0:
update1.run()
else:
update2.run()
var0_np, m0, v0 = adam_update_numpy(
var0_np, grads0_np, t, m0, v0)
var1_np, m1, v1 = adam_update_numpy(
var1_np, grads1_np, t, m1, v1)
# Validate updated params
self.assertAllCloseAccordingToType(var0_np,
self.evaluate(var0))
self.assertAllCloseAccordingToType(var1_np,
self.evaluate(var1))
def test_serialization():
optimizer = lazy_adam.LazyAdam()
config = tf.keras.optimizers.serialize(optimizer)
new_optimizer = tf.keras.optimizers.deserialize(config)
assert new_optimizer.get_config() == optimizer.get_config()