def testCheckpointOptimizer(self):
x = tf.Variable([[1.0, 2.0], [3.0, 4.0]], dtype=tf.float32)
lr_schedule = learning_rate_schedule.ExponentialDecay(
initial_learning_rate=1e-2, decay_steps=10000, decay_rate=0.9)
optimizer_1 = adam_new.Adam(learning_rate=lr_schedule,
beta_1=0.8,
beta_2=0.888)
grads = tf.convert_to_tensor([[1.0, 2.0], [3.0, 4.0]])
for _ in range(1):
optimizer_1.apply_gradients(zip([grads], [x]))
# Then save the variable and optimizer to a checkpoint.
checkpoint_1 = tf.train.Checkpoint(var=x, optimizer=optimizer_1)
checkpoint_path = checkpoint_1.save(self.get_temp_dir())
# Create a new optimizer and call restore on it (and x)
x2 = tf.Variable([[0.0, 0.0], [0.0, 0.0]], dtype=x.dtype)
optimizer_2 = adam_new.Adam(learning_rate=0.02,
beta_1=0.7,
beta_2=0.777)
optimizer_2.build([x2])
checkpoint_2 = tf.train.Checkpoint(var=x2, optimizer=optimizer_2)
checkpoint_2.restore(checkpoint_path)
self.assertTrue((self.evaluate(
optimizer_1._momentums._storage[0]) == self.evaluate(
optimizer_2._momentums._storage[0])).all())
self.assertEqual(
self.evaluate(optimizer_1._iterations),
self.evaluate(optimizer_2._iterations),
)
def testContinuous(self, serialize):
self.evaluate(tf.compat.v1.global_variables_initializer())
step = 5
decayed_lr = learning_rate_schedule.ExponentialDecay(0.05, 10, 0.96)
decayed_lr = _maybe_serialized(decayed_lr, serialize)
expected = 0.05 * 0.96**(5.0 / 10.0)
self.assertAllClose(self.evaluate(decayed_lr(step)), expected, 1e-6)
def testSetLearningRate(self):
optimizer = adam_new.Adam(learning_rate=1.0)
self.assertIsInstance(optimizer._learning_rate, tf.Variable)
self.assertEqual(self.evaluate(optimizer.learning_rate), 1.0)
optimizer.learning_rate = 2.0
self.assertEqual(self.evaluate(optimizer.learning_rate), 2.0)
# Test the legacy setter.
optimizer.lr = 3.0
self.assertEqual(self.evaluate(optimizer.learning_rate), 3.0)
lr_schedule = learning_rate_schedule.ExponentialDecay(
initial_learning_rate=1e-2, decay_steps=10000, decay_rate=0.9)
optimizer = adam_new.Adam(learning_rate=lr_schedule)
self.assertIsInstance(optimizer._learning_rate,
learning_rate_schedule.ExponentialDecay)
self.assertEqual(optimizer.learning_rate, 0.01)
# Test the legacy property.
self.assertEqual(optimizer.lr, 0.01)
x = tf.Variable([1.0, 2.0], dtype=tf.float32)
grads = tf.convert_to_tensor([1.0, 2.0])
for _ in range(2):
optimizer.apply_gradients(zip([grads], [x]))
self.assertTrue(optimizer.learning_rate < 0.01
and optimizer.learning_rate > 0.00999)
with self.assertRaisesRegex(TypeError,
"This optimizer was created with*"):
optimizer.learning_rate = 2.0
def testVariables(self, serialize):
# TODO(tanzheny, omalleyt): Fix test in eager mode.
with tf.Graph().as_default():
step = tf.Variable(1)
assign_1 = step.assign(1)
assign_2 = step.assign(2)
assign_100 = step.assign(100)
decayed_lr = learning_rate_schedule.ExponentialDecay(
.1, 3, 0.96, staircase=True)
decayed_lr = _maybe_serialized(decayed_lr, serialize)
self.evaluate(tf.compat.v1.global_variables_initializer())
# No change to learning rate
self.evaluate(assign_1.op)
self.assertAllClose(self.evaluate(decayed_lr(step)), .1, 1e-6)
self.evaluate(assign_2.op)
self.assertAllClose(self.evaluate(decayed_lr(step)), .1, 1e-6)
# Decayed learning rate
self.evaluate(assign_100.op)
expected = .1 * 0.96**(100 // 3)
self.assertAllClose(self.evaluate(decayed_lr(step)), expected, 1e-6)
def testStaircase(self, serialize):
if tf.executing_eagerly():
step = tf.Variable(0)
self.evaluate(tf.compat.v1.global_variables_initializer())
decayed_lr = learning_rate_schedule.ExponentialDecay(
.1, 3, 0.96, staircase=True)
decayed_lr = _maybe_serialized(decayed_lr, serialize)
# No change to learning rate due to staircase
expected = .1
self.evaluate(step.assign(1))
self.assertAllClose(self.evaluate(decayed_lr(step)), expected, 1e-6)
expected = .1
self.evaluate(step.assign(2))
self.assertAllClose(self.evaluate(decayed_lr(step)), expected, 1e-6)
# Decayed learning rate
expected = .1 * 0.96 ** (100 // 3)
self.evaluate(step.assign(100))
self.assertAllClose(self.evaluate(decayed_lr(step)), expected, 1e-6)
def natural_exp_decay(
learning_rate,
global_step,
decay_steps,
decay_rate,
staircase=False,
name=None,
):
"""Applies natural exponential decay to the initial learning rate.
When training a model, it is often recommended to lower the learning rate as
the training progresses. This function applies an exponential decay function
to a provided initial learning rate. It requires an `global_step` value to
compute the decayed learning rate. You can just pass a TensorFlow variable
that you increment at each training step.
The function returns the decayed learning rate. It is computed as:
```python
decayed_learning_rate = learning_rate * exp(-decay_rate * global_step /
decay_step)
```
or, if `staircase` is `True`, as:
```python
decayed_learning_rate = learning_rate * exp(-decay_rate * floor(global_step /
decay_step))
```
Example: decay exponentially with a base of 0.96:
```python
...
global_step = tf.Variable(0, trainable=False)
learning_rate = 0.1
decay_steps = 5
k = 0.5
learning_rate = tf.compat.v1.train.natural_exp_decay(learning_rate,
global_step,
decay_steps, k)
# Passing global_step to minimize() will increment it at each step.
learning_step = (
tf.compat.v1.train.GradientDescentOptimizer(learning_rate)
.minimize(...my loss..., global_step=global_step)
)
```
Args:
learning_rate: A scalar `float32` or `float64` `Tensor` or a Python number.
The initial learning rate.
global_step: A Python number. Global step to use for the decay computation.
Must not be negative.
decay_steps: How often to apply decay.
decay_rate: A Python number. The decay rate.
staircase: Whether to apply decay in a discrete staircase, as opposed to
continuous, fashion.
name: String. Optional name of the operation. Defaults to
'ExponentialTimeDecay'.
Returns:
A scalar `Tensor` of the same type as `learning_rate`. The decayed
learning rate.
Raises:
ValueError: if `global_step` is not supplied.
@compatibility(eager)
When eager execution is enabled, this function returns a function which in
turn returns the decayed learning rate Tensor. This can be useful for changing
the learning rate value across different invocations of optimizer functions.
@end_compatibility
"""
natural_exp_rate = tf.exp(tf.negative(decay_rate))
decayed_lr = learning_rate_schedule.ExponentialDecay(
learning_rate,
decay_steps,
natural_exp_rate,
staircase=staircase,
name=name,
)
if not tf.executing_eagerly():
decayed_lr = decayed_lr(global_step)
else:
decayed_lr = functools.partial(decayed_lr, global_step)
return decayed_lr
def exponential_decay(
learning_rate,
global_step,
decay_steps,
decay_rate,
staircase=False,
name=None,
):
"""Applies exponential decay to the learning rate.
When training a model, it is often recommended to lower the learning rate as
the training progresses. This function applies an exponential decay function
to a provided initial learning rate. It requires a `global_step` value to
compute the decayed learning rate. You can just pass a TensorFlow variable
that you increment at each training step.
The function returns the decayed learning rate. It is computed as:
```python
decayed_learning_rate = learning_rate *
decay_rate ^ (global_step / decay_steps)
```
If the argument `staircase` is `True`, then `global_step / decay_steps` is an
integer division and the decayed learning rate follows a staircase function.
Example: decay every 100000 steps with a base of 0.96:
```python
...
global_step = tf.Variable(0, trainable=False)
starter_learning_rate = 0.1
learning_rate = tf.compat.v1.train.exponential_decay(starter_learning_rate,
global_step,
100000, 0.96, staircase=True)
# Passing global_step to minimize() will increment it at each step.
learning_step = (
tf.compat.v1.train.GradientDescentOptimizer(learning_rate)
.minimize(...my loss..., global_step=global_step)
)
```
Args:
learning_rate: A scalar `float32` or `float64` `Tensor` or a Python number.
The initial learning rate.
global_step: A scalar `int32` or `int64` `Tensor` or a Python number. Global
step to use for the decay computation. Must not be negative.
decay_steps: A scalar `int32` or `int64` `Tensor` or a Python number. Must
be positive. See the decay computation above.
decay_rate: A scalar `float32` or `float64` `Tensor` or a Python number.
The decay rate.
staircase: Boolean. If `True` decay the learning rate at discrete intervals
name: String. Optional name of the operation. Defaults to
'ExponentialDecay'.
Returns:
A scalar `Tensor` of the same type as `learning_rate`. The decayed
learning rate.
Raises:
ValueError: if `global_step` is not supplied.
@compatibility(eager)
When eager execution is enabled, this function returns a function which in
turn returns the decayed learning rate Tensor. This can be useful for changing
the learning rate value across different invocations of optimizer functions.
@end_compatibility
"""
decayed_lr = learning_rate_schedule.ExponentialDecay(learning_rate,
decay_steps,
decay_rate,
staircase=staircase,
name=name)
if not tf.executing_eagerly():
decayed_lr = decayed_lr(global_step)
else:
decayed_lr = functools.partial(decayed_lr, global_step)
return decayed_lr
