Adafactor
classkeras.optimizers.Adafactor(
learning_rate=0.001,
beta_2_decay=-0.8,
epsilon_1=1e-30,
epsilon_2=0.001,
clip_threshold=1.0,
relative_step=True,
weight_decay=None,
clipnorm=None,
clipvalue=None,
global_clipnorm=None,
use_ema=False,
ema_momentum=0.99,
ema_overwrite_frequency=None,
loss_scale_factor=None,
gradient_accumulation_steps=None,
name="adafactor",
**kwargs
)
Optimizer that implements the Adafactor algorithm.
Adafactor is commonly used in NLP tasks, and has the advantage of taking less memory because it only saves partial information of previous gradients.
The default argument setup is based on the original paper (see reference). When gradients are of dimension > 2, Adafactor optimizer will delete the last 2 dimensions separately in its accumulator variables.
Arguments
keras.optimizers.schedules.LearningRateSchedule
instance, or
a callable that takes no arguments and returns the actual value to
use. The learning rate. Defaults to 0.001
.beta_2
.clipnorm
,
clipvalue
, and global_clipnorm
.True
. If learning_rate
is a
constant and relative_step=True
, learning rate will be adjusted
based on current iterations. This is a default learning rate decay
in Adafactor.False
.
If True
, exponential moving average
(EMA) is applied. EMA consists of computing an exponential moving
average of the weights of the model (as the weight values change
after each training batch), and periodically overwriting the
weights with their moving average.use_ema=True
.
This is the momentum to use when computing
the EMA of the model's weights:
new_average = ema_momentum * old_average + (1 - ema_momentum) *
current_variable_value
.use_ema=True
. Every ema_overwrite_frequency
steps of iterations,
we overwrite the model variable by its moving average.
If None, the optimizer
does not overwrite model variables in the middle of training,
and you need to explicitly overwrite the variables
at the end of training by calling
optimizer.finalize_variable_values()
(which updates the model
variables in-place). When using the built-in fit()
training loop,
this happens automatically after the last epoch,
and you don't need to do anything.None
. If a float, the scale factor will
be multiplied the loss before computing gradients, and the inverse
of the scale factor will be multiplied by the gradients before
updating variables. Useful for preventing underflow during
mixed precision training. Alternately,
keras.optimizers.LossScaleOptimizer
will
automatically set a loss scale factor.None
. If an int, model & optimizer
variables will not be updated at every step; instead they will be
updated every gradient_accumulation_steps
steps, using the average
value of the gradients since the last update. This is known as
"gradient accumulation". This can be useful
when your batch size is very small, in order to reduce gradient
noise at each update step. EMA frequency will look at "accumulated"
iterations value (optimizer steps // gradient_accumulation_steps).
Learning rate schedules will look at "real" iterations value
(optimizer steps).Reference