|View source on GitHub|
Base class for all TF1 loss scales.
See Migration guide for more details.
This is an abstract base class, so you cannot instantiate it directly. Instead, use one of its concrete subclasses:
Loss scaling is a process that multiplies the loss by a multiplier called the loss scale, and divides each gradient by the same multiplier. The pseudocode for this process is:
loss = ... loss *= loss_scale grads = gradients(loss, vars) grads /= loss_scale
Mathematically, loss scaling has no effect, but can help avoid numerical underflow in intermediate gradients when float16 tensors are used for mixed precision training. By multiplying the loss, each intermediate gradient will have the same multiplier applied.
Instances of this class represent a loss scale. Calling instances of this
class returns the loss scale as a scalar float32 tensor, while method
update() updates the loss scale depending on the values of the gradients.
Optimizers use instances of this class to scale loss and gradients.
In most functions that accept a LossScale, you can also pass an int (such as
8) to create a
FixedLossScale or the string
"dynamic" to create a dynamic
from_config( config )
Creates the LossScale from its config.
Returns the config of this loss scale.
update( grads )
Updates the value of the loss scale.
The loss scale will be potentially updated, based on the value of
The tensor returned by calling this class is only updated when this function
In eager mode, this directly updates the loss scale, so that calling
__call__ will return the newly updated loss scale. In graph mode,
this returns an op that, when evaluated, updates the loss scale.
This function also returns a
should_apply_gradients bool. If False,
gradients should not be applied to the variables that step, as nonfinite
gradients were found, and the loss scale has been be updated to reduce the
chance of finding nonfinite gradients in the next step. Some loss scale
classes will always return True, as they cannot adjust themselves in
response to nonfinite gradients.
When a DistributionStrategy is used, this function may only be called in a cross-replica context.
||A nested structure of unscaled gradients, each which is the gradient of the loss with respect to a weight. The gradients should have already been divided by the loss scale being before passed to this function. 'None' gradients are accepted, and are ignored.|