Abstract loss scale manager class.
Loss scale managers with a different strategy should subclass this class. Loss scaling is a process that:
1) Applies a multiplier on the loss before computing gradients, and 2) Applies the reciprocal of the multiplier on the gradients before they are applied on variables.
This class is used together with
tf.contrib.mixed_precision.LossScaleOptimizer for mixed precision training
(float32 variables and float16 ops) on Nvidia GPUs in order to achieve the
same model quality as single precision training, with the benefits of
potential higher throughput.
tf.contrib.mixed_precision.LossScaleOptimizer for more details.
Returns the loss scale as a scalar
Updates loss scale based on if gradients are finite in current step.
finite_grads: bool scalar tensor indicating if all gradients are finite (i.e., not inf or nan).
An op, when executed updates the loss scale. If eager execution is enabled, does not return anything.