Keras Loss scale optimizer does not increment iterations

Loss scale optimizer does not automatically increment iteration counter when apply method is called.

As a side effect, this breaks evaluation_loss_vs_iterations metric in TensorBoard when mixed precision is enabled.

See example here: colab

Comment From: jeffcarp

Thanks for the report, confirmed this fixes your original colab. As a workaround before this is released you can monkey patch the fix:

@property
def new_iterations(self):
  return self.inner_optimizer.iterations

keras.optimizers.LossScaleOptimizer.iterations = new_iterations

Comment From: google-ml-butler[bot]

Are you satisfied with the resolution of your issue? Yes No