Optimizators Step Tracking

Seems the optimizers do not expose the step they are in during their loops, for example in the inner function of the BFGS optimizer.

This value is required to use the tf.summary API

 tf.summary.scalar('loss', avg_loss.result()))

The specific error is:

/usr/local/lib/python3.6/dist-packages/tensorflow/python/ops/summary_ops_v2.py in write(tag, tensor, step, metadata, name)
    647       step = get_step()
    648       if step is None:
--> 649         raise ValueError("No step set via 'step' argument or "
    650                          "tf.summary.experimental.set_step()")
    651     if metadata is None:

ValueError: No step set via 'step' argument or tf.summary.experimental.set_step()

Maybe for the common use cases, the optimization loop is so fast that it would not be very useful, but I would like to hear your thoughts.

Start by locating the BFGS optimizer's inner function and comparing how optimizer loops track iterations. Review how the tf.summary API receives its step value and determine the intended scope across optimizers. Done means the relevant optimization loops expose a usable step so tf.summary no longer raises the reported missing-step error.