I guess I am trying to understand the relationship of this callback to that of the optimizers. Yes it looks to be the same signature as a scipy optimizer one but otherwise this is just a callback, True if you are using SPSA and have a callback for that, since it has a different signature you could not use it for the callback here in this context. The test which is failing is using SPSA, which presumably worked before, so I am having a hard time figuring things.

I guess I am trying to understand the relationship of this callback to that of the optimizers. Yes it looks to be the same signature as a scipy optimizer one but otherwise this is just a callback, True if you are using SPSA and have a callback for that, since it has a different signature you could not use it for the callback here in this context. The test which is failing is using SPSA, which presumably worked before, so I am having a hard time figuring things.

Maybe, changing optimizers such as SPSA and ADAM might be a better option

Maybe, changing optimizers such as SPSA and ADAM might be a better optio

I guess I am still confused to some degree. The trainable model has its own callback that it defines right - it may have the same signature as scipy optimizer one but that is the only correlation as far as I know. The test was working with SPSA, which itself has a different callback. The issue really did not provide enough information for me to understand besides the notion that something mismatched - no code sample to reproduce or whatever.

Hi Steve,

I was not clear enough on the issue, maybe this might be useful:

SPSA has its own callback function since it is not based on SciPyOptimizer and it is called inside its own loop:

But trainable_model makes calls itself, and I think it is only compatible with SciPyOptimizer:

and my suggested change is:

  def objective(objective_weights):
      objective_value = function.objective(objective_weights)
      if isinstance(self._optimizer, SciPyOptimizer):
          self._callback(objective_weights, objective_value)
      return objective_value

This way we can allow custom callback functions and clear potential ambiguity when the function is called (inside or outside of the optimisation loop for custom callback functions).

Please note that SPSA is used in tutorials with QKT, not with trainable models:

I think trainable models must be adaptive if SciPyOptimizer is not used.

Pull Request Test Coverage Report for Build 12356721976

Details

• 3 of 3 (100.0%) changed or added relevant lines in 1 file are covered.
• 1 unchanged line in 1 file lost coverage.
• Overall coverage increased (+0.002%) to 90.635%

Totals
Change from base Build 12356603561:	0.002%
Covered Lines:	4365
Relevant Lines:	4816

💛 - Coveralls