TAP works much slower than standard ArgumentParser about typed-argument-parser HOT 4 OPEN

Hi @Enolerobotti,

Thank you for bringing this to our attention! Yes, the slow part is the source code inspection in order to extract the help strings from comments. In our machine learning workflows, we find this overhead to be acceptable, but we can imagine that certain applications may need a faster argument parser. In the future, we will add a flag to disable this. Concretely, we would skip this call:

Best,
Kyle and Jesse

from typed-argument-parser.

Hi @martinjm97,

Thank you for your reply!
Indeed, the _get_class_variables method works slow. The problem is that when I try to use ChemProp in my workflow, I need some default arguments (which are defined inside a TAP instance) to be loaded several times. I have to initiate TAP based classes several times due to an assertion (btw, for me it looks redundant). So, for ensemble of 10 models, I spend additional 0.5 sec per molecule to predict. For list longer than 1000 molecules it would be worth for me to get rid of redundant operations.

Cheers,
Artem

from typed-argument-parser.

@Enolerobotti,

That makes sense! I didn't think a Tap would be in an inner loop. We'll add the flag soon and then you can save the 500s for your 1000 molecules! Thanks for letting us know and for the clear explanation.

--Jesse

from typed-argument-parser.

Encountering the same issue, but now it seems like the class variables are being used on line 72 in _add_argument:

            # Description
            if variable in self.class_variables:
                kwargs['help'] += ' ' + self.class_variables[variable]['comment']

and on line 298 in _add_arguments:

        # Add class variables (in order)
        for variable in self.class_variables:

As that variable is now also being used for ordering, is it now not possible to make this change as noted above?

from typed-argument-parser.