Make map conversions more efficient.

[groutr]

Changes

• Make convert_mapping_values work more efficiently (for a dict of length 1000, runtime drops from 471us to 294us on my machine)

Notes

Since the numpy dtypes are C functions, pushing the loop into C (by using map) will be far more efficient.

There was a type mismatch between the type of mapping and the return of the function. Updated the type of mapping to be MutableMapping.

[hellkite500]

@groutr just a quick glance through the test logs:

_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

mapping = true_positive     1
false_positive    2
false_negative    3
true_negative     4
dtype: object
converter = <class 'numpy.float64'>

    def convert_mapping_values(
        mapping: MutableMapping[str, npt.DTypeLike],
        converter: np.dtype = np.float64
        ) -> MutableMapping:
        """Convert mapping values to a consistent type. Primarily used to validate
        contingency tables.
    
        Parameters
        ----------
        mapping: dict-like, required
            Input mapping with string keys and values that can be coerced into a
            numpy data type.
        converter: numpy.dtype, optional, default numpy.float64
            Converter data type or function used to convert mapping values to a
            consistent type.
    
        Returns
        -------
        converted_mapping: dict-like, same type as mapping
            New mapping with converted values.
    
        """
        # Build object of same type of mapping
        d = type(mapping)()
>       d.update(zip(mapping.keys(), np.fromiter(mapping.values(), dtype=converter)))
E       TypeError: 'numpy.ndarray' object is not callable

/opt/hostedtoolcache/Python/3.7.12/x64/lib/python3.7/site-packages/hydrotools/metrics/metrics.py:262: TypeError

Seems like a similar failure causes a few of the tests to fail. Thoughts?

[jarq6c]

The issue could be related to assumptions made in the original method. The original (slow) method was meant to work with both Python dict and pandas.Series. For dict values is a Callable. However, for pandas.Series values is a non-callable property.

An additional nuance is that pandas.Series.update takes an object that must be coercible into a pandas.Series. I'm unsure of what Series will do when feeding it zip directly like this.

[jarq6c]

Actually come to think of it, it may be easier (and faster) to offload this task to pandas.

return pd.Series(mapping, dtype=converter)

[groutr]

Actually come to think of it, it may be easier (and faster) to offload this task to pandas.

return pd.Series(mapping, dtype=converter)

That was my initial thought, however it seemed like there was effort made to make this general to any mapping type. If a pandas series is assumed, this is going to be the best way.

I was adhering to the type information in the function signature. pd.Series is not a Mapping type (though it sort of mimics the python dictionary behavior)

isinstance(pd.Series(), Mapping)  # False

[jarq6c]

Thus far, this method is only used internally to validate contingency tables. @groutr were you calling this method directly? If not, this might be a good argument for replacing this method with something more strict, but hopefully faster (that only returns pandas.Series).

[groutr]

@jarq6c I am not calling the method directly. Replacing it with something more strict seems like a good idea to me.

[jarq6c]

Superseded by #191