Discussion: input and output argument type casting

[AartGoossens]

@sladkovm This is the discussion we started before. Shall we continue here?

For algorithms it is possible to accept multiple input argument types and cast them to a desirable type and cast the output to a type similar to the input type (example of a method that does this in the vmpy source code).

Benefits:

• Usage of the algorithms is more straightforward when you don't have to worry about the input types
• ...

Drawbacks:

• Input argument type is part of the interface specification. Allowing multiple input types might make the interface vague
• It is not possible to accept all input types. How do we communicate what is/is not possible?
• Unittesting for multiple input types is a hassle
• ...

I have a preference for accepting (i.e. developing for) one input type.

Alternative solution: It might be possible to add input and output argument type casting by decorating the algorithms "on the fly" by calling a helper function after import (sort of similar to this). It's not the cleanest way to do it but it might work, I'll try to make a POC for this. For me this would solve the drawbacks and it would still offer input/output type casting for users that do want it.

[sladkovm]

I have myself second thoughts about supporting all array-like inputs - it makes type transformation very opaque for the user. To put it differently, I'm convinced now that converting lists to ndarrays and back should be a user's headache. If we can design a decorator for it - it would be nice, but not a must.

More interesting question - could we make the ndarray as a default input type, but than still design all algorithms/ function to work directly with pd.Series as well? Basically, repeat what NumPy does already.

[AartGoossens]

pandas.Series is in a way a wrapper around numpy.ndarray so my expectation is that in most cases it will just work. E.g.

>>> import numpy as np
>>> import pandas as pd
>>> a = pd.Series([1, 2, 3, 4])
>>> np.power(a, 2)
0     1
1     4
2     9
3    16
dtype: int64
>>>

I want to prevent if-statements in the algorithms like:

if isinstance(power, pd.Series):
    # do this
else:
    # do that