Derive dtype from np.dtype in CastTransformer/Regressor #927

Hi!

Thanks for the great work.

In numpy, any function that receives a dtype can also receive a string representation of that dtype:

import numpy as np

assert np.zeros(3, dtype="float64").dtype == np.zeros(3, dtype=np.float64).dtype
assert np.zeros(3, dtype="float64").dtype != np.zeros(3, dtype=np.float32).dtype

I expected the same behavior in CastTransformer which led to a weird error. The CastTransformer seemed to interpret my string representation of the dtype as a literal string type (i.e., it expected a string as input).

As a fix, we could replace the setting of the dtype in the constructor with an inferred dtype from numpy.

i.e.,

def __init__(self, *, dtype=np.float32):
        self.dtype = np.dtype(dtype)

This looks up the appropriate dtype, and is also a no-op if the passed value is already a dtype (and not a string). Alternatively, we could catch whether the passed value is actually a dtype.

def __init__(self, *, dtype=np.float32):
        if not _is_dtype(dtype):
             raise ValueError("informative message")
        self.dtype = dtype

We would have to implement _is_dtype somehow though.

I fully concede that this was user error on my part, but given this is how numpy behaves, maybe the behavior could be copied.

Happy to contribute a PR.