select_dtypes fails on float16 typecode

[jseabold]

I'll see if it's an easy fix.

import pandas as pd
import numpy as np

FLOAT_TYPES = list(np.typecodes['AllFloat'])

df = pd.util.testing.makeDataFrame()
df.select_dtypes(FLOAT_TYPES)

FLOAT_TYPES.remove('e')
df.select_dtypes(FLOAT_TYPES)

[jseabold]

Oh getattr(np, dtype) doesn't work for e, obviously.

[jseabold]

I'm not sure what is handled by getattr(np, dtype) that isn't handled by np.dtype(dtype), actually.

[jreback]

have you had other issues with float16 btw? its not support for some of the numeric ops (and will get upcast). what are you doing with it?

[jreback]

xref #9220, #10382

[jseabold]

I'm not doing anything with float16. I was just lazily trying to differentiate between numeric, integer, and object types to do e.g. continuous vs categorical data handling without more specific types than numpy primitives, and I noticed this.