BUG: Allow typecode of 'e' to be passed to select_dtypes

[jseabold]

Closes #11990.

[jreback]

hmm why does his raise a TypeError? float16 is platform specific I think ?

[jseabold]

np.e is defined and it's a float :)

In [1]: np.e
Out[1]: 2.718281828459045

[jreback]

hmm, I think this should be changed to:

np.dtype(dtype) as that wontt have the issue, and catch TypeError and not AttributeError

and is more correct

[jseabold]

On Thu, Jan 7, 2016, 1:09 PM Jeff Reback [email protected] wrote:

In pandas/core/common.py
#11991 (comment):

@@ -1482,10 +1482,11 @@ def _get_dtype_from_object(dtype):

     try:
         return _get_dtype_from_object(getattr(np,dtype))

hmm, I think this should be changed to:

np.dtype(dtype) as that wontt have the issue, and catch TypeError and not
AttributeError

and is more correct

I tried that. There was a test for dtype='integer' that fails. If you think
it shouldn't be supported, I would agree. But I don't feel strongly about
it.

[jreback]

how about (just trying to be somewhat generic)

try:
    np.typeDict.get(dtype) or np.typecodes[dtype]``
except KeyError:
    pass

[jseabold]

This would still fail the dtype='integer' case, and I'm not sure the latter would work at all without more processing. It's the use case I'm solving with the test.

What case are you trying to solve with the above?

[jreback]

all right, this ok I guess

[jreback]

add a whatsnew note as well

[jreback]

double backticks around np.float16

[jreback]

minor change, pls squash, ping when green.

[jreback]

can you update

[jseabold]

Green.

[jreback]

I find your patch fine. But isn't this odd?

In [11]: df = DataFrame(np.random.randn(10,2))

In [12]: df[0] = df[0].astype('float16')

In [13]: df.dtypes
Out[13]: 
0    float16
1    float64
dtype: object

In [14]: df.select_dtypes(include=['float'])
Out[14]: 
          1
0 -0.458946
1  0.456845
2  1.246123
3  1.296991
4  1.445800
5  1.082351
6 -0.895652
7  0.031383
8  1.320909
9 -1.783126

In [15]: df.select_dtypes(include=['float16'])
Out[15]: 
          0
0 -0.501465
1  0.623047
2  0.207397
3 -0.251465
4  1.599609
5 -2.359375
6  0.354004
7 -0.191650
8 -1.067383
9 -0.903320

[jreback]

merged via 138b8ba

[jseabold]

That would be my expectation since np.float64 == np.dtype(float), platform dependent. It's one of the reasons I ran into this, trying to use all float types to capture all eventualities.

[jseabold]

Actually, it might be nice to just accept the all typecode aliases too.

On Tue, Jan 12, 2016, 5:48 PM Jeff Reback [email protected] wrote:

Closed #11991 #11991.

—
Reply to this email directly or view it on GitHub
#11991 (comment).

[jreback]

would accept PR for that addition