CLN: Make ufunc works for Index

[sinhrks]

closes #9966, #9974 (PR)

I understand these are never used because Index is no longer the subclass of np.array. Correct?

• Add tests for all almost ufuncs
  • http://docs.scipy.org/doc/numpy/reference/ufuncs.html
• CategoricalIndex: Needs to be done separately to fix Categorical, because number of categories can be changed.

np.sin(pd.Categorical([1, 2, 3]))
array([ 0.84147098,  0.90929743,  0.14112001])

• MultiIndex: Raise TypeError or AttributeError, as ufuncs are performed to array of tuples.

[jreback]

yep I think you could take them out. as np.array(index) will coerce to a ndarray (and not an index sub-class). This is done by _shallow_copy in any event.

[shoyer]

No, we need this. Otherwise np.sin(idx) returns an ndarray.

It's not documented, but __array_finalize__ works even on non-subclasses.

[jreback]

ok let's add some tests and clean up the code
I suspect this should be in core/base then

eg to support the example above

[jreback]

IIRC there is another issue where for example (#9966 with a PR #9974) :

In [22]: np.sin(Index(range(5)))
Out[22]: Int64Index([0.0, 0.841470984808, 0.909297426826, 0.14112000806, -0.756802495308], dtype='float64')

which is wrong, this needs to be created w/o slightly differently, e.g.

Index(func(...), **self._get_attributes_dict())

so that the attributes get propogates AND the correct type gets inferred (e.g. this would return a FloatI64Index in this case)

maybe you want to supersede #9974

[sinhrks]

Thanks. I'll fix Index.__array_wraps__ to make ufunc work.

https://github.com/pydata/pandas/blob/master/pandas/core/index.py#L253

Am I correct that defining __array_finalize__ is not needed if __array_wraps__ is defined? Looks __array_finalize__ is not defined in normal Index.

http://docs.scipy.org/doc/numpy/reference/arrays.classes.html

[shoyer]

This is correct -- array_wraps is what we need.

On Tue, Jul 21, 2015 at 5:44 AM, Sinhrks notifications@github.com wrote:

Thanks. I'll fix Index.__array_wraps__ to make ufunc work.
https://github.com/pydata/pandas/blob/master/pandas/core/index.py#L253
Am I correct that defining __array_finalize__ is not needed if __array_wraps__ is defined? Looks __array_finalize__ is not defined in normal Index.

http://docs.scipy.org/doc/numpy/reference/arrays.classes.html

Reply to this email directly or view it on GitHub:
#10638 (comment)

[sinhrks]

Let me confirm 2 points:

1. Invalid ops for datetime-like indexes.

   I understand numeric ufuncs (like np.sin) for datetime-like is almost meaningless. PeriodIndex is likely to return unexpected result because of current __array__ definition. One idea is to raise error when PeriodIndex get float (invalid) result.

   import pandas as pd
   
   # OK
   np.sin(pd.DatetimeIndex(['2015-07', '2015-08']))
   # TypeError: ufunc 'sin' not supported for the input types, and the inputs could not be safely coerced to   any supported types according to the casting rule ''safe''
   
   # NG
   pidx = pd.PeriodIndex(['2015-07', '2015-08'], freq='M')
   np.sin(pidx)
   # Float64Index([-0.594884318359, 0.354966539136], dtype='float64')
   
   # ufunc is applied to i8 values
   pidx.__array__()
   # array([546, 547])
   

2. Some ufunc returns bool dtype. In such cases, it should return bool array as it is (like Index.duplicated)

   np.isnan(pidx)
   # Index([False, False], dtype='object')
   

[jreback]

@sinhrks both of those points look right

[shoyer]

Yes, agreed

On Thu, Jul 23, 2015 at 4:14 AM, jreback notifications@github.com wrote:

@sinhrks both of those points look right

Reply to this email directly or view it on GitHub:
#10638 (comment)

[jreback]

doesn't this need to go in tseries/base, e.g. ufuncs are not valid for any datetimelike?

[shoyer]

timedelta64 and datetime64 are real numpy dtypes, so ufuncs will already choke on them:

In [1]: import pandas as pd

In [2]: i = pd.date_range('2000-01-01', periods=5)

In [3]: import numpy as np

In [4]: np.square(i)
TypeError: ufunc 'square' not supported for the input types, and the inputs could not be safely coerced to any supported types according to the casting rule ''safe''

In [5]: np.square(i.to_period())
Out[5]:
PeriodIndex(['330671-10-02', '330731-10-03', '330791-10-05', '330851-10-09',
             '330911-10-16'],
            dtype='int64', freq='D')

[shoyer]

I agree that tests for these other index types would be a good idea, though.

Also, I think we should probably prohibit all ufuncs on PeriodIndex, not just those that return float. For example, it's not valid to square periods, even though that returns integers.

[sinhrks]

Raising TypeError in __array_wrap__ affects to arithmetic using np.array. Currently, it works:

pd.PeriodIndex(['2011', '2012'], freq='A') + np.array([1, 2])
# PeriodIndex(['2012', '2014'], dtype='int64', freq='A-DEC')

So it can't be prohibited?

[shoyer]

PeriodIndex + integer ndarray should not be expected to work -- the array
does not have the right type!

On Mon, Aug 17, 2015 at 3:35 PM, Sinhrks notifications@github.com wrote:

In pandas/tseries/period.py
#10638 (comment):

@@ -281,6 +281,16 @@ def contains(self, key):
return False
return key.ordinal in self._engine

• def array_wrap(self, result, context=None):

Raising TypeError in array_wrap affects to arithmetic using np.array.
Currently, it works:

pd.PeriodIndex(['2011', '2012'], freq='A') + np.array([1, 2])

PeriodIndex(['2012', '2014'], dtype='int64', freq='A-DEC')

So it can't be prohibited?

—
Reply to this email directly or view it on GitHub
https://github.com/pydata/pandas/pull/10638/files#r37245206.

[jreback]

actually I disagree
add integers to period works and makes sense as its a freq op

[sinhrks]

Yes, because this is a valid freq shift, np.array should work as the same.

pd.PeriodIndex(['2011', '2012'], freq='A') + 1
# PeriodIndex(['2012', '2013'], dtype='int64', freq='A-DEC')

And #10744 has been done expecting arithmetic using np.array works.

[sinhrks]

Reconsidered this, and defining _add_ndarray for PeriodIndex may be an alternative. I think options are:

1. Always raise TypeError in __array_wraps__. Support ndarray ops in another methods (PeriodIndex must be lhs).
2. Raise TypeError if __array_wraps__ gets non-integers. Some ufunc which return int outputs meaningless results (like square)

NOTE: Maybe not intentional, PeriodIndex.shift works for ndarray. Thus we can use shift for ndarray ops.

[jreback]

@sinhrks remember that ATM PeriodIndex is an odd duck. Its a real int64 dtyped array, that happens to have boxing. So it allows any integer-like ops (as opposed to DatetimeIndex which prohibits a lot more because its M8[ns]).

So its needs handling to error on prohibited ops. So on on raising in __array_wraps__ for prohibited ops, but this might require some dispatching mechanism to the sub-class to determine what is allowed. E.g. we do it similarly like this in the _add_numeric_ops and like routines.

[jreback]

@sinhrks you might want to steal tests from #9974

[jreback]

@sinhrks how's this coming along?

[sinhrks]

Should be TypeError, but cannot be raised from here because numpy catches.

[jreback]

@shoyer you know any better way to override the ufunc calling scheme? (e.g. as __numpy_ufunc__ doesn't exist yet). or is this best way?

[shoyer]

I think this sort of thing is the best we can do for now

[jreback]

ok let's add a note that eventually we could replace this with numpy_ufunc

[sinhrks]

Updated and now green.

[jreback]

why the different exception behavior?

[sinhrks]

TypeError raised from __array_wrap__ seems to be catchrd by numpy and return unintended result.

[jreback]

this doesn't seeem to close #9966 ?

In [9]: df = pd.DataFrame({'i': np.linspace(0, 5, 6).astype(np.int64), 
               'j': np.linspace(2, 3, 6)}).set_index('i')

In [10]: df
Out[10]: 
     j
i     
0  2.0
1  2.2
2  2.4
3  2.6
4  2.8
5  3.0

In [11]: df.index
Out[11]: Int64Index([0, 1, 2, 3, 4, 5], dtype='int64', name=u'i')

In [12]: df.index = df.index / 10.

In [13]: df.index
Out[13]: Int64Index([0.0, 0.1, 0.2, 0.3, 0.4, 0.5], dtype='float64', name=u'i')

[sinhrks]

@jreback Standard arithmetic is handled by _numeric_binop. Now fixed.

df.index / 2.
# Float64Index([0.0, 0.5, 1.0, 1.5, 2.0, 2.5], dtype='float64', name=u'i')

np.divide(df.index, 2.)
# Float64Index([0.0, 0.5, 1.0, 1.5, 2.0, 2.5], dtype='float64', name=u'i')

[sinhrks]

Failed in Timedelta ops. On current master, mult and div doesn't affect to existing freq, but it should?

import pandas as pd
idx = pd.timedelta_range('2 days', periods=3, freq='2D')

idx * 2
# TimedeltaIndex(['4 days', '8 days', '12 days'], dtype='timedelta64[ns]', freq='2D')
# -> freq should be "4D"?

idx / 2
# TimedeltaIndex(['1 days', '2 days', '3 days'], dtype='timedelta64[ns]', freq='2D')
# -> freq should be "D"?

[jreback]

@sinhrks hmm I think you need to set the freq=None so it get's reinferred (or maybe can et it to 'infer')

[jreback]

@sinhrks thanks!