regression: NotImplementedError: operator '*' not implemented for bool dtypes

[yarikoptic]

seemed to work just fine (with 0.13.1) and I do not see why it should not if you are decorating numpy arrays here:

  File "/home/neurodebian/proj/numpy-vbench/vbench/vbench/analysis.py", line 83, in __call__
    Tts, Tps = ss.ttest_1samp(test_samples.timing, results.timing)
  File "/usr/lib/python2.7/dist-packages/scipy/stats/stats.py", line 2887, in ttest_1samp
    t = np.where((d==0)*(v==0), 1.0, t) #define t=0/0 = 1, identical mean, var
  File "/home/neurodebian/deb/gits/pkg-exppsy/pandas/pandas/core/ops.py", line 505, in wrapper
    return left._constructor(wrap_results(na_op(lvalues, rvalues)),
  File "/home/neurodebian/deb/gits/pkg-exppsy/pandas/pandas/core/ops.py", line 445, in na_op
    raise_on_error=True, **eval_kwargs)
  File "/home/neurodebian/deb/gits/pkg-exppsy/pandas/pandas/computation/expressions.py", line 196, in evaluate
    _bool_arith_check(op_str, a, b)
  File "/home/neurodebian/deb/gits/pkg-exppsy/pandas/pandas/computation/expressions.py", line 177, in _bool_arith_check
    "dtypes" % op_str)
NotImplementedError: operator '*' not implemented for bool dtypes

[jreback]

on purpose numexpr doesn't allow this and since that is heavily used are banning it.

should be using & for anding

[jreback]

discussions here: #7011, #6762, #7015

[jorisvandenbossche]

Could that be added in the error message? (that you can use & instead)

[jreback]

sure...I don't think that's hard, FYI I don't think their is an analog for |, is there?

@cpcloud can you change that error message a bit?

[cpcloud]

Dure

[cpcloud]

I mean sure

[yarikoptic]

oh well -- just don't get surprised then whenever people with stock scipy would complain that they can't even do a ttest now (whenever it was working before)... FWIW -- scipy has refactored that code I believe in v0.11.0b1~20^2, so only wheezy and precise (among sensible) would be affected.

[dsm054]

"sure...I don't think that's hard, FYI I don't think their is an analog for |, is there?"

Isn't that simply +?

[jreback]

its not +

True | True == True
True + True == 2

[cpcloud]

aha but:

In [2]: np.array([True]) + np.array([True])
Out[2]: array([ True], dtype=bool)

[jreback]

interesting.

@cpcloud another possibility is to try to detect these condition and then just do the eval in python space (which will work I guess), maybe with a warning?

[cpcloud]

i pretty much do that now ... but i raise instead of continuing on

[cpcloud]

so i'll change it

[dsm054]

I had the dtype=bool behaviour in mind, but even apart from that, while True != 2, bool(True) == bool(2) which (as we say here) is close enough for government work.

[jorisvandenbossche]

What do we do with this one? Ideally this would be handled for 0.14 (as this is introduced in 0.14)

[cpcloud]

@jorisvandenbossche agreed i'll have something up tonight

[jreback]

maybe easiest to do a translation, warning then evaluate it?

[cpcloud]

yep something like:

if bool arith op: warn("don't do this") -> do_it_anyway()

[jreback]

@cpcloud PR for this?

[cpcloud]

Yep sorry got a bit sidetracked practicing my talk. :) should have something by lunch.

[cpcloud]

so really this should just be *, + and -, since none of the others really make sense

[jreback]

yep

[cpcloud]

interesting never really figured out what x - y for bool does, it's actually xor:

In [58]: x, y = rand(10) > 0.5, rand(10) > 0.5

In [59]: np.all((x | y) & ~(x & y) == x - y)
Out[59]: True

[jreback]

x-y == x^y as well (which is the xor operator)

[cpcloud]

ah right ... thanks that helps actually (simplifies the error message)

[jreback]

thanks @cpcloud !

[cpcloud]

np happy to help