difference in promotion between py2 and py3 #41
Description
Here's a weird one. I have a histogram class that overloads += operator:
bp::class_<hist_t> ("histogram", bp::init<double,double,double,bool> (
(bp::arg ("min"),
bp::arg ("max"),
bp::arg ("delta"),
bp::arg ("clip")=true
)))
.def (bp::init<nd::Array<int64_t,1,1>,double,double,double,bool>())
.def (bp::self += hist_t::value_type())
.def (bp::self += nd::Array<double,1,0>())
.def ("__call__", &hist_t::operator())
.def_readwrite ("results", &hist_t::buckets)
.add_property ("buckets", &get_buckets<hist_t,nd::Array<double,1,1> >)
.def ("mean", &hist_t::Mean)
.def ("rms", &hist_t::RMS)
.def ("cumulative", &hist_t::Cumulative<nd::Array<double,1,1> >)
.def ("inv_cumulative", &hist_t::OneMinusCumulative<nd::Array<double,1,1> >)
.def_pickle(hist_pickle_suite<hist_t>())
.def_readonly ("min", &hist_t::min)
.def_readonly ("max", &hist_t::max)
.def_readonly ("delta", &hist_t::delta)
.add_property ("rc", &getrc<hist_t>) // just for info
;I call it:
remod_out_hist += np.sum (np.max (remod_out_cnts, axis=1))
Now if we print out the type (np.sum (np.max ...)) on
py2: <type 'numpy.int64'>
but on py3: <class 'numpy.int64'>
When run on py2, the type numpy.int64 seems to be promoted to double, which matches the 1st self += (value_type should be double).
But when run on py3, the promotion fails:
Boost.Python.ArgumentError: Python argument types in
histogram.iadd(histogram, numpy.int64)
did not match C++ signature:
iadd(boost::python::back_reference<Histogram<long, double, ndarray::Array<long, 1, 1> >&>, ndarray::Array<double, 1, 0>)
iadd(boost::python::back_reference<Histogram<long, double, ndarray::Array<long, 1, 1> >&>, double)
If I explicitly write
remod_out_hist += float(np.sum (np.max (remod_out_cnts, axis=1)))
then it runs on py3
I'm wondering if the difference in type() on py2 and py3 might explain the different behavior.