You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Bottleneck uses an array copy to do median, and accesses array elements with an integer indexer, requiring 1d input. I suppose we can do a copy if we allow the start of the gufunc to drop out of nopython mode, but it would be unfortunate to only be able to support aggregating over single axes at once. On the other hand, numpy doesn't have no-copy support for aggregating over multiple axes at once, either, and np.nanmedian finally arrived in numpy 1.9. If Numbagg won't be much faster, there may be no point in supplying either median or nanmedian ourselves.
move_median is even trickier. Bottleneck uses a C library that implements an appropriate data structure. Pandas has its own Cython code for an efficient data structure. In principle, I support we could interface with C from Numba to do this in Numbagg, but that is not so elegant.
Insight or ideas from others would be greatly appreciated here :).
The text was updated successfully, but these errors were encountered:
Bottleneck uses an array copy to do median, and accesses array elements with an integer indexer, requiring 1d input. I suppose we can do a copy if we allow the start of the gufunc to drop out of nopython mode, but it would be unfortunate to only be able to support aggregating over single axes at once. On the other hand, numpy doesn't have no-copy support for aggregating over multiple axes at once, either, and
np.nanmedian
finally arrived in numpy 1.9. If Numbagg won't be much faster, there may be no point in supplying either median or nanmedian ourselves.move_median
is even trickier. Bottleneck uses a C library that implements an appropriate data structure. Pandas has its own Cython code for an efficient data structure. In principle, I support we could interface with C from Numba to do this in Numbagg, but that is not so elegant.Insight or ideas from others would be greatly appreciated here :).
The text was updated successfully, but these errors were encountered: