Numpy 2.x Compatibility Issue

[JBorrow]

We've tracked down an incompatibility of pixell and Numpy 2.x to this line:

pixell/pixell/utils.py

Line 2359 in 73d3f2b

pre, mid, tail = np.split(a, [off>0,(off>0)+nwhole], axis)

Here a boolean is being used as an index into the array being split into blocks. That is not compatible with numpy 2.x.

I am a little puzzled by this function; I presume off is the offset at which you start blocking the array from. Given that, should this line not read pre, mid, tail = np.split(a, [off,off+nwhole], axis) with some assertion that off must be positive?

[cpvargas]

It should be added at least in the readme file that pixell currently doesn't support numpy>=2.0

[JBorrow]

It does, though, anything that doesn't work still is a bug that should be fixed.

[iparask]

Can you paste the error? My understanding is that boolean to integer conversion is an intrinsic feature to python. This line should be equivalent to:

pre, mid, tail = np.split(a, [0,0+nwhole], axis) 

or

pre, mid, tail = np.split(a, [1,1+nwhole], axis) 

depending on the result of the check.

[JBorrow]

Repro:

henry product download 695d412851912051c46b8128 --slugify
python3 matched_filter_error.py

matched_filter_error.py:

from astropy import units as u
from pixell import enmap

from sotrplib.filters.filters import matched_filter_depth1_map

map_path = "deep56-lat-iso-phase-2-i1-f090-coadd-20251119/map/sky_map_i1_f090_coadd_20251119.fits"
ivar_path = "deep56-lat-iso-phase-2-i1-f090-coadd-20251119/ivar/sky_ivar_i1_f090_coadd_20251119.fits"

map = enmap.read_map(map_path)
ivar = enmap.read_map(map_path)

matched_filter_depth1_map(
    map,
    ivar,
    band_center=90.0 * u.GHz,
    beam_fwhm=1.0 * u.arcmin,
    band_height=1.0 * u.deg,
)

Traceback:

/Users/borrow-adm/Documents/Projects/sotrplib/sotrplib/filters/filters.py:139: RuntimeWarning: invalid value encountered in sqrt
  white = map * ivar**0.5
Traceback (most recent call last):
  File "/Users/borrow-adm/Documents/Projects/sotrplib/matched_filter_error.py", line 12, in <module>
    matched_filter_depth1_map(
  File "/Users/borrow-adm/Documents/Projects/sotrplib/sotrplib/filters/filters.py", line 347, in matched_filter_depth1_map
    iC = build_ips2d_udgrade(
         ^^^^^^^^^^^^^^^^^^^^
  File "/Users/borrow-adm/Documents/Projects/sotrplib/sotrplib/filters/filters.py", line 148, in build_ips2d_udgrade
    ps2d = enmap.upgrade(
           ^^^^^^^^^^^^^^
  File "/Users/borrow-adm/Documents/Projects/sotrplib/.venv/lib/python3.12/site-packages/pixell/enmap.py", line 2027, in upgrade
    omap = utils.block_expand(emap, factor[0], oshape[-2], off=off[0], axis=-2, inclusive=inclusive)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/borrow-adm/Documents/Projects/sotrplib/.venv/lib/python3.12/site-packages/pixell/utils.py", line 2234, in block_expand
    pre, mid, tail = np.split(a, [off>0,(off>0)+nwhole], axis)
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/borrow-adm/Documents/Projects/sotrplib/.venv/lib/python3.12/site-packages/numpy/lib/_shape_base_impl.py", line 886, in split
    return array_split(ary, indices_or_sections, axis)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/borrow-adm/Documents/Projects/sotrplib/.venv/lib/python3.12/site-packages/numpy/lib/_shape_base_impl.py", line 801, in array_split
    sub_arys.append(_nx.swapaxes(sary[st:end], axis, 0))
                                 ~~~~^^^^^^^^
  File "/Users/borrow-adm/Documents/Projects/sotrplib/.venv/lib/python3.12/site-packages/pixell/enmap.py", line 143, in __getitem__
    return ndmap(np.ndarray.__getitem__(self, sel), wcs)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: slice indices must be integers or None or have an __index__ method

Slice indices on numpy arrays cannot be boolean.

[mhasself]

I just spent a minute looking at this and I'm confident that you can replace (off>0) with int(off>0). In block_expand, it's used to say "if off > 0, we need to assign the first element from the block-reduced computation (blocks[0]) into an initial bit of the output (output[:ofs]).

[mhasself]

".