Operations on transposed (or similar) binned data return incorrect values in bin-element coords/masks/attrs #3277
Description
Consider
scipp/lib/dataset/variable_instantiate_bin_elements.cpp
Lines 90 to 109 in 5b12dc3
Note in particular the first TODO. This means
- We cannot apply operations to a slice of binned data (without copying first), instead we get an exception since the buffer length is not matching up. This is not nice, but at least correct.
- If indices are not in order, coords (and masks) are copied directly into the output buffer without reordering. The implicit length check is not a sufficient safe guard here. This is a bad bug.
Example:
import scipp as sc
table = sc.data.table_xyz(10)
da = table.bin(x=2, y=2)
print('data:')
print(da['x', 1]['y', 0].bins.data.value.values)
print((da.transpose() * 1)['x', 1]['y', 0].bins.data.value.values)
print((da.transpose().copy())['x', 1]['y', 0].bins.data.value.values)
print('coords:')
print(da['x', 1]['y', 0].bins.coords['x'].value.values)
print((da.transpose() * 1)['x', 1]['y', 0].bins.coords['x'].value.values)
print((da.transpose().copy())['x', 1]['y', 0].bins.coords['x'].value.values)Output:
data (correct):
[1.08584904 1.00868831]
[1.08584904 1.00868831]
[1.08584904 1.00868831]
coords (broken!):
[0.97669977 0.96407925]
[0.38019574 0.26169242]
[0.97669977 0.96407925]
Note the copy() uses a different code path and is thus correct.
Metadata
Metadata
Assignees
Type
Projects
Status