FixedScaleOffset flattens arrays, breaking further image codecs

[raphaeljolivet]

I've just noticed that FixedScaleOffset flattens arrays on encoding.

This is not stated in the documentation

This breaks further images encoding in the pipeline.

Minimal test to reproduce:

import numpy as np
from numcodecs import FixedScaleOffset
shape = (8, 4, 5)

x = (200.0 + np.arange(np.prod(shape), dtype=np.float32).reshape(shape) * 0.001)
fso = FixedScaleOffset(offset=200.0, scale=1000.0, dtype="<f4", astype="<u2")

# Expected  (8, 4, 5)
# Actuel (160, )
print(fso.shape)

[d-v-b]

that's an unfortunate bug! how are you using FixedScaleOffset? If you are working with zarr v3 data, then there are alternatives that don't have this problem (the cast_value and scale_offset codecs, which are defined here: https://github.com/zarr-developers/cast-value.py)

[raphaeljolivet]

Thanks, I didn't notice there were several scal/offset codec.
I am working with both zarr2 and zarr3 actually.

I will try wih this one.

I still think this should be fixed, or stated in the documentation.
Maybe an option like "keep_dimensions" would make the change backward compatible.

[d-v-b]

Maybe an option like "keep_dimensions" would make the change backward compatible.

This makes sense, i don't have time to add this but if you want to open a PR i can review it