`ZEISSDataReader` wrongly produces 64-bit data

[JeppeKlitgaard]

Using ZEISSDataReader on data gathered from a ZEISS X-Ray system at my university incorrectly produces arrays with float64 dtypes, which leads to a whole host of issues when the AcquisitionData is processed using ASTRA-based optimisers.

A stop-gap measure is to recast the data as 32-bit floats:

data = ZEISSDataReader(file_name=...).read()
data.array = data.array.astype('float32')
data.geometry.dtype = 'float32'

Note that the geometry dtype is overridden by the data dtype as indicated by warning UserWarning: Over-riding geometry.dtype with data.dtype

I believe this is an upstream issue, but thought it would be useful to track here as well. The stop-gap could potentially be implemented directly in ZEISSDataReader.

Upstream bug

In https://github.com/data-exchange/dxchange/blob/10c9d870bd403f1ea5fbba4c8194b9c8d9295aca/dxchange/reader.py#L529C8-L540C73, the data is clearly read as 32-bit floats as given by the f in the format string, but when the array is initalised in https://github.com/data-exchange/dxchange/blob/10c9d870bd403f1ea5fbba4c8194b9c8d9295aca/dxchange/reader.py#L1149-L1156 no dtype is given and thus the default float64 dtype is used.

I would be happy to report the bug in dxchange and submit a PR either there or here in ZEISSDataReader, but would appreciate input as to whether my diagnosis is sensible first.

Progress

• Submit PR fix to dxchange: PR
• PR Merged
• dxchange release cut
• Optional: Update dxchange dependency in CIL