Round tripping non-float Quantities from QTable to fits converts all Quantities to floats

Description

Trying to serialise a QTable to and from fits the non-floats are always converted to floats rather then keeping the original dtypes

Expected behavior

The created QTable to have the same dtype as the underling fits column.

Actual behavior

The created QTable has converted the uint and int to float64

Steps to Reproduce

import astropy.units as u
import numpy as np

from astropy.table import QTable
from astropy.io import fits

my_table = QTable()

my_table['uint'] = (list(range(10)) * u.m).astype(np.uint16)
my_table['int'] = (list(range(10)) * u.m).astype(np.int16)
my_table['float32'] = (list(range(10)) * u.m).astype(np.float32)

my_table.dtype
# dtype([('uint', '<u2'), ('int', '<i2'), ('float32', '<f4')])

my_table.write('test.fits')

table_fromfits = QTable.read('test.fits')
table_fromfits.dtype
#dtype([('uint', '<f8'), ('int', '<f8'), ('float32', '>f4')])

A simple but not elegant solution is to use the dtype info from the HDUL to cast the Quantity back to the desired dtype This actually doesn’t work for the uint case or cases with mixed quantities an non-quantities dtype([('uint', '<i2'), ('int', '<i2'), ('float32', '<f4')]) should be dtype([('uint', '<u2'), ('int', '<i2'), ('float32', '<f4')])

hdul = fits.open('test.fits')

hdul[1].data.columns

#ColDefs(
#    name = 'uint'; format = 'I'; unit = 'm'; bzero = 32768
#    name = 'int'; format = 'I'; unit = 'm'
#    name = 'float32'; format = 'E'; unit = 'm'
#)

fixed_table = QTable([u.Quantity(table_fromfits[c.name], dtype=c.dtype) for c in hdul[1].data.columns])

fixed_table.dtype
# dtype([('uint', '<i2'), ('int', '<i2'), ('float32', '<f4')])

System Details

Linux-5.4.0-33-generic-x86_64-with-glibc2.29 Python 3.8.10 (default, Sep 28 2021, 16:10:42) [GCC 9.3.0] Numpy 1.20.2 pyerfa 1.7.3 astropy 4.3.1 Scipy 1.6.3 Matplotlib 3.4.1

[samaloney]

Sorry for the delay @dhomeier I’ve just loaded some of my fits files using your branch seems to work as expected the dtypes are respected.

[dhomeier]

@samaloney I have pushed a fix for correctly reading back the mixins in #12505 ; if you have chance to test if this addresses your problems, feedback would be great.

For a temporary workaround, as you noted that your use cases typically involve reading in large volumes of data, I strongly recommend to use getheader and getdata to read the tables in their original dtypes and then extract the units to directly construct u.Quantity(tab[col], dtype=tab[col].dtype, name=col, unit=u.Unit(header.get('TUNITn'), format='fits')). It’s a bit cumbersome, but would avoid conversion losses and in particular the temporary creation of arrays 4x the original size or larger, in particular since the type conversion does not let you take advantage of memmap=True.