Open
Description
We used to the following property: if x is a dtype then numpy.dtype(x.descr) is the same dtype. This property is broken in NumPy 1.8 when one specifies metadata to dtype. Is it possible to at least fix that descr returned by dtype can always be evaluated back to a dtype?
Related issues: h5py/h5py#1307
Reproducing code example:
import numpy as np
# 1. metadata is lost for simple dtype
s_dt = np.dtype('S8', metadata={'msg': 'Hello'})
print(s_dt.descr)
# prints [('', '|S8')]
# 2. metadata is returned for structured dtype, but it cannot be evaluated into dtype
struct_dtype = np.dtype([('a', s_dt)])
print(struct_dtype.descr)
# prints [('a', ('|S8', {'msg': 'Hello'}))]
# next line raises a ValueError exception
# ValueError: invalid shape in fixed-type tuple.
np.dtype(struct_dtype.descr)Numpy/Python version information:
1.18.1 3.8.1 | packaged by conda-forge | (default, Jan 29 2020, 14:55:04)
[GCC 7.3.0]