When loading tensor map that requires use_numpy=True the error message is unclear

[agoscinski]

When I load the tensor map

In [1]: import metatensor
In [2]: metatensor.load("./python/metatensor-operations/tests/data/qm7-power-spectrum.npz")
---------------------------------------------------------------------------
MetatensorError                           Traceback (most recent call last)
Cell In[2], line 1
----> 1 metatensor.load("./python/metatensor-operations/tests/data/qm7-power-spectrum.npz")

File /opt/anaconda3/envs/metatensor-dev/lib/python3.11/site-packages/metatensor/io.py:72, in load(file, use_numpy)
     70 else:
     71     if isinstance(file, (str, pathlib.Path)):
---> 72         return load_custom_array(file, create_numpy_array)
     73     else:
     74         # assume we have a file-like object
     75         buffer = file.read()

File /opt/anaconda3/envs/metatensor-dev/lib/python3.11/site-packages/metatensor/io.py:118, in load_custom_array(path, create_array)
    114     path = bytes(path)
    116 ptr = lib.mts_tensormap_load(path, mts_create_array_callback_t(create_array))
--> 118 return TensorMap._from_ptr(ptr)

File /opt/anaconda3/envs/metatensor-dev/lib/python3.11/site-packages/metatensor/tensor.py:121, in TensorMap._from_ptr(ptr)
    118 @staticmethod
    119 def _from_ptr(ptr):
    120     """Create a tensor map from a pointer owning its data"""
--> 121     _check_pointer(ptr)
    122     obj = TensorMap.__new__(TensorMap)
    123     obj._lib = _get_library()

File /opt/anaconda3/envs/metatensor-dev/lib/python3.11/site-packages/metatensor/status.py:48, in _check_pointer(pointer)
     46     raise MetatensorError(last_error()) from e
     47 else:
---> 48     raise MetatensorError(last_error())

MetatensorError: serialization format error: unsupported Zip archive: Compression method not supported: at 'keys.npy'

Correct usage

metatensor.load("./python/metatensor-operations/tests/data/qm7-power-spectrum.npz", use_numpy=True)

But I cannot understand from the first error message that this is the issue. I am not sure what the right solution would be, but anything would help. For example, checking the file ending and add something to the error message if npz is identified (or just always add this to the message). Or if this fails trying to also load it with use_numpy option to give user more informative error message.

[Luthaf]

Right. I think the best solution is to add support for DEFLATE storage in the non-numpy implementation. We currently only support STORED (which should enable memory-mapping), but giving both options would be good!