EMD reader promotes integer data to float64

[mkuehbach]

Observed on a Velox EMD file where images were stored as chunked 16bit unsigned int datasets, i.e. "u2" rosettasciio returns these under the "data" entry in the parsing results dictionary as a numpy array whose values are promoted to np.float64. While this is not necessarily in this particular case a loss in precision but in practice resulting in an eventually unnecessary quadrupling of the memory consumption.

Is there are chance to have the individual entries under the "data" entry read and loaded such that these are represented with the same datatype as in the source or is there some normalization required for reasons and context that I am missing, that demands a promotion to np.float64 ?

[ericpre]

Yes, it would make sense to keep it in integer type and it may a matter of specifying the dtype in the right place? Can you provide more details about what kind of data (image, stack of images, EDS spectrum, EDS stream) you observed this? Or can you share a file?

If this is for EDS stream data, it may be more complicated because there may be the risk that some value are clipped - it may have been discussed before.

[mkuehbach]

Agree, that should be fetched with some sort of np.asarray(), assuming h5py under is used that payload should not get promoted by h5py automatically, at least not happened in cases I saw with several h5py versions.

For context how I saw this.

self.objs = emd.file_reader(self.file_path)

Then for this objs's individual object obj dictionaries we observe

for idx, obj in enumerate(self.objs):
    # inspect obj["data"]

Need to ask the original authors first, if we can share that specific file.