Unable to read uints DType as f64

[Shatur]

I discovered that I can't read files with uints as f64.
Can we make this conversion to work since it should be lossless?

[ExpHP]

Can you clarify, are you wanting to cast the integers (1 becomes 1.0), or to reinterpret them (0x3FF0000000000000 becomes 1.0)? Is the source type np.uint32 or np.uint64?

[Shatur]

Can you clarify, are you wanting to cast the integers (1 becomes 1.0)

This one.

Is the source type np.uint32 or np.uint64

In my case it's np.uint32.

[ExpHP]

Okay, this could be considered a possible extension to #25. Currently, npyz does not even support reading u32s as u64s; this was originally planned, but resulted in a complicated implementation with unclear performance implications. (basically it adds a level of indirection to every individual item read, even in cases where we are trivially reading uint32 as u32)

In the meanwhile, if you know in advance that the array is going to be of dtype uint32, you can deserialize as u32 and then map the iterator:

npy.data::<u32>()?
    .map(|res| res.map(|value| value as f64))
    // this is now an iterator of Result<f64, std::io::Error>

If you know the array might be uint32 or float64 (but aren't sure which), you can use .try_data to attempt deserializing more than one type. Here's an example of how to collect into a Vec:

let floats: Vec<f64> = match npy.try_data::<u32>() {
    Ok(ints) => {
        ints.map(|res| res.map(|value| value as f64))
            .collect::<Result<_, _>>()?
    },
    Err(npy) => npy.data::<f64>()?.collect::<Result<_, _>>()?,
};

[Shatur]

Okay, thanks for the info!