ndarray equivalent of numpy.astype?

[drewm1980]

How does one convert the type of an array? We're trying to go from an owned Array2<u32> to an owned Array2<usize>.

We've tried to_owned(), into_owned(), and from(), and can't find anything that works. Do we need to resort to iterators for this?

As a documentation feature request, could you please add the answer to the "ndarray_for_numpy_users" page?

The wider context is that we're trying to get our data out of python using the rust-numpy crate:

https://github.com/rust-numpy/rust-numpy

but their example doesn't get the data all the way into an owned ndarray; it's still a view into a custom type wrapping data that is still owned by a python interpreter.

It already took us hours to get this far:

    let a = pyarray.as_array().unwrap();
    let b:ArrayView2<i32> = a.into_dimensionality::<Ix2>().unwrap();
    let c:ndarray::Array2<i32> = b.to_owned();

I'm hoping you can get us the rest of the way there, or maybe even point us to a conversion path that only incurs one copy rather than two...

[jturner314]

A safe way to perform the conversion is arr.mapv(|elem| elem as usize), which copies the data. Instead of copying the data, you could use unsafe code to perform an in-place cast of the array, but getting it correct without undefined behavior is tricky.

For your specific use-case (getting an array of usize from rust-numpy), I recommend a different approach: add support for usize directly to rust-numpy instead of trying to cast an array of u32 to an array of usize. This avoids the need for a copy and doesn't require any tricky casts. At first glance, it looks like all that's necessary is a few lines at the end of rust-numpy/src/types.rs to implement TypeNum for usize.

[jturner314]

By the way, depending on what you're trying to do, you may also be interested in ndarray-npy. For example, I like to do my calculations with ndarray and dump the results to .npy/.npz files with ndarray-npy. I can then easily read the results with NumPy and plot with Maptlotlib.

[drewm1980]

Thanks for the response; we will try mapv out tomorrow. I was expecting a copy anyway since usize is 64 bit on our current target, but maybe the view pyarray returns will work with mapv. Thanks also for the recommendation for ndarray-npy; with a rapidly evolving ecosystem recommendations for what works are always valuable!

[jturner314]

I was expecting a copy anyway since usize is 64 bit on our current target

Yeah, if the new element type is larger than the old element type, you won't be able to avoid a copy.