Safe conversion of numpy array to ndarray without copying data

[peendebak]

From Python I am trying to convert an object into the following structure:

pub enum ArrayLike {
    Float(f64),
    Int(isize),
    FloatArray(ArrayD<f64>), // todo: replace with ArrayViewD<a', f64>
}

I have this working, with the following definition and the standard pyo3 setup:

impl<'source> FromPyObject<'source> for ArrayLike {
    fn extract(ob: &'source PyAny) -> PyResult<Self> {
        if let Ok(f) = <PyFloat as pyo3::PyTryFrom>::try_from(ob) {
            return Ok(ArrayLike::Float(f.extract()?));
        } else if let Ok(i) = <PyLong as pyo3::PyTryFrom>::try_from(ob) {
            return Ok(ArrayLike::Float(i.extract()?));
        } else if let Ok(a) = <PyArrayDyn<f64> as pyo3::PyTryFrom>::try_from(ob) {
            let b = unsafe { a.as_array() };
            return Ok(ArrayLike::FloatArray(b.to_owned()));
        } else {
             ....

Note that there is an unsafe block for the numpy array conversion. The documentation of as_array suggests to either use a PyReadonlyArray or the to_array method.

1. I cannot find the to_array method on the documentation for latest. Does it exist, or is the documentation old?
2. I tried to use the PyReadonlyArray by changing the code to

        } else if let Ok(a) = <PyReadonlyArrayDyn<f64> as pyo3::PyTryFrom>::try_from(ob) {

but this results in the trait 'PyNativeType' is not implemented for 'PyReadonlyArray<'_, f64, Dim<IxDynImpl>>'. Any suggestions on how to use PyReadonlyArray?

[adamreichold]

I think this discussion is probably more relevant to the rust-numpy crate providing the bindings between NumPy and ndarray than to ndarray itself.

In any case, maybe the example provided here is helpful for you as an example of how to implement polymorphism like use the derive macro for FromPyObject.

Concerning the question on PyReadonlyArray, it is a wrapper around PyArray and hence not itself a native Python type. It does implement FromPyObject, i.e. you should be able to use instead of FromPyObject::extract instead of PyTryFrom::try_from. Or something like

#[derive(FromPyObject)]
pub enum ArrayLike<'a> {
    Float(f64),
    Int(isize),
    FloatArray(PyReadonlyArrayD<'a, f64>),
}