[Basic Question] What is the difference in usage between numpy and ndarray key crates?

[rkshthrmsh]

I have been working with the ndarray crate to achieve the Python Numpy features I wanted in Rust so far. I stumbled upon this numpy crate while looking for some of the advanced features such boolean array indexing which the Python version offers.

Now, my questions is, what are the key differences in usage between ndarray (which, I understand is re-exported in this crate) and this numpy crate? This was not clear to me from the documentation. What are situations in which one would use numpy over ndarray?

[adamreichold]

This crate integrates NumPy and ndarray and it is used to implement Python extensions using the PyO3 bindings which provide their Python API in terms of NumPy types but internally work on these arrays using ndarray.

The key functionality to enable this is that your #[pyfunction] can take for example a PyReadonlyArray2<f64> to reference the NumPy array passed in on the Python side and use PyReadonlyArray::as_array to turn this into a ArrayView2<f64> without copying the underlying data which you can operate on using the functionality provided by ndarray.

[rkshthrmsh]

Hi @adamreichold, am I correct in understanding that passing a Rust Array (ArrayBase<OwnedRepr<A>, D>) for computation to Numpy on the Python side requires either consumption of the array (into_pyarray()) or copying the entire array (to_pyarray())?

Is there a way to avoid consumption or copying?

[adamreichold]

Is there a way to avoid consumption or copying?

No there is not as Python cannot safely model borrowing, it always uses shared ownership and hence we need to either produce a completely independent NumPy array by copying or wrap up the owned Array in an immutable object we put onto the Python heap which then backs a NumPy array.