TypeError: argument 'input': 'ndarray' object is not an instance of 'ndarray'

[mikelync]

When trying to capture an Array from Python using the following input scheme:

fn _set_node_vector<'py>(
    &mut self, 
    py: Python<'py>, 
    input: Bound<'py, numpy::PyArray1<Py<PyAny>>>,
) -> PyResult<()>

I have three cases:
a) Valid input: the array is correctly specified: v1 = np.array([1.0, 2.0], dtype=object)
b) Invalid i1 = np.array([1.0, 2.0]) (possibly because the dtype is float)
c) Invalid: i2 = np.array([[1.0, 2.0]], dtype=object) (shape is wrong)

In both invalid cases the error returned is:

TypeError: argument 'input': 'ndarray' object is not an instance of 'ndarray'

Bit more steering can be helpful...

[davidhewitt]

Thanks for the report. Agreed this message should be better, but I'm not sure there's an easy way to fix this.

The blanket FromPyObject implementation for Bound uses .cast() internally, which will always produce an error of this form. I guess if PyTypeInfo / PyTypeCheck implementations were removed from PyArray1, the blanket would not apply and a custom FromPyObject implementation could be used.

[asjad2401]

Related to this: I filed #549 which describes the same root cause from a different angle — PyTypeCheck not checking element dtype means cast_exact::<PyArrayDyn>() silently succeeds on a float32 array, and there's currently no non-deprecated way to do dtype-aware dispatch.

@davidhewitt's suggestion of removing PyTypeCheck and using a custom FromPyObject implementation would fix both issues. It would give correct dtype dispatch and produce a meaningful error message.