Marking a pybind11::array as read-only

[stukowski]

The method pybind11::array::writeable() tells whether the array's data is mutable. It would be nice to have a corresponding method for marking the data in an array as read-only. Unless I overlooked something, it seems the only way to do this is with a hack like this:

const double* data = ...
size_t num_elements = ...

py::array_t<double> array({ num_elements }, data);

reinterpret_cast<py::detail::PyArray_Proxy*>(array.ptr())->flags &= ~py::detail::npy_api::NPY_ARRAY_WRITEABLE_;

My suggestion is to add a method to the pybind11::array class that allows modifying the array flags (either all or specifically the writable flag).

Background: I would like to implement a getter method binding that returns a NumPy array pointing to an internal C++ data buffer (without making a copy of the data). The Python side should not be able to modify the data, so I would like to mark the array as read-only.

[aldanor]

IIRC when constructing an array, the NPY_ARRAY_WRITEABLE flag is a writeability requirement; so, if this flag is missing and the input array is writeable, it won't make it non-writeable. (we could change the semantics of that, I guess).

Yes, probably makes sense to allow setting writeability, either through some setter or the ctor. I'll add this to the pending flags rework PR.

[aldanor]

@wjakob I was thinking, for this to work properly (i.e. to roundtrip buffer -> array -> buffer), it would make sense to add a readonly (or writeable) flag to buffer_info.

Python buffer interface already supports it: https://docs.python.org/3/c-api/buffer.html#c.Py_buffer.readonly and I believe NumPy arrays would propagate this flag (needs checking).

This way you could e.g. expose an STL vector (as proposed in #488) as a readonly-only buffer, and the corresponding numpy arrays would have writeable set to false by default.

[wjakob]

@aldanor This sounds good to me -- feel free to add such a flag.

[chaoran]

It would be nice to have this feature. I was looking for a way to set an array non writable. I think array::ensure would do the trick but it doesn't.

Also, it would be nice to set other flags on an array too, such as NPY_ARRAY_ALIGNED.

[chaoran]

In addition, I found out if I clear an array's writable flag in C++. Then return that array to python side. The array I got in python is still writable, meaning arr.flags.writable is True.

[jiapei100]

@stukowski
Hey, did you find a solution? I'm having the same issue, please take a look here: #3181

[stukowski]

@stukowski
Hey, did you find a solution? I'm having the same issue, please take a look here: #3181

I am still using the workaround I described in my initial question to modify the flags of the PyArray_Proxy structure and clear the NPY_ARRAY_WRITEABLE bit. But I am not sure, maybe PR #1466 enables a better way to do this nowadays.

[jiapei100]

@stukowski

Great... Thank you so much... This problem is NOW resolved. Thanks.
However, I've got some new issues. Can you please help to take a look at fwilliams/numpyeigen#54 ?