Graceful handling of a numpy matrix

[gustavla]

If you give quiver numpy matrices, it will crash. It raises a ValueError for a reshape, which can be hard to realize was caused by the matrices (potentially to great frustration for the user).

The source of the problem boils down to ravel() for np.matrix creating a row matrix of shape (1, N), instead of an ndarray with shape (N,).

I naively tried fixing it (pull request: #1546) by throwing in a call to np.asarray, but it was rightly closed because it would break support for masked arrays. It was also pointed out that it was likely to be a problem beyond just quiver.

As I see it, the issue is either resolved by adding support for numpy.matrix, or better information to the users of the source of the error.

[WeatherGod]

Wouldn't doing an np.squeeze() on the result of ravel() do the trick? Not
sure about the implications with masked arrays and such, but just a thought.

[gustavla]

That handles masked arrays well, but it still retains the shape (1, N) after the squeeze.

[tacaswell]

From the other thread, we need to make sure the docs are clear that mpl does not support matrix.

[anntzer]

The correct way to fix this is probably to figure out each call site of np.asanyarray (there's like ~25 of them) to see whether they should be replaced by calls to np.asarray. The trick here is that you probably still want masked arrays to pass through, so perhaps you really need a helper function that behaves like asarray, except that masked arrays are passed through.

This helper function should probably(?) then be used in most places where either asarray or asanyarray is used.

[dopplershift]

Instead of only whitelisting masked arrays, it should blacklist matrix. Otherwise, you needlessly preclude other ndarray subclasses.

[efiring]

On 2016/10/11 9:17 AM, Ryan May wrote:

Instead of only whitelisting masked arrays, it should blacklist matrix.
Otherwise, you needlessly preclude other |ndarray| subclasses.

Right. The matrix subclass is an anomaly that causes more problems than
it is worth; it alters some of the most fundamental ndarray behavior.
In most discussions I have seen, the consensus advice is to discourage
its use.

[anntzer]

Agreed, just blacklisting matrix and matrix subclasses seems better.