ENH: Add ndmax param for numpy.array

[pdnm]

This PR implements the proposal in #5933.

The parameter makes it possible to prevent array from digging deeper into the input, even if the lengths at the the next level are all equal.

Example:

np.array([[1, 2], [3, 4]], ndmax=1)
# array([list([1, 2]), list([3, 4])], dtype=object)

[eric-wieser]

Thanks for the patch. I'm not yet sure whether a ndmax argument to array is the right solution, but this patch does a nice job of showing it's a straighforward one.

At some point I'll read through the original issue and come back to this.

[eric-wieser]

What's the difference between maxdims and max_depth?

[pdnm]

max_depth is like a guard, if the discovered depth is more than that, the creation fails

[seberg]

Thanks for the PR. Sorry for being late to the party, but I realize there is not actually much overlap with my PR, so its fine.

We have to figure out what the exact API is that we want, e.g. I could imagine just fixing ndim=3 instead of ndmax (but then you have to decide whether you allow broadcasting the input to the output).

One thing that probably currently fails is if the input is already an array(-like), i.e. we should probably pass in max_depth and maxdims from python.

[seberg]

One API consideration. If we add this argument (whatever it is) we should deprecate ragged arrays for objects as well probably (although that can be a followup of course). That may be argument for a "maximum" as opposed to a fixed number of dims though.

[pdnm]

Thanks for the PR. Sorry for being late to the party, but I realize there is not actually much overlap with my PR, so its fine.

We have to figure out what the exact API is that we want, e.g. I could imagine just fixing ndim=3 instead of ndmax (but then you have to decide whether you allow broadcasting the input to the output).

One thing that probably currently fails is if the input is already an array(-like), i.e. we should probably pass in max_depth and maxdims from python.

@seberg Indeed it currently fails if the input is an array-like.
I imagine ndim=3 could be achieved by the users with np.array(obj, ndmin=3, ndmax=3)

What "ragged arrays for objects" do you mean?

[mattip]

What "ragged arrays for objects" do you mean?

See NEP 34

[mattip]

Very outdated, and mostly covered by NEP 34. We may want to continue this as an idea for consideration at some point, but would need to expose the discovery function.