Differentiation with np.select (piecewise functions) about autograd HOT 8 CLOSED

I think the problem was that np.select was returning a 0-dimensional array (like np.array(FloatNode)) and those are a pain to handle. I think we just have to call wrap_if_nodes_inside. (My first fix unpacked the 0d array into a scalar type, but I realized that constitutes a change in behavior from the unwrapped numpy function.)

Reopen if still busted!

from autograd.

@mattjj Are you sure you don't want to add a test? A while ago I added a systematic test for np.select but left it commented out:
https://github.com/HIPS/autograd/blob/master/tests/test_systematic.py#L176

from autograd.

Yeah yeah okay :)

That test is actually for the gradient of select, right? My fix didn't implement a gradient for select so much as it made select work as control flow (as it would have done automatically if not for the zero-dimensional array issue, afaict).

from autograd.

Except it only works in the scalar case (i.e. when condlist and choicelist are lists of scalars)... reopening!

from autograd.

Okay, I undid the change because making np.select actually work in general seems like a bit of a mess and I can't think of a quick fix. Any ideas?

from autograd.

My quick-fix thinking was that we could just use the fact that np.select should work on object ndarrays, then just wrap its output back to an ArrayNode. I'm kind of confused at the moment as to why that's not working.

from autograd.

Okay, took another quick stab at this one. The basic strategy is for autograd.numpy.select to be a non-primitive function that unboxes its arguments (from ListNode to list, or just keeps them as lists if they come in as lists), uses the underlying np.select on the enclosed object ndarrays (containing FloatNodes and stuff), and then reboxes the result.

Maybe this is shady! But it passes both David's test and the OP's test case.

Since I haven't thought too hard about how shady this might be, I'm leaving it in a branch named issue36 for now.

from autograd.

The consensus is that 2ae6bed is a good fix!

from autograd.