ENH: add a neigborwise function #9428
Conversation
|
What about a general purpose moving window function which would apply a function to |
For large arrays, the speed is the same. The overhead of building the slices should be small. |
|
New public APIs need a mailing list post. The more general/conventional way of naming this pattern is moving/sliding window, as @bashtage notes. This is something that pandas has some specialty in – might be worth checking out what they do. |
|
Aren't a lot of these operations also much faster using convolutions? |
| for ax in axis: | ||
| x_begin = x[slices[:ax] + begin_sl + slices[ax+1:]] | ||
| x_end = x[slices[:ax] + end_sl + slices[ax+1:]] | ||
| x = func(x_begin, x_end) |
There was a problem hiding this comment.
To be used as a backend for diff, this would need to be flipped
There was a problem hiding this comment.
Thanks. This is probably better than adding a method to ufuncs. It will also be much easier to introduce this to the ma module than a ufunc method.
There was a problem hiding this comment.
Why does the ma module need its own version? What would be different
There was a problem hiding this comment.
It may not. I am not 100% sure that it will combine the masks correctly otherwise though.
|
|
||
| def neighborwise(x, func, axis=None): | ||
| """ | ||
| Apply a function to pairs of neigbours along the given axes |
There was a problem hiding this comment.
I am pretty sure a period is required at the end of this line.
| Can be used to resample a grid at the center of each cell, useful for | ||
| matploblib's `pcolor`: | ||
|
|
||
| >>> midpoint = lambda a, b: (a + b) / 2 |
There was a problem hiding this comment.
This could be done with just np.add if you multiply the result by 0.5.
| func : function | ||
| A function taking two arguments, to call on the (vectorized) pairs of | ||
| neighbours. The first argument is the values at the lower indices | ||
| axis : int or sequence of int |
There was a problem hiding this comment.
It was my impression that None usually means to operate on a raveled version.
There was a problem hiding this comment.
Technically any iterable seems to work here.
| >>> midpoint = lambda a, b: (a + b) / 2 | ||
| >>> x, y = np.indices((3, 4)) * 2 | ||
| >>> x | ||
| array([[0, 0, 0, 0], |
There was a problem hiding this comment.
This should be (0, 2, 4) repeated, or remove the * 2 in the line above.
| [1, 1, 1, 1], | ||
| [2, 2, 2, 2]]) | ||
| >>> y | ||
| array([[0, 1, 2, 3], |
There was a problem hiding this comment.
This should be (0, 2, 4, 6) repeated.
|
I did not mean that this function can always be replaced by a convolution. What I meant was that I do not quite like that in many cases where the function might seem like "the right tool", it may actually not be. And if those cases are a big portion of the likely use cases, I get unsure. |
|
@charris how about we also add a label like "Good start" or so? Just to make it a bit easier to find these to reopen them (I am not sure that "Inactive" quite covers that part). |
|
Sure, "Inactive" was an existing label, so I reused it. Maybe "Good idea"? Someone might be motivated to follow up. |
|
"Good idea" sounds good. Will add it later if you don't and then tag some of these at least (and anyone should feel free to tag more). |
seberg
added
the
64 - Good Idea
This ends up being something I write a lot. It would be handy for matplotlib/matplotlib#6404 and
axes.pcolor.Naming and argument order perhaps needs work. Suggestions?