Repository navigation
ENH: Augment ufunc.reduce with state #8773
Description
Right now, reduce can be roughly described as:
it = iter_along_axis(arr) #pseudocode
result = func.identity or next(it)
for i in it:
result = func(result, i)
return resultOr diagramatically:
x2 x3 x3 xn
| | | |
V V V V
x1 -> (f) -> (f) -> (f) ... (f) --> result
What I'm proposing is a more powerful variant that allows extra state to be preserved between each function call:
it = iter_along_axis(arr) #pseudocode
result = func.identity or next(it)
state = None
for i in it:
result, state = func(result, i, reduce_state=state)
return result x2 x3 x3 xn
| | | |
V V V V
x1 -> ( ) -> ( ) -> ( ) ... ( ) --> result
(f) (f) (f) (f)
None -> ( ) -> ( ) -> ( ) ... ( )
This makes ufuncs capable of describing:
argminargmax
*gcd(ENH: Integer gcd function #8772)
Activity
eric-wieser commentedon Mar 12, 2017
An alternative way to implement this might be to just allow the ufunc to modify its
reduce_stateparameter directly, like anoutargumentnjsmith commentedon Mar 12, 2017
eric-wieser commentedon Mar 12, 2017
@njsmith : I was wrong about gcd -
ufunc.reduceis already able to handle this.Yes,
gufuncs do provide a more general solution, but they don't support anaxis=argument for the basic case, nor can they beaccumulated.mhvk commentedon Oct 21, 2017
This makes me realize I should get back to #8819, implementing
axesfor gufuncs...hameerabbasi commentedon Feb 27, 2018
This would also make things like
all_equalrelatively simple without an intermediate array. For example, thestatecould beTrueby default and thenfwould dostate = state && (x_i == x_{i-1}.You would have to output the
statein this case though.mhvk commentedon Feb 28, 2018
Looking at this again, the logical path would seem to be a
gufunc, withreducejust being a special case of one. #8819 is about to provide the ability to work along a single axis. Combined axes oraxes=Noneis less trivial, though (might would need atransposeorravelpreprocessing step).