Should to_vec(::Integer) return an empty vector #188

Activity

1. willtebbutt commented on Aug 4, 2021

   

   willtebbutt

   on Aug 4, 2021

   Member

   More actions

   This would make sense to me.
2. oxinabox commented on Oct 1, 2021

   

   oxinabox

   on Oct 1, 2021

   Last edited by oxinabox

   Member

   More actions

   The practical problem with this, is that sometimes people want to test special case methods that take integers.
   Which is just all round hard, since when finite diferencing they don't get integers.

   But they might still want to test it against custom rules that do.
   So I wonder if we really need to handle this better in CRTU or somewhere in FiniteDifferences.

   Right now we often (always?) error for it
3. willtebbutt commented on Oct 1, 2021

   

   willtebbutt

   on Oct 1, 2021

   Member

   More actions

   Could you provide an example @oxinabox ? I'm not really clear what sort of thing you're imagining.
4. oxinabox commented on Oct 1, 2021

   

   oxinabox

   on Oct 1, 2021

   Member

   More actions

   For example

   function rrule(::typeof(sinpi), x)
       sinpi_pullback(dy) = -pi*cospi(x)
       foo(x), sinpi_pullback
   end
   
   function rrule(::typeof(sinpi), x::Integer)
       sinpi_pullback(dy) = Zero()
       foo(x), sinpi_pullback
   end
   
5. willtebbutt commented on Oct 1, 2021

   

   willtebbutt

   on Oct 1, 2021

   Member

   More actions

   And what we're proposing to do here would solve this problem, yes?
6. oxinabox commented on Oct 1, 2021

   

   oxinabox

   on Oct 1, 2021

   Member

   More actions

   Would it? Cool!
7. mcabbott commented on Oct 1, 2021

   

   mcabbott

   on Oct 1, 2021

   Member

   More actions

   Isn't this example the thing which already works? Integer taken to mean non-differentiable:

   julia> mysinpi(x) = sinpi(x)
   mysinpi (generic function with 1 method)
   
   julia> function ChainRulesCore.rrule(::typeof(mysinpi), x)
              sinpi_pullback(dy) = NoTangent(), pi*cospi(x)*dy
              sinpi(x), sinpi_pullback
          end
   
   julia> function ChainRulesCore.rrule(::typeof(mysinpi), x::Integer)
              sinpi_pullback(dy) = NoTangent(), NoTangent()
              sinpi(x), sinpi_pullback
          end
   
   julia> test_rrule(mysinpi, 1.23)
   Test Summary:                  | Pass  Total
   test_rrule: mysinpi on Float64 |    7      7
   Test.DefaultTestSet("test_rrule: mysinpi on Float64", Any[], 7, false, false)
   
   julia> test_rrule(mysinpi, 1)
   Test Summary:                | Pass  Total
   test_rrule: mysinpi on Int64 |    6      6
   Test.DefaultTestSet("test_rrule: mysinpi on Int64", Any[], 6, false, false)
   
8. willtebbutt commented on Oct 1, 2021

   

   willtebbutt

   on Oct 1, 2021

   Member

   More actions

   Oh, yeah, this particular example is. I think stuff would need to be nested in order to break it.