Closed sunxd3 closed 1 year ago
I'm not sure if this is a good idea, the main point of the ADTypes
extension is to allow users to pass around all information in form of the types from ADTypes
. To me the main point is to not have to specify keyword arguments - otherwise you could also just use ADgradient
.
I also think that the Turing issue that motivated this PR should be fixed in a different way in Turing, to avoid that users have to deal with this problem.
Make sense to me
By default, compilation of ReverseDiff uses zeros as input, but this can cause incorrect compile tape given other inputs. The PR allows user to specify inputs with ADTypes' AutoReverseDiff.