# ReverseDiff not using custom derivative from ChainRules
**URL:** https://discourse.julialang.org/t/reversediff-not-using-custom-derivative-from-chainrules/89888
**Category:** General Usage
**Tags:** reversediff, autodiff, chainrulescore
**Created:** [November 7, 2022, 3:19pm UTC](https://discourse.julialang.org/t/reversediff-not-using-custom-derivative-from-chainrules/89888 "2022-11-07T15:19:10Z")
**Posts on this page:** 4
**Page:** 1
### Author:  [@mrVeng](https://discourse.julialang.org/u/mrVeng)
#### Post date: [November 7, 2022, 3:19pm UTC](https://discourse.julialang.org/t/reversediff-not-using-custom-derivative-from-chainrules/89888/1 "2022-11-07T15:19:10Z")
Hi there,
I would like to use a custom derivative for a function - the cdf of a T distribution wrt to the data. I keep the parameter of this distribution fixed, so NoTangent for the parameter is enough for me here. Here is my current progress:
```julia
using Distributions, ChainRulesCore, ForwardDiff, ReverseDiff
import Distributions: cdf, TDist
import ChainRulesCore: rrule
function mytargetfunction(data::AbstractVector)
function obtaingradient(ν::AbstractVector{R}) where {R<:Real}
dist = Distributions.TDist(1.0)
_data = data ./ sum(ν)
data_uniform = [cdf(dist, _data[iter]) for iter in eachindex(data)] #Line of error
return sum( logpdf(dist, data_uniform[i]) for i in eachindex(data_uniform) )
end
end
data = randn(1000)
ν = [1., 2., 3., 4., 5.]
target = mytargetfunction(data)
target(ν)
ReverseDiff.gradient(target, ν) # ArgumentError: Converting an instance of ReverseDiff.TrackedReal{Float64, Float64, ReverseDiff.TrackedArray{Float64, Fl ....
```
Now I believe ReverseDiff does not know how to take the derivative wrt to `cdf(TDist(1.0), x)`, so I want to write a custom rule:
```julia
function ChainRulesCore.rrule(::typeof(cdf), d::T, x
) where {T<:Distributions.TDist}
∇cdf(x) = Distributions.pdf(d, x) # Gradient w.r.t. x
val = cdf(d, x)
function _pullback(Δy)
# Only return differential w.r.t. x, keep d parameter as NoTangent
return NoTangent(), NoTangent(), Δy * ∇cdf(x)
end
return val, _pullback
end
target = mytargetfunction(data)
target(ν)
ReverseDiff.gradient(target, ν) # ArgumentError: Converting an instance of ReverseDiff.TrackedReal{Float64, Float64, ReverseDiff.TrackedArray{Float64, Fl ....
```
Unfortunately, I still seem to get the same error for some reason, and ReverseDiff does not seem to use my custom rule? I tried to check this and the rrule seems to work:
```julia
_val = cdf(TDist(10.), 1.)
_deriv = pdf(TDist(10.), 1.)
_RDval, _RDpullback = ChainRulesCore.rrule(Distributions.cdf, TDist(10.), 1.0)
_, _, _RDderiv = _RDpullback(1.0)
isequal(_val, _RDval) #true
isequal(_deriv, _RDderiv) #true
```
I would be grateful if someone could provide some guidance here. I have seen that one needs to opt-in for ReverseDiff to use ChainRules, but I dont know how to do that unfortunately.
---
### Author:  [@StevenWhitaker](https://discourse.julialang.org/u/StevenWhitaker)
#### Post date: [November 7, 2022, 4:35pm UTC](https://discourse.julialang.org/t/reversediff-not-using-custom-derivative-from-chainrules/89888/2 "2022-11-07T16:35:44Z")
You need to use [`ReverseDiff.@grad_from_chainrules`](https://github.com/JuliaDiff/ReverseDiff.jl/blob/df0067465ca436b05377054ac7a184da691db1a3/src/macros.jl#L299). So adding
```julia
ReverseDiff.@grad_from_chainrules cdf(d, x::TrackedReal)
```
should help.
---
### Author:  [@mrVeng](https://discourse.julialang.org/u/mrVeng)
#### Post date: [November 7, 2022, 5:05pm UTC](https://discourse.julialang.org/t/reversediff-not-using-custom-derivative-from-chainrules/89888/3 "2022-11-07T17:05:48Z")
Thanks a lot!
Is there a way that this macro also works with types outside ReverseDiff? I.e.:
```julia
ReverseDiff.@grad_from_chainrules cdf(d::D, x::R) where {D<:Distributions.TDist, R<:TrackedReal} #LoadError: `@grad_from_chainrules` has to be applied to a function signature
ReverseDiff.@grad_from_chainrules cdf(d::Distributions.TDist, x::TrackedReal) #UndefVarError: Distributions not defined
```
I guess I could make this otherwise work with closures or structs but this would be more convenient. Just using `cdf(d, x::TrackedReal)` would be ambigous for all the methods defined in Distributions itself as TrackedReal should be a subtype of Real.
---
### Author:  [@StevenWhitaker](https://discourse.julialang.org/u/StevenWhitaker)
#### Post date: [November 7, 2022, 6:02pm UTC](https://discourse.julialang.org/t/reversediff-not-using-custom-derivative-from-chainrules/89888/4 "2022-11-07T18:02:18Z")
I’m not exactly sure what causes the second error (maybe something along the lines of the macro expanding in the context of module `ReverseDiff`, which doesn’t have `Distributions` loaded). But try `Main.Distributions.TDist` instead and see if that works. ~~(Of course, replace `Main` with whatever module you are working in.)~~
EDIT: If you are working in module `MyModule`, you could try `Main.MyModule.Distributions.TDist`.
EDIT: It appears this approach doesn’t work if you are creating a package (there are errors during precompilation about `MyModule` not being defined). I’m not sure what to do in this case.