Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Fix and add more AD tests #625

Merged
merged 3 commits into from
Dec 14, 2018
Merged

Fix and add more AD tests #625

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
b42a7d8
add binomlogpdf ForwardDiff.Dual workaround
mohamed82008 Dec 13, 2018
6ebdf12
fix broken AD tests and add some more
mohamed82008 Dec 13, 2018
85ef546
remove FDM from REQUIRE
mohamed82008 Dec 14, 2018
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 7 additions & 0 deletions src/core/ad.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -134,6 +134,13 @@ Tracker.@grad function poislogpdf(v::Tracker.TrackedReal, x::Int)
Δ->(Δ * (x/v - 1), nothing)
end

function binomlogpdf(n::Int, p::ForwardDiff.Dual{T}, x::Int) where {T}
FD = ForwardDiff.Dual{T}
val = ForwardDiff.value(p)
Δ = ForwardDiff.partials(p)
return FD(binomlogpdf(n, val, x), Δ * (x / val - (n - x) / (1 - val)))
end

function poislogpdf(v::ForwardDiff.Dual{T}, x::Int) where {T}
FD = ForwardDiff.Dual{T}
val = ForwardDiff.value(v)
Expand Down
28 changes: 26 additions & 2 deletions test/ad.jl/AD_compatibility_with_distributions.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,6 +37,12 @@ let
rtol=1e-8,
atol=1e-8,
)
@test isapprox(
Tracker.gradient(foo, 0.5)[1],
ForwardDiff.derivative(foo, 0.5);
rtol=1e-8,
atol=1e-8,
)

bar = p->logpdf(Binomial(10, p), 3)
@test isapprox(
Expand All @@ -45,22 +51,40 @@ let
rtol=1e-8,
atol=1e-8,
)
@test isapprox(
Tracker.gradient(bar, 0.5)[1],
ForwardDiff.derivative(bar, 0.5),
rtol=1e-8,
atol=1e-8,
)
end

let
foo = p->poislogpdf(1, p)
foo = p->Turing.poislogpdf(p, 1)
@test isapprox(
Tracker.gradient(foo, 0.5)[1],
central_fdm(5, 1)(foo, 0.5);
rtol=1e-8,
atol=1e-8,
)
@test isapprox(
Tracker.gradient(foo, 0.5)[1],
ForwardDiff.derivative(foo, 0.5);
rtol=1e-8,
atol=1e-8,
)

bar = p->logpdf(Poisson(1), 3)
bar = p->logpdf(Poisson(p), 3)
@test isapprox(
Tracker.gradient(bar, 0.5)[1],
central_fdm(5, 1)(bar, 0.5);
rtol=1e-8,
atol=1e-8,
)
@test isapprox(
Tracker.gradient(bar, 0.5)[1],
ForwardDiff.derivative(bar, 0.5);
rtol=1e-8,
atol=1e-8,
)
end
6 changes: 3 additions & 3 deletions test/ad.jl/adr.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,8 +23,8 @@ mvn = collect(Iterators.filter(vn -> vn.sym == :m, keys(vi)))[1]
_s = getval(vi, svn)[1]
_m = getval(vi, mvn)[1]

x = map(_->Float64(_), vi[nothing])
∇E = gradient_reverse(x, vi, ad_test_f)
x = map(x->Float64(x), vi[nothing])
∇E = gradient_reverse(x, vi, ad_test_f)[2]
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is the [2] on purpose?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, the first value seems to be negative the logp, whereas the second is the gradient. This is defined here.

# println(vi.vns)
# println(∇E)
grad_Turing = sort(∇E)
Expand All @@ -37,7 +37,7 @@ function logp(x::Vector)
# s = invlink(dist_s, s)
m = x[1]
lik_dist = Normal(m, sqrt(s))
lp = logpdf(dist_s, s, false) + logpdf(Normal(0,sqrt(s)), m, false)
lp = logpdf(dist_s, s) + logpdf(Normal(0,sqrt(s)), m)
lp += logpdf(lik_dist, 1.5) + logpdf(lik_dist, 2.0)
lp
end
Expand Down
2 changes: 0 additions & 2 deletions test/ad.jl/skip_tests
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1 @@
# tests to skip
adr.jl
AD_compatibility_with_distributions.jl