Fix Dirichlet logpdf_with_trans to work with a Vector{Real}

src/Bijectors.jl

@@ -164,7 +164,7 @@ function logpdf_with_trans(d::Distribution, x, transform::Bool)
  if ispd(d)
  return pd_logpdf_with_trans(d, x, transform)
  elseif isdirichlet(d)
- l = logpdf(d, x .+ eps(eltype(x)))
+ l = logpdf(d, x .+ _eps(eltype(x)))
  else
  l = logpdf(d, x)
  end

test/transform.jl

@@ -42,28 +42,51 @@ function single_sample_tests(dist)
 
  # Check that invlink is inverse of link.
  x = rand(dist)
+ _single_sample_tests_inner(dist, x, ϵ)
 
+ # If the sample is a vector of scalars, check that we can run the tests even if the
+ # vector has the abstract element type Real. Skip type stability tests though.
+ if x isa Vector{<:Real}
+ _single_sample_tests_inner(dist, Vector{Real}(x), ϵ, false)
+ end
+end
+
+function _single_sample_tests_inner(dist, x, ϵ, test_type_stability=true)
  if dist isa LKJCholesky
  x_inv = @inferred Cholesky{Float64,Matrix{Float64}} invlink(
  dist, link(dist, copy(x))
  )
  @test x_inv.UL ≈ x.UL atol = 1e-9
  else
- @test @inferred(invlink(dist, link(dist, copy(x)))) ≈ x atol = 1e-9
+ x_reconstructed = if test_type_stability
+ @inferred invlink(dist, link(dist, copy(x)))
+ else
+ invlink(dist, link(dist, copy(x)))
+ end
+ @test x_reconstructed ≈ x atol = 1e-9
  end
 
  # Check that link is inverse of invlink. Hopefully this just holds given the above...
- y = @inferred(link(dist, x))
+ y = if test_type_stability
+ @inferred(link(dist, x))
+ else
+ link(dist, x)
+ end
+ y_reconstructed = if test_type_stability
+ @inferred(link(dist, invlink(dist, copy(y))))
+ else
+ link(dist, invlink(dist, copy(y)))
+ end
  if dist isa Dirichlet
  # `logit` and `logistic` are not perfect inverses. This leads to a diversion.
  # Example:
  # julia> logit(logistic(0.9999999999999998))
  # 1.0
  # julia> logistic(logit(0.9999999999999998))
  # 0.9999999999999998
- @test @inferred(link(dist, invlink(dist, copy(y)))) ≈ y atol = 0.5
+ @test y_reconstructed ≈ y atol = 0.5
  else
- @test @inferred(link(dist, invlink(dist, copy(y)))) ≈ y atol = 1e-9
+ @test y_reconstructed ≈ y atol = 1e-9
  end
  if dist isa SimplexDistribution
  # This should probably be exact.