JuliaDiff
Comparison tests ZeroTangent with 0-like numbers should be true
Was just testing something boiling down to:
ZeroTangent() ≈ 0
Should this be accepted? Along with arrays for which the norm is 0?
Yes, though we use test_approx in ChainRulesTestUtils
julia> using ChainRulesTestUtils
julia> test_approx(ZeroTangent(), 0)
Test Passed
julia> test_approx(ZeroTangent(), zeros(2, 3))
Test Passed
indeed good point. Could it be worth it to have it somewhere close to the ZeroTangent doc?
I would argue it belongs to the ChainRulesTestUtils docs, but we could certainly add a link to those.
What's the argument against defining these?
Besides being one less thing to know about, @test x ≈ 0.0 will I think give more useful information when it fails, as the macro records where it was called.
Maybe array ones are weird, in that @test x ≈ [0 0 0] normally implies you check the size as well as the contents, whereas Zero doesn't have any size of course.
Finally, x ≈ 0 is the same as x == 0 for numbers, I think. Unless you supply the tolerance. But ZeroTangent() == 0 is false, not undefined.
I think the reason we have test_approx is to better display the information when we have thunking involved.
compare:
julia> using ChainRulesCore
julia> using ChainRulesTestUtils
julia> using Test
julia> struct Foo end
julia> unfoo(x) = x
unfoo (generic function with 1 method)
julia> unfoo(f::Foo) = 0.0
unfoo (generic function with 2 methods)
julia> Base.isapprox(f::Foo, b) = isapprox(unfoo(f), b)
julia> @test Foo() ≈ 0.1
Test Failed at REPL[11]:1
Expression: Foo() ≈ 0.1
Evaluated: Foo() ≈ 0.1
ERROR: There was an error during testing
to
julia> using ChainRulesCore
julia> using ChainRulesTestUtils
julia> using Test
julia> struct Foo end
julia> unfoo(f::Foo) = 0.0
unfoo (generic function with 1 method)
julia> unfoo(x) = x
unfoo (generic function with 2 methods)
julia> ChainRulesTestUtils.test_approx(f::Foo, b) = test_approx(unfoo(f), b)
julia> test_approx(Foo(), 0.1)
Test Failed at /Users/mzgubic/JuliaEnvs/ChainRulesTestUtils.jl/src/check_result.jl:24
Expression: isapprox(actual, expected; kwargs...)
Evaluated: isapprox(0.0, 0.1)
ERROR: There was an error during testing
We used to define ≈.
It was surprisingly awkward.
And it gave surprisingly bad error messages.
I suspect it is also pretty invalidating, if we define it in CRC, but I haven't tested that.
