Fix @test_broken problems

[simsurace]

Fixes part of #526

[devmotion]

There are more bugs regarding @test_broken such as

KernelFunctions.jl/test/transform/selecttransform.jl

Line 107 in f71cfe0

@test_broken ga = gradient(AD, A) do a

: @test_broken should only applied to expressions that evaluate to a Bool (which in the broken case is false but true if the problem is fixed).

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (e5696d7) 67.00% compared to head (6412453) 0.44%.
Report is 1 commits behind head on master.

❗ Current head 6412453 differs from pull request most recent head 50de544. Consider uploading reports for the commit 50de544 to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #546       +/-   ##
==========================================
- Coverage   67.00%   0.44%   -66.56%     
==========================================
  Files          52      52               
  Lines        1373    1342       -31     
==========================================
- Hits          920       6      -914     
- Misses        453    1336      +883     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[willtebbutt]

This is great. Please merge if you're happy.

[simsurace]

There are more bugs regarding @test_broken such as

KernelFunctions.jl/test/transform/selecttransform.jl

Line 107 in f71cfe0

@test_broken ga = gradient(AD, A) do a

: @test_broken should only applied to expressions that evaluate to a Bool (which in the broken case is false but true if the problem is fixed).

I see. Why don't they show up as errors though?
What is the test supposed to do anyway? It is just evaluating a gradient and assigning it to ga. What is broken about this?

[devmotion]

Probably it throws an error currently which is handled by @test_broken as well. So that case is currently fine - but it's still wrong since the tests will error as soon as the gradient computation does not throw an error anymore. See https://discourse.julialang.org/t/psa-when-tests-fail-due-to-improvements-in-type-inference/102001

[simsurace]

Would it be reasonable to replace these by an approximate equality comparing the chosen AD backend to :FiniteDiff?

[devmotion]

I guess these are the intended tests:

KernelFunctions.jl/test/transform/selecttransform.jl

Lines 76 to 82 in f71cfe0

	gx = gradient(AD, X) do x
	testfunction(tx_row, x, 2)
	end
	ga = gradient(AD, A) do a
	testfunction(ta_row, a, 2)
	end
	@test gx ≈ ga

So for the ReverseDiff we could just not create the intermediate variables and directly @test_broken the approximate equality of the two gradients?

[simsurace]

This should do it.