Michael Abbott

Yes, the gradients are computed by very similar loops, details from `verbose=true`: ``` ┌ Info: symbolic gradients │ inbody = │ 2-element Array{Any,1}: │ :(𝛥a[m, i, batch] = 𝛥a[m, i,...

The naiive expectation (for square matrices) would be 3x, as `gradient` first runs the forward pass, and then the reverse pass has two similar multiplications, one to make `𝛥a` and...

Implementing them here (like Base) sounds fine, but is testing with Quaternions going to quadruple the time for tests to run? Would be nice to avoid that if possible. This...

~~Test failure on master is that `"3-element ForwardDiff.Partials{3, Int32}"` now has a space in it, see #476.~~ ~~Test failure on 1.0 might be my fault?~~ Edit: Locally tests pass on...

Here's something this PR breaks: ```julia julia> ForwardDiff.gradient(x -> sum(abs2, unique(x)), [1,2,1,2,3]) |> println [2, 4, 0, 0, 6] # before [2, 4, 2, 4, 6] # after julia> hash(ForwardDiff.Dual(1,0))...

Latest commit proposes to make this a bugfix, since apparently it's what the authors of all the above issues expected, and making a breaking release sounds like it'll take another...

The integer complaint is this, which is unchanged by this PR: ```julia julia> using ForwardDiff: Dual julia> isinteger(Dual(1.0, 0.2)) # is this wrong? true julia> convert(Int, Dual(1.0, 0.2)) ERROR: InexactError:...

This PR is narrower than that, it fixes only `==` (and `isequal`) not all possible measure-zero branches. To start somewhere, and because equality is what seem to cause problems in...

Shall we merge this, before it goes stale again?

Thanks! Test failures on master appear to be small changes in random numbers? This (copied from CI logs) would pass with slightly larger tolerance: ``` julia> @test [1.7010290555795968e-6, 1.063143159737248e-6, 8.505145277897984e-7]...