> > Are there any benchmark results for the "more performant and accurate" bit? >
Very valid question! There are some benchmarks here, near the bottom of this file <https://github.com/mlubin/EuroAD2015/blob/master/forwarddiff.ipynb>. The code used to run those benchmarks can be found here <https://github.com/JuliaDiff/ForwardDiff.jl/tree/master/benchmarks>. Those benchmarks are actually old, though - the current release should actually be quite a bit faster than the version we ran those benchmarks against (they were actually taken right before we fixed a major performance regression - updating them is on my to-do list). You may notice that we only really compare against other automatic differentiation (AD) tools. That's because AD encompasses a whole different class of techniques then what's referred to as "symbolic" or "numeric" differentiation; results of AD algorithms are generally within machine epsilon precision of the "real" answers, and these algorithms have provably faster runtime complexities (e.g. potentially O(1) evaluations of the target function using AD vs O(n) evaluations of the target function using finite differencing). Thus, the benchmarks don't compare against non-AD methods as it's not a very useful comparison for how fast the package is; the non-AD methods are generally provably slower. That being said, we should probably start benchmarking against traditional methods just to track performance regressions, and to win over people who haven't yet encountered AD. In that spirit, here's a benchmark comparing Calculus's gradient method with ForwardDiff's: julia> function ackley(x) a, b, c = 20.0, -0.2, 2.0*π len_recip = inv(length(x)) sum_sqrs = zero(eltype(x)) sum_cos = sum_sqrs for i in x sum_cos += cos(c*i) sum_sqrs += i^2 end return (-a * exp(b * sqrt(len_recip*sum_sqrs)) - exp(len_recip*sum_cos) + a + e) end ackley (generic function with 1 method) julia> calc_g = Calculus.gradient(ackley); julia> @time calc_g(x); 4.379173 seconds (69.50 k allocations: 1.137 MB) julia> fd_g = ForwardDiff.gradient(ackley, chunk_size=10); julia> @time fd_g(x); 0.523414 seconds (2.02 k allocations: 266.266 KB) Best, Jarrett On Friday, September 4, 2015 at 5:27:27 AM UTC-4, Johan Sigfrids wrote: > > Are there any benchmark results for the "more performant and accurate" bit? > > On Thursday, September 3, 2015 at 11:25:01 PM UTC+3, Jarrett Revels wrote: >> >> I'm proud to announce that we've tagged and released a new version >> ForwardDiff.jl (https://github.com/JuliaDiff/ForwardDiff.jl). >> >> ForwardDiff.jl is a package for performing automatic differentiation >> <https://en.wikipedia.org/wiki/Automatic_differentiation> on native >> Julia functions/callable objects. The techniques used by this package *are >> more performant and accurate than other standard algorithms for >> differentiation*, so if taking derivatives is something you're at all >> interested in, I suggest you give ForwardDiff.jl a try! >> >> If you don't already have the package, you can install it with Julia's >> package manager by running the following: >> >> julia> Pkg.update(); Pkg.add("ForwardDiff") >> >> If you already have the old version of ForwardDiff.jl, you can update it >> to the new one by simply running Pkg.update(). >> >> Note that *the new version of ForwardDiff.jl only supports Julia v0.4. >> *Julia >> v0.3 users will have to stick to the old version of the package. Also note >> that *the new version introduces some breaking changes*, so you'll >> probably have to rewrite any old ForwardDiff.jl code you have (I promise >> it'll be worth it). >> >> I've spent a good chunk of the summer overhauling it as part of my Julia >> Summer of Code project, so I hope other folks will find it useful. As >> always, opening issues and pull requests in ForwardDiff.jl's GitHub repo is >> very welcome. >> >> I'd like to thank Julia Computing, NumFocus, and The Betty and Gordon >> Moore Foundation for putting JSoC together. And, of course, I thank Miles >> Lubin, Theo Papamarkou, and a host of other Julians for their invaluable >> guidance and mentorship throughout the project. >> >> Best, >> Jarrett >> >> >>
