A differential eDSL that can run faster than light and go back to the past.
243 Stars
Updated Last
1 Year Ago
Started In
November 2019

NiLang.jl (逆lang), is a reversible domain-specific language (DSL) that allow a program to go back to the past.

  • Requires Julia version >= 1.3,

NiLang features:

  • any program written in NiLang is differentiable,
  • a reversible language with abstraction and arrays,
  • complex values
  • reversible logarithmic number system

CI codecov

The main docs can be found here:

There are also some Pluto-based notebooks:

The strangeness of reversible computing is mainly due to our lack of experience with it.—Henry Baker, 1992

To Start

pkg> add NiLang

An example: Compute the norm of a vector

julia> using NiLang

julia> @i function f(res, y, x)
           for i=1:length(x)
               y += x[i] ^ 2
           res += sqrt(y)

julia> res_out, y_out, x_out = f(0.0, 0.0, [1, 2, 3.0])
(3.7416573867739413, 14.0, [1.0, 2.0, 3.0])

julia> (~f)(res_out, y_out, x_out)  # automatically generated inverse program.
(0.0, 0.0, [1.0, 2.0, 3.0])
julia> ∂res, ∂y, ∂x = NiLang.AD.gradient(Val(1), f, (0.0, 0.0, [1, 2, 3.0])) 
    # automatic differentiation, `Val(1)` means the first argument of `f` is the loss.
(1.0, 0.1336306209562122, [0.2672612419124244, 0.5345224838248488, 0.8017837257372732])

The performance of reversible programming automatic differentiation is much better than most traditional frameworks. Here is why, and how it works,

how it works

Check our paper

    title={Differentiate Everything with a Reversible Programming Language},
    author={Jin-Guo Liu and Taine Zhao},