Is it possible to unnormalise data in Flux.jl?

New to Julia
,
1

Hi,

I’ve seen a convenient normalisation function in Flux.jl, Flux.normalise.
However, what if I want to unnormalise data (recover the original data)? Is there such functionality in Flux.jl?

Thanks.

2

you will have to store the mean and variance by yourself to perform the inverse transform, otherwise that information is lost

3

Nice thing about Flux is that it works with all julia packages.
I am surprised it actually has a normalize of it’s own.

Something like StatsBase can do this.

julia> using StatsBase

julia> x = randn(10) .+ 5
10-element Vector{Float64}:
 5.964172883616336
 4.116442184696866
 6.605584680364782
 3.519368666428235
 6.201987836586334
 4.446148516233912
 6.684526916570591
 4.791433884353377
 3.683098936166946
 4.955188515988089

julia> t = fit(ZScoreTransform, x)
ZScoreTransform{Float64, Vector{Float64}}(1, 1, [5.0967953021005465], [1.1905029391356516])

julia> px = StatsBase.transform(t, x)
10-element Vector{Float64}:
  0.7285807980831507
 -0.823478115993105
  1.267354601711165
 -1.3250086025134726
  0.9283408701940692
 -0.5465310201913723
  1.3336645902132718
 -0.2564978276902644
 -1.1874782660847485
 -0.11894702772869165

julia> StatsBase.reconstruct
reconstruct  reconstruct!
julia> StatsBase.reconstruct(t, px)
10-element Vector{Float64}:
 5.964172883616336
 4.116442184696866
 6.605584680364782
 3.5193686664282353
 6.201987836586334
 4.446148516233912
 6.684526916570591
 4.791433884353377
 3.683098936166946
 4.955188515988089

Another option is FeatureTransforms.jl

julia> using FeatureTransforms

julia> t = MeanStdScaling(x)
MeanStdScaling(5.0967953021005465, 1.1905029391356516)

julia> xp = FeatureTransforms.apply(x, t)
10-element Vector{Float64}:
  0.7285807980831507
 -0.823478115993105
  1.267354601711165
 -1.3250086025134726
  0.9283408701940692
 -0.5465310201913723
  1.3336645902132718
 -0.2564978276902644
 -1.1874782660847485
 -0.11894702772869165

julia> FeatureTransforms.apply(x, t; inverse=true)
10-element Vector{Float64}:
 12.197160649558949
  9.997431821764149
 12.960763278784253
  9.286614043385278
 12.48028005004014
 10.389948178510636
 13.05474424300921
 10.801011424097393
  9.481535410734686
 10.995961794355594

I haven’t actually tested these with Flux.
They should work, though they might give mutation errors.

4 Likes
4

I’ve tested StatsBase.jl with Flux.jl as you suggested but it does not work properly.
It seems that StatsBase.jl mutates arrays, which is not supported by Zygote.jl for auto-differentiation.

5

Did you try the second package @oxinabox mentioned? FeatureTransforms should be relatively AD-safe since it doesn’t use mutation at all.

6

Not yet.
Since the package is relatively young, I thought there is a risk to use it.
But it would be worth to try it. Thanks :slight_smile:

7

NB I work for Invenia, so I am in no way unbiased.
But you can trust most things we make.
We do this as a job, we have higher standards than anyone else for continued maintainance.
That one is being used in like 6+ components of our production system.

2 Likes