# Gumbel Softmax in Julia
**URL:** https://discourse.julialang.org/t/gumbel-softmax-in-julia/26190
**Category:** General Usage
**Tags:** question, package
**Created:** [July 9, 2019, 3:12pm UTC](https://discourse.julialang.org/t/gumbel-softmax-in-julia/26190 "2019-07-09T15:12:08Z")
**Posts on this page:** 5
**Page:** 1
### Author:  [@Elita\_Lobo](https://discourse.julialang.org/u/Elita_Lobo)
#### Post date: [July 9, 2019, 3:12pm UTC](https://discourse.julialang.org/t/gumbel-softmax-in-julia/26190/1 "2019-07-09T15:12:08Z")
Can someone please point me to some implementation of gumbel softmax in julia or an equivalent of the following function?
[https://pytorch.org/docs/master/tensors.html#torch.Tensor.scatter\_](https://pytorch.org/docs/master/tensors.html#torch.Tensor.scatter_)
---
### Author:  [@robsmith11](https://discourse.julialang.org/u/robsmith11)
#### Post date: [July 9, 2019, 5:43pm UTC](https://discourse.julialang.org/t/gumbel-softmax-in-julia/26190/2 "2019-07-09T17:43:14Z")
Isn’t the Gumbel softmax trick just the argmax of the sum of the log-probabilities and samples from the Gumbel distribution?
You should be able to something like
```julia
julia> using Distributions
julia> p = 0.8
0.8
julia> lps = log.([1.0 - p, p])
2-element Array{Float64,1}:
-1.6094379124341005
-0.2231435513142097
julia> ixs = map(_ -> argmax(lps .+ rand(Gumbel(), 2)), 1:10^6);
julia> mean(ixs .- 1)
0.799612
```
---
### Author:  [@cscherrer](https://discourse.julialang.org/u/cscherrer)
#### Post date: [July 9, 2019, 8:31pm UTC](https://discourse.julialang.org/t/gumbel-softmax-in-julia/26190/3 "2019-07-09T20:31:14Z")
> [@robsmith11](#):
>
> Isn’t the Gumbel softmax trick just the argmax of the sum of the log-probabilities and samples from the Gumbel distribution?
You need _softmax_ in place of _argmax_. There are some details here:
> **[The Gumbel-Softmax Trick for Inference of Discrete Variables | Columbia...](https://casmls.github.io/general/2017/02/01/GumbelSoftmax.html)**
>
> This week we scrutinized, in a discussion led by Shizhe Chen, two recent papers: “The Concrete Distribution: a Continuous Relaxation of Discrete Random Varia...
There are implementations of softmax in Flux, Knet, or NNLib, or @Joshua_Bowles has a straightforward implementation with a nice description here:
> **[M2.1 Softmax in Julia](https://medium.com/@bowlescompling/m2-1-softmax-in-julia-1498901f741c)**
>
> The goal of this post is not to motivate and explain softmax, nor to provide the best softmax function for Julia, nor to introduce the…
EDIT: Neither of these give direct access to the log-density; it would be really nice to have this directly available in Distributions.jl
---
### Author:  [@Elita\_Lobo](https://discourse.julialang.org/u/Elita_Lobo)
#### Post date: [July 11, 2019, 6:05pm UTC](https://discourse.julialang.org/t/gumbel-softmax-in-julia/26190/4 "2019-07-11T18:05:00Z")
Thanks! I got it working!
---
### Author:  [@cscherrer](https://discourse.julialang.org/u/cscherrer)
#### Post date: [July 12, 2019, 4:13pm UTC](https://discourse.julialang.org/t/gumbel-softmax-in-julia/26190/5 "2019-07-12T16:13:29Z")
Great! Is this in a public repo? I’d love to check it out