Feature request: add embedding layer about flux.jl

[MikeInnes]

This is trivial with normal slicing:

julia> W = randn(3, 10)
3×10 Array{Float64,2}:
  1.52921     0.34536   -1.54119   …  -0.303022    0.123047    0.00684176
 -0.138502    0.4285    -0.841863      0.0720752  -2.19672    -0.0968025 
 -0.00694323  0.916636   0.834962     -0.239865    0.0237941   0.0236618 

julia> W[:, 5]
3-element Array{Float64,1}:
 -0.972017
  0.293653
 -1.20708 

julia> W[:, [5, 6]]
3×2 Array{Float64,2}:
 -0.972017   2.58503
  0.293653  -2.29474
 -1.20708    1.82632

julia> W[:, [5 6; 7 8]]
3×2×2 Array{Float64,3}:
[:, :, 1] =
 -0.972017  -0.914234
  0.293653   0.333886
 -1.20708   -0.386469

[:, :, 2] =
  2.58503  -0.303022 
 -2.29474   0.0720752
  1.82632  -0.239865

There may be some benefit to having a convenience wrapper though. Should be easy to put together if you want to set up a PR.

[oxinabox]

Better still is to use the OneHot magic, I think.
This is like expressly what it is for?
Since onehot encoded value takes up no more space than an Int, and onehot multiplication is slicing under the hood.
http://fluxml.github.io/Flux.jl/latest/data/onehot.html

[ngphuoc]

Thanks. I've tried onehot and got the follow error:

julia> x = onehot(1, 1:10)
10-element Flux.OneHotVector:
  true
 false
 false
 false
 false
 false
 false
 false
 false
 false

julia> m = Chain(Dense(10, 5), Dense(5, 2))
Chain(Dense(10, 5), Dense(5, 2))

julia> m(x)
ERROR: MethodError: *(::TrackedArray{…,Array{Float64,2}}, ::Flux.OneHotVector) is ambiguous. Candidates:
  *(A::AbstractArray{T,2} where T, b::Flux.OneHotVector) in Flux at /home/phuoc/.julia/v0.6/Flux/src/onehot.jl:10
  *(a::Flux.Tracker.TrackedArray{T,2,A} where A where T, b::AbstractArray{T,1} where T) in Flux.Tracker at /home/phuoc/.julia/v0.6/Flux/src/tracker/lib.jl:67
Possible fix, define
  *(::Flux.Tracker.TrackedArray{T,2,A} where A where T, ::Flux.OneHotVector)
Stacktrace:
 [1] (::Flux.Dense{Base.#identity,TrackedArray{…,Array{Float64,2}},TrackedArray{…,Array{Float64,1}}})(::Flux.OneHotVector) at /home/phuoc/.julia/v0.6/Flux/src/layers/basic.jl:61
 [2] mapfoldl_impl(::Base.#identity, ::Flux.##45#46, ::Flux.OneHotVector, ::Array{Any,1}, ::Int64) at ./reduce.jl:43
 [3] (::Flux.Chain)(::Flux.OneHotVector) at /home/phuoc/.julia/v0.6/Flux/src/layers/basic.jl:28

[ngphuoc]

I also tried to add the following model but failed to train since there was no parameters returned by params. Did I miss something?:

julia> using Flux
julia> using Flux: onehotbatch, unstack, truncate!, throttle, logloss, initn
julia> using Flux.Tracker: param, back!, data, grad
julia> struct LangModel{E,R,F}
         emb::E
         rnn::R
         fc::F
       end

julia> LangModel(v::Int, e::Int, h::Int) = LangModel(param(initn(e, v)),
                                      LSTM(e, h),
                                      Dense(h, v))
LangModel

julia> Flux.children(m::LangModel) = (m.emb, m.rnn, m.fc,)

julia> (m::LangModel)(x) = softmax(m.fc(m.rnn(m.emb[:,x])))
julia> m = LangModel(2,3,4)
LangModel{TrackedArray{…,Array{Float64,2}},Flux.Recur{Flux.LSTMCell{Flux.Dense{NNlib.#σ,TrackedArray{…,Array{Float64,2}},TrackedArray{…,Array{Float64,1}}},Flux.Dense{Base.#tanh,TrackedArray{…,Array{Float64,2}},TrackedArray{…,Array{Float64,1}}},TrackedArray{…,Array{Float64,1}}}},Flux.Dense{Base.#identity,TrackedArray{…,Array{Float64,2}},TrackedArray{…,Array{Float64,1}}}}(param([0.00883202 -0.00318895; 0.0140951 0.0126121; -0.0082283 -0.00736651]), Recur(LSTMCell(3, 4)), Dense(4, 2))

julia> params(m)
0-element Array{Any,1}

[MikeInnes]

The ambiguity issue is one I can fix.

If you're on the latest Flux you need to change the Flux.children line to Flux.treelike(LangModel).

[ngphuoc]

Thank you. Flux.treelike(LangModel) works perfectly.

[MikeInnes]

Closing this for now as I think matmul is fine for embeddings, unless anyone has a specific proposal. I've noted the ambiguity issue though so I'll fix that ASAP.

[datnamer]

How about a layer that would provide a wrapper to import weights for pretrained embeddings? https://discuss.pytorch.org/t/can-we-use-pre-trained-word-embeddings-for-weight-initialization-in-nn-embedding/1222

And then a way to freeze them during training.

[oxinabox]

@datnamer it is gloriously trivial

Load the pretrained weights into a matrix W from Embeddings.jl.
(Code to do the import lives there)

Then use the onehot product discussed above.
e.g for word with one hot vector ei.

If you want to allow the embedding to fine tune use Param(W)*ei,
if you want to freeze it, use W*ei.

[datnamer]

@oxinabox thanks that looks great. Would it be similiairly easy with graph embedding? I'm just starting to play around but there aren't a lot of tutorials for that.

[oxinabox]

I'm not sure. I don't know that pretrained graph embeddings are a thing.