Using users provided weight matrix to build LSTM layers

[Chengfeng-Jia]

Describe the potential feature

In this demo for Dense layer, user can use provided weight matrix to build Dense layer. d1 = Dense(ones(2, 5), false, tanh). When I want to build a RNN/LSTM layer, it failed Lstm = Flux.LSTMCell(W2_1,W2_2,b2,s2_1) Lstm(data[1]).

The above W2_1,W2_2,b2,s2_1 are pre-defined random weight matrix. The error is MethodError: objects of type Flux.LSTMCell{Matrix{Float64}, Vector{Float64}, Vector{Float64}} are not callable

Motivation

No response

Possible Implementation

No response

[ToucheSir]

The state you initialize the LSTMCell with needs to be a 2-tuple of AbstractMatrix (e.g. (ones(1, 1), zeros(1, 1)) and not a single array as you described. Otherwise

Flux.jl/src/layers/recurrent.jl

Line 308 in b26113d

function (m::LSTMCell{I,H,V,<:NTuple{2,AbstractMatrix{T}}})((h, c), x::Union{AbstractVecOrMat{T},OneHotArray}) where {I,H,V,T}

won't work. You can see how the main constructor does this initialization in

Flux.jl/src/layers/recurrent.jl

Lines 302 to 303 in b26113d

	init_state = zeros32)
	cell = LSTMCell(init(out * 4, in), init(out * 4, out), initb(out * 4), (init_state(out,1), init_state(out,1)))

.

[Chengfeng-Jia]

Thank you so much! It works.