Added the wrapper Bidirectional for RNN layers

src/layers/recurrent.jl

@@ -430,3 +430,61 @@ Recur(m::GRUv3Cell) = Recur(m, m.state0)
 @adjoint function Broadcast.broadcasted(f::Recur, args...)
   Zygote.∇map(__context__, f, args...)
 end
+
+
+"""
+    Bidirectional{A,B} 
+
+A wrapper layer that allows the use of [bidirectional](https://ieeexplore.ieee.org/document/650093) layers. It contains two parts that are Flux layers: `forward` and `backward` where 
+the forward layer weights are concatenated with the reversed order of the backward layer weights.
+
+It is intended to be used with recurrent layers such as `LSTM`, `GRU` or `RNN` to benefit from the sequential information that recurrent 
+layers have, but it will not raise an error if used with a different layer such as `Dense`, as long as the layer is compatible with the concatenation function `vcat`.
+
+# Examples
+```jldoctest
+julia> BLSTM = Bidirectional(LSTM, 3, 5)
+Bidirectional(
+  Recur(
+    LSTMCell(3, 5),                     # 190 parameters
+  ),
+  Recur(
+    LSTMCell(3, 5),                     # 190 parameters
+  ),
+)         # Total: 10 trainable arrays, 380 parameters,
+          # plus 4 non-trainable, 20 parameters, summarysize 2.141 KiB.
+julia> Bidirectional(LSTM, 3, 5)(rand(Float32, 3)) |> size
+(10,)
+
+julia> model = Chain(Embedding(10000, 200), Bidirectional(LSTM, 200, 128), Dense(256, 5), softmax)
+Chain(
+  Embedding(10000, 200),                # 2_000_000 parameters
+  Bidirectional(
+    Recur(
+      LSTMCell(200, 128),               # 168_704 parameters
+    ),
+    Recur(
+      LSTMCell(200, 128),               # 168_704 parameters
+    ),
+  ),
+  Dense(256, 5),                        # 1_285 parameters
+  NNlib.softmax,
+)         # Total: 13 trainable arrays, 2_338_693 parameters,
+          # plus 4 non-trainable, 512 parameters, summarysize 8.922 MiB.
+```
+"""
+struct Bidirectional{A,B} 
+  forward::A
+  backward::B
+end
+
+# Constructor that creates a bidirectional with the same layer for forward and backward
+Bidirectional(rnn, a...; ka...) = Bidirectional(rnn(a...; ka...), rnn(a...; ka...))
+
+
+# Concatenate the forward and reversed backward weights
+function (m::Bidirectional)(x::Union{AbstractVecOrMat{T},OneHotArray}) where {T}
+  return vcat(m.forward(x), reverse(m.backward(reverse(x; dims=1)); dims=1))
+end
+
+@functor Bidirectional

src/layers/show.jl

@@ -1,6 +1,6 @@
 
 for T in [
-    :Chain, :Parallel, :SkipConnection, :Recur  # container types
+    :Chain, :Parallel, :SkipConnection, :Recur, :Bidirectional  # container types
   ]
   @eval function Base.show(io::IO, m::MIME"text/plain", x::$T)
     if get(io, :typeinfo, nothing) === nothing  # e.g. top level in REPL