feature requests for RNNs

[CarloLucibello]

After the redesign in #2500, here is a list of potential improvements for recurrent layers and recurrent cells

• add an option in constructors to have trainable initial state let's keep it simple (and also follow pytorch) by not having this, it can be part of the outer model
• use cuDNN
• implement the num_layers argument for stacked RNNs (issue Stacked RNN in Flux.jl?  #2452) (see feature requests for RNNs #2514 (comment))
• add dropout (issue recurrent dropout  #1040) (see feature requests for RNNs #2514 (comment))
• add bidirectional option (old PR Added the wrapper Bidirectional for RNN layers #1790)
• add initialstates function. It could be useful in the LSTM case where the initial state is more complicated (two vectors). (done in Adding initialstates function to RNNs #2541)
• distinguish init for kernel and recurrent (Distinct init for kernel and recurrent #2522)
• define a generic wrapper around cells to take entire sequences as input. It could be named like the old Recur (but maybe this is confusing) or Recurrence as in Lux. (Recurrence layer #2549)
• add an option in the RNN/GRU/LSTM constructor to return the last state, along the output

[MartinuzziFrancesco]

Could it be possible to add to the list the option to use different initializers for the input matrix and recurrent matrix? This is provided by both Keras/TF and Flax.

This should be as straightforward as

function RNNCell((in, out)::Pair, σ=relu;
    kernel_init = glorot_uniform,
    recurrent_kernel_init = glorot_uniform,
    bias = true)
    Wi = kernel_init(out, in)
    U = recurrent_kernel_init(out, out)
    b = create_bias(Wi, bias, size(Wi, 1))
    return RNNCell(σ, Wi, U, b)
end

I can also open a quick PR on this if needed

[CarloLucibello]

yes! PR welcome

[MartinuzziFrancesco]

Following up on this, should we also have an option to choose the init for the bias?

[CarloLucibello]

We don't do it for feedforward layers, if someone wants a non-zero bias can just change it manually in the constructor, layer.bias .= ...

[MartinuzziFrancesco]

Ehi, a couple of questions on this features request again. Would the initialstates function be like the current function (rnn::RNN)(x::AbstractVecOrMat) but returning just the state?

would something simple like

function initialstates(rnn::RNNCell; init_state = zeros)
    state = init_state(size(rnn.Wh, 2))
    return state
end

function initialstates(lstm::LSTMCell; init_state = zeros, init_cstate = zeros)
    state = init_state(size(lstm.Wh, 2))
    cstate = init_cstate(size(lstm.Wh, 2))
    return state, cstate
end

suffice or were you looking for something more? Maybe more control on the type would be needed

[CarloLucibello]

I would just have

initialstates(c::RNNCell) = zeros_like(c.Wh, size(c.Wh, 2)))
initialstates(c:: LSTMCell) = zeros_like(c.Wh, size(c.Wh, 2)), zeros_like(c.Wh, size(c.Wh, 2))

If different initializations are needed, we could add an init_state to the constructor, but maybe better let the user handle it as a part of the model, for simplicity and flexibility (e.g. making the initial state trainable). I don't have a strong opinion though.

[MartinuzziFrancesco]

so this way we would simply do

function (rnn::RNNCell)(inp::AbstractVecOrMat)
    state = initialstates(rnn)
    return rnn(inp, state)
end

to keep compatibility for the current version, right?

I think your point is good, additionally no other library provides a specific init for the state so that's probably a little overkill. I'll push something along these lines later

[MartinuzziFrancesco]

Trying to tackle adding num_layers and dropout, with Bidirectional as well once I figured out the general direction. I just wanted to ask if the current approach is in line with what you had in mind:

struct TestRNN{A, B}
    cells::A
    dropout_layer::B
end

Flux.@layer :expand TestRNN

function TestRNN((in_size, out_size)::Pair;
    n_layers::Int=1,
    dropout::Float64=0.0,
    kwargs...)
    cells = []

    for i in 1:n_layers
        tin_size = i == 1 ? in_size : out_size
        push!(cells, RNNCell(tin_size => out_size; kwargs...))
    end

    if dropout > 0.0
        dropout_layer = Dropout(dropout)
    else
        dropout_layer = nothing
    end

    return TestRNN(cells, dropout_layer)
end

function (rnn::TestRNN)(inp, state)
    @assert ndims(inp) == 2 || ndims(inp) == 3
    output = []
    num_layers = length(rnn.cells)

    for inp_t in eachslice(inp, dims=2)
        new_states = []
        for (idx_cell, cell) in enumerate(rnn.cells)
            new_state = cell(inp_t, state[:, idx_cell])
            new_states = vcat(new_states, [new_state])
            inp_t = new_state

            if rnn.dropout_layer isa Dropout && idx_cell < num_layers - 1
                inp_t = rnn.dropout_layer(inp_t)
            end
        end
        state = stack(new_states)
        output = vcat(output, [inp_t])
    end
    output = stack(output, dims=2)
    return output, state
end

[CarloLucibello]

I think we don't need this additional complexity. A simple stacked rnn can be constructed as a chain:

stacked_rnn = Chain(LSTM(3 => 3), Dropout(0.5), LSTM(3 => 3))

If control of the initial states is also needed, it is not hard to define a custom struct the job:

struct StackedRNN{L,S}
    layers::L
    states0::S
end

function StackedRNN(d, num_layers)
    layers = [LSTM(d => d) for _ in num_layers]
    states0 = [Flux.initialstates(l) for l in layers]
    return StackedRNN(layers, states0)
end

function (m::StackedRNN)(x)
   for (layer, state0) in zip(rnn.layers, rnn.states0)
       x = layer(x, state0) 
   end
   return x
end

I think it is enough to document this in the guide
https://github.com/FluxML/Flux.jl/blob/master/docs/src/guide/models/recurrence.md