Merge #1367

1367: RNN update to drop CUDNN, fix LSTM bug and output type stability r=CarloLucibello a=jeremiedb

PR related to #1114 #1360 #1365 

Some experiment for RNN handling. 

Hidden state of each cell structure was dropped as they weren't needed (AFAIK, only needed for size inference for CUDNN, but bias size could be used as a substitute to cells' `h` there as well). 

Looked to drop dependence on CUDNN entirely, so it's a pure Flux/CUDA.jl. File `src/cuda/curnnjl` no longer used. No  modifications were made to the cell computations. Initial test seems to show decent performance, but yet to benchmark. 

Pending issue: despite having dropped completely the CUDNN dependency, there's still an instability issue that seems present when running on GPU. This is illustrated in the test at lines 1-50 of file `test\rnn-test-jdb.jl`. If that test runs on CPU, it goes well thorugh the 100 iterations. However, the same on GPU will thow NAs after couple dozens of iterations. 
My only hypothesis so far: when performing the iteration over the sequence through `m.(x)` or `map(rnn, x)`, is the order of the execution safe? Ie: is it possible that there isn't a `sync()` on the CUDA side between those seq steps, which may mess up the state?

### PR Checklist

- [x] Tests are added
- [ ] Entry in NEWS.md
- [ ] Documentation, if applicable
- [ ] Final review from `@dhairyagandhi96` (for API changes).


Co-authored-by: jeremiedb <[email protected]>
Co-authored-by: jeremie.db <[email protected]>

Author: 3 people

.gitignore

@@ -5,4 +5,5 @@
 docs/build/
 docs/site/
 deps
-# Manifest.toml
+.vscode
+# Manifest.toml

src/cuda/cuda.jl

@@ -1,9 +1,12 @@
 module CUDAint
 
 using ..CUDA
-
 using CUDA: CUDNN
-include("curnn.jl")
+
+import ..Flux: Flux
+import Zygote
+using Zygote: @adjoint
+
 include("cudnn.jl")
 
 end

src/cuda/curnn.jl

@@ -1,3 +1,4 @@
+
 gate(h, n) = (1:h) .+ h*(n-1)
 gate(x::AbstractVector, h, n) = @view x[gate(h,n)]
 gate(x::AbstractMatrix, h, n) = x[gate(h,n),:]
@@ -24,21 +25,19 @@ rnn.(1:10)  # apply to a sequence
 rnn.state   # 60
 ```
 """
-mutable struct Recur{T}
+mutable struct Recur{T,S}
   cell::T
-  init
-  state
+  state::S
 end
 
-Recur(m, h = hidden(m)) = Recur(m, h, h)
-
 function (m::Recur)(xs...)
   h, y = m.cell(m.state, xs...)
   m.state = h
   return y
 end
 
-@functor Recur cell, init
+@functor Recur
+trainable(a::Recur) = (a.cell,)
 
 Base.show(io::IO, m::Recur) = print(io, "Recur(", m.cell, ")")
 
@@ -52,34 +51,30 @@ Assuming you have a `Recur` layer `rnn`, this is roughly equivalent to:
 rnn.state = hidden(rnn.cell)
 ```
 """
-reset!(m::Recur) = (m.state = m.init)
+reset!(m::Recur) = (m.state = m.cell.state)
 reset!(m) = foreach(reset!, functor(m)[1])
 
 flip(f, xs) = reverse(f.(reverse(xs)))
 
 # Vanilla RNN
 
-mutable struct RNNCell{F,A,V}
+struct RNNCell{F,A,V,S}
   σ::F
   Wi::A
   Wh::A
   b::V
-  h::V
+  state::S
 end
 
-RNNCell(in::Integer, out::Integer, σ = tanh;
-        init = glorot_uniform) =
-  RNNCell(σ, init(out, in), init(out, out),
-          init(out), zeros(out))
+RNNCell(in::Integer, out::Integer, σ=tanh; init=Flux.glorot_uniform, initb=zeros, init_state=zeros) = 
+  RNNCell(σ, init(out, in), init(out, out), initb(out), init_state(out,1))
 
 function (m::RNNCell)(h, x)
   σ, Wi, Wh, b = m.σ, m.Wi, m.Wh, m.b
   h = σ.(Wi*x .+ Wh*h .+ b)
   return h, h
 end
 
-hidden(m::RNNCell) = m.h
-
 @functor RNNCell
 
 function Base.show(io::IO, l::RNNCell)
@@ -94,22 +89,23 @@ end
 The most basic recurrent layer; essentially acts as a `Dense` layer, but with the
 output fed back into the input each time step.
 """
+Recur(m::RNNCell) = Recur(m, m.state)
 RNN(a...; ka...) = Recur(RNNCell(a...; ka...))
 
 # LSTM
 
-mutable struct LSTMCell{A,V}
+struct LSTMCell{A,V,S}
   Wi::A
   Wh::A
   b::V
-  h::V
-  c::V
+  state::S
 end
 
 function LSTMCell(in::Integer, out::Integer;
-                  init = glorot_uniform)
-  cell = LSTMCell(init(out * 4, in), init(out * 4, out), init(out * 4),
-                  zeros(out), zeros(out))
+                  init = glorot_uniform,
+                  initb = zeros,
+                  init_state = zeros)
+  cell = LSTMCell(init(out * 4, in), init(out * 4, out), initb(out * 4), (init_state(out,1), init_state(out,1)))
   cell.b[gate(out, 2)] .= 1
   return cell
 end
@@ -126,8 +122,6 @@ function (m::LSTMCell)((h, c), x)
   return (h′, c), h′
 end
 
-hidden(m::LSTMCell) = (m.h, m.c)
-
 @functor LSTMCell
 
 Base.show(io::IO, l::LSTMCell) =
@@ -142,20 +136,22 @@ recurrent layer. Behaves like an RNN but generally exhibits a longer memory span
 See [this article](https://colah.github.io/posts/2015-08-Understanding-LSTMs/)
 for a good overview of the internals.
 """
+# Recur(m::LSTMCell) = Recur(m, (zeros(length(m.b)÷4), zeros(length(m.b)÷4)),
+#   (zeros(length(m.b)÷4), zeros(length(m.b)÷4)))
+Recur(m::LSTMCell) = Recur(m, m.state)
 LSTM(a...; ka...) = Recur(LSTMCell(a...; ka...))
 
 # GRU
 
-mutable struct GRUCell{A,V}
+struct GRUCell{A,V,S}
   Wi::A
   Wh::A
   b::V
-  h::V
+  state::S
 end
 
-GRUCell(in, out; init = glorot_uniform) =
-  GRUCell(init(out * 3, in), init(out * 3, out),
-          init(out * 3), zeros(out))
+GRUCell(in, out; init = glorot_uniform, initb = zeros, init_state = zeros) =
+  GRUCell(init(out * 3, in), init(out * 3, out), initb(out * 3), init_state(out,1))
 
 function (m::GRUCell)(h, x)
   b, o = m.b, size(h, 1)
@@ -167,8 +163,6 @@ function (m::GRUCell)(h, x)
   return h′, h′
 end
 
-hidden(m::GRUCell) = m.h
-
 @functor GRUCell
 
 Base.show(io::IO, l::GRUCell) =
@@ -183,6 +177,7 @@ RNN but generally exhibits a longer memory span over sequences.
 See [this article](https://colah.github.io/posts/2015-08-Understanding-LSTMs/)
 for a good overview of the internals.
 """
+Recur(m::GRUCell) = Recur(m, m.state)
 GRU(a...; ka...) = Recur(GRUCell(a...; ka...))
 
 @adjoint function Broadcast.broadcasted(f::Recur, args...)

src/layers/recurrent.jl

@@ -8,7 +8,7 @@ using Flux: pullback
   Flux.reset!(m)
   θ = gradient(() -> sum(m(x)), params(m))
   @test x isa CuArray
-  @test_broken θ[m.cell.Wi] isa CuArray
+  @test θ[m.cell.Wi] isa CuArray
   @test_broken collect(m̄[].cell[].Wi) == collect(θ[m.cell.Wi])
 end
 
@@ -20,17 +20,15 @@ end
     Flux.reset!(rnn)
     Flux.reset!(curnn)
     x = batch_size == 1 ?
-      rand(10) :
-      rand(10, batch_size)
+      rand(Float32, 10) :
+      rand(Float32, 10, batch_size)
     cux = gpu(x)
 
     y, back = pullback((r, x) -> r(x), rnn, x)
     cuy, cuback = pullback((r, x) -> r(x), curnn, cux)
 
     @test y ≈ collect(cuy)
 
-    @test haskey(Flux.CUDAint.descs, curnn.cell)
-
     ȳ = randn(size(y))
     m̄, x̄ = back(ȳ)
     cum̄, cux̄ = cuback(gpu(ȳ))

test/cuda/curnn.jl

@@ -1,6 +1,6 @@
 # Ref FluxML/Flux.jl#1209
 @testset "BPTT" begin
-  seq = [rand(2) for i = 1:3]
+  seq = [rand(Float32, (2,1)) for i = 1:3]
   for r ∈ [RNN,]
     rnn = r(2,3)
     Flux.reset!(rnn)
@@ -11,7 +11,7 @@
     bptt = gradient(Wh->sum(tanh.(rnn.cell.Wi * seq[3] + Wh *
                                   tanh.(rnn.cell.Wi * seq[2] + Wh *
                                         tanh.(rnn.cell.Wi * seq[1] +
-                                            Wh * rnn.init
+                                            Wh * rnn.cell.state
                                         + rnn.cell.b)
                                   + rnn.cell.b)
                             + rnn.cell.b)),

test/layers/recurrent.jl

@@ -101,16 +101,16 @@ end
   m = Dense(10, 5)
   @test size.(params(m)) == [(5, 10), (5,)]
   m = RNN(10, 5)
-  @test size.(params(m)) == [(5, 10), (5, 5), (5,), (5,)]
+  @test size.(params(m)) == [(5, 10), (5, 5), (5,), (5, 1)]
 
   # Layer duplicated in same chain, params just once pls.
   c = Chain(m, m)
-  @test size.(params(c)) == [(5, 10), (5, 5), (5,), (5,)]
+  @test size.(params(c)) == [(5, 10), (5, 5), (5,), (5, 1)]
 
   # Self-referential array. Just want params, no stack overflow pls.
   r = Any[nothing,m]
   r[1] = r
-  @test size.(params(r)) == [(5, 10), (5, 5), (5,), (5,)]
+  @test size.(params(r)) == [(5, 10), (5, 5), (5,), (5, 1)]
 end
 
 @testset "Basic Stacking" begin