replace old shape inference system

examples/MNIST.jl

@@ -4,7 +4,7 @@ data = [(trainfeatures(i), onehot(trainlabel(i), 0:9)) for i = 1:60_000]
 train = data[1:50_000]
 test = data[50_001:60_000]
 
-m = Chain(
+m = @Chain(
   Input(784),
   Affine(128), relu,
   Affine( 64), relu,

examples/integration-mx.jl

@@ -3,17 +3,16 @@ using Flux, MXNet
 Flux.loadmx()
 
 conv1 = Chain(
-  Input(28,28),
   Conv2D((5,5), out = 20), tanh,
   MaxPool((2,2), stride = (2,2)))
 
 conv2 = Chain(
-  conv1,
   Conv2D((5,5), in = 20, out = 50), tanh,
   MaxPool((2,2), stride = (2,2)))
 
-lenet = Chain(
-  conv2,
+lenet = @Chain(
+  Input(28,28,1),
+  conv1, conv2,
   flatten,
   Affine(500), tanh,
   Affine(10), softmax)

examples/integration-tf.jl

@@ -1,17 +1,17 @@
 using Flux, Juno
 
 conv1 = Chain(
-  Reshape(28,28,1),
   Conv2D((5,5), out = 20), tanh,
   MaxPool((2,2), stride = (2,2)))
 
 conv2 = Chain(
-  Input(12,12,20),
   Conv2D((5,5), in = 20, out = 50), tanh,
   MaxPool((2,2), stride = (2,2)))
 
-lenet = Chain(
-  conv1, conv2, flatten,
+lenet = @Chain(
+  Input(28,28,1),
+  conv1, conv2,
+  flatten,
   Affine(500), tanh,
   Affine(10), softmax)
 

src/Flux.jl

@@ -7,6 +7,7 @@ using DataFlow: graphm, syntax, prewalk!, postwalk!, prewalk, postwalk,
   iscyclic, Constant, constant, isconstant, group, Split, splitnode,
   detuple, value, inputs, thread!, value, inputs, Split, splitnode, inputnode,
   spliceinputs, bumpinputs, Line, Frame, applylines
+using MacroTools: @q
 using Juno: Tree, Row
 
 # Zero Flux Given
@@ -27,7 +28,6 @@ include("compiler/shape.jl")
 include("layers/affine.jl")
 include("layers/activation.jl")
 include("layers/recurrent.jl")
-include("layers/shape.jl")
 include("layers/chain.jl")
 include("layers/shims.jl")
 

src/compiler/shape.jl

@@ -2,6 +2,8 @@ using DataFlow.Interpreter
 
 export @shapes
 
+Dims{N} = NTuple{N,Int}
+
 struct Hint
   typ
 end
@@ -41,7 +43,9 @@ shapes(args...) = shapesv(args...) |> syntax |> applylines |> (x->prettify(x, li
 
 infer(f, args...) = graph(f) == nothing ? nothing : gethint(shapesv(f, args...))
 
-function infer(::typeof(*), a::NTuple{2}, b::NTuple{2})
+infer(::typeof(identity), x) = x
+
+function infer(::typeof(*), a::Dims{2}, b::Dims{2})
   a[2] == b[1] || return nothing
   (a[1], b[2])
 end
@@ -55,3 +59,17 @@ macro shapes(ex)
   @capture(ex, f_(args__)) || error("@shapes f(args...)")
   :(shapes($(esc(f)), mapt(size, ($(map(esc, args)...),))...))
 end
+
+# Shim for kicking off shape inference
+
+export Input
+
+struct Input{N} <: Model
+  dims::Dims{N}
+end
+
+Input(i...) = Input((i...,))
+
+(::Input)(x) = x
+
+inferchain(f::Input, xs) = (-1, f.dims...)

src/layers/activation.jl

@@ -8,11 +8,14 @@ relu(x) = max(0, x)
 
 back!(::typeof(relu), Δ, x) = Δ .* (x .< 0)
 
+# TODO: correct behaviour with batches
 softmax(xs) = exp.(xs) ./ sum(exp.(xs))
 
+# TODO: correct behaviour with batches
 flatten(xs) = reshape(xs, length(xs))
 
-shape(::typeof(flatten), in) = prod(in)
-
 infer(::typeof(softmax), x) = x
+infer(::typeof(tanh), x) = x
 infer(::typeof(σ), x) = x
+
+infer(::typeof(flatten), x::Dims) = (x[1], prod(x[2:end])...)

src/layers/affine.jl

@@ -8,3 +8,6 @@ end
 
 Affine(in::Integer, out::Integer; init = initn) =
   Affine(init(in, out), init(1, out))
+
+inferred(::Type{Affine}, in::Tuple{Dims{2}}, out::Integer) =
+  Affine(in[1][2], out)

src/layers/chain.jl

@@ -1,26 +1,11 @@
-export Chain
-
-function inferchain(ms)
-  chain = []
-  sh = nothing
-  for m in ms
-    m = init(m, single(sh))
-    sh = shape(m, sh)
-    push!(chain, m)
-  end
-  return chain, sh
-end
+export Chain, @Chain
 
 type Chain <: Model
   layers::Vector{Any}
-  shape
-  function Chain(ms...)
-    ms, shape = inferchain(ms)
-    return new(ms, shape)
-  end
+  Chain(xs...) = new([xs...])
 end
 
-@forward Chain.layers Base.getindex, Base.first, Base.last, Base.endof
+@forward Chain.layers Base.getindex, Base.first, Base.last, Base.endof, Base.push!
 @forward Chain.layers Base.start, Base.next, Base.done
 
 (s::Chain)(x) = foldl((x, m) -> m(x), x, s.layers)
@@ -30,6 +15,25 @@ update!(s::Chain, η) = foreach(l -> update!(l, η), s.layers)
 graph(s::Chain) =
   foldl((v, m) -> vertex(m, v), constant(inputnode(1)), s.layers)
 
-shape(c::Chain, in) = c.shape
-
 Base.getindex(c::Chain, i::AbstractArray) = Chain(c.layers[i]...)
+
+# Chain Macros
+
+inferred(f, in, args...; kws...) = f(args...; kws...)
+
+# `inferchain` allows for overriding inference behaviour for convenience.
+# For example, `infer(Affine(10, 20), nothing)` would normally return a shape
+# error, but for the interface we just ignore any errors and return (1, 20).
+inferchain(f, xs...) = infer(f, xs...)
+
+macro Chain(x, xs...)
+  inferconstructor(x) =
+    @capture(x, f_(xs__)) ? :(inferred($(esc(f)), (shape,), $(esc.(xs)...))) : esc(x)
+  @q let
+    shape = nothing
+    c = Chain($(esc(x)))
+    $([:(shape = inferchain(c.layers[end], shape);
+         push!(c, $x)) for x in inferconstructor.(xs)]...)
+    c
+  end
+end

src/layers/shape.jl

@@ -1,47 +0,0 @@
-export Input
-
-Dims{N} = NTuple{N,Int}
-
-dims(d::Dims) = d
-
-dims(i...) = (i...,)
-
-single(i) = i
-single(i::Dims) = length(i) == 1 ? first(i) : i
-
-# Shim for kicking off shape inference
-
-struct ShapeError <: Exception
-  layer
-  shape
-end
-
-struct Input{N} <: Model
-  dims::Dims{N}
-end
-
-Input(i...) = Input(dims(i...))
-
-(::Input)(x) = x
-back!(::Input, Δ, x) = Δ
-
-# Initialise placeholder
-
-struct Init{F}
-  f::F
-end
-
-init(i::Init, input...) = i.f(input...)
-init(m, input...) = m
-
-# Shape inference API
-
-shape(x, in) = in
-
-shape(i::Input, _) = i.dims
-
-# Implementation for bundled layers
-
-shape(d::Affine, _) = length(state(d.b)) # TODO: could perhaps infer this
-
-Affine(out::Integer) = Init(in::Integer -> Affine(in, out))

src/layers/shims.jl

@@ -8,11 +8,10 @@ end
 Conv2D(size; in = 1, out = 1, stride = (1,1), init = initn) =
   Conv2D(param(initn(size..., in, out)), stride)
 
-shape(c::Conv2D, in::Dims{2}) =
-  (map(i -> (in[i]-size(c.filter,i))÷c.stride[i]+1, (1,2))..., size(c.filter, 4))
+infer(c::Conv2D, in::Dims{4}) =
+  (in[1], map(i -> (in[i+1]-size(c.filter,i))÷c.stride[i]+1, (1,2))..., size(c.filter, 4))
 
-shape(c::Conv2D, in::Dims{3}) =
-  shape(c, (in[1],in[2]))
+# TODO: many of these should just be functions
 
 for Pool in :[MaxPool, AvgPool].args
   @eval begin
@@ -24,11 +23,8 @@ for Pool in :[MaxPool, AvgPool].args
     $Pool(size; stride = (1,1)) =
       $Pool(size, stride)
 
-    shape(c::$Pool, in::Dims{2}) =
-      map(i -> (in[i]-c.size[i])÷c.stride[i]+1, (1,2))
-
-    shape(c::$Pool, in::Dims{3}) =
-      (shape(c, (in[1],in[2]))..., in[3])
+    infer(c::$Pool, in::Dims{4}) =
+      (in[1], map(i -> (in[i+1]-c.size[i])÷c.stride[i]+1, (1,2))..., in[4])
 
     shape(c::$Pool) = nothing
   end
@@ -40,9 +36,4 @@ end
 
 Reshape(dims::Integer...) = Reshape(dims)
 
-function shape(r::Reshape, dims)
-    prod(dims) == prod(r.dims) || throw(ShapeError(r, dims))
-    return r.dims
-end
-
 shape(r::Reshape, ::Void) = r.dims