reshape layer

src/backend/tensorflow/tensorflow.jl

@@ -36,6 +36,7 @@ graph(::typeof(tanh), x) = tanh(x)
 # reshape hack due to https://github.com/malmaud/TensorFlow.jl/issues/79
 batchsize(x::Tensor) = reduce_sum(slice(TensorFlow.shape(x), [0], [1]))
 graph(::typeof(flatten), x) = reshape(x, pack([batchsize(x), Int32(-1)]))
+graph(r::Reshape, x) = reshape(x, pack([batchsize(x), map(Int32, r.dims)...]))
 
 graph(::Input, x) = x
 

src/layers/shims.jl

@@ -1,4 +1,4 @@
-export Conv2D, MaxPool
+export Conv2D, MaxPool, Reshape
 
 type Conv2D <: Model
   filter::Param{Array{Float32,4}} # [height, width, inchans, outchans]
@@ -29,3 +29,16 @@ shape(c::MaxPool, in::Dims{3}) =
   (shape(c, (in[1],in[2]))..., in[3])
 
 shape(c::MaxPool, in) = throw(ShapeError(c, in))
+
+immutable Reshape{N}
+  dims::Dims{N}
+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