Float32 by default

examples/MNIST.jl

@@ -1,6 +1,6 @@
 using Flux, MNIST
 
-data = [(trainfeatures(i), Vector{Float64}(onehot(trainlabel(i), 0:9))) for i = 1:60_000]
+data = [(trainfeatures(i), Vector{Float32}(onehot(trainlabel(i), 0:9))) for i = 1:60_000]
 train = data[1:50_000]
 test = data[50_001:60_000]
 

examples/mnist-conv.jl

@@ -40,8 +40,8 @@ using TensorFlow
 
 sess = Session(Graph())
 
-x  = placeholder(Float64)
-y′ = placeholder(Float64)
+x  = placeholder(Float32)
+y′ = placeholder(Float32)
 
 y = Tensor(lenet, x)
 

src/backend/tensorflow/tensorflow.jl

@@ -65,7 +65,7 @@ Media.render(::Juno.Clipboard, ::Model) = "Flux.TF.Model()"
 
 function tf(model)
   sess = Session(Graph())
-  input = placeholder(Float64)
+  input = placeholder(Float32)
   g = graph(model, input)
   run(sess, initialize_all_variables())
   Model(sess, [input], g, gradients(g, input))
@@ -90,7 +90,7 @@ function Flux.train!(m::Model, train, test=[]; epoch = 1, η = 0.1,
                      loss = (y, y′) -> reduce_sum((y - y′).^2)/2,
                      opt = TensorFlow.train.GradientDescentOptimizer(η))
   i = 0
-  Y = placeholder(Float64)
+  Y = placeholder(Float32)
   Loss = loss(m.graph, Y)
   minimize_op = TensorFlow.train.minimize(opt, Loss)
   for e in 1:epoch

src/layers/shims.jl

@@ -1,7 +1,7 @@
 export Conv2D, MaxPool
 
 type Conv2D <: Model
-  filter::Param{Array{Float64,4}} # [height, width, outchans, inchans]
+  filter::Param{Array{Float32,4}} # [height, width, outchans, inchans]
   stride::Dims{2}
 end
 

src/utils.jl

@@ -5,7 +5,7 @@ const AArray = AbstractArray
 onehot(label, labels) = [i == label for i in labels]
 onecold(pred, labels = 1:length(pred)) = labels[findfirst(pred, maximum(pred))]
 
-initn(dims...) = randn(dims...)/1000
+initn(dims...) = randn(Float32, dims...)/1000
 
 function train!(m, train, test = []; epoch = 1, batch = 10, η = 0.1)
     i = 0