sampling + tweaks

examples/char-rnn.jl

@@ -11,11 +11,25 @@ Xs, Ys = getbatches(input, alphabet), getbatches(input[2:end], alphabet)
 
 model = Chain(
   Input(N),
-  Recurrent(N, 128, N),
+  Recurrent(N, 128),
+  Dense(128, N),
   softmax)
 
 m = tf(unroll(model, 50))
 
+# Flux.train!(m, take(Xs,100), take(Ys,100),
+#             η = 0.1, epoch = 1)
 Flux.train!(m, Xs, Ys, η = 0.2e-3, epoch = 1)
 
-map(c->onecold(c, alphabet), m(first(first(first(train)))))
+string(map(c -> onecold(c, alphabet), m(first(first(Xs))))...)
+
+function sample(model, n)
+  s = [rand(alphabet)]
+  m = tf(unroll(model, 1))
+  for i = 1:n
+    push!(s, onecold(m(Seq((onehot(Float32, 'b', alphabet),)))[1], alphabet))
+  end
+  return string(s...)
+end
+
+sample(model, 100) |> println

src/utils.jl

@@ -2,7 +2,7 @@ export AArray
 
 const AArray = AbstractArray
 
-initn(dims...) = randn(Float32, dims...)/1000
+initn(dims...) = randn(Float32, dims...)/10
 
 function train!(m, train, test = []; epoch = 1, batch = 10, η = 0.1)
     i = 0