WIP: Simple Regression with ADAM

examples/mnist_adam.jl

@@ -0,0 +1,28 @@
+using Flux, Functors, Optimisers, XLA
+using Flux.Data.MNIST
+using Flux: onehotbatch, crossentropy
+
+i = rand(784, 2)
+t = rand(10, 2)
+
+opt = Optimisers.ADAM(0.001, (0.9,0.99))
+
+function step(m)
+  m̄, = gradient(m) do m
+    @show Flux.mse(m(i), t)
+  end
+  return opt2(m, m̄)
+end
+
+m = Chain(Dense(28^2, 32, relu), Dense(32, 10), softmax) |> f64
+
+xstep = xla(step)
+
+function train(m)
+  for _ = 1:100
+    m = xstep(m)
+  end
+  m
+end
+
+train(m)