Our network will be as simple as having input data followed by FullyConnected layers and SoftmaxOutput. There's no need to add any complexity as the features are already precalculated:
arch = @mx.chain mx.Variable(:data) => mx.FullyConnected(num_hidden=length(CLASSES)) => mx.SoftmaxOutput(mx.Variable(:label))nnet_new = mx.FeedForward(arch, context = mx.gpu())optimizer = mx.ADAM()