Don't ignore MNIST training test in MNISTSpec

[kostaleonard]

As a machine learning engineer, I want the library to work on MNIST so that I know it can at least get good results on the Hello World of machine learning.

The MNIST training test in MNISTSpec (which also appears in the new main README) is too slow and for that reason we have ignored it. That test needs to work for the first release. Here is the test in question, for reference.

  // We ignore this test because it is too slow to converge, but we keep it as
  // our goal for the initial release.
  ignore should "be easy to train a model on MNIST" in {
    val xTrain = dataset._1.reshape(Array(60000, 28 * 28)).toFloat / 255
    val yTrain = dataset._2.toCategorical().toFloat
    assert(xTrain.flatten().forall(pixel => pixel >= 0 && pixel <= 1))
    assert(yTrain.flatten().forall(label => label == 0 || label == 1))
    val input = Input[Float]("X", Array(None, Some(28 * 28)))
    val inputLayer = InputLayer(input)
    val dense1 = Dense.withRandomWeights(inputLayer, 128)
    val activation1 = Sigmoid(dense1)
    val dense2 = Dense.withRandomWeights(activation1, 10)
    val activation2 = Sigmoid(dense2)
    val model = Model(activation2)
    val inputs = Map(input -> xTrain)
    val lossFunctionBefore = Mean(
      Square(Subtract(model.outputLayer.getComputationGraph, Constant(yTrain)))
    )
    val lossBefore = lossFunctionBefore.compute(inputs).flatten().head
    val fittedModel = model.fit(inputs, yTrain, 10)
    val lossFunctionAfter = Mean(
      Square(
        Subtract(fittedModel.outputLayer.getComputationGraph, Constant(yTrain))
      )
    )
    val lossAfter = lossFunctionAfter.compute(inputs).flatten().head
    assert(lossAfter < lossBefore)
  }