Use mnist validation or test set when querying the served model

[mathiasburger]

When evaluating the model, the validation or test set should be used because classifiers may overfit and yield perfect results on the training set.

I would suggest changing
batch_xs, batch_ys = mnist.train.next_batch(1)
to
batch_xs, batch_ys = mnist.test.next_batch(1)
in the section about querying the model.