train and test loss

[soroushr]

How is train/test splitting handled and how can it be stored in tf.summary logs (for visualizing in tensorboard)?

Several open issues (see #130, #163, #203) have discussed similar problems but I cannot find any straightforward answer to it.

The batch_x and batch_y that are used in verification are apparently already included in the training set, so this seems like an inappropriate test set (@jakeret correct me if I'm wrong).

Thanks!