Evaluation before batch training, training on batch and post training returns different loss values.
pre_train_loss = model.evaluate(batch_x, batch_y, verbose=0)
