Strategies in picking training data for 3D convolutional neural networks in stratigraphic interpretation