Add classification capacity about geo-deep-learning

[epeterson12]

There seem to be a few main ways that labels are associated with images for classification purposes:

1. A relational-db-like structure using csv or text files (WordNet, DeepSat SAT6, Overhead Imagery Research Data Set, OpenImages)
2. A file system tree (ImageNet, Caltech 101)
3. A custom file format (MNIST)
4. Pre-made datasets (CIFAR10 and CIFAR100)

In pytorch, there is a useful generic dataset class called ImageFolder. The data must be stored in a file system tree structure like ImageNet. I suggest that we take advantage of this for our classification feature. We would be able to skip the create_samples.py step when doing classification if we use this class.

For our net, we could use an existing net model from Torchvision such as a ResNet or inception3 which both are good for classification. I will be looking in to what needs to be changed in our existing code to add a classification feature using these methods.