Better use of validation split

[JohnGiorgi]

Overview

Previously, only one of k_folds or validation_split could be used, with k_folds taking precedence over validation_split. Now, they can be used together.

• To use k-fold cross-validation, provide only a train.* file under dataset_folder. Choose the number of folds with the k_folds argument. Optionally, specify a proportion of examples to hold-out at random in each fold as a validation set with validation_split.
• To create a validation split from the train set, provide a train.* file (and optionally, a test.* file) under dataset_folder and specify the proportion of training examples to hold-out for a validation set with validation_split.

Otherwise, provide the partitions yourself with the files train.*, valid.* and test.* under dataset_folder and leave k_folds and validation_split equal to 0.

E.g.

.
├── NCBI_Disease
│   └── train.tsv
│   └── valid.tsv
│   └── test.tsv

k_folds will be ignored if either a valid.* or test.* file is found under dataset_folder. Both arguments k_folds and validation_split will be ignored if a valid.* file is found under dataset_folder.

TODOs

• [x] Update tests for this new functionality.
• [x] Update docs to reflect this new scheme

Closes

Closes #154.