A library and a set of commands to simplify transformer-based machine learning experimentation.
The following commands are provided:
farm-estimate: perform crossvalidation or holdout-estimation. This supports stratified sampling for both the dev set and the training/test sets.
farm-hsearch: perform hyperparameter search via iterated estimation
farm-train: train a model on a training set
farm-apply: apply a trained model to new data
All the training/estimation commands automatically create per-run directories with detailed logs and data about the experiment so that all information is available later. Hyperparameters and other configuration data can be provided via command line arguments or in configuration files.