-
Batch - Most atomic step of the learning process. Usually a single forward-backward propagation followed by a model weight update.
-
Callback - A custom piece of code that can be injected into various stages of training to extend the training loop.
-
Epoch - A collection of batches over which we aggregate training metrics over. In supervised learning epoch is often equal to iterating over the whole dataset.
-
Metric - A variable that we track through the learning process for introspection and analysis purposes.
-
Storage - Object responsible for persistence of training outputs. Mostly used for storing trained models and metrics collected during training.