Components of DL frameworks
Since the configuration of model training follows the same process regardless of the underlying tasks, many engineers and researchers have put together the common building blocks into frameworks. Most of the frameworks simplify DL model development by keeping data loading logic and model definitions independent from the training logic.
The data loading logic
Data loading logic includes everything from loading the raw data in memory to preparing each sample for training and evaluation. In many cases, data for the train set, validation set, and test set are stored in separate locations, so that each of them requires a distinct loading and preparation logic. The standard frameworks keep these logics separate from the other building blocks so that the model can be trained using different datasets in a dynamic way with minimal changes on the model side. Furthermore, the frameworks have standardized the way that these logics are defined to improve reusability...