Structure of a training program
The structure of a training program always consists of the following steps:
Set the script environment: Such as package imports, the use of the GPU, and so on.
Load data: A data loader class to access the data during training, usually in a random order to avoid too many similar examples of the same class, but sometimes in a precise order, for example, in the case of curriculum learning with simple examples first and complex ones last.
Preprocess the data: A set of transformations, such as swapping dimensions on images, adding blur or noise. It is very common to add some data augmentation transformations, such as random crop, scale, brightness, or contrast jittering to get more examples than the original ones, and reduce the risk of overfitting on data. If the number of free parameters in the model is too important with respect to the training dataset size, the model might learn from the available examples. Also, if the dataset is too small and too many iterations...
