Training

Training in nnodely aims to identify model parameters by minimizing the discrepancy between predicted and target outputs using user-defined loss functions and evaluation metrics.

The training process is managed through the Training module, which provides utilities such as addMinimize, removeMinimize, and trainModel. Training status and diagnostics can be accessed through getTrainingInfo.

The framework supports standard gradient-based optimization methods and multi-objective loss functions through optional weighting. Optimization algorithms are provided by the Optimizer module, including SGD and Adam.

Regularization mechanisms such as dropout and early stopping are also available to improve generalization and prevent overfitting. Early stopping strategies are implemented in the Early Stopping module through functions such as early_stop_patience, mean_stopping, standard_early_stopping, and select_best_model.

To handle temporal heterogeneity across samples, training uses a dedicated batching strategy that preserves temporal coherence within each batch. This ensures that time-dependent relationships are correctly maintained during optimization.

• Training module
• Optimizer module
• Early Stopping module