Module training 

FastGRNN training pipeline with knowledge distillation

This module provides a complete training infrastructure for the FastGRNN model:

• Adam optimizer implementation
• Binary Cross-Entropy loss with gradient computation
• Backpropagation Through Time (BPTT)
• Mini-batch training with validation split
• Early stopping and learning rate scheduling
• Knowledge distillation from teacher models
• Progress reporting and metrics tracking

Structs

BatchIterator

Batch iterator for training

Trainer

FastGRNN trainer

TrainingConfig

Training hyperparameters

TrainingDataset

Training dataset with features and labels

TrainingMetrics

Training metrics

Functions

generate_teacher_predictions

Generate teacher predictions for knowledge distillation

temperature_softmax

Temperature-scaled softmax for knowledge distillation with numerical stability