Skip to content

architectures

Overview

The architectures subpackage provides a framework for defining, training, and retraining various neural network architectures tailored for biosignal processing. The package supports:

✔ Training from scratch and retraining with configurable hyperparameters.

✔ Model checkpointing & hyperparameter optimization via grid search with PyTorch Lightning.

✔ Integration with TensorBoard for visualization of training metrics.

Package Structure

  1. Architecture Class and related functions
    • Definition of Architecture class, which is extended by all neural network models in the package.
    • Utility functions for checkpoint management, hyperparameter extraction, training and retraining, and testing.
  2. Biosignals Architectures
    • Implements neural network architectures designed for biosignal processing tasks, including GRU-based models (e.g., GRUseq2seq, GRUEncoderDecoder) and Transformer-based models (Transformerseq2seq). These architectures are optimized for sequence modeling tasks such as signal denoising or timeseries classification.
  3. Training and Deployment Workflow in NeuralLib
    • Provides high-level functions for training and retraining models dynamically based on user-defined parameters.
  4. upload_to_hugging : enables uploading trained models to Hugging Face for production.
  5. post_process_fn: contains post-processing utilities for handling predictions.

Hands on

Check Tutorials:

  1. Tutorial #1: Train, retrain, and test a biosignals model.

    Key methods: train_from_scratch(), retrain() and test_on_test_set()

  2. Tutorial #2: Hyperparameter optimization using grid search. Key methods: run_grid_search, get_hparams_from_checkpoints, test_on_single_signal

  3. Tutorial #3: Convert a trained model into a production-ready model. Key method: upload_production_model()

Best Practices

  1. Directory Management:
    • Ensure checkpoint directories are accessible and writable.
  2. Dataset formatting:
    • Ensure datasets are structured correctly: input (path_x) and output (path_y) each with ‘train’, ‘test’, and ‘val’ subfolders.
    • (optional) Add the dataset directory to config.file
    • biosignal datasets should be in .npy files: each signal in a separate file.
  3. Logging:
    • Enable TensorBoard logging for detailed monitoring of training progress.