Tutorials#
Our model is developed based on three major Python packages: PyTorch, PyTorch Lightning, and Hydra.
PyTorch provides the fundamental components for model implementation and training, such as neural network layers, optimizers, and learning rate schedulers. With its flexible API and efficient GPU support, it serves as the backbone of deep learning development.
PyTorch Lightning simplifies the training process, especially for distributed and multi-GPU training. Many mainstream models are now implemented with Lightning because it standardizes the training pipeline while keeping research code clean and modular.
Hydra enables configuration management through
YAMLfiles. It allows users to dynamically instantiate any Python class, function, or method from external packages with minimal effort. This makes the entire training and evaluation pipeline highly customizable. For example, users can freely use their own optimizers, learning rate schedulers, etc., without modifying the source code of the software.