Poutyne

Poutyne is an open source Python framework that removes much of the boilerplate involved in training neural networks with PyTorch. A developer writes a PyTorch module as usual and passes it to the Poutyne Model, which then runs the training loop, tracks statistics and applies callbacks. The design is openly inspired by Keras, so the interface will feel familiar to anyone who has used its fit style of training. The library is aimed at researchers and engineers who train their own deep learning models and would rather not rewrite the same loop, validation and logging code for each project. Typical use includes training a network with a chosen optimizer and loss, tracking batch and epoch metrics such as accuracy, F1 and AUROC through torchmetrics, and using callbacks to save the best model or stop training early. It runs on a GPU when one is available and falls back to the CPU otherwise. The documentation shows a short end to end example and gives a citation for academic use. Poutyne is compatible with the latest version of PyTorch and requires Python 3.10 or later. The documentation site covers installation and the API. The project publishes no pricing, and the library is distributed for developers to install and use directly.