Looking at the machine learning landscape, one of the major trends is the proliferation of projects focused on applying software engineering principles to machine learning. Cortex, for example, recreates the experience of deploying serverless functions, but with inference pipelines. DVC, similarly, implements modern version control and CI/CD pipelines, but for ML.

PyTorch Lightning has a similar philosophy, only applied to training. The frameworks provides a Python wrapper for PyTorch that lets data scientists and engineers write clean, manageable, and performant training code.

As people who built an entire deployment platform in part because we hated writing boilerplate, we’re huge fans of PyTorch Lightning. In that spirit, I’ve put together this guide to deploying PyTorch Lightning models to production. In the process, we’re going to look at a few different options for exporting PyTorch Lightning models for inclusion in your inference pipelines.

Every way to deploy a PyTorch Lightning model for inference

There are three ways to export a PyTorch Lightning model for serving:

• Saving the model as a PyTorch checkpoint
• Converting the model to ONNX
• Exporting the model to Torchscript

We can serve all three with Cortex.

#deep-learning #programming #machine-learning #devops #pytorch