How to Build a Custom Optimizer in PyTorch: 5 Simple Steps

Author(s): Benjamin Bodner

Originally published on Towards AI.

Leverage Pytorch’s powerful optimization framework to implement your algorithm

This member-only story is on us. Upgrade to access all of Medium.

Optimization is key in deep learning. It is the learning part of deep learning!

If you are reading this post, you likely already know that the optimization process is typically controlled by rule-based weight update algorithms called optimizers. A wide variety of them are available in Pytorch torch.optim.

However, in certain cases, you may find that Pytorch may not have the exact algorithm you need. Whether you are working on a research project or just trying to get a better “under-the-hood” understanding of the Pytorch library, you might consider implementing your own optimizer.

Though it is rarely done, creating a custom optimizer in PyTorch is pretty simple and can be done in just a few small steps.

Start by creating a class that inherits from the base PyTorch optimizer torch.optim.Optimizer. This ensures your custom optimizer will have the basic structure PyTorch needs. It also provides your optimizer with a lot of functionality, such as: hook management, statedict management, a basic zero grad operation, adding parameter groups, and more.

If none of these are familiar to you, don’t worry! We’ll only dive into the relevant parts in this… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI