EVE: The EVer Evolving Deep Learning Optimizer

EVE is a new optimizer library built on top of PyTorch that combines the best of multiple state-of-the-art optimizer algorithms into one flexible, infinitely customizable super-optimizer.

The goal of EVE is not to provide one final, static optimizer, but rather an interface to a PyTorch optimizer that will continue to implement the latest, well-tested methods from modern research.

In preliminary testing, the current implementation of EVE was able to beat Adam and other near state-of-the-art optimizers without a significant increase in compute time. Here are some inital results from training a ResNet18 on the ImageNette (subset of ImageNet that encompasses a few hard to classify classes) 5 epoch challenge.

Adam (Final Accuracy = 40.00%)

EVE (Final Accuracy = 70.62%)

Here are a few animations demonstrating EVE's convergence properties on simple functions:

2D Convex Surface	2D Non-Convex Surface	3D Surface with Saddle Point

Installation and Getting Started

The simplest way to use EVE in your PyTorch models is to install it using pip:

pip install eve-optimizer

Then, the main EVE optimizer can be imported as follows:

from eve.optimizers import eveo3

This will import a function that returns a torch.optim.Optimizer object, which can be used in the usual way.

The EVE library also provides a direct interface to other optimizers (like Ranger, RAdam, etc.) that were used in part or were built upon to create the main EVE optimizer. These can also be accessed from eve.optimizers in the same way.

What Exactly is EVE?

At present, EVE implements (and combines) the following algorithms:

• Adam
• RAdam
• LARS
• LAMB
• LookAhead
• DiffGrad

We are currently working on adding in the following variants as well:

• AdaMod
• DeepMemory
• Marina