Prodigy: An Expeditiously Adaptive Parameter-Free Learner

This is the official repository used to run the experiments in the paper that proposed the Prodigy optimizer. The optimizer is implemented in PyTorch. There is also a JAX version of Prodigy in Optax, which currently does not have the slice_p argument.

Prodigy: An Expeditiously Adaptive Parameter-Free Learner
K. Mishchenko, A. Defazio
Paper: https://arxiv.org/pdf/2306.06101.pdf

Installation

To install the package, simply run pip install prodigyopt

How to use

Let net be the neural network you want to train. Then, you can use the method as follows:

from prodigyopt import Prodigy
# choose weight decay value based on your problem, 0 by default
# set slice_p to 11 if you have limited memory, 1 by default
opt = Prodigy(net.parameters(), lr=1., weight_decay=weight_decay, slice_p=slice_p)

Note that by default, Prodigy uses weight decay as in AdamW. If you want it to use standard $\ell_2$ regularization (as in Adam), use option decouple=False. We recommend using lr=1. (default) for all networks. If you want to force the method to estimate a smaller or larger learning rate, it is better to change the value of d_coef (1.0 by default). Values of d_coef above 1, such as 2 or 10, will force a larger estimate of the learning rate; set it to 0.5 or even 0.1 if you want a smaller learning rate.
Standard values of weight_decay to try are 0 (default in Prodigy), 0.001, 0.01 (default in AdamW), and 0.1.
Use values of slice_p larger than 1 to reduce the memory consumption. slice_p=11 should give a good trade-off between accuracy of estimate learning rate and memory efficiency.

Scheduler

As a rule of thumb, we recommend either using no scheduler or using cosine annealing with the method:

# n_epoch is the total number of epochs to train the network
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=total_steps)

We do not recommend using restarts in cosine annealing, so we suggest setting T_max=total_steps, where total_steps should be the number of times scheduler.step() is called. If you do use restarts, we highly recommend setting safeguard_warmup=True.

Extra care should be taken if you use linear warm-up at the beginning: The method will see slow progress due to the initially small base learning rate, so it might overestimate d. To avoid issues with warm-up, use option safeguard_warmup=True.

Diffusion models

Based on the interaction with some of the users, we recommend setting safeguard_warmup=True, use_bias_correction=True, and weight_decay=0.01 when training diffusion models. Sometimes, it is helpful to set betas=(0.9, 0.99).
If the model is not training, try to keep track of d and if it remains too small, it might be worth increasing d0 to 1e-5 or even 1e-4. That being said, the optimizer was mostly insensitive to d0 in our other experiments.

Examples of using Prodigy

See this Colab Notebook for a toy example of how one can use Prodigy to train ResNet-18 on Cifar10 (test accuracy 80% after 20 epochs).
If you are interested in sharing your experience, please consider creating a Colab Notebook and sharing it in the issues.

How to cite

If you find our work useful, please consider citing our paper.

@inproceedings{mishchenko2024prodigy,
    title={Prodigy: An Expeditiously Adaptive Parameter-Free Learner},
    author={Mishchenko, Konstantin and Defazio, Aaron},
    booktitle={Forty-first International Conference on Machine Learning},
    year={2024},
    url={https://openreview.net/forum?id=JJpOssn0uP}
}