/cgds-package

Package for CGD and ACGD optimizers

Primary LanguagePythonMIT LicenseMIT

CGDs

Overview

CGDs is a package implementing optimization algorithms including CGD and ACGD in Pytorch with Hessian vector product and conjugate gradient.
CGDs is for minimax optimization problem such as generative adversarial networks (GANs) as follows: $$ \min_{\mathbf{x}} \max_{\mathbf{y}} f(\mathbf{x}, \mathbf{y}) $$

Update: ACGD now supports distributed training. Set backward_mode=True to enable.

Warning: This implementation is only for zero sum game setting because it relies on conjugate gradient method to solve matrix inversion efficiently, which requires the matrix to be positive definite. If you are using competitive gradient descent (CGD) algorithm for non-zero sum games, please check more details in CGD paper https://arxiv.org/abs/1905.12103. For example, GMRES (the generalized minimal residual) algorithm can be a solver for non-zero sum setting.

Installation

CGDs can be installed with the following pip command. It requires Python 3.6+.

pip3 install CGDs

You can also directly download the CGDs directory and copy it to your project.

Package description

The CGDs package implements the following optimization algorithms with Pytorch:

How to use

Quickstart with notebook: Examples of using ACGD.

Similar to Pytorch package torch.optim, using optimizers in CGDs has two main steps: construction and update steps.

Construction

To construct an optimizer, you have to give it two iterables containing the parameters (all should be Variables). Then you need to specify the device, learning rates.

Example:

from src import CGDs
import torch
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
optimizer = CGDs.ACGD(max_param=model_G.parameters(), min_params=model_D.parameters(), 
                      lr_max=1e-3, lr_min=1e-3, device=device)
optimizer = CGDs.BCGD(max_params=[var1, var2], min_params=[var3, var4, var5], 
                      lr_max=0.01, lr_min=0.01, device=device)

Update step

Both two optimizers have step() method, which updates the parameters according to their update rules. The function can be called once the computation graph is created. You have to pass in the loss but do not have to compute gradients before step() , which is different from torch.optim.

Example:

for data in dataset:
    optimizer.zero_grad()
    real_output = model_D(data)
   	latent = torch.randn((batch_size, latent_dim), device=device)
    fake_output = D(G(latent))
    loss = loss_fn(real_output, fake_output)
    optimizer.step(loss=loss)

Changelog

  • version 0.0.2: adjust the stopping criterion of CG for better stability

Citation

Please cite it if you find this code useful.

@misc{cgds-package,
  author = {Hongkai Zheng},
  title = {CGDs},
  year = {2020},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/devzhk/cgds-package}},
}