Implementation for ACProp ( Momentum centering and asynchronous update for adaptive gradient methdos, NeurIPS 2021)

This repository contains code to reproduce results for submission NeurIPS 2021, "Momentum Centering and Asynchronous Update for Adaptive Gradient Methods".

This repo heavily depends on the official implementation of AdaBound: https://github.com/Luolc/AdaBound and AdaBelief: https://github.com/juntang-zhuang/Adabelief-Optimizer.

Dependencies

python 3.7 pytorch 1.1.0 torchvision 0.3.0 jupyter notebook AdaBelief (Please install by "pip install adabelief-pytorch==0.2.0")

Visualization of pre-trained curves

Please use the jupyter notebook "visualization.ipynb" to visualize the training and test curves of different optimizers.

Training and evaluation code

(1) train network with CUDA_VISIBLE_DEVICES=0 python main.py --model vgg --optim acprop --lr 1e-3 --eps 1e-8 --beta1 0.9 --beta2 0.999 --momentum 0.9

--model: name of model, choices include ['vgg','resnet','densenet'] --optim: name of optimizers, choices include ['acprop','adashift','sgd', 'adam', 'adamw', 'adabelief', 'radam',] --lr: learning rate --eps: epsilon value used for optimizers. Note that Yogi uses a default of 1e-03, other optimizers typically uses 1e-08 --beta1, --beta2: beta values in adaptive optimizers --momentum: momentum used for SGD.

(2) visualize using the notebook "visualization.ipynb"