Code for the paper Layerwise Proximal Replay: A Proximal Point Method for Online Continual Learning, Jason Yoo, Yunpeng Liu, Frank Wood, Geoff Pleiss, ICML 2024 arxiv.
This repository is based on the OCL Survey repository.
Clone this repository
git clone [email protected]:plai-group/LPR.git
Create a new environment with python 3.10
conda create -n lpr python=3.10
conda activate lpr
Install specific ocl_survey repo dependencies
pip install -r requirements.txt
Set your PYTHONPATH as the root of the project
conda env config vars set PYTHONPATH=$(pwd -P)
In order to let the scripts know where to fetch and log data, you should also create a deploy config, indicating where the results should be stored and the datasets fetched. Either add a new one or change the content of config/deploy/default.yaml.
LPR can be enabled by setting the flag strategy.lpr=true in the run command. Most of LPR's implementation is located at src/strategies/lpr_plugin.py.
Sample Run Commands
Split-CIFAR100
cd experiments
python main.py strategy=er experiment=split_cifar100 optimizer.lr=0.1 strategy.lpr=true strategy.lpr_kwargs.preconditioner.omega_0=4 strategy.lpr_kwargs.preconditioner.beta=2 strategy.lpr_kwargs.update.every_iter=30
Split-TinyImageNet
cd experiments
python main.py strategy=er experiment=split_tinyimagenet optimizer.lr=0.01 strategy.lpr=true strategy.lpr_kwargs.preconditioner.omega_0=0.25 strategy.lpr_kwargs.preconditioner.beta=2 strategy.lpr_kwargs.update.every_iter=90
Online CLEAR
Download the dataset from this link and unzip it at ./data.
Then, run
cd experiments
python main.py strategy=er experiment=clear optimizer.lr=0.01 strategy.lpr=true strategy.lpr_kwargs.preconditioner.omega_0=1. strategy.lpr_kwargs.preconditioner.beta=1. strategy.lpr_kwargs.update.every_iter=100
The code is structured as follows:
├── avalanche.git # Avalanche-Lib code
├── config # Hydra config files
│ ├── benchmark
│ ├── best_configs # Best configs found by main_hp_tuning.py are stored here
│ ├── deploy # Contains machine specific results and data path
│ ├── evaluation # Manage evaluation frequency and parrallelism
│ ├── experiment # Manage general experiment settings
│ ├── model
│ ├── optimizer
│ ├── scheduler
│ └── strategy
├── experiments
│ ├── main_hp_tuning.py # Main script used for hyperparameter optimization
│ ├── main.py # Main script used to launch single experiments
│ └── spaces.py
├── notebooks
├── results # Exemple results structure containing results for ER
├── scripts
└── get_results.py # Easily collect results from multiple seeds
├── src
│ ├── factories # Contains the Benchmark, Method, and Model creation
│ ├── strategies # Contains code for additional strategies or plugins
│ └── toolkit
└── tests
LPR specific hyperparameters can be found at config/strategy/method_defaults.yaml.
To launch an experiment, start from the default config file and change the part that needs to change
python main.py strategy=er_ace experiment=split_cifar100 evaluation=parallel
It's also possible to override more fine-grained arguments
python main.py strategy=er_ace experiment=split_cifar100 evaluation=parallel strategy.alpha=0.7 optimizer.lr=0.05
Before running the script, you can display the full config with "-c job" option
python main.py strategy=er_ace experiment=split_cifar100 evaluation=parallel -c job
Results will be saved in the directory specified in results.yaml. Under the following structure:
<results_dir>/<strategy_name>_<benchmark_name>/<seed>/
If you use this repo for a research project please use the following citation:
@misc{yoo2024layerwise,
title={Layerwise Proximal Replay: A Proximal Point Method for Online Continual Learning},
author={Jason Yoo and Yunpeng Liu and Frank Wood and Geoff Pleiss},
year={2024},
eprint={2402.09542},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2402.09542},
}