Codebase for Continual Prototype Evolution (CoPE) to attain perpetually representative prototypes for online and non-stationary datastreams. Includes implementation of the Pseudo-Prototypical Proxy (PPP) loss.
Continual Prototype Evolution (CoPE) establishes online adaptation of class-representative prototypes in non-stationary data streams, exploiting latent space representations in the novel PPP-loss to enhance the state-of-the-art in continual learning.
This codebase contains the original PyTorch implementation of CoPE, along with the Split-MNIST, Split-CIFAR10, Split-CIFAR100 benchmarks. The benchmarks have both a balanced and highly imbalanced variant, resembling more real-life settings. Included baselines outperformed in these settings are: CoPE-CrossEntropy, GEM, iCaRL, GSS, reservoir sampling, finetuning, online iid, offline iid.
Keywords: continual learning, prototypical learning, online learning, incremental learning, deep learning, representation learning, catastrophic forgetting, concept drift
Main scripts main_MNIST.sh, main_CIFAR10.sh, main_CIFAR100.sh contain fully automatic pipeline (auto datapreparation), with hyperparameter configs for all of the experiments in the main paper.
The balanced setups contain:
The imbalanced setups contain (averaged over 5 different choices of dominant task):
Use environment.yml to create anaconda environment:
conda env create -f environment.yml # Env named 'cope'
conda activate cope
Or manually, as in:
# Create and activate environment
conda create -n <name> python=3.7
conda activate <name>
# Pytorch (e.g. for CUDA 10.2)
conda install pytorch==1.5.0 torchvision==0.6.0 cudatoolkit=10.2 -c pytorch
# Optional
conda install -c conda-forge matplotlib=3.1.3 # T-SNE plots
conda install -c conda-forge scikit-learn=0.22.1
conda install -c omnia quadprog # GEM baseline
This final code-base is validated to produce similar results to the original results reported in the paper.
Although the data streams are divided into tasks to compare with task and class-incremental learning alorithms (iCaRL, GEM), in CoPE the continual learner is unaware of tasks or task transitions. This means CoPE can learn from any labeled data stream, without the bias of hand-designed task boundaries within the stream.
The learner-evaluator framework defined in the paper, explicitly models all the requirements of the continual learning system.
We define the learner here for CoPE:
With the evaluator:
Consider citing our work upon using this repo.
@InProceedings{De_Lange_2021_ICCV,
author = {De Lange, Matthias and Tuytelaars, Tinne},
title = {Continual Prototype Evolution: Learning Online From Non-Stationary Data Streams},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
month = {October},year = {2021}, pages = {8250-8259}
}
CoPE has been made available in the Avalanche framework (free to use under MIT license)!
Thanks to the following repositories:
This source code is released under a Attribution-NonCommercial 4.0 International license, find out more about it in the LICENSE file.