DER++ implementation

Implementation and test of DER++ Continual Learning baseline from the article Dark Experience for General Continual Learning: a Strong, Simple Baseline, Buzzega et al. (https://arxiv.org/pdf/2004.07211.pdf).

The method was implemented mainly using Avalanche and Pytorch libraries. Tests and comparison of the model were done mainly on Split-Cifar100 dataset, which was not explored in the original paper

Project structure

• der.py contains the implementation of DER/DER++ as an Avalanche plugin + a custom reservoir buffer.
• preliminary_analysis_SCIFAR10.ipynb contains a comparison on Split-CIFAR10 of DER++, Experience Replay and Naive (standard finetuning), trying to reproduce the orignal paper.
• grid_search_der.ipynb contains the grid search code and analysis over DER++ hyperparameters on Split-CIFAR100.
• grid_search_replay_naive.ipynb contains the grid search code over Experience Replay and Naive hyperparameters on Split-CIFAR100.
• final_retrain contains the retrain on the entire training stream of Split-CIFAR100 for DER++, Replay and Naive. Contains also comparisons between the 3 methods final results.
• results folder containing the results of the various model runs.