A Simple Recipe to Meta-Learn Forward and Backward Transfer

<div> <img src="images/sim4c_overview.png" alt="Overview of meta-pretraining with SiM4C" width="100% align="middle"> </div>

This repository contains the code for the main meta pre-training experiments with the simple meta-learning algorithm for continual learning (SiM4C) from the ICCV paper A Simple Recipe to Meta-Learn Forward and Backward Transfer. The code to download and preprocess the Omniglot dataset is based on the prior meta pre-training implementation from Javed et al. 2019.

Installation

We provide a configuration file to easily install dependencies via conda:

conda env create -f conda_env.yml
conda activate metaL

Replicating the results with SiM4C

<div> <img src="images/omniglot_meta_test.png" alt="Meta pre-training omniglot results" width="100% align="middle"> </div>

We provide a simple script to replicate our main experiments with SiM4C:

To perform both meta pre-training and meta-testing, run:

./scripts/train_sim4c.sh

To re-run meta-testing with different configurations, check the location of the experiment folder (ending in the current date) and run:

python eval_all.py experiment_path=path/to/experiment/folder

To run and evaluate alternative baselines execute train.py and eval.py/eval_all.py overriding the appropriate arguments (see hydra for details).

Reference

To reference this work in future research, you can use the following:

@InProceedings{Cetin_2023_ICCV,
    author    = {Cetin, Edoardo and Carta, Antonio and Celiktutan, Oya},
    title     = {A Simple Recipe to Meta-Learn Forward and Backward Transfer},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {18732-18742}
}