FETCH: A Memory-Efficient Replay Approach for Continual Learning in Image Classification

This repository contains the source code for our IDEAL2023 Paper.

[Paper Website]

If you have any questions please open an issue or find my email adress on this website.

How to use this code

Tested with python 3.10.11

1. Create a Virtual Environment

python3 -m venv venv
source venv/bin/activate

2. Install the dependencies from requirements.txt:

pip3 install -r requirements.txt

3. To download the TinyImagenet-Dataset, have a look at utils/download_TinyImagenet.py. Change the configuration of this script at the beginning of the script.
4. To replicate the results from the thesis, have a look at notebooks/-directory. The whole codebase is structured as follows: a jupyter-notebook creates a shell script with the configuration of the experiments. The shell scripts start the exeriments and which create log-files. After that you can visualize the logs in another jupyter-notebook
5. Open a notebook and configure the experiment (Especially the parameters LOG_DIR and DATA_DIR).
6. Run the first part of the notebook to produce a shell-script.
7. Run the shell script
8. If you want to read the code, src/main.py is a good place to start

Example

Replace YOUR_DATA_DIR with the directory where pytorch should download the datasets. Run the code from the project root. It should take ~5 minutes to train on a normal laptop. You should find the logs in logs/demo/. I get an accuracy of ~0.2

python3 src/main.py \
    --dataset CIFAR10 \
    --num_classes_per_task 2 \
    --num_tasks 5 \
    --seed 0 \
    --memory_size 64 \
    --num_passes 4 \
    --sampler greedy_sampler \
    --encoder cutr_cifar \
    --encoding_block 3 \
    --compressor quantization \
    --n_states 4 \
    --strategy cifar10_transfer \
    --backbone resnet18_cifar \
    --backbone_block 3 \
    --data_dir YOUR_DATA_DIR \
    --log_dir logs \
    --device cpu \
    --exp_name demo

How to replicate results from the paper

Have a look at the following notebooks in the notebooks directory. Adapt the parameters to your needs

figure	related files
2	cifar10_rescon.ipynb, cifar100_rescon.ipynb, plot_comb.ipynb
3	setup_wts.ipynb, plot_wts.ipynb
4	cifar10_fixed_slots.ipynb, cifar100_fixed_slots.ipynb, plot_eoc.ipynb
5	cifar10_fixed_memory.ipynb, cifar100_fixed_memory.ipynb, plot_fixb.ipynb
6	cifar10_fixed_memory.ipynb, cifar100_fixed_memory.ipynb, plot_vs_acae.ipynb

Command Line Options

See src/opts.py for some of the command line options. Each of the encoders, compressors and classification heads (backbones) have their own set of options. To see them, have a look at the implementation in src/encoders, src/compressors and src/backbones respectively.

Acknowledgment

This code is based on GDumb:

Prabhu, A, Torr, P, Dokania, PGDumb: A Simple Approach that Questions Our Progress in Continual Learning. In The European Conference on Computer Vision (ECCV) 2020.