CFA - Class-Incremental Learning via Knowledge Amalgamation

Official repository of Class-Incremental Learning via Knowledge Amalgamation

Citing this work

To be updated

Setting up a CONDA environment

Execute line by line

conda create -n CFA python=3.8
conda activate CFA
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
conda install tqdm matplotlib
pip install avalanche-lib

Setting up a PIP environment

Execute line by line

pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu113
pip install tqdm
pip install matplotlib
pip install avalanche-lib

Running

For a list of commands:

python cfa.py --help

For MNIST

python cfa.py --dataset mnist --memory_budget 1000 --memory_strategy fixed
python cfa.py --dataset mnist --memory_budget 1000 --memory_strategy grow

For CIFAR10

python cfa.py --dataset cifar10 --memory_budget 1000 --memory_strategy fixed
python cfa.py --dataset cifar10 --memory_budget 1000 --memory_strategy grow

For CIFAR100

python cfa.py --dataset cifar100 --memory_budget 1000 --memory_strategy fixed
python cfa.py --dataset cifar100 --memory_budget 1000 --memory_strategy grow

For Tiny ImageNet

python cfa.py --dataset tiny10 --memory_budget 1000 --memory_strategy fixed
python cfa.py --dataset tiny10 --memory_budget 1000 --memory_strategy grow

Tip 1

If you are not intersted in evaluating the BWT and FWT metrics, just the ACC, modify the line 721 from:

        for n_task in range(2, n_tasks + 1, 1):

to

        for n_task in range(n_tasks, n_tasks + 1, 1):

In order to calculate BWT and FWT, we need to run multiple CFA experiments, which can be time-consuming. By making this change, you force the algorithm to just run a full amalgamation of all teachers. This will give you the ACC metric, but BWT and FWT will not be valid.

Tip 2

CFA accuracy (the student model accuracy) is really dependent on the performance of the teacher models.