New Insights on Reducing Drastic Representation Drift in Online Continual Learning

Avoiding disrupting learned representations when new classes are introduced

(key) Requirements

• Python 3.8
• Pytorch 1.6.0

Structure

├── method.py           # Implementation of proposed methods and reported baselines
├── buffer.py           # Basic buffer implementation 
├── data.py             # DataLoaders and Datasets
├── main.py             # main file for ER
├── model.py            # model defitinition and cosine cross-entropy head
├── losses.py           # custom loss functions
├── utils.py            # utilities, from logging to data download

Running Experiments

python main.py --dataset <dataset> --method <method> --mem_size <mem_size> 

ER-ACE example:

python main.py --dataset split_cifar10 --method er_ace --mem_size 20 

ER-AML example:

python main.py --dataset split_cifar10 --method er_aml --mem_size 20 --supcon_temperature 0.2

Logging

Logging is done with Weights & Biases and can be turned on like this:

python -dataset <dataset> --method <method> --wandb_log online --wandb_entity <wandb username>