Full Project Report

Please see AI Project 2 Report

By the end of the project due, we rank 2nd of the new classes task in this competition!

Summary

We followed the idea from Overcoming catastrophic forgetting in neural networks (EWC) and conducted experiments on both MINST and CORe50 datasets. Results shows that our model beats the Resnet-18 baseline by 2% on CORe50.

Running on MNIST and Fashion MNIST dataset

python ewc_mnist.py

Running on CoRe50 dataset

Setup

Follow the CVPR Starter Instruction

sh fetch_data_and_setup.sh
conda env create -f environment.yml
conda activate clvision-challenge

Run

python main_ewc.py --scenario="multi-task-nc" --sub_dir="ewc3000tt" [ --epochs=EPOCH ] [ --ewc_weight=EWC_WEIGHT ] [ --ewc_explosion_multr_cap=EWC_EXPLOSION_MULTR_CAP ]