1'st Place approach for CVPR 2020 Continual Learning Challenge
Zheda Mai(University of Toronto), Hyunwoo Kim(LG Sciencepark), Jihwan Jeong (University of Toronto), Scott Sanner (University of Toronto, Vector Institute)
Contact: [email protected]
Final Ranking: https://sites.google.com/view/clvision2020/challenge/challenge-winners
Paper: http://arxiv.org/abs/2007.05683
Continual learning is a branch of deep learning that seeks to strike a balance between learning stability and plasticity. The CVPR 2020 CLVision Continual Learning for Computer Vision challenge is dedicated to evaluating and advancing the current state-of-the-art continual learning methods.
The challenge will be based on the CORe50 dataset and composed of three tracks:
Each solution will be evaluated across a number of metrics:
Final aggregation metric (CL_score): weighted average of the 1-5 metrics (0.3, 0.1, 0.15, 0.125, 0.125 respectively
Our approach is based on Experience Replay, a memory-based continual learning method that has been proved effective in various continual learning problems. The details of the approach can be found in our paper.
Download the dataset and related utilities:
sh fetch_data_and_setup.sh
Setup the conda environment:
conda env create -f environment.yml
conda activate clvision-challenge
sh create_submission.sh
The parameters for the final submissions:
config/final/nc.yml
config/final/ni.yml
config/final/nic.yml
The detailed explanation of these parameters can be found in general_main.py
The starting code of this repository is from the official starting repository.