/AGS-CL

Primary LanguagePython

Continual Learning with Node Importance based Adaptive Group Sparse Regularization (AGS-CL)

Execution Details

1. Supervised Learning

Requirements

  • Python 3
  • Cifar100/Cifar10/100 : GPU 1080Ti / Pytorch 1.3.1+cu9.2 / CUDA 9.2
  • Omniglot : GPU Titan RTX / Pytorch 1.3.1 / CUDA 10.0
  • CUB200 : GPU 1080Ti / Pytorch 1.0.0+cu9.2 / CUDA 9.2

1) Download dataset

2) Execution command

# Cifar100
$ python3 ./main.py --experiment split_cifar100 --approach gs --lamb 400 --mu 10 --rho 0.3 --eta 0.9 

# Cifar10/100
$ python3 ./main.py --experiment split_cifar10_100 --approach gs --lamb 7000 --mu 20 --rho 0.2 --eta 0.9 

# Omniglot
$ python3 ./main.py --experiment omniglot --approach gs --lamb 1000 --mu 7 --rho 0.5 --eta 0.9 

# CUB200
$ cd LargeScale_AGS
$ python3 ./main.py --dataset CUB200 --trainer gs --lamb 1.5 --mu 0.5 --rho 0.1 --eta 0.9

3) Result(Average accuracy)

CIFAR100 CIFAR-10/100 Omniglot CUB-200
AGS-CL 64.2 76.0 82.2 82.9

2. Reinforcement Learning (Atari)

Requirements

  • Python 3.6
  • Pytorch 1.2.0+cu9.2 / CUDA 9.2
  • OpenAI Gym, Baselines

Notes

The experimental environment for reinforcement learning is built based on pytorch-a2c-ppo-acktr-gaail

1) Install OpenAI Gym, Baselines

​ Follow below links for installation

​ OpenAI Gym, Baselines

2) Execution command

# Fine-tuning
$ CUDA_VISIBLE_DEVICES=0 python3 main_rl.py --approach 'fine-tuning' --seed 0 --date 200605  

# EWC
$ CUDA_VISIBLE_DEVICES=0 python3 main_rl.py --approach 'ewc' --seed 0 --date 200605 

# AGS-CL
$ CUDA_VISIBLE_DEVICES=0 python3 main_rl.py --approach 'gs' --seed 0 --date 200605 --gs-mu 0.1 --gs-lamb 1000