RactoFit

(CoLLAs 2024) Replaying with Realistic Latent Vectors in Generative Continual Learning

Running the code

1. Preliminary

• Prepare dataset for training
  cd Data
python main_data.py --task disjoint --dataset cifar10 --n_tasks 10 --dir ../Archives
python main_data.py --task disjoint --upperbound True --dataset cifar10 --n_tasks
• Prepare pre-trained model for FID (expert)
  download from here

2. Training

• MerGAN

  python main.py --method Generative_Replay --dataset cifar10 --train_G True
  

• (0.8%) Rehearsal

  python main.py  --method Rehearsal --dataset cifar10 --train_G True  --nb_samples_rehearsal 50
  

• (0%) RactoFit

  python main.py  --method Ractofit_0 --dataset cifar10 --train_G True
  

• (0.8%) RactoFit

  python main.py --method Ractofit --dataset cifar10 --train_G True --num_z 1200
  

3. Evaluation

• FID (expert) python main.py --method Generative_Replay --dataset cifar10 --FID True
• fitting capacity python main.py --method Generative_Replay --dataset cifar10 --Fitting_capacity True

Requirements

python=3.8.8
pytorch=1.13.1
scipy
matplotlib
tqdm
imageio
scikit-learn
lpips

Acknowledgment

Our code is based on the implementations of Generative_Continual_Learning