Continual Learning for Blind Image Quality Assessment

The codebase of Continual Learning for Blind Image Quality Assessment

Requirement

torch 1.8+ torchvision Python 3 scikit-learn scipy

Usage

Replay-free training:

(1) Using LwF for continual learning of a model BIQA on six tasks:

Modify Line 192 - Line 193 in BIQA_CL.py to :

method = 'LwF'  
training = True  

Then run in terminal: python BIQA_CL.py

(2) Using other continual learning methods:

Modify Line 192 - Line 193 in BIQA_CL.py to :

method = 'EWC' /  'SI' / 'MAS'   
training = True  

Set appropriate regularization weight by modifying Line 84 in BIQA_CL.py:

1000 for si, 10 for mas, 10000 for ewc

Then run in terminal: python BIQA_CL.py

(3) Baselines:

Modify Line 192 - Line 193 in BIQA_CL.py to :

method = 'SL'  / 'SH-CL' / 'MH-CL'
training = True  

Then run in terminal: python BIQA_CL.py

Replay-based training:

(1) Using iCaRL for continual learning of a model BIQA on six tasks:

Modify Line 192 - Line 193 in BIQA_CL.py to :

method = 'LwF-Replay'
training = True

Set Line 98 in BIQA_CL.py

new_replay = False %for iCaRL-v1  
or  
new_replay = True %for iCaRL-v2

Then run in terminal: python BIQA_CL.py

(2) Using other replay methods:

Modify Line 192 - Line 193 in BIQA_CL.py to :

method = 'SH-CL-Replay' / 'MH-CL-Replay'
training = True

Then run in terminal: python BIQA_CL.py

Joint Learning:

Modify Line 192 - Line 193 in BIQA_CL.py to :

method = 'JL'
training = True

Then run in terminal: python BIQA_CL.py

Inference:

(1) Using the weighted quality predictions for inference, can be used with models trained by LwF / Reg-CL / MH-CL / MH-CL-Replay / iCaRL-v1 / iCaRL-v2:

Modify Line 193 - Line 194 in BIQA_CL.py to :

training = False
head_usage = 2

Then run in terminal: python BIQA_CL.py

(2) Using task-oracle information for inference, can be used with models trained by LwF / Reg-CL / MH-CL / MH-CL-Replay / iCaRL-v1 / iCaRL-v2:

Modify Line 193 - Line 194 in BIQA_CL.py to :

training = False
head_usage = 1

Then run in terminal: python BIQA_CL.py

(3) Using the prediction head trained in the latest task for inference, can be used with models trained by LwF / Reg-CL / MH-CL / MH-CL-Replay / iCaRL-v1 / iCaRL-v2:

Modify Line 193 - Line 194 in BIQA_CL.py to :

training = False
head_usage = 0

Then run in terminal: python BIQA_CL.py

(4) Using the single head for inference, can be used with models trained by SL / SH-CL / SH-CL-Replay / JL:

Modify Line 193 - Line 194 in BIQA_CL.py to :

training = False
head_usage = 3

Then run in terminal: python BIQA_CL.py

Citation

Should you find this repo useful to your research, we sincerely appreciate it if you cite our paper :blush: :

@article{zhang2023continual,
  title={Continual Learning for Blind Image Quality Assessment},
  author={Zhang, Weixia and Li, Dingquan and Ma, Chao and Zhai, Guangtao and Yang, Xiaokang and Ma, Kede},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  month={Mar.},
  volume={45},
  issue={3},
  pages={2864 - 2878},
  year={2023}
}