Awesome

BASQ: Branch-wise Activation-clipping Search Quantization for Sub-4-bit Neural Networks

Paper | Supplementary

BASQ: Branch-wise Activation-clipping Search Quantization for Sub-4-bit Neural Networks, ECCV 2022 <br> Han-Byul Kim¹, Eunhyeok Park², Sungjoo Yoo¹ <br> ¹Seoul National University, ²POSTECH

What is BASQ?

Accurate 2-bit to 4-bit uniform quantization method for both MobileNets and ResNets working with

① Quantization hyper-parameter (clipping value with L2-decay weight) search with architecture search method

② Block structure for removing unstable effect from proposed quantization hyper-parameter search

The search process of BASQ is composed of three steps as below.

01. Supernet training

$ python 01train_supernet.py --data {your_imagenet_dataset_folder}

02. Architecture search

$ python 02kfold_evolution.py --data {your_imagenet_dataset_folder}

03. Finetuning

Before running finetuning, please attach the result of architecture search as below.

For example, our results of architecture search are same as below.

===total-result===
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
fold_accs: [67.62, 66.2, 64.86, 68.94, 55.18, 58.96, 54.82, 56.46, 53.54, 63.48]
full_accs: [59.30222222222222, 59.29333333333334, 59.26222222222222, 58.67777777777778, 59.733333333333334, 59.15777777777778, 60.01111111111111, 60.45333333333333, 60.10888888888889, 59.53333333333333]
archs: [(5, 3, 2, 2, 1, 1, 6, 6, 1, 2, 4, 6, 0, 3, 3, 5, 5), (1, 4, 5, 0, 5, 2, 3, 6, 4, 5, 6, 6, 6, 6, 3, 2, 5), (1, 4, 5, 0, 2, 5, 5, 2, 5, 2, 1, 1, 2, 2, 6, 4, 5), (1, 4, 5, 1, 1, 2, 4, 6, 2, 1, 2, 6, 4, 2, 6, 4, 5), (1, 4, 5, 5, 1, 5, 1, 5, 6, 5, 4, 1, 6, 4, 3, 2, 3), (5, 3, 2, 2, 6, 6, 5, 0, 5, 5, 6, 6, 4, 6, 5, 0, 2), (1, 4, 5, 1, 6, 2, 6, 4, 2, 1, 1, 6, 4, 0, 6, 4, 5), (5, 3, 2, 2, 1, 1, 4, 2, 1, 2, 4, 6, 0, 0, 3, 6, 5), (1, 4, 5, 5, 6, 5, 2, 3, 5, 6, 3, 6, 5, 2, 6, 0, 2), (5, 3, 2, 2, 4, 1, 4, 2, 1, 4, 6, 6, 0, 4, 3, 2, 5)]

Paste it to line 66 of 03finetuning.py.

FOLD_ACCS = [67.62, 66.2, 64.86, 68.94, 55.18, 58.96, 54.82, 56.46, 53.54, 63.48]
FULL_ACCS = [59.30222222222222, 59.29333333333334, 59.26222222222222, 58.67777777777778, 59.733333333333334, 59.15777777777778, 60.01111111111111, 60.45333333333333, 60.10888888888889, 59.53333333333333]
BEST_ARCHS = [(5, 3, 2, 2, 1, 1, 6, 6, 1, 2, 4, 6, 0, 3, 3, 5, 5), (1, 4, 5, 0, 5, 2, 3, 6, 4, 5, 6, 6, 6, 6, 3, 2, 5), (1, 4, 5, 0, 2, 5, 5, 2, 5, 2, 1, 1, 2, 2, 6, 4, 5), (1, 4, 5, 1, 1, 2, 4, 6, 2, 1, 2, 6, 4, 2, 6, 4, 5), (1, 4, 5, 5, 1, 5, 1, 5, 6, 5, 4, 1, 6, 4, 3, 2, 3), (5, 3, 2, 2, 6, 6, 5, 0, 5, 5, 6, 6, 4, 6, 5, 0, 2), (1, 4, 5, 1, 6, 2, 6, 4, 2, 1, 1, 6, 4, 0, 6, 4, 5), (5, 3, 2, 2, 1, 1, 4, 2, 1, 2, 4, 6, 0, 0, 3, 6, 5), (1, 4, 5, 5, 6, 5, 2, 3, 5, 6, 3, 6, 5, 2, 6, 0, 2), (5, 3, 2, 2, 4, 1, 4, 2, 1, 4, 6, 6, 0, 4, 3, 2, 5)]

Run finetuning.

$ python 03finetune.py --data {your_imagenet_dataset_folder}

Final results are printed as below. Bnfull_accs are finetuned results.

===final-result===
fold_accs: [67.62, 66.2, 64.86, 68.94, 55.18, 58.96, 54.82, 56.46, 53.54, 63.48]
full_accs: [59.30222222222222, 59.29333333333334, 59.26222222222222, 58.67777777777778, 59.733333333333334, 59.15777777777778, 60.01111111111111, 60.45333333333333, 60.10888888888889, 59.53333333333333]
**bnfull_accs: [63.95, 64.02, 64.28, 63.89, 65.46, 65.1, 65.23, 65.32, 65.4, 64.49]

Checkpoint

Supernet training checkpoint is in train-supernet folder with named as 'model_actwbin_best_466000it.pth'.

Citation

@inproceedings{basq,
  title={BASQ: Branch-wise Activation-clipping Search Quantization for Sub-4-bit Neural Networks},
  author={Han-Byul Kim and Eunhyeok Park and Sungjoo Yoo},
  year={2022},
  booktitle={ECCV},
}