BitSplit

BitSplit Post-trining Quantization

Code for papers:

• 'Towards Accurate Post-training Network Quantization via Bit-Split and Stitching', ICML 2020

Bit-split is a novel post-training network quantization framework where no finetuning is needed.

A8W4 model for ResNet-18:

BaiduCloud Extraction Code: bsci

GoogleDrive

Files:

• main_quant_resnet18_twostep: slow version, separately extract features for different layers.
• main_quant_resnet18_twostep_fast: fast version, extract features for all layers in parallel (should be accomplished in around 15 min).

Train:

CUDA_VISIBLE_DEVICES=0 python main_quant_resnet18_twostep_fast.py -a resnet18_quan --pretrained ~/data/cnn_models/pytorch/resnet/resnet18-5c106cde.pth --act-bit-width 8 --weight-bit-width 4

Test:

CUDA_VISIBLE_DEVICES=0 python main_quant_resnet18_twostep_fast.py -a resnet18_quan --pretrained ./resnet18_quan/A8W4/state_dict.pth --scales resnet18_quan/A8W4/act_8_scales.npy --act-bit-width 8 --weight-bit-width 4 --evaluate 

Results:

\* Acc@1 69.146 Acc@5 88.670

Derivation of Eq.(5):

Related Papers

Please cite our paper if it helps your research:

@InProceedings{Wang_2020_ICML,
    author = {Wang, Peisong, Qiang Chen, Xiangyu He, and Cheng, Jian},
    title = {Towards Accurate Post-training Network Quantization via Bit-Split and Stitching},
    booktitle = {Proceedings of the 37nd International Conference on Machine Learning (ICML)},
    month = {July},
    pages = {243--252},
    year = {2020}
}