PalQuant

This is the offical implementation of "PalQuant: Accelerating High-precision Networks on Low-precision Accelerators" in ECCV 2022

Installation

git clone git@github.com:huqinghao/PalQuant.git
cd PalQuant
pip install -r requirements.txt

Training

To train resnet-18 with 4-bit weights and activations, PalQuant trains a wide 2-bit resnet-18 with group=2:

python main.py --data your-imagenet-data-path --visible_gpus '0,1,2,3' --workers 20  \
--arch 'resnet18_quant' --epochs 90 --groups 2 --weight_levels 4 --lr_m 0.1 --lr_q 0.0001 \
-b 256  --act_levels 4 --log_dir "../results/resnet-18/W2A2G2/"

Here weight_levels and activations levels equals $2^{bit}$

Testing

We have uploaded the training checkpoint to the BaiduNetdisk and Google Storge. To test the pre-trained model, run:

python main.py --data your-imagenet-data-path --visible_gpus '0' --workers 20 \
--arch 'resnet18_quant'  --groups 2  --weight_levels 4 -b 256  --act_levels 4 \
--evaluate  --model the-model-to-eval

Model	Weight Bits	Act Bits	Groups	DownloadUrl	TensorboardLog
ResNet18	2	2	2	BaiduNetDisk	log
ResNet18	2	2	3	BaiduNetDisk	log
ResNet18	2	2	4	BaiduNetDisk	log

extract code: quan

Citation

@inproceedings{PalQuant2021,
  author  = {Qinghao Hu and Gang Li and Qiman Wu and Jian Cheng},
  title   = {PalQuant: Accelerating High-precision Networks on Low-precision Accelerators},
  year    = {2022},
  booktile={European Conference on Computer Vision},
  organization={Springer}
}

Acknowledgement

The code base is origined from EWGS, we thank their awesome work.