Code for papers:

• 'Fixed-Point Factorized Networks', CVPR 2017
• 'Unsupervised Network Quantization via Fixed-point Factorization', TNNLS

Fixed-point Factorized Network (FFN) is a novel network ternarization approach, i.e., it turns all weights into ternary values {-1, 0, 1}. FFN works well in both training-aware and post-training quantization schemes. It can achieve negligible degradation even without any supervised finetuning on labeled data.

Factorized VGG16 model: BaiduCloud

python main_ffn_vgg.py --pretrained <path-to-pretrained-model>

python main_ffn_vgg_test.py --pretrained vgg16_bn/vgg16_ternary_final.pth

* Acc@1 70.810 Acc@5 90.050

Please cite our paper if it helps your research:

@InProceedings{Wang_2017_CVPR,
  author = {Wang, Peisong and Cheng, Jian},
  title = {Fixed-Point Factorized Networks},
  booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {July},
  year = {2017}
}