Loss-aware-weight-quantization

Implementation of ICLR 2018 paper "Loss-aware Weight Quantization of Deep Networks", tested with GTX TITAN X, python 2.7, theano 0.9.0 and lasagne 0.2.dev1.

This repository is divided in two subrepositories:

• FNN: enables the reproduction of the FNN results(on MNIST, CIFAR-10, CIFAR-100, SVHN) reported in the article

• RNN: enables the reproduction of the RNN results(on War and Peace, Linux Kernel, PTB) reported in the article

Requirements This software is implemented on top of the implementation of BinaryConnect and has all the same requirements.

Example training command on War and Peace dataset:

• training using approximate ternarization method LATa

python warpeace.py --method="LATa" --lr_start=0.002  --len=100

• training using 3-bit linear quantization method LAQ_linear

python warpeace.py --method="LAQ_linear" --lr_start=0.002  --len=100

If you find loss-aware weight quantization useful in your research, please consider citing the the paper

@InProceedings{hou2017loss,
	title={Loss-aware Binarization of Deep Networks},
	author={Hou, Lu and Yao, Quanming and Kwok, James T.},
	booktitle={International Conference on Learning Representations},
	year={2017}
}

@InProceedings{hou2018loss,
	title={Loss-aware Weight Quantization of Deep Networks},
	author={Hou, Lu and Kwok, James T.},
	booktitle={International Conference on Learning Representations},
	year={2018}
}

@InProceedings{hou2019analysis,
	title={Analysis of Quantized Models},
	author={Hou, Lu and Zhang, Ruiliang and Kwok, James T.},
	booktitle={International Conference on Learning Representations},
	year={2019}
}