- TL;DR: We train accurate fully quantized networks using a loss function maximizing full precision model accuracy and minimizing the difference between the full precision and quantized networks.
- Abstract: Network quantization is a model compression and acceleration technique that has become essential to neural network deployment. Most quantization methods per- form fine-tuning on a pretrained network, but this sometimes results in a large loss in accuracy compared to the original network. We introduce a new technique to train quantization-friendly networks, which can be directly converted to an accurate quantized network without the need for additional fine-tuning. Our technique allows quantizing the weights and activations of all network layers down to 4 bits, achieving high efficiency and facilitating deployment in practical settings. Com- pared to other fully quantized networks operating at 4 bits, we show substantial improvements in accuracy, for example 66.68% top-1 accuracy on ImageNet using ResNet-18, compared to the previous state-of-the-art accuracy of 61.52% Louizos et al. (2019) and a full precision reference accuracy of 69.76%. We performed a thorough set of experiments to test the efficacy of our method and also conducted ablation studies on different aspects of the method and techniques to improve training stability and accuracy. Our codebase and trained models are available on GitHub.
- Code: https://gofile.io/?c=GEvKoF
- Keywords: Network quantization, Efficient deep learning