Lattice Quantization

Abstract

Low bit quantization of weights in increasingly large deep convolutional neural networks (DCNNs) can be critical for their implementation in memory constrained hardware systems. Post-training quantization consists in quantizing a model without retraining, which is user-friendly, fast and data frugal. In this paper, we propose LatticeQ, a new post-training weight quantization method designed for DCNNs. Instead of the standard scalar rounding widely used in state-of-the-art quantization methods, LatticeQ uses a quantizer based on lattices - discrete algebraic structures - which we show are able to exploit the inner correlations between the model parameters. LatticeQ allows us to achieve state-of-the-art results in post-training quantization, enabling us to approach full precision accuracies for bitwidths previously not accessible to post-training quantization methods. In particular, we achieve ImageNet classification results close to full precision on the popular Resnet-18/50, with only 0.5% and 5% accuracy drop for the 4-bit weights and 3-bit weights model architectures respectively.

Tasks

Reproductions