Abstract
Deep neural networks (DNNs) are nowadays ubiquitous in the computer vision landscape. However, they suffer from high computational costs in inference, particularly when evaluated on edge devices. This problem is generally addressed via post-hoc quantization, i.e. converting the DNN values (weights and inputs) from floating point into e.g. int8, int4 or ternary quantization. In this paper, we propose REx, a data-free quantization algorithm for pre-trained models that is compliant with data protection regulations, convenient and fast to execute. First, we improve upon the naive linear quantization operator by decomposing the weights as an expansion of residual quantization errors. Second, we propose a budgeted group-sparsity formulation to achieve better accuracy vs. number of bit-wise operation trade-offs with sparse, higher expansion orders. Third, we show that this sparse expansion can be approximated by an ensemble of quantized neural networks to dramatically improve the evaluation speed through more efficient parallelization. We provide theoretical guarantees of the efficiency of REx as well as a thorough empirical validation on several popular DNN architectures applied to multiple computer vision problems, e.g. ImageNet classification, object detection as well as semantic segmentation. In particular, we show that REx significantly outperforms existing state-of-the-art data-free quantization techniques.
Abstract (translated)
URL
https://arxiv.org/abs/2203.14645
