Fixed Point Quantization of Deep Convolutional Networks

In recent years increasingly complex architectures for deep convolution networks (DCNs) have been proposed to boost the performance on image recognition tasks. However, the gains in performance have come at a cost of substantial increase in compute resources, the model size and processing speed of the network for training and evaluation. Fixed point implementation of these networks has the potential to alleviate some of the burden of these additional complexities. In this paper, we propose a quantizer design for fixed point implementation for DCNs. We then formulate an optimization problem to identify optimal fixed point bit-width allocation across DCN layers. We perform experiments on a recently proposed DCN architecture for CIFAR-10 benchmark that generates test error of less than 7%. We evaluate the effectiveness of our proposed fixed point bit-width allocation for this DCN. Our experiments show that in comparison to equal bit-width settings, the fixed point DCNs with optimized bit width allocation offer >20% reduction in the model size without any loss in performance. We also demonstrate that fine tuning can further enhance the accuracy of fixed point DCNs beyond that of the original floating point model. In doing so, we report a new state-of-the-art fixed point performance of 6.78% error-rate on CIFAR-10 benchmark.