| Title |
Implementation of a Hardware-friendly Quantized CNN using Integer Arithmetic Only |
| Authors |
김재명(Jaemyung Kim) ; 김용우(Yongwoo Kim) |
| DOI |
https://doi.org/10.5573/ieie.2020.57.12.60 |
| Keywords |
Deep learning; Image classification; Quantization; Integer arithmetic; Hardware-friendly |
| Abstract |
Recently, in the field of computer vision, research using CNN has established itself as a universal research method as it shows excellent performance. However, to obtain high performance, a CNN model with a large number of parameters and high computational complexity is required. This model is not suitable in an environment where hardware resource usage is limited. Therefore, while maintaining the structure of the CNN, quantization research that can optimize computational complexity and memory usage by lowering the precision is being actively conducted. However, most quantization techniques are not hardware-friendly methods were used, such as using a floating-point scale factor to maintain performance. In this paper, quantization aware training(QAT) was performed by applying a hardware-friendly quantization technique and floating-point parameters learned by QAT were converted into integer parameters. In addition, we proposed a method for constructing a CNN that combines basic operation blocks using integer arithmetic only operation and a method for verifying the precision of each layer. As a result of evaluating the performance of the proposed CNN using 8-bit quantized parameters with the CIFAR-10 dataset, it was confirmed that the number of parameters compared to floating-point was reduced to 1/4, but the accuracy drop was less than 0.71%. Besides, it was confirmed that the computational complexity could be reduced by using only an integer operator instead of an additional floating-point operator, and the accuracy drop was less than 0.13% when compared with the conventional quantization aware inference method. |