Performance Analysis of Network Pruning for Deep Learning based Age-Gender Estimation
Autumn Knight, Byeong Kil Lee · 2020
With the advance of visual AI technology, age-gender estimation plays a fundamental role in identifying individuals. As deep learning technologies are emerging, identification schemes show significant progress and can handle many challenges of unconstrained imaging conditions. Research on age-gender estimation has begun applying deep convolutional neural networks (CNN) as a framework. However, due to large memory footprints and computational workloads, deep neural networks are hard to apply to on-device training and inference for embedded devices which have limited hardware resources. To solve this issue, network model pruning has been proposed as an efficient approach to reduce the model redundancy without significant degradation of the performance. In this paper, we modeled and characterized several pre-training models with architecture variations on baseline age-gender estimation before applying pruning schemes. For each of the models, three types of pruning comprised of weight, layer and filter pruning are applied and the pruning results are analyzed in terms of complexity and accuracy to find optimal pruning conditions. Combined schemes of pre-training models and network pruning techniques are discussed, and their results are compared with the original model's. Based on our experiments, the actual size of a fully trained prediction model can be reduced by as much as 90% with an accuracy loss of 2%~9%.