Deep Model Compression via Two-Stage Deep Reinforcement Learning

Besides accuracy, the model size of convolutional neural networks (CNN) models is another important factor considering limited hardware resources in practical applications. For example, employing deep on mobile systems requires design accurate yet fast CNN for low latency classification and object detection. To fulfill need, we aim at obtaining with both high testing accuracy small to address resource constraints many embedded devices. In particular, this paper focuses proposing a generic reinforcement learning-based compression approach two-stage pipeline: pruning quantization. The first stage compression, i.e., pruning, achieved via exploiting learning (DRL) co-learn FLOPs updated after layer-wise channel element-wise variational information dropout. second stage, quantization, similar DRL but optimal bits representation individual layers. We further conduct experimental results CIFAR-10 ImageNet datasets. dataset, proposed method can reduce VGGNet by \(9\times \) from 20.04 MB 2.2 slight increase. VGG-16 \(33\times 138 4.14 no loss.