Compression of Deep Convolutional Neural Network Using Additional Importance-Weight-Based Filter Pruning Approach

The success of the convolutional neural network (CNN) comes with a tremendous growth diverse CNN structures, making it hard to deploy on limited-resource platforms. These over-sized models contain large amount filters in layers, which are responsible for almost 99% computation. key question here arises: Do we really need all those filters? By removing entire filters, computational cost can be significantly reduced. Hence, this article, filter pruning method, process discarding subset unimportant or weak from original model, is proposed, alleviates shortcomings architectures at storage space and time. proposed strategy adopted compress model by assigning additional importance weights filters. help each learn its responsibility contribute more efficiently. We different initialization strategies about aspects prune accordingly. Furthermore, unlike existing approaches, method uses predefined error tolerance level instead rate. Extensive experiments two widely used image segmentation datasets: Inria AIRS, known segmentation: TernausNet standard U-Net, verify that our approach efficiently negligible no loss accuracy. For instance, could reduce 85% floating point operations (FLOPs) drop 0.32% validation This compressed six-times smaller seven-times faster (on cluster GPUs) than TernausNet, while accuracy less 1%. Moreover, reduced FLOPs 84.34% without deteriorating output performance AIRS dataset TernausNet. effectively number parameters retaining compact deployed any embedded device specialized hardware. show pruned very similar unpruned model. also report numerous ablation studies validate approach.