Poster: Fast GPU Inference with Unstructurally pruned DNNs for Explainable DOC

We have developed a code compiler to compress unstructurally pruned DNN models and demonstrated inference time less than 1 msec with AUC accuracy over 90 % for an anomaly detection task using MVTec AD dataset and edge Graphics Processing Unit (GPU) devices. Reduced RepVGG convolutional neural network (CNN) architecture is applied to an explainable deep one-class classification (XDOC) algorithm and such fast inference is obtained without sacrificing the accuracy by using a training scheme, CutPaste, to keep the accuracy high under an extremely higher pruning rate condition.