Abstract: Disclosed is a method for enhancing memory utilization and throughput of a computing platform in training a deep neural network (DNN). The critical features of the method includes: calculating a memory size for every operation in a computational graph, storing the operations in the computational graph in multiple groups with the operations in each group being executable in parallel and a total memory size less than a memory threshold of a computational device, sequentially selecting a group and updating a prefetched group buffer, and simultaneously executing the group and prefetching data for a group in the prefetched group buffer to the corresponding computational device when the prefetched group buffer is update. Because of group execution and data prefetch, the memory utilization is optimized and the throughput is significantly increased to eliminate issues of out-of-memory and thrashing.

Abstract: Disclosed is a method for training a graphics processing neural network with a patch-based approach, which involves calculating an overlapping size and an invalid size of an output of each of at least one of multiple feature extraction layers of the graphic processing neural network according to a predetermined cropping scheme, dividing an input image into first patches in the forward pass and the first gradients into second patches in the backward pass to run streamline operation of the first and second patches. Before training, each first patch overlaps neighboring first patches at adjacent edges. In the forward pass, an invalid portion of the output of each of the at least one of the feature extraction layers cropped out based on the predetermined cropping scheme and a corresponding invalid size. Such method secures streamline operation in favor of enhanced memory utilization in training and accurate model prediction.

Abstract: Disclosed is a method for enhancing memory utilization and throughput of a computing platform in training a deep neural network (DNN). The critical features of the method includes: calculating a memory size for every operation in a computational graph, storing the operations in the computational graph in multiple groups with the operations in each group being executable in parallel and a total memory size less than a memory threshold of a computational device, sequentially selecting a group and updating a prefetched group buffer, and simultaneously executing the group and prefetching data for a group in the prefetched group buffer to the corresponding computational device when the prefetched group buffer is update. Because of group execution and data prefetch, the memory utilization is optimized and the throughput is significantly increased to eliminate issues of out-of-memory and thrashing.

Abstract: Disclosed are an object detection method and a convolution neural network. The method is performed through hierarchical architecture of the CNN and includes extracting groups of augmented feature maps from an input image through a backbone and two other groups of feature maps, identifying positive and negative samples with an IOU-based sampling scheme to be proposals for foreground and background through a proposal-sampling classifier, mapping the proposals to regions on the groups of augmented feature maps through the region proposal module, pooling the regions to fixed scale feature maps based on ROI aligning, fusing the fixed scale feature maps, and flattening the fused feature maps to generate an ROI feature vector through an ROI aligner for object classification and box regression. Because extracted features in the groups of augmented feature maps range from spatially-rich features to semantically-rich features, enhanced performance in object classification and box regression can be secured.