A method and data processing system for resampling a set of samples

• Main Authors: Aria Ahmadi, Cagatay Dikici
• Format: Patent
• Language: English
• Published: 09.10.2024
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; IMAGE DATA PROCESSING OR GENERATION, IN GENERAL; PHYSICS

Resampling, using neural network accelerator hardware, comprising: receiving first samples arranged in a tensor that extends in a first dimension; determining a resampling factor a1/b1 and offset d1 (resampling parameters) for the first dimension, wherein a1 and b1 are integers greater than zero and either a1 and b1 are not equal or d1 is non-zero; convolving the first samples with a number of kernels to produce output tensors wherein the convolutions traverse the first dimension with a stride that is greater than 1; performing a depth-to-space operation to arrange the output tensor values to produce second samples that are offset relative to the first samples in the first dimension by d1. First samples may be padded by inserting rows or columns. The depth to space operation may involve flattening to reduce the dimensionality of the output tensors, eg. from 3D to 2D. The number of kernels may be based on the number of first samples such that one kernel interpolates a value at one sampling position. Sampling location coordinates of first samples may be periodic such that kernels traverse the first dimension with a stride that matches the periodicity, ensuring that each sampling point is only convolved with one kernel.