US 12,141,681 B2
Merge operations for darts
Xiaoxing Wang, Beijing (CN); Chao Xue, Beijing (CN); Yonggang Hu, Richmond Hill (CA); and Ke Wei Sun, Beijing (CN)
Assigned to International Business Machines Corporation, Armonk, NY (US)
Filed by International Business Machines Corporation, Armonk, NY (US)
Filed on Dec. 21, 2020, as Appl. No. 17/128,290.
Prior Publication US 2022/0198248 A1, Jun. 23, 2022
Int. Cl. G06N 3/063 (2023.01); G06F 16/2455 (2019.01); G06N 3/04 (2023.01)
CPC G06N 3/063 (2013.01) [G06F 16/2455 (2019.01); G06N 3/04 (2013.01)] 3 Claims
OG exemplary drawing
 
1. A computer-implemented method (CIM) comprising:
receiving a plurality of convolutions, with each given convolution of the plurality of convolutions being characterized by a same kernel size of n*n and including a kernel and a corresponding weight tensor that includes a plurality of depth-wise weights and a plurality of point-wise weights, and wherein the plurality of convolutions comprises normal convolution and separable convolution;
merging the plurality of convolutions into a single merged convolution the being characterized by a kernel size of m*m, where m is greater than n, with a kernel of the single merged convolution being calculated based on a weighted average of kernels for the plurality of convolutions, by utilizing weighting factors from the weight tensors of the plurality of convolutions so that depth-wise and point-wise weights are reused in the merging operation; and
creating a one-shot neural architecture search method based on the single merged convolution to reduce search cost.