US 11,928,582 B1
System, media, and method for deep learning
Piyush Pathak, Fremont, CA (US); Haoyu Yang, Ma Liu Shui (HK); Frank E. Gennari, Campbell, CA (US); and Ya-Chieh Lai, Mountain View, CA (US)
Assigned to Cadence Design Systems, Inc., San Jose, CA (US)
Filed by Cadence Design Systems, Inc., San Jose, CA (US)
Filed on Dec. 31, 2018, as Appl. No. 16/237,524.
Int. Cl. G06N 3/08 (2023.01); G06F 30/392 (2020.01); G06N 3/042 (2023.01)
CPC G06N 3/08 (2013.01) [G06F 30/392 (2020.01); G06N 3/042 (2023.01)] 20 Claims
OG exemplary drawing
 
1. A method for losslessly encoding portions of circuit layouts for machine learning using a tensor processing machine learning module, comprising:
generating a plurality of tensors from a circuit layout, the plurality of tensors having one or more sizes and losslessly encoding a plurality of portions of the circuit layout, a tensor comprising at least a bitmap having a plurality of rows and column, and a width for each row and each column;
determining that at least some of the plurality of tensors generated from the circuit layout have dimensions less than one or more target dimensions, the target dimensions comprising one or more corresponding dimensions of tensors stored in a training database, wherein the tensors stored in the training database were used to train a machine learning model;
modifying, in response to said determination, at least some of the plurality of tensors generated from the circuit layout having dimensions less than one or more target dimensions at least by scaling the at least some of the plurality of tensors, wherein scaling comprises at least splitting an existing row or column in a corresponding tensor into two rows or columns and where a width of the existing row or column is equal to a sum of a width of the two rows or columns; and
processing the plurality of tensors using the trained machine learning model.