首页> 外国专利> CNN 3x3 1x3 3x1 CNN LOW OVERHEAD IMPLEMENTATION OF WINOGRAD FOR CNN WITH 3x3 1x3 AND 3x1 FILTERS ON WEIGHT STATION DOT-PRODUCT BASED CNN ACCELERATORS

CNN 3x3 1x3 3x1 CNN LOW OVERHEAD IMPLEMENTATION OF WINOGRAD FOR CNN WITH 3x3 1x3 AND 3x1 FILTERS ON WEIGHT STATION DOT-PRODUCT BASED CNN ACCELERATORS

机译:CNN 3X3 1X3 3X1 CNN用于CNN的WinoGrad的WinoGrad的低开销实施,3x3 1x3和3x1滤波器在基于CNN加速器的CNN加速器上

摘要

A system and method for forming an output feature map (OFM) is disclosed. Activation values of the input feature map (IFM) are selected and transformed on-the-fly into the Winograd domain. Elements of the Winograd filter each corresponding to the transformed activation values are selected. The transformed activation value is multiplied by the corresponding element of the Winograd filter to form the corresponding product value in the Winograd domain. Activation values are iteratively selected, transformed and multiplied by the corresponding element of the Winograd filter to form the corresponding product value in the Winograd domain until all activation values of the IFM are transformed and multiplied by the corresponding element. The product values are summed in the Winograd domain to form an element of the feature map in the Winograd domain. Elements of the feature map in the Winograd domain are transformed on the fly inverse-winograd to form an OFM.
机译:公开了一种用于形成输出特征图(OFM)的系统和方法。 输入功能映射(IFM)的激活值被选中并在-WOR-Flow中转换为WinoGrad域。 选择WinoGrad滤波器的元素,每个滤波器对应于转换的激活值。 转换的激活值乘以WinoGrad滤波器的相应元素,以在WinoGrad域中形成相应的产品值。 激活值迭代地选择,转换和乘以WinoGrad滤波器的相应元素,以在WinoGrad域中形成相应的产品值,直到IFM的所有激活值被转换并乘以相应的元素。 产品值在WinoGRAD域中求和,以在WinoGRAD域中形成要素映射的元素。 WinoGRAD域中的特征映射的元素在Fly Inverse-Winograd上转换为ofm。

著录项

相似文献

  • 专利
  • 外文文献
  • 中文文献
获取专利

客服邮箱:kefu@zhangqiaokeyan.com

京公网安备:11010802029741号 ICP备案号:京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有
  • 客服微信

  • 服务号