Method and apparatus for maintaining statistical inference accuracy with 8-bit wobbling convolution

摘要

Methods and apparatus maintain statistical inference accuracy in an 8-bit wobbled convolution.A pre trained CNN including a calibration data set and a 32-bit floating point value may be sampled to generate an input activation tensor and a weight tensor.The converted input activation tensor can be generated by multiplying the input activation tensor and the input matrix to generate the transformed input activation tensor.By multiplying the weight tensor and the weight matrix, the converted weight tensor can be generated.The scale factor may be calculated for each converted tensor.An 8-bit CNN model including scale factors may be generated.