首页>
外国专利>
In order to quantize CNN parameters and feature values used for optimizing hardware applicable to mobile devices or high-precision small networks, FL (Fractional Length) values using weight quantization loss values are used. how to determine and apparatus {mETHOD aND dEVICE fOR dETERMINING FL VALUE BY USING WEIGHTED QUANTIZATION LOSS VALUES tO THEREBY QUANTIZE CNN PARAMETERS aND FEATURE VALUES tO bE USED fOR OPTIMIZING HARDWARE APPLICABLE tO MOBILE dEVICES oR COMPACT NETWORKS wITH HIGH PRECISION}
In order to quantize CNN parameters and feature values used for optimizing hardware applicable to mobile devices or high-precision small networks, FL (Fractional Length) values using weight quantization loss values are used. how to determine and apparatus {mETHOD aND dEVICE fOR dETERMINING FL VALUE BY USING WEIGHTED QUANTIZATION LOSS VALUES tO THEREBY QUANTIZE CNN PARAMETERS aND FEATURE VALUES tO bE USED fOR OPTIMIZING HARDWARE APPLICABLE tO MOBILE dEVICES oR COMPACT NETWORKS wITH HIGH PRECISION}
PROBLEM TO BE SOLVED: To optimize hardware applicable to a mobile device or a high-precision small network. SOLUTION: In the FL parameter determination corresponding to the size of the LSB included in the quantized value and corresponding to the size of the FL value, an original vector including the original value is acquired, and the BW value (quantized value A quantization operation is applied to the original value with reference to each of the BW parameter corresponding to the number of bits used to indicate each) and the FL candidate value, and the quantized vector (corresponding to each FL candidate value, (Including quantized values) and for each difference between the quantized value contained in each quantized vector and the original value contained in the original vector, The weight quantization loss calculation is applied to generate the weight quantization loss value corresponding to each FL candidate value, and the FL value is determined from the FL candidate values with reference to the weight quantization loss value. [Selection diagram] None
展开▼