DEEP NEURAL NETWORK-BASED METHOD AND DEVICE FOR QUANTIFYING ACTIVATION AMOUNT

摘要

A deep neural network-based method and device for quantifying an activation amount, the method comprising: acquiring an activation amount of a network layer in a deep neural network (S101), the elements in the activation amount being arranged in a height direction, a width direction and a depth direction; dividing, along the depth direction of the activation amount, depths, in which the difference between the element features in the activation amount is less than a preset threshold, into one same slice group, so as to obtain a plurality of slice groups (S102); and using the quantization parameter, which corresponds to each slice group and is obtained by a quantization formula, to quantize each slice group, respectively (S103). The quantization error can be reduced by the method.