CPC G06F 16/2272 (2019.01) [G06F 16/2455 (2019.01)] | 32 Claims |
1. A method of a hardware accelerator, comprising:
loading a lookup table; and
generating an output result of at least one layer of a neural network model by:
generating input data values by respectively normalizing output values, from among the at least one layer of the neural network model, corresponding to input data being input to the at least one layer, resulting in all generated input data values having positive values, and
generating output data values with respect to the output result by mapping each of the input data values to respective indexes of the lookup table based on differences between a maximum input data value of the input data values and each of the input data values, including mapping a sum of the output data values to indexes in a compensation coefficient lookup table, including
obtaining compensation coefficients corresponding to the input data values using the compensation coefficient lookup table, and
obtaining normalized output data values corresponding to the input data values based on the compensation coefficient,
wherein the output data values are proportional to corresponding softmax values of the input data values.
|