| CPC G06N 3/08 (2013.01) [G06F 9/50 (2013.01)] | 14 Claims |
|
1. A processor-implemented method, comprising:
calculating required resource information of each of a plurality of layers of a neural network model;
selecting, a layer from among the plurality of layers as a target layer, based on the required resource information and hardware information of a processing device including one or more processors;
generating reconfigured layers including first and second reconfigured layers, of the target layer, based on a portion of the calculated required resource information corresponding to the target layer and the hardware information, the first reconfigured layer is generated based on different required resource information than the second reconfigured layer;
selecting one layer from among the target layer, the first reconfigured layer, and the second reconfigured layer, based on performance information of the target layer and performance information of each of the first and second reconfigured layers; and
implementing the selected one layer using the processing device,
wherein the generating of the reconfigured layers comprises:
generating the first reconfigured layer by applying first multiple paths to the target layer based on required operation information related to the target layer and processing core information in the hardware information; and
generating the second reconfigured layer by applying second multiple paths to the target layer based on required capacity information related to the target layer and memory information in the hardware information, and
wherein a total number of the first multiple paths is determined based on the required operation information and a number of processing cores and a total number of the second multiple paths is determined based on the required capacity information and a local memory size.
|