| With the development of deep learning,the network structure is becoming more and more complex.The optimization of super parameters and network structure parameters of deep neural network model has become a hot spot.Neural architecture search(NAS)makes it possible to automatically create neural network architecture.Learning the high-performance model structure in this systematic and automatic way increases the odds for us to explore the whole network structure space and find the best model.In recent years,many effective artificial models have emerged in the task of speech music separation.However,these models are based on intensive human labor,which requires constant trial and error to improve the performance of the model.In order to break through the limitations of human thinking and find a better architecture organization,this paper studies the search algorithm of network architecture for voice music separation task.The following research work has been completed on the task of music source separation:(1)Considering the importance of network benchmark for architecture search,this work improves UMX.A new model m-umx is proposed.It greatly solves the problem of poor performance of LSTM in the face of irrelevant context.On the musdb187 s dataset,UMX and m-umx use the same super parameters to train the network model.The experimental results show that the m-umx model has better separation and generalization ability,and effectively improves the performance of music source separation.(2)This paper adopts the differential search structure(darts)algorithm to consider the search efficiency of the network.Firstly,the stacking mode of the outermost network is predefined,and a unit level search space is constructed.This paper uses the search algorithm and optimization strategy of darts.After the structure is obtained by convergence,the searched structure is trained and evaluated to obtain the final separation result. |
