System and method for automating design of sound source separation deep learning model

1 . A method of automating a design of a sound source separation deep learning model, which is performed by a design automation system, the method comprising:

automatically searching for a combination of hyper parameters of a separation model constructed in a sound source separation deep learning model by using a neural architecture search (NAS) algorithm; and

reconstructing the sound source separation deep learning model based on the retrieved, combination of the hyper parameters of the separation model,

wherein searching for the combination comprises forming a search space of the separation model based on a kernel size P and channel expansion ratio H of the separation model,

wherein searching for the combination comprises:

constructing one layer by using candidate operations in the formed search space of the separation model and an architecture parameter that assigns a weight to each of the candidate operations;

learning a weight and the architecture parameter of the candidate operation in a model architecture generated by using the constructed one layer, by using a training set and a validation set, respectively; and

obtaining a model architecture by selecting only one candidate operation in which results of the architecture parameter are a highest as the training of the generated model architecture is completed by using the constructed one layer, and

wherein searching for the combination comprises:

extracting a feature vector for each repeat constructed in the separation model through an auxiliary loss operation;

calculating a loss by the number of plurality of repeats constructed in the separation model by calculating a loss for the extracted feature vector; and

calculating a final loss by calculating an average of results of the calculated loss.

2 . The method of claim 1 , wherein:

the sound source separation deep learning model is a fully convolutional time domain voice separation network (Conv-TasNet),

the separation model constructed in the sound source separation deep learning model comprises a plurality of repeats,

each of the plurality of repeats comprises a plurality of TCN blocks, and

the plurality of TCN blocks comprises a layer (Bottleneck) that adjusts the number of channels of an input on which convolution is to be performed by a D-CNN and a layer (depth-wise convolutional network (CNN)) that performs a convolution operation.

3 . The method of claim 1 , wherein searching for the combination comprises performing NAS by adjusting the number of layers for each repeat based on a search space of the formed separation model.

4 . The method of claim 1 , wherein searching for the combination comprises obtaining a model architecture through training using any one parameter, among the weight parameter or the architecture parameter of the candidate operation, by using a binary gate algorithm that forces only one path to be recorded on a GPU memory.

5 . A non-transitory computer-readable recording medium storing instructions that, when executed by a processor, cause the processor to perform the method of claim 1 .

6 . A design automation system comprising:

a hyper parameter combination search unit configured to automatically search for a combination of hyper parameters of a separation model constructed in a sound source separation deep learning model by using a neural architecture search (NAS) algorithm; and

a model reconstruction unit configured to reconstruct the sound source separation deep learning model based on the retrieved combination of the hyper parameters of the separation model,

wherein the hyper parameter combination search unit forms a search space of the separation model based on a kernel size P and channel expansion ratio H of the separation model,

wherein the hyper parameter combination search unit

constructs one layer by using candidate operations in the formed search space of the separation model and an architecture parameter that assigns a weight to each of the candidate operations,

learns a weight and the architecture parameter of the candidate operation in a model architecture generated by using the constructed one layer, by using a training set and a validation set, respectively, and

obtains a model architecture by selecting only one candidate operation in which results of the architecture parameter are a highest as the training of the generated model architecture is completed by using the constructed one layer,

wherein the hyper parameter combination search unit

constructs one layer by using candidate operations in the formed search space of the separation model and an architecture parameter that assigns a weight to each of the candidate operations,

learns a weight and the architecture parameter of the candidate operation in a model architecture generated by using the constructed one layer, by using a training set and a validation set, respectively, and

obtains a model architecture by selecting only one candidate operation in which results of the architecture parameter are a highest as the training of the generated model architecture is completed by using the constructed one layer, and

wherein the hyper parameter combination search unit

extracts a feature vector for each repeat constructed in the separation model through an auxiliary loss operation,

calculates a loss by the number of plurality of repeats constructed in the separation model by calculating a loss for the extracted feature vector, and

calculates a final loss by calculating an average of results of the calculated loss.

7 . The design automation system of claim 6 , wherein:

the sound source separation deep learning model is a fully convolutional time domain voice separation network (Conv-TasNet),

the separation model constructed in the sound source separation deep learning model comprises a plurality of repeats,

each of the plurality of repeats comprises a plurality of TCN blocks, and

the plurality of TCN blocks comprises a layer (Bottleneck) that adjusts the number of channels of an input on which convolution is to be performed by a D-CNN and a layer (depth-wise convolutional network (CNN)) that performs a convolution operation.

8 . The design automation system of claim 6 , wherein the hyper parameter combination search unit performs NAS by adjusting the number of layers for each repeat based on a search space of the formed separation model.

9 . The design automation system of claim 6 , wherein the hyper parameter combination search unit obtains a model architecture through training using any one parameter, among the weight parameter or the architecture parameter of the candidate operation, by using a binary gate algorithm that forces only one path to be recorded on a GPU memory.