Neural network model apparatus and compressing method of neural network model

1. An electronic apparatus comprising:

a memory storing computer-readable instructions; and

a processor which when executing the computer-readable instructions is configured to:

obtain an original neural network model;

compress the original neural network model according to a compression algorithm, to obtain a compressed neural network model;

generate training data pairs in the form of input-output data for training the compressed neural network model, based on the original neural network model; and

train the compressed neural network model using the training data pairs, to obtain a compressed and trained neural network model,

wherein the processor when executing the computer-readable instructions to generate the training data pairs is further configured to:

acquire an objective function based on a specific output value of the original neural network model;

acquire an input value from the original neural network model based on the objective function; and

acquire the training data pairs comprising the input value and the specific output value.

2. The electronic apparatus of claim 1 , wherein the processor when executing the computer-readable instructions to generate the training data pairs is further configured to:

acquire an output value of the original neural network model based on the input value;

acquire an output value of the objective function by applying the output value of the original neural network model to the objective function; and

acquire the input value based on whether the output value of the objective function satisfies a predetermined condition.

3. The electronic apparatus of claim 2 , wherein the processor when executing the computer-readable instructions to generate the training data pairs is further configured to, based on the output value of the objective function satisfying the predetermined condition, acquire the input value and the output value from the original neural network model as the training data pairs.

4. The electronic apparatus of claim 2 , wherein the processor when executing the computer-readable instructions to generate the training data pairs is further configured to:

based on the output value of the objective function not satisfying the predetermined condition, change the input value to a changed input value;

acquire the output value of the original neural network model;

acquire the output value of the objective function by applying the output value of the original neural network model according to the changed input value to the objective function; and

based on the output value of the objective function satisfying the predetermined condition, acquire the changed input value and the output value acquired from the original neural network model based on the changed input value as the training data pairs.

5. The electronic apparatus of claim 2 , wherein the predetermined condition comprises a condition of the output value of the objective function being a minimum value or a maximum value.

6. The electronic apparatus of claim 2 , wherein the processor when executing the computer-readable instructions to generate the training data pairs is further configured to, based on whether the output value of the objective function based on each of a plurality of input values satisfies the predetermined condition, acquire input values corresponding to each of the plurality of input values.

7. The electronic apparatus of claim 1 , wherein the processor when executing the computer-readable instructions is further configured to:

acquire a plurality of objective functions based on a plurality of specific output values of the original neural network model; and

acquire an input value from the original neural network model based on each of the plurality of objective functions.

8. The electronic apparatus of claim 1 , claim 2 , wherein the training data pairs comprise a first input value and a first output value of the original neural network model acquired based on the first input value, and

wherein the processor when executing the computer-readable instructions to train the compressed neural network model is further configured to:

acquire a second input value based on the first input value;

acquire the second input value and different training data pairs comprising a second output value of the original neural network model acquired based on the second input value; and

train the compressed neural network model based on the training data pairs and the different training data pairs.

9. The electronic apparatus of claim 1 , further comprising:

a communication interface,

wherein the processor when executing the computer-readable instructions is further configured to, based on a request being received from an external apparatus through the communication interface, transmit the compressed and trained neural network model to the external apparatus through the communication interface.

10. A method of compressing a neural network model by an electronic device, the method comprising:

generating a compressed neural network model by compressing an original neural network model using a compression algorithm;

generating training data pairs in the form of input-output data based on the original neural network model; and

generating a compressed and trained neural network model by training the compressed neural network model based on the training data pairs,

wherein the generating the training data pairs comprises:

acquiring an objective function based on a specific output value of the original neural network model;

acquiring an input value from the original neural network model based on the objective function; and

acquiring the training data pairs comprising the input value and the specific output value.

11. The method of claim 10 , wherein the generating the training data pairs comprises:

acquiring an output value of the original neural network model based on the input value;

acquiring an output value of the objective function by applying the output value of the original neural network model to the objective function; and

acquiring the input value based on whether the output value of the objective function satisfies a predetermined condition.

12. The method of claim 11 , wherein the generating the training data pairs comprise:

based on the output value of the objective function satisfying the predetermined condition, acquiring the input value and the output value from the original neural network model as the training data pairs.

13. The method of claim 11 , wherein the generating the training data pairs comprise:

based on the output value of the objective function not satisfying the predetermined condition, changing the input value to a changed input value;

acquiring the output value of the original neural network model based on the changed input value;

applying the output value of the original neural network model according to the changed input value to the objective function to acquire the output value of the objective function; and

based on the output value of the objective function satisfying the predetermined condition, acquiring the training data pairs comprising the changed input value.

14. The method of claim 11 , wherein the generating the training data pairs comprises, based on whether the output value of the objective function based on each of a plurality of input values satisfies the predetermined condition, acquiring input values corresponding to each of the plurality of input values and acquiring the training data pairs comprising the input values.

15. The method of claim 10 , wherein the generating the training data pairs comprises:

acquiring a plurality of objective functions based on a plurality of specific output values of the original neural network model; and

acquiring an input value from the original neural network model based on each of the plurality of objective functions.

16. The method of claim 10 , wherein the training data pairs comprise a first input value and a first output value of the original neural network model acquired based on the first input value, and

wherein the method further comprises:

acquiring a second input value based on the first input value;

acquiring the second input value and a different training data pairs comprising a second output value of the original neural network model acquired based on the second input value; and

training the compressed neural network model based on the training data pairs and the different training data pairs.

17. The method of claim 10 further comprising, based on receiving a request from an external apparatus, transmitting the compressed and trained neural network model to the external apparatus.

18. A server apparatus comprising:

a memory storing computer-readable instructions;

a communicator; and

a processor which when executing the computer-readable instructions is configured to:

compress an original neural network model according to a compression algorithm, to acquire a compressed neural network model;

generate training data pairs in the form of input-output data for training the compressed neural network model based on the original neural network model;

train the compressed neural network model based on the training data pairs, to obtain a compressed and trained neural network model; and

output the compressed and trained neural network model through the communicator,

wherein the processor when executing the computer-readable instructions to generate the training data pairs is further configured to:

acquire an objective function based on a specific output value of the original neural network model;

acquire an input value from the original neural network model based on the objective function; and

acquire the training data pairs comprising the input value and the specific output value.