Method and apparatus for optimizing quantization model, electronic device, and computer storage medium

1. An optimization method applicable to an electronic device, the method comprising:

determining jump ratios of embedding layer parameters of a trained quantization model in a first predetermined time range, each jump ratio corresponding to a parameter jump of the embedding layer parameters of the trained quantization model in each predetermined time interval within the first predetermined time range, the quantization model being a neural network model obtained after quantization processing on the embedding layer parameters;

determining a jump curve in the first predetermined time range according to the jump ratios;

fitting the jump curve to obtain a corresponding time scaling parameter;

optimizing an initial algorithm of the quantization model based on the time scaling parameter to obtain an optimized target optimization algorithm;

training the quantization model based on the target optimization algorithm to obtain an optimized quantization model, wherein the time scaling parameter is configured to adjust a convergence speed and a precision of the quantization model;

acquiring user behavior data associated with a user at a terminal in a second predetermined time range;

learning the user behavior data using the optimized quantization model;

determining a user behavior feature corresponding to the user behavior data;

determining target recommendation information based on the user behavior feature; and;

providing the target recommendation information to the user at the terminal.

2. The method according to claim 1 , wherein the quantization processing comprises N-valued quantization processing on the embedding layer parameters, wherein N is an integer greater than one.

3. The method according to claim 1 , wherein:

a respective jump ratio of a parameter jump of the embedding layer parameters in any predetermined time interval is determined by:

determining a quantity of parameters that jump in the embedding layer parameters in the any predetermined time interval, relative to the embedding layer parameters in a previous predetermined time interval of the any predetermined time interval; and

determining the jump ratio of the parameter jump of the embedding layer parameters in the any predetermined time interval according to the quantity of parameters and a total quantity of embedding layer parameters.

4. The method according to claim 1 , wherein fitting the jump curve further comprises:

determining a curve type of the jump curve, and determining a fitting function corresponding to the curve type; and

fitting the jump curve according to the fitting function to obtain the corresponding time scaling parameter.

5. The method according to claim 1 , wherein optimizing the initial algorithm of the quantization model further comprises:

optimizing a learning rate parameter in the initial algorithm according to the time scaling parameter.

6. The method according to claim 5 , wherein optimizing the learning rate parameter in the initial algorithm further comprises:

updating the learning rate parameter in the initial algorithm to a product of the learning rate parameter and the time scaling parameter.

7. The method according to claim 1 , wherein training the quantization model further comprises:

increasing the time scaling parameter when the convergence speed of the quantization model is less than a predetermined speed threshold; and

decreasing the time scaling parameter when the precision of the quantization model is less than a predetermined precision threshold.

8. An electronic device, comprising:

one or more processors; and

memory storing one or more programs that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:

determining jump ratios of embedding layer parameters of a trained quantization model in a first predetermined time range, each jump ratio corresponding to a parameter jump of the embedding layer parameters of the trained quantization model in each predetermined time interval within the first predetermined time range, the quantization model being a neural network model obtained after quantization processing on the embedding layer parameters;

determining a jump curve in the first predetermined time range according to the jump ratios;

fitting the jump curve to obtain a corresponding time scaling parameter;

optimizing an initial algorithm of the quantization model based on the time scaling parameter to obtain an optimized target optimization algorithm;

training the quantization model based on the target optimization algorithm to obtain an optimized quantization model, wherein the time scaling parameter is configured to adjust a convergence speed and a precision of the quantization model;

acquiring user behavior data associated with a user at a terminal in a second predetermined time range;

learning the user behavior data using the optimized quantization model;

determining a user behavior feature corresponding to the user behavior data;

determining target recommendation information based on the user behavior feature; and;

providing the target recommendation information to the user at the terminal.

9. The electronic device according to claim 8 , wherein the quantization processing comprises N-valued quantization processing on the embedding layer parameters, wherein N is an integer greater than one.

10. The electronic device according to claim 8 , wherein:

a respective jump ratio of a parameter jump of the embedding layer parameters in any predetermined time interval is determined by:

determining a quantity of parameters that jump in the embedding layer parameters in the any predetermined time interval, relative to the embedding layer parameters in a previous predetermined time interval of the any predetermined time interval; and

determining the jump ratio of the parameter jump of the embedding layer parameters in the any predetermined time interval according to the quantity of parameters and a total quantity of embedding layer parameters.

11. The electronic device according to claim 8 , wherein fitting the jump curve further comprises:

determining a curve type of the jump curve, and determining a fitting function corresponding to the curve type; and

fitting the jump curve according to the fitting function to obtain the corresponding time scaling parameter.

12. The electronic device according to claim 8 , wherein optimizing the initial algorithm of the quantization model further comprises:

optimizing a learning rate parameter in the initial algorithm according to the time scaling parameter.

13. The electronic device according to claim 12 , wherein optimizing the learning rate parameter in the initial algorithm further comprises:

updating the learning rate parameter in the initial algorithm to a product of the learning rate parameter and the time scaling parameter.

14. The electronic device according to claim 8 , wherein training the quantization model further comprises:

increasing the time scaling parameter when the convergence speed of the quantization model is less than a predetermined speed threshold; and

decreasing the time scaling parameter when the precision of the quantization model is less than a predetermined precision threshold.

15. A non-transitory computer readable storage medium storing instructions that, when executed by one or more processors of an electronic device, cause the one or more processors to perform operations comprising:

determining jump ratios of embedding layer parameters of a trained quantization model in a first predetermined time range, each jump ratio corresponding to a parameter jump of the embedding layer parameters of the trained quantization model in each predetermined time interval within the first predetermined time range, the quantization model being a neural network model obtained after quantization processing on the embedding layer parameters;

determining a jump curve in the first predetermined time range according to the jump ratios;

fitting the jump curve to obtain a corresponding time scaling parameter;

optimizing an initial algorithm of the quantization model based on the time scaling parameter to obtain an optimized target optimization algorithm;

training the quantization model based on the target optimization algorithm to obtain an optimized quantization model, wherein the time scaling parameter is configured to adjust a convergence speed and a precision of the quantization model;

acquiring user behavior data associated with a user at a terminal in a second predetermined time range;

learning the user behavior data using the optimized quantization model:

determining a user behavior feature corresponding to the user behavior data;

determining target recommendation information based on the user behavior feature; and;

providing the target recommendation information to the user at the terminal.

16. The non-transitory computer readable storage medium according to claim 15 , wherein the quantization processing comprises N-valued quantization processing on the embedding layer parameters, wherein N is an integer greater than one.

17. The non-transitory computer readable storage medium according to claim 15 , wherein:

a respective jump ratio of a parameter jump of the embedding layer parameters in any predetermined time interval is determined by:

determining a quantity of parameters that jump in the embedding layer parameters in the any predetermined time interval, relative to the embedding layer parameters in a previous predetermined time interval of the any predetermined time interval; and

determining the jump ratio of the parameter jump of the embedding layer parameters in the any predetermined time interval according to the quantity of parameters and a total quantity of embedding layer parameters.

18. The non-transitory computer readable storage medium according to claim 15 , wherein fitting the jump curve further comprises:

determining a curve type of the jump curve, and determining a fitting function corresponding to the curve type; and

fitting the jump curve according to the fitting function to obtain the corresponding time scaling parameter.

19. The non-transitory computer readable storage medium according to claim 15 , wherein optimizing the initial algorithm of the quantization model further comprises:

optimizing a learning rate parameter in the initial algorithm according to the time scaling parameter.