Method, device, and program product for determining model compression rate

1. A method for determining a model compression rate, comprising:

determining a near-zero importance value subset from an importance value set associated with a machine learning model, a corresponding importance value in the importance value set indicating an importance degree of a corresponding input of a processing layer of the machine learning model, importance values in the near-zero importance value subset being closer to zero than other importance values in the importance value set;

determining a target importance value from the near-zero importance value subset, the target importance value corresponding to a turning point of a magnitude of the importance values in the near-zero importance value subset;

determining a proportion of importance values in the importance value set less than the target importance value in the importance value set; and

determining the model compression rate for the machine learning model based on the determined proportion.

2. The method according to claim 1 , wherein the determining the near-zero importance value subset comprises:

determining distribution of the importance value set, the distribution presenting a number of occurrences of a given importance value in the importance value set; and

determining the near-zero importance value subset based on the distribution.

3. The method according to claim 1 , wherein the determining the target importance value comprises:

obtaining a corresponding logarithmic transformation value set by performing logarithmic transformation on the importance values in the near-zero importance value subset;

determining a target logarithmic transformation value from the logarithmic transformation value set, the target logarithmic transformation value corresponding to a turning point of a magnitude of logarithmic transformation values in the logarithmic transformation value set; and

determining an importance value in the importance value set corresponding to the target logarithmic transformation value for use as the target importance value.

4. The method according to claim 3 , wherein a first difference between a first logarithmic transformation value less than the target logarithmic transformation value and a second logarithmic transformation value greater than the target logarithmic transformation value among the logarithmic transformation values is greater than two differences below:

a second difference between two logarithmic transformation values less than the target logarithmic transformation value in the logarithmic transformation value set; and

a third difference between two logarithmic transformation values greater than the target logarithmic transformation value in the logarithmic transformation value set.

5. The method according to claim 3 , wherein performing logarithmic transformation comprises utilizing large scale logarithmic projection (LSLP).

6. The method according to claim 3 , wherein the determining a target logarithmic transformation value from the logarithmic transformation value set comprises:

sorting the logarithmic transformation values in the logarithmic transformation value set in ascending or descending order;

determining a change trend of the sorted logarithmic transformation values; and

determining the target logarithmic transformation value based on the change trend.

7. The method according to claim 3 , wherein the determining a target logarithmic transformation value from the logarithmic transformation value set comprises:

determining the target logarithmic transformation value based on an adaptive threshold segmentation algorithm.

8. The method according to claim 1 , further comprising:

compressing the machine learning model by model pruning based on the model compression rate.

9. An electronic device, comprising:

at least one processor; and

at least one memory storing computer program instructions, the at least one memory and the computer program instructions being configured, together with the at least one processor, to cause the electronic device to perform actions for determining a model compression rate, the actions comprising:

determining a near-zero importance value subset from an importance value set associated with a machine learning model, a corresponding importance value in the importance value set indicating an importance degree of a corresponding input of a processing layer of the machine learning model, importance values in the near-zero importance value subset being closer to zero than other importance values in the importance value set;

determining a target importance value from the near-zero importance value subset, the target importance value corresponding to a turning point of a magnitude of the importance values in the near-zero importance value subset;

determining a proportion of importance values in the importance value set less than the target importance value in the importance value set; and

determining the model compression rate for the machine learning model based on the determined proportion.

10. The device according to claim 9 , wherein the determining the near-zero importance value subset comprises:

determining distribution of the importance value set, the distribution presenting a number of occurrences of a given importance value in the importance value set; and

determining the near-zero importance value subset based on the distribution.

11. The device according to claim 9 , wherein the determining the target importance value comprises:

obtaining a corresponding logarithmic transformation value set by performing logarithmic transformation on the importance values in the near-zero importance value subset;

determining a target logarithmic transformation value from the logarithmic transformation value set, the target logarithmic transformation value corresponding to a turning point of a magnitude of logarithmic transformation values in the logarithmic transformation value set; and

determining an importance value in the importance value set corresponding to the target logarithmic transformation value for use as the target importance value.

12. The device according to claim 11 , wherein a first difference between a first logarithmic transformation value less than the target logarithmic transformation value and a second logarithmic transformation value greater than the logarithmic transformation value among the logarithmic transformation values is greater than two differences below:

a second difference between two logarithmic transformation values less than the target logarithmic transformation value in the logarithmic transformation value set; and

a third difference between two logarithmic transformation values greater than the target logarithmic transformation value in the logarithmic transformation value set.

13. The device according to claim 11 , wherein performing logarithmic transformation comprises utilizing large scale logarithmic projection (LSLP).

14. The device according to claim 11 , wherein the determining a target logarithmic transformation value from the logarithmic transformation value set comprises:

sorting the logarithmic transformation values in the logarithmic transformation value set in ascending or descending order;

determining a change trend of the sorted logarithmic transformation values; and

determining the target logarithmic transformation value based on the change trend.

15. The device according to claim 11 , wherein the determining a target logarithmic transformation value from the logarithmic transformation value set comprises:

determining the target logarithmic transformation value based on an adaptive threshold segmentation algorithm.

16. The device according to claim 9 , wherein the actions further comprise:

compressing the machine learning model by model pruning based on the model compression rate.

17. A computer program product, the computer program product being tangibly stored in a non-volatile computer-readable medium and comprising machine-executable instructions, the machine-executable instructions for determining a model compression rate, when executed by at least one processing device, causing the at least one processing device to:

determine a near-zero importance value subset from an importance value set associated with a machine learning model, a corresponding importance value in the importance value set indicating an importance degree of a corresponding input of a processing layer of the machine learning model, importance values in the near-zero importance value subset being closer to zero than other importance values in the importance value set;

determine a target importance value from the near-zero importance value subset, the target importance value corresponding to a turning point of a magnitude of the importance values in the near-zero importance value subset;

determine a proportion of importance values in the importance value set less than the target importance value in the importance value set; and

determine the model compression rate for the machine learning model based on the determined proportion.

18. The computer program product according to claim 17 , wherein the machine-executable instructions, when executed, cause the device to determine the near-zero importance value subset by:

determining distribution of the importance value set, the distribution presenting a number of occurrences of a given importance value in the importance value set; and

determining the near-zero importance value subset based on the distribution.

19. The computer program product according to claim 17 , wherein the machine-executable instructions, when executed, cause the device to determine the target importance value by:

obtaining a corresponding logarithmic transformation value set by performing logarithmic transformation on the importance values in the near-zero importance value subset;

determining a target logarithmic transformation value from the logarithmic transformation value set, the target logarithmic transformation value corresponding to a turning point of a magnitude of logarithmic transformation values in the logarithmic transformation value set; and

determining an importance value in the importance value set corresponding to the target logarithmic transformation value for use as the target importance value.

20. The computer program that according to claim 19 , wherein a first difference between a first logarithmic transformation value less than the target logarithmic transformation value and a second logarithmic transformation value greater than the logarithmic transformation value among the logarithmic transformation values is greater than two differences below:

a second difference between two logarithmic transformation values less than the target logarithmic transformation value in the logarithmic transformation value set; and

a third difference between two logarithmic transformation values greater than the target logarithmic transformation value in the logarithmic transformation value set.