Granted Patent B2

US 12,205,003 · App. 17/361,633 · Granted Jan 21, 2025

Inventors: Yao Zhang (Shanghai, CN); Guang Jiang (Shanghai, CN); Xishan Zhang (Shanghai, CN); Shiyi Zhou (Shanghai, CN); Di Huang (Shanghai, CN); Chang Liu (Shanghai, CN); Jiaming Guo (Shanghai, CN)

Assignee: SHANGHAI CAMBRICON INFORMATION TECHNOLOGY CO., LTD

G06N20/00G06F15/7832G06N3/045H03M7/3059H03M7/6041G06N3/04H03M7/70

Abstract

Claims (60)

1. A method for processing data, comprising:

obtaining a plurality of pieces of data for a machine learning model, wherein the data is represented in a first data format;

distributing the plurality of pieces of data in the first group of intervals based on the value of the plurality of pieces of data, wherein the number of the first group of intervals is predetermined;

determining a first average value of the plurality of pieces of data;

mapping the data distributed in the first group of intervals into a second group of intervals according to a plurality of different mapping units so as to determine a second average value of the mapped data according to a corresponding mapping unit, wherein the mapping unit includes a predetermined number of consecutive intervals starting from the first interval of the first group of intervals, and the ending interval of the mapping unit is distributed with the plurality of pieces of data that is larger than or located in the end interval;

the number of intervals in the second group of intervals is related to the second data format, and the accuracy of the second data format is lower than that of the first data format; and

processing the plurality of pieces of data based on the first average value and a plurality of second average values, so that the processed plurality of pieces of data is represented in the second data format.

2. The method of claim 1 , wherein the mapping the data distributed in the first group of intervals to the second group of intervals according to the plurality of different mapping units includes:

redistributing data in the plurality of pieces of data that is larger than the end interval of the mapping unit to the end interval of the mapping unit; and

mapping a plurality of mapping units to the second group of intervals respectively to determine the second average value of the mapped data according to the corresponding mapping unit.

3. The method of claim 1 , wherein the processing the plurality of pieces of data includes:

determining a plurality of differences related to the plurality of mapping units based on the first average value and the plurality of second average values;

determining a truncation threshold used for the plurality of pieces of pieces of data based on the plurality of differences;

truncating the obtained plurality of pieces of data based on the truncation threshold; and

determining a scale value related to the mapping unit based on the truncation threshold.

4. The method of claim 3 , wherein the processing the plurality of pieces of data further comprises:

scaling the truncated plurality of pieces of data based on the scale value.

5. The method of claim 1 , further comprising:

determining the first group of intervals related to the plurality of pieces of data based on a maximum absolute value of the data.

6. The method of claim 1 , wherein the determining the first average value of the plurality of pieces of data includes:

determining the absolute value of the data in the plurality of pieces of data and the number of the plurality of pieces of data; and

determining the average value based on the absolute value of the data in the plurality of pieces of data and the number of the plurality of pieces of data.

7. The method of claim 1 , wherein the determining the first average value of the plurality of pieces of data includes:

determining an intermediate value of each interval in the first group of intervals and the amount of data in each interval; and

determining the average value based on the intermediate value and the number.

8. The method of claim 1 , wherein the determining the first average value of the plurality of pieces of data includes:

determining the amount of data in each interval of the mapping unit and the intermediate value of each interval; and

determining the first average value based on the number and the intermediate value.

9. The method of claim 1 , wherein the number of intervals in the second group of intervals is determined based on the number of bits related to the second data format.

10. An artificial intelligence chip, comprising:

a processor and

a memory, wherein the memory stores computer program instructions, and the processor runs the computer program instructions in the memory to control the artificial intelligence chip to execute the method of claim 1 .

11. A device for processing data, comprising:

an acquisition circuit configured to acquire the plurality of pieces of data for a machine learning model, wherein the data is represented in a first data format;

a distribution circuit configured to distribute the plurality of pieces of data in the first group of intervals based on values of the plurality of pieces of data, wherein the number of the first group of intervals is predetermined;

a first average value determining circuit configured to determine the first average value of the plurality of pieces of data;

a first mapping circuit configured to map the data distributed in the first group of intervals to the second group of intervals according to a plurality of different mapping circuits to determine a second average value of the mapped data according to the corresponding mapping circuit, wherein the mapping circuit includes a predetermined number of consecutive intervals starting from the first interval of the first group of intervals, and the ending interval of the mapping circuit is distributed with the plurality of pieces of data that is larger than or located in the end interval; the number of intervals in the second group of intervals is related to the second data format, and the accuracy of the second data format is lower than that of the first data format; and

an execution circuit configured to process on the plurality of pieces of data based on the first average value and the plurality of second average values, so that the processed plurality of pieces of data is represented in the second data format.

12. The device of claim 11 , wherein the first mapping circuit includes:

a redistribution circuit configured to redistribute data in the plurality of pieces of data that is greater than the end interval of the mapping circuit to the end interval of the mapping circuit; and

a second mapping circuit configured to map the plurality of different mapping circuits to the second group of intervals respectively to determine the second average value of the mapped data according to a corresponding mapping circuit.

13. The device of claim 11 , wherein the execution circuit includes:

a difference determining circuit configured to determine a plurality of difference values related to the plurality of mapping circuits based on the first average value and the plurality of second average values;

a truncation threshold determining circuit configured to determine a truncation threshold for the plurality of pieces of data based on the plurality of differences;

a truncation circuit configured to truncate the obtained plurality of pieces of data based on the truncation threshold; and

a scale value determining circuit configured to determine a scale value related to the mapping circuit based on the truncation threshold.

14. The device of claim 13 , wherein the execution circuit further includes

a scale circuit configured to scale the truncated plurality of pieces of data based on the scale value.

15. The device of claim 11 , further comprising

an interval determining circuit configured to determine the first group of intervals related to the plurality of pieces of data based on the maximum absolute value of the data.

16. The device of claim 11 , wherein the first average value determining circuit includes:

an absolute value and number determining circuit configured to determine the absolute value of the data in the plurality of pieces of data and the number of the plurality of pieces of data; and

a second average value determining circuit configured to determine the average value based on the absolute value of the data in the plurality of pieces of data and the number of the plurality of pieces of data.

17. The device of claim 11 , wherein the first average value determining circuit includes:

a first intermediate value and number determining circuit configured to determine an intermediate value of each interval in the first group of intervals and the amount of data in each interval; and

a third average value determining circuit configured to determine the average value based on the intermediate value and the number.

18. The device of claim 11 , wherein the first average value determining device includes:

a second intermediate value and number determining circuit configured to determine the number of data located in each interval of the mapping circuit and the intermediate value of each interval; and

a fourth average value determining circuit configured to determine the first average value based on the number and the intermediate value.

19. The device of claim 11 , wherein the number of intervals in the second group of intervals is determined based on the number of bits related to the second data format.

Assignments (1)

Priority Claims (1)

Continuity (2)

References Cited (264)