Decompression of model parameters using functions based upon cumulative count distributions

Patent metadata

Field	Value
Publication number	US-10938412-B2
Application number	US-202016865207-A
Country	US
Kind code	B2
Filing date	May 1, 2020
Priority date	Sep 15, 2017
Publication date	Mar 2, 2021
Grant date	Mar 2, 2021

Abstract

A predictive model utilizes a set of coefficients for processing received input data. To reduce memory usage storing the coefficients, a compression circuit compresses the set of coefficients prior to storage by generating a cumulative count distribution of the coefficient values, and identifying a distribution function approximating the cumulative count distribution. Function parameters for the determined function are stored in a memory and used by a decompression circuit to apply the function the compressed coefficients to determine the decompressed component values. Storing the function parameters may consume less memory in comparison to storing a look-up table for decompression, and may reduce an amount of memory look-ups required during decompression.

First claim

What is claimed is:

1. A processor, comprising:
a memory storing:
compressed coefficient data corresponding to a set of coefficients associated with a predictive model;
a set of function parameters associated with the compressed coefficient data;

a decompression circuit comprising:
a first function calculation circuit associated with a first function type; and
a second function calculation circuit associated with a second function type,

wherein the decompression circuit is configured to:
receive the set of function parameters from the memory;
select between the first function calculation circuit and the second function calculation circuit based on the received set of function parameters; and
decompress the compressed coefficient data using the selected function calculation circuit to generate a set of decompressed coefficients.

2. The processor of claim 1 , further comprising an arithmetic circuit unit to receive a set of input data and the set of decompressed coefficients, and to execute one or more arithmetic operations based on the set of input data and the set of decompressed coefficients to generate a set of output values.

3. The processor of claim 1 , wherein the set of function parameters comprises a first parameter indicating a function type, and at least one additional parameter corresponding to a function coefficient, and wherein the decompression circuit selects between the first function calculation circuit and the second function calculation circuit based upon the first parameter indicating the function type.

4. The processor of claim 1 , wherein the compressed coefficient data is compressed from the set of coefficients using a compression function selected based upon a cumulative distribution of values of the set of coefficients, and wherein the set of function parameters correspond to the selected compression function.

5. The processor of claim 1 , wherein values of the set of coefficients are determined through a model training process.

6. The processor of claim 1 , wherein the set of function parameters correspond to a function type selected from at least one of a polynomial function, a bimodal distribution function, a Gaussian distribution function, or a Poisson distribution function.

7. The processor of claim 1 , wherein the set of coefficients is compressed using arithmetic or Huffman coding to generate the compressed coefficient data.

8. The processor of claim 1 , wherein the decompression circuit is configured to apply the set of function parameters to at least a portion of the compressed coefficient data corresponding to a first compressed coefficient to determine a first decompressed coefficient value of the set of decompressed coefficients.

9. The processor of claim 8 , wherein the decompression circuit is further configured to:
receive a sequence of one or more bits of the first compressed coefficient from the compressed coefficient data;
generate first and second extended bit sequences based upon the received sequence of bits;
apply the set of function parameters to the first and second extended bit sequences to determine first and second respective coefficient values; and
in response to a determination that the first and second coefficient values are the same, output the first coefficient value as the first decompressed coefficient value.

10. The processor of claim 9 , wherein the decompression circuit is further configured to:
in response to a determination that the first and second coefficient values are not the same, receive at least one additional bit of the compressed coefficient data appended to the sequence of one or more bits to generate an updated sequence of bits; and
generate updated first and second extended bit sequences based upon the updated sequence of bits.

11. A method, comprising:
receiving, at a decompression circuit comprising a first function calculation circuit associated with a first function type and a second function calculation circuit associated with a second function type, compressed coefficient data from a memory, the compressed coefficient data corresponding to a set of coefficients associated with a predictive model;
receiving, at the decompression circuit, a set of function parameters from the memory, the set of function parameters associated with the compressed coefficient data;
selecting between the first function calculation circuit and the second function calculation circuit for decompressing the compressed coefficient data, based on the received set of function parameters; and
decompressing the compressed coefficient data using the selected function calculation circuit to generate a set of decompressed coefficients.

12. The method of claim 11 , further comprising:
receiving, at an arithmetic circuit unit, a set of input data and the set of decompressed coefficients; and
executing, at the arithmetic circuit unit, one or more arithmetic operations based on the set of input data and the set of decompressed coefficients to generate a set of output values.

13. The method of claim 11 , wherein the set of function parameters comprises a first parameter indicating a function type, and at least one additional parameter corresponding to a function coefficient, and wherein the decompression circuit selects between the first function calculation circuit and the second function calculation circuit based upon the first parameter indicating the function type.

14. The method of claim 11 , wherein the compressed coefficient data is compressed from the set of coefficients using a compression function selected based upon a cumulative distribution of values of the set of coefficients, and wherein the set of function parameters correspond to the selected compression function.

15. The method of claim 11 , wherein values of the set of coefficients are determined through a model training process.

16. The method of claim 11 , wherein the set of function parameters correspond to a function type selected from at least one of a polynomial function, a bimodal distribution function, a Gaussian distribution function, or a Poisson distribution function.

17. The method of claim 11 , wherein the set of coefficients is compressed using arithmetic or Huffman coding to generate the compressed coefficient data.

18. The method of claim 11 , wherein decompressing the compressed coefficient data using the selected function calculation circuit comprises applying the set of function parameters to at least a portion of the compressed coefficient data corresponding to a first compressed coefficient to determine a first decompressed coefficient value of the set of decompressed coefficients.

19. The method of claim 18 , wherein decompressing the compressed coefficient data using the selected function calculation circuit comprises:
receiving a sequence of one or more bits of the first compressed coefficient from the compressed coefficient data;
generating first and second extended bit sequences based upon the received sequence of bits;
applying the set of function parameters to the first and second extended bit sequences to determine first and second respective coefficient values; and
in response to a determination that the first and second coefficient values are the same, outputting the first coefficient value as the first decompressed coefficient value.

20. The method of claim 19 , further comprising:
in response to a determination that the first and second coefficient values are not the same, receiving at least one additional bit of the compressed coefficient dat

Assignees

• Groq Inc

Inventors

• Ross Jonathan Alexander
• Abts Dennis Charles

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/0495

  Quantised networks; Sparse networks; Compressed networks · CPC title

• H03M7/425 Primary

  for the decoding process only · CPC title

• H03M7/4037

  Prefix coding · CPC title