Data compression method, data compression system and operation method of deep learning acceleration chip
A data compression method, a data compression system and an operation method of a deep learning acceleration chip are provided. The data compression method includes the following steps. A filter coefficient tensor matrix of a deep learning model is obtained. A matrix decomposition procedure is performed according to the filter coefficient tensor matrix to obtain a sparse tensor matrix and a transformation matrix, which is an orthonormal matrix. The product of the transformation matrix and the filter coefficient tensor matrix is the sparse tensor matrix. The sparse tensor matrix is compressed. The sparse tensor matrix and the transformation matrix, or the sparse tensor matrix and a restoration matrix, are stored in a memory. A convolution operation result is obtained by the deep learning acceleration chip using the sparse tensor matrix. The convolution operation result is restored by the deep learning acceleration chip using the restoration matrix.
1 . A data compression method of a deep learning acceleration chip, comprising:
obtaining a filter coefficient tensor matrix of a deep learning model;
using a relationship that product of a transformation matrix and the filter coefficient tensor matrix is a sparse tensor matrix to convert the filter coefficient tensor matrix to the sparse tensor matrix and the transformation matrix is an orthonormal matrix;
compressing the sparse tensor matrix; and
storing the sparse tensor matrix and the transformation matrix, or storing the sparse tensor matrix and a restoration matrix, in a memory, wherein after the deep learning acceleration chip obtains a recognition data, the deep learning acceleration chip uses a relationship that product of the sparse tensor matrix and the recognition data is the convolution operation result and a relationship that product of the restoration matrix and the convolution operation result is a final convolution operation result, to obtain the final convolution operation result.
2 . The data compression method of the deep learning acceleration chip according to claim 1 , wherein the restoration matrix is a transpose matrix of the transformation matrix.
3 . The data compression method of the deep learning acceleration chip according to claim 1 , wherein the convolution operation result is restored without loss of information.
4 . The data compression method of the deep learning acceleration chip according to claim 1 , wherein in the step of decomposing the filter coefficient tensor matrix, the filter coefficient tensor matrix is partitioned into M parts, the at least one sparse tensor matrix has a quantity of M, the at least one transformation matrix has a quantity of M, and M is a natural number.
5 . The data compression method of the deep learning acceleration chip according to claim 4 , wherein M is 2N, and N is a natural number.
6 . The data compression method of the deep learning acceleration chip according to claim 4 , wherein the filter coefficient tensor matrix is equally partitioned.
7 . The data compression method of the deep learning acceleration chip according to claim 4 , wherein the sparse tensor matrixes have the same size.
8 . The data compression method of the deep learning acceleration chip according to claim 4 , wherein the transformation matrixes have the same size.
9 . A data compression system of a deep learning acceleration chip, wherein the data compression system is used to reduce data movement for a deep learning model and comprises:
a decomposition unit configured to use a relationship that product of a transformation matrix and the filter coefficient tensor matrix is a sparse tensor matrix to convert the filter coefficient tensor matrix to the sparse tensor matrix and the transformation matrix is an orthonormal matrix;
a compression unit configured to compress the sparse tensor matrix; and
a transfer unit configured to store the sparse tensor matrix and the transformation matrix, or store the sparse tensor matrix and a restoration matrix, in a memory, wherein after the deep learning acceleration chip obtains a recognition data, the deep learning acceleration chip uses a relationship that product of the sparse tensor matrix and the recognition data is the convolution operation result and a relationship that product of the restoration matrix and the convolution operation result is a final convolution operation result, to obtain the final convolution operation result.
10 . The data compression system of the deep learning acceleration chip according to claim 9 , wherein the restoration matrix is a transpose matrix of the transformation matrix.
11 . The data compression system of the deep learning acceleration chip according to claim 9 , wherein the convolution operation result is restored without loss of information.
12 . The data compression system of the deep learning acceleration chip according to claim 9 , wherein the decomposition unit further partitions the filter coefficient tensor matrix into M parts, the at least one sparse tensor matrix has a quantity of M, the at least one transformation matrix has a quantity of M, and M is a natural number.
13 . The data compression system of the deep learning acceleration chip according to claim 12 , wherein M is 2N, and N is a natural number.
14 . The data compression system of the deep learning acceleration chip according to claim 12 , wherein the filter coefficient tensor matrix is equally partitioned.
15 . The data compression system of the deep learning acceleration chip according to claim 12 , wherein the sparse tensor matrixes have the same size.
16 . The data compression system of the deep learning acceleration chip according to claim 12 , wherein the transformation matrixes have the same size.