CN103177414B - A kind of node of graph similarity parallel calculating method of structure based - Google Patents

Abstract

The invention discloses a kind of node of graph similarity parallel calculating method of structure based.Said method comprising the steps of: CPU end reads in multiple story text or image as host side, sets up graph model, obtains the adjacency matrix of figure; GPU end, as equipment end, receives the adjacency matrix of CPU end transmission, and GPU end calculates adjacency matrix; GPU end obtains adjacency matrix, and transfers to CPU end.Utilize parallel method to substantially increase the speed of Similarity Measure by this method, and higher computational accuracy can be ensured, meet efficiency and the accuracy requirement of voluminous media computing application; Experimental result shows, under the prerequisite of similarity precision, accelerating algorithm proposed by the invention achieves the speed-up ratio of average more than 100 times of more existing algorithm.

Description

A kind of node of graph similarity parallel calculating method of structure based

Technical field

The present invention relates to media computation field, particularly a kind of node of graph similarity parallel calculating method of structure based.

Background technology

At present in media computation field, when the problem such as solve Iamge Segmentation, content retrieval and mate, by design of graphics model, spread based on the similarity between node and obtain corresponding result.Simple, node of graph Similarity Measure is used to a kind of means of evaluation map interior joint (such as: super-pixel) structural similarity.

Descriptor in prior art usually between employing node goes the similarity between tolerance two nodes, carries out similarity diffusion based on the similarity relation between nodes neighbors and syntople.

Inventor is realizing in process of the present invention, finds at least there is following shortcoming and defect in prior art:

Along with the increase of figure scale, the computing time of similarity diffusion can greatly increase, and add the complexity of calculating, complexity even can reach O(kn ⁴), the needs in practical application cannot be met.

Summary of the invention

The invention provides a kind of node of graph similarity parallel calculating method of structure based, present approach reduces complexity and the computing time of calculating, meet the needs in practical application, described below:

A node of graph similarity parallel calculating method for structure based, said method comprising the steps of:

(1) CPU end reads in multiple story text or image as host side, sets up graph model, obtains the adjacency matrix of figure;

(2) GPU end is as equipment end, receives the adjacency matrix of CPU end transmission, and GPU end calculates adjacency matrix;

(3) GPU end obtains adjacency matrix, and transfers to CPU end.

When CPU end reads in multiple story text as host side, the step that described GPU end calculates adjacency matrix W is specially: described GPU holds calculating first adjacency matrix, that is,

1) by the location index computing node of node a and b in described first adjacency matrix to (a, b) corresponding in grid block index and thread index, wherein grid is the grid of GPU kernel function, block is the thread block in grid, and thread is the thread in thread block;

2) GPU end be in the first adjacency matrix each node to (a, b) Similarity Measure between distributes corresponding thread, that is: by block index and thread index search node to (a, b) corresponding thread, hold by corresponding thread computes node the similarity of (a, b) at GPU.

When CPU end reads in multiple image as host side, the step that described GPU end calculates adjacency matrix W is specially: described GPU holds calculating second adjacency matrix, comprising:

1) location matrix P during kth-1 iteration is searched for _k-1middle nonzero value, charges to row respectively by the line index of nonzero element, column index and respective value, in col, value tri-arrays;

2) by location matrix P _k-1calculate the location matrix P of the K time iteration _k:

3) calculate diagonal element and;

4) M current iteration obtained _kadd in S (a, b): S (a, b)=S (a, b)+M _k.

When CPU end reads in multiple image as host side, described method also comprises:

Described CPU end obtains transition matrix T, and described GPU end receives described transition matrix T as equipment end.

When CPU end reads in multiple image as host side, described method also comprises: the structure that described transition matrix T stores with row compression is stored as sparse matrix, the step that described GPU end calculates adjacency matrix W is specially: described GPU holds calculating second adjacency matrix, comprising:

1) CPU holds circulation to call GPU for K time and holds kernel function parallel computation similarity;

2) result of calculation is passed back CPU end by GPU end;

3) GPU end calculates location matrix P _kdiagonal line and M _k:

4) GPU end calculates value S (a, b): S (a, b)=S (a, the b)+M of the similarity of corresponding element in similarity matrix s _k.

The circulation of described CPU end is called GPU for K time and is held kernel function parallel computation similarity specifically to comprise:

A) T is calculated _iin nonzero value index x;

B) T is calculated _jin nonzero value index y;

C) similarity of manipulative indexing is calculated;

D) calculating crunode is to (a, b) position in location matrix;

E) location matrix P is upgraded _k:

The beneficial effect of technical scheme provided by the invention is: CPU end reads in multiple story text or image as host side, sets up graph model, obtains the adjacency matrix of figure; GPU end calculates adjacency matrix, and transfers to CPU end; Improve the precision of calculating chart node similarity by this method, reduce complexity and the computing time of calculating, meet the needs in practical application; Experimental result shows, under the prerequisite of similarity precision, accelerating algorithm proposed by the invention achieves the speed-up ratio of average more than 100 times.

Accompanying drawing explanation

Fig. 1 (a) and Fig. 1 (b) is original graph;

The result that the conspicuousness that Fig. 1 (c) and Fig. 1 (d) calculates for this method detects;

The result that the conspicuousness that Fig. 1 (e) and Fig. 1 (f) calculates for prior art detects;

Fig. 2 is a kind of process flow diagram of node of graph similarity parallel calculating method of structure based;

Fig. 3 is another process flow diagram of a kind of node of graph similarity parallel calculating method of structure based;

Fig. 4 is another process flow diagram of a kind of node of graph similarity parallel calculating method of structure based.

Embodiment

For making the object, technical solutions and advantages of the present invention clearly, below in conjunction with accompanying drawing, embodiment of the present invention is described further in detail.

In CUDA programming model, there are main frame and equipment two concepts.Computer CPU is commonly called host side (Host), and its GPU is commonly called equipment end (Device).A main frame and multiple equipment can be had in General System.

Use CUDA programming model to programme, can host side be assigned the task to and equipment end completes separately.Wherein host side is responsible for process function and the logic of things and the calculating of applicable serial, and equipment end then may be used for the thread-level task performing applicable highly-parallel.CPU, GPU have separate memory address space separately: the internal memory of host side and the video memory of equipment end.The operation of CUDA to internal memory uses former C language to operate same syntax of figs and the function of internal memory, operate video memory then to need to call function relevant with memory management in CUDAAPI, these operations include application, release and initialization video memory space, and the copy of data between internal memory and video memory etc.After having analyzed in program the part being applicable to walking abreast, these tasks that can walk abreast just can be considered to transfer to GPU to calculate.

In order to reduce complexity and the computing time of calculating, meet the needs in practical application, embodiments provide a kind of node of graph similarity parallel calculating method of structure based, the method comprises the following steps:

Embodiment 1

101:CPU end reads in multiple story text as host side, sets up graph model, obtains the first adjacency matrix W of figure;

Word in node on behalf story in figure, the similarity between the limit representation node between node, and according to first similarity tolerance rule, the relation on limit is set up in graph model to word between story inside word and story, obtain the first adjacency matrix W of figure.

Wherein, first similarity tolerance rule sets according to the needs in practical application, such as: the frequency that the frequency that the similarity measurement between the inner word of story is occurred by word A, word B occur and word A and word B occur jointly and the number of times that distance between two words is less than preset value is determined jointly; The frequency that between story, the similarity measurement of word is occurred by word A and the frequency that word B occurs are determined, preset value is set by the needs in practical application.

102:GPU end, as equipment end, receives the first adjacency matrix W of CPU end transmission, and GPU holds calculating first adjacency matrix W;

Wherein, the first adjacency matrix W is determined by the similarity that nodes all in matrix are right, and this step specifically comprises:

1) be GPU end line journey Distribution Calculation task: by the location index computing node of node a and b in the first adjacency matrix W to (a, b) corresponding in grid block index and thread index, wherein grid is the grid of GPU kernel function, block is the thread block in grid, and thread is the thread in thread block;

2) GPU end be in the first adjacency matrix W each node to (a, b) Similarity Measure between distributes corresponding thread, that is: by block index and thread index search node to (a, b) corresponding thread, GPU end by corresponding thread computes node to the similarity of (a, b):

For a given digraph G, make the similarity that s (a, b) comes between representation node a and node b, then these two node SimRank similarities are defined as follows:

For the situation of a=b, s (a, b)=1; For the situation of a ≠ b, s (a, b) is calculated as follows:

$S(a,b)=\frac{C}{\left|I\left(a\right)\right|\left|I\left(b\right)\right|}\underset{i=1}{\overset{\left|I\left(a\right)\right|}{\mathrm{\Σ}}}\underset{j=1}{\overset{\left|I\left(b\right)\right|}{\mathrm{\Σ}}}S(I_i\left(a\right),I_j\left(b\right))$

Wherein, C is between 0, the constant coefficient between 1; | I (a) | with | I (b) | what represent a and b respectively enters neighbours' number; I _ia i-th of () representation node a enters neighbours, I _jb the jth of () representation node b enters neighbours.

This full point in algorithm, calculate simultaneously all nodes between similarity use for next iterative computation.Make R _k(a, b) represents the SimRank similarity of the kth time iteration between (a, b), then S (a, b)=lim _{k → ∞}r _k(a, b).

For R _kthe calculating of (a, b), uses iterative manner.During initialization, when a is not equal to b, R (a, b)=0, otherwise it equals 1, then carries out iteration by following formula:

$R_{k+1}(a+b)=\frac{C}{\left|I\left(a\right)\right|\left|I\left(b\right)\right|}\underset{i=1}{\overset{\left|I\left(a\right)\right|}{\mathrm{\Σ}}}\underset{j=1}{\overset{\left|I\left(b\right)\right|}{\mathrm{\Σ}}}{R}_k(I_i\left(a\right),I_j\left(b\right))$

103:GPU end obtains the first adjacency matrix W, and transfers to CPU end.

Embodiment 2

201:CPU end inputs multiple images as host side, sets up graph model, obtains the second adjacency matrix W and the transition matrix T of figure;

Super-pixel in node on behalf image in figure, similarity between limit representation node between node, and according to second similarity tolerance rule between the super-pixel of an image inside and set up the relation on limit between different images super-pixel in graph model, finally calculate the second adjacency matrix W, and calculate transition matrix T by the second adjacency matrix W, and read in algorithm parameter constant attenuation factor C and error e rr.

Wherein, second similarity tolerance rule sets according to the needs in practical application, such as: calculate each block super-pixel and obtain range descriptors, calculate its similarity between two in image between super-pixel by range descriptors; Its similarity is calculated by range descriptors between super-pixel between two between image.

During specific implementation, CPU end also needs:

1) initialization similarity matrix and location matrix: similarity matrix is initialized as zero, s (a, b)=0; Location matrix is initialized as 1,

Due to a kind of character that this algorithm has, that is: the internodal similarity of a, b two is equal to two random walk persons respectively since a, b set out, the possibility of meeting for the first time during random advance in converse digraph. when having represented to walk k step, the location matrix that two random walk persons a, b were not meeting before, then namely the element [i, j] in matrix represents the probability arriving i, j after migration person a, b have walked k step respectively.

2) number of times K, K=(int) log is calculated repeatly _Δc.

Wherein, C represents decay factor: C=0.5, and Δ represents error: Δ=0.01.

202:GPU end is as equipment end, and the second adjacency matrix W and transition matrix T, the GPU that receive the transmission of CPU end hold calculating second adjacency matrix W;

Wherein, this step specifically comprises:

1) P is searched for _k-1nonzero value in matrix (location matrix during kth-1 iteration), charges to row respectively by the line index of nonzero element, column index and respective value, in col, value tri-arrays;

At search P _k-1during matrix, when running into nonzero element, by line number stored in array row, row are number stored in array col, and respective value is stored in array value.

2) P is passed through _k-1the location matrix P of matrix computations the K time iteration _k: by following formulae discovery location matrix P _k:

$P_k=C\·\underset{i=1}{\overset{\left|V\right|}{\mathrm{\Σ}}}\underset{j=1;j\≠i}{\overset{\left|V\right|}{\mathrm{\Σ}}}{\left(P_{k-1}\right)}_{\mathrm{ij}}\·\left({T_i}^{\text{'}}{T}_j\right)$

Wherein, T _i, T _jfor the row vector of transition matrix T, T _i' be T _itransposition, | V| is neighbours' number of node.

During specific implementation, the algorithm of corresponding kernel function is as follows: first, calculates the horizontal ordinate and ordinate that will calculate in location matrix according to built-in variable; Then, initialized location matrix P _k, make P _k[a, b]=0, go to calculate according to following formula finally by circulation n time (n is the number of figure node):

P _k[a,b]=P _k[a,b]+T[i,a]*T[j,b]*value[m]

Wherein, T [i, a] represents i-th row of transition matrix T, a column element; T [j, b] represents the jth row of transition matrix T, b column element; Value [m] is P _k-1value in [a, b].

2) calculate diagonal element and;

$M_k=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{P}_k[i,i]$

3) M current iteration obtained _kadd in S (a, b): S (a, b)=S (a, b)+M _k.

203:GPU end obtains the second adjacency matrix W, and transfers to CPU end.

Embodiment 3

When figure larger and more sparse time, in order to improve the speed of calculating second adjacency matrix, this method can also calculate the second adjacency matrix by step 302.

301:CPU end inputs multiple images as host side, sets up graph model to it, obtains the second adjacency matrix W and the transition matrix T of figure, and transition matrix T is stored as sparse matrix with CRS (CompressedRowStorage, compressed line stores) structure;

This file layout use continuous print core position store following vector: val line number group with the main order of behavior to store non-zero matrix element, the column index of each element in col storage of array val array, rowptr vector stores the element numbers starting a line in val array.

Super-pixel in node on behalf image in figure, similarity between limit representation node between node, and according to second similarity tolerance rule between the super-pixel of an image inside and set up the relation on limit between different images super-pixel in graph model, finally calculate the second adjacency matrix W, and calculate transition matrix T by the second adjacency matrix W, and read in algorithm parameter constant C and error e rr.

Wherein, second similarity tolerance rule sets according to the needs in practical application, such as: calculate each block super-pixel and obtain range descriptors, calculate its similarity between two in image between super-pixel by range descriptors; Its similarity is calculated by range descriptors between super-pixel between two between image.

During specific implementation, CPU end also needs:

1) initialization similarity matrix and location matrix: similarity matrix is initialized as zero, S (a, b)=0; Location matrix is initialized as 1,

Wherein, due to a kind of character that this algorithm has, that is: the internodal similarity of a, b two is equal to two random walk persons respectively since a, b set out, the possibility of meeting for the first time during random advance in converse digraph. when having represented to walk k step, the location matrix that two random walk persons a, b were not meeting before, then namely the element [i, j] in matrix represents the probability arriving i, j after migration person a, b have walked k step respectively.

2) number of times K, K=(int) log is calculated repeatly _Δc.

302:GPU end is as equipment end, and the second adjacency matrix W and transition matrix T, the GPU that receive the transmission of CPU end hold calculating second adjacency matrix W;

Wherein, this step specifically comprises:

1) CPU holds circulation to call GPU for K time and holds kernel function parallel computation similarity;

To each nonzero element corresponding in transition matrix T, call kernel function, a corresponding startup thread, corresponding nonzero value is calculated rear coal addition position matrix P by this thread _kmiddle correspondence position, wherein, mainly comprises the following steps:

A) T is calculated _iin nonzero value index x: calculate and represent this thread and need to use T _ia middle xth nonzero value.

B) T is calculated _jin nonzero value index y: calculate and represent this thread and need to use T _jin y nonzero value.

C) similarity of manipulative indexing is calculated: according to x, y, by corresponding T _i, T _jin value fetch calculating, similarity is designated as s.

D) calculating crunode is to (a, b) position in location matrix: according to index x, y, calculates the index in the location matrix that s should insert;

Wherein, the result after transition matrix T and the converse digraph adjacency matrix column criterion of carrying out.

E) location matrix P is upgraded _k:

2) result of calculation is passed back CPU end by GPU end: by P _kmatrix passes CPU end back;

3) GPU end calculates location matrix P _kdiagonal line and M _k:

4) GPU end calculates value S (a, b): S (a, b)=S (a, the b)+M of the similarity of corresponding element in similarity matrix s _k.

303:GPU end obtains the second adjacency matrix W, and transfers to CPU end.

The feasibility of this method is verified below with concrete experiment, described below:

One, story text

Choose five story text, the corresponding theme id of each story text, be respectively: 20001,20015,20039,20070 and 20076, prior art and this method is adopted to calculate respectively above-mentioned five story text, obtain corresponding adjacency matrix, and then obtain the similarity of each story text, then calculate average similarity and average operating time in average similarity between theme, theme, be respectively shown in table 1 and table 2.

Table 1

Table 2

Known by the contrast of table 1 and table 2, in the theme adopting prior art and this method to get, on average between similarity and theme, the ratio of average similarity is more or less the same, but the average operating time of this method is far smaller than the average operating time of prior art.

Two, many saliencies cooperation detection

Adopt prior art and this method to process Fig. 1 (a) and Fig. 1 (b) simultaneously, get the result that the conspicuousness calculated for this algorithm in Fig. 1 (c) and Fig. 1 (d) detects, and the result of the conspicuousness detection that Fig. 1 (e) and Fig. 1 (f) calculates for original serial algorithm.As can be seen from the figure, when conspicuousness result of calculation is more or less the same, the working time of this method is far smaller than the working time of prior art, see table 3.

Table 3

	Working time
		Serial Simrank	1.98minutes
Parallel Simrank	1.27minutes

It will be appreciated by those skilled in the art that accompanying drawing is the schematic diagram of a preferred embodiment, the invention described above embodiment sequence number, just to describing, does not represent the quality of embodiment.

The foregoing is only preferred embodiment of the present invention, not in order to limit the present invention, within the spirit and principles in the present invention all, any amendment done, equivalent replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (3)

1. a node of graph similarity parallel calculating method for structure based, is characterized in that, said method comprising the steps of:

(1) CPU end reads in multiple story text or image as host side, sets up graph model, obtains the adjacency matrix of figure;

(2) GPU end is as equipment end, receives the adjacency matrix of CPU end transmission, and GPU end calculates adjacency matrix;

(3) GPU end obtains adjacency matrix, and transfers to CPU end;

Wherein, when CPU end reads in multiple story text as host side, the step that described GPU end calculates adjacency matrix W is specially: described GPU holds calculating first adjacency matrix, that is,

1) by the location index computing node of node a and b in described first adjacency matrix to (a, b) corresponding in grid block index and thread index, wherein grid is the grid of GPU kernel function, block is the thread block in grid, and thread is the thread in thread block;

2) GPU end be in the first adjacency matrix each node to (a, b) Similarity Measure between distributes corresponding thread, that is: by block index and thread index search node to (a, b) corresponding thread, hold by corresponding thread computes node the similarity of (a, b) at GPU;

Wherein, when CPU end reads in multiple image as host side, the step that described GPU end calculates adjacency matrix W is specially: described GPU holds calculating second adjacency matrix, comprising:

1) location matrix P during kth-1 iteration is searched for _k-1middle nonzero value, charges to row respectively by the line index of nonzero element, column index and respective value, in col, value tri-arrays;

2) by location matrix P _k-1calculate the location matrix P of the K time iteration _k:

3) diagonal element and M is calculated _k;

4) diagonal element current iteration obtained and M _kadd in S (a, b), S (a, b) is the value of the similarity of corresponding element in GPU end calculating similarity matrix s, S (a, b)=S (a, b)+M _k.

2. a node of graph similarity parallel calculating method for structure based, is characterized in that, said method comprising the steps of:

(1) CPU end reads in multiple story text or image as host side, sets up graph model, obtains the adjacency matrix of figure;

(2) GPU end is as equipment end, receives the adjacency matrix of CPU end transmission, and GPU end calculates adjacency matrix;

(3) GPU end obtains adjacency matrix, and transfers to CPU end;

When CPU end reads in multiple image as host side, described method also comprises:

Described CPU end obtains transition matrix T, and described GPU end receives described transition matrix T as equipment end;

When CPU end reads in multiple image as host side, described method also comprises: the structure that described transition matrix T stores with row compression is stored as sparse matrix, the step that described GPU end calculates adjacency matrix W is specially: described GPU holds calculating second adjacency matrix, comprising:

1) CPU holds circulation to call GPU for K time and holds kernel function parallel computation similarity;

2) result of calculation is passed back CPU end by GPU end;

3) GPU end calculates location matrix P _kdiagonal line and M _k: n is the number of figure node;

4) GPU end calculates value S (a, b): S (a, b)=S (a, the b)+M of the similarity of corresponding element in similarity matrix s _k.

3. the node of graph similarity parallel calculating method of a kind of structure based according to claim 2, is characterized in that, the circulation of described CPU end is called GPU for K time and held kernel function parallel computation similarity specifically to comprise:

A) T is calculated _iin nonzero value index x;

B) T is calculated _jin nonzero value index y;

C) similarity of manipulative indexing is calculated;

D) calculating crunode is to (a, b) position in location matrix;

E) location matrix P is upgraded _k: