CN110427960A - A kind of restructural multi-category support vector machines system - Google Patents
A kind of restructural multi-category support vector machines system Download PDFInfo
- Publication number
- CN110427960A CN110427960A CN201910528797.5A CN201910528797A CN110427960A CN 110427960 A CN110427960 A CN 110427960A CN 201910528797 A CN201910528797 A CN 201910528797A CN 110427960 A CN110427960 A CN 110427960A
- Authority
- CN
- China
- Prior art keywords
- classification
- module
- kernel function
- support vector
- computing module
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000012706 support-vector machine Methods 0.000 title claims abstract description 45
- 239000013598 vector Substances 0.000 claims abstract description 48
- 238000003860 storage Methods 0.000 claims abstract description 39
- 238000000034 method Methods 0.000 claims abstract description 18
- 238000012360 testing method Methods 0.000 claims abstract description 17
- 238000010276 construction Methods 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 4
- 238000013500 data storage Methods 0.000 claims description 2
- 230000006870 function Effects 0.000 description 41
- 238000013461 design Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 5
- 238000012549 training Methods 0.000 description 4
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 3
- 238000010224 classification analysis Methods 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000000611 regression analysis Methods 0.000 description 2
- 238000009825 accumulation Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000001186 cumulative effect Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000005265 energy consumption Methods 0.000 description 1
- 238000013100 final test Methods 0.000 description 1
- 230000014759 maintenance of location Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Complex Calculations (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention relates to restructural multi-category support vector machines system, which includes: main control module, storage control module, kernel function computing module, classification computing module and result comparison module.The main control module provides control information and reconfiguration information for entire decision process；The storage of the storage control module control data；The kernel function computing module calculates the kernel function between test data and supporting vector；The classification computing module calculates decision value and class categories；The decision value that the more different models of result comparison module are calculated, obtains the final classification result of test data.It is compared with the traditional method, the present invention makes full use of the concurrency of hardware, accelerates the arithmetic speed of support vector cassification, and kernel function computing module and classification computing module share computing resource, it supports hardware reconfiguration, there is considerable flexibility for the sample of different characteristic number.
Description
Technical field
The invention belongs to the hardware realization field of machine learning algorithm, more particularly to a kind of restructural more classification support to
Amount machine system.
Background technique
Support vector machines (support vector machine, SVM) is a kind of supervised machine learning algorithm, can be with
For analyzing data, recognition mode, data classification and regression analysis, utilization is very extensive.Algorithm of support vector machine can be divided into
Trained and decision two parts, wherein training is partially due to calculation is complicated, it is computationally intensive, usually in high-performance server
Upper progress；Decision part carries out classification or regression analysis to sample data using trained model, and calculating process includes big
The multiply-add operation of amount.Note input sample is x, and i-th of supporting vector is xi, corresponding coefficient is αi, corresponding label is
yi, it is biased to b, the kernel function between two vectors is k (xi,xj), then decision function are as follows:
The above process is directed to two classification problems.When carrying out data classification using support vector machines in realistic task, it can meet
To more classification problems.Currently, there are two main classes for the method for construction SVM multi-categorizer: one kind is direct method, directly constructs one
Model carries out more classification, and this method seems simply, but computation complexity is high, realizes difficult.Therefore it is generally adopted by indirectly
Method: this method is trained multiple two classification SVM, carries out more categorised decisions further according to certain mechanism.Common method has:
One-to-many method (one-versus-rest, OVR), one-to-one method (one-versus-one, OVO), Binomial Trees (Binary
Tree, BT), directed acyclic graph (Directed Acyclic Graph, DAG).
With the rapid development of IC industry, high-performance and real-time are the unremitting pursuits of built-in field.When
Before, for the hardware realization of support vector machines, largely it is all based on CPU, GPU platform is designed.Exist in CPU a large amount of
Cache and control unit, for the ALU negligible amounts of calculating, and GPU has a large amount of computing resource and storage resource, control
System is simple, therefore for computation-intensive algorithm, GPU ratio CPU has faster arithmetic speed, but the area of GPU is larger, energy
Consumption is very high, is not suitable for embedded real-time operation.
Summary of the invention
Present invention aims to overcome that in above-mentioned background technique support vector machines implementation deficiency, for decision portion
Point, more classification are carried out using one-to-one method, a kind of hardware realization of restructural multi-category support vector machines is proposed, supports meter
Calculation resource is restructural, and hardware can be cut, and parallel and water operation is supported, to improve calculating speed.Specifically by following technical side
Case is realized:
The restructural multi-category support vector machines system, comprising:
Main control module controls entire calculation process, provides the reconfiguration information of computing resource；
Storage control module controls the data storage of each computing module；
Kernel function computing module receives test data and supporting vector, and calculates between test data and supporting vector
Kernel function, and obtain operation result；
Classification computing module, according to the operation result, calculating kernel function multiplies with corresponding Lagrange multiplier coefficient
Accumulation result, then result will be multiplied accumulating plus biasing, decision value and classification are obtained, and corresponding classification poll is added 1；
As a result comparison module, the poll of more each classification obtain the final classification result of test data.
The further design of the restructural multi-category support vector machines system is that main control module receives system and opens
Dynamic signal, completes the sort operation of getting tickets of single vector machine, the operation specifically: starting kernel function computing module, kernel function fortune
Start classification computing module after the completion of calculating, classification is calculated, and pass through storage control module after respective classes poll is added 1
Storage, whether then main control module judgement is currently the last one support vector machines, if it is starts result and compares mould
Block obtains classification results to the end；Otherwise the sort operation of getting tickets of single vector machine is repeated.
The restructural multi-category support vector machines system it is further design be, the kernel function computing module and
Classification computing module shares computing resource, the topological structure and interconnected relationship of the basic processing unit by changing computing resource,
It is respectively formed kernel function computing module and classification computing module.
The further design of the restructural multi-category support vector machines system is that the basic processing unit includes
Adder, multiplier, accumulator and exponential operator.
The further design of the restructural multi-category support vector machines system is, in the kernel function computing module,
Computing resource is reconstructed into multiple parallel computation units, and each parallel computation unit is by adder, multiplier, accumulator and refers to
Number arithmetic unit composition, each parallel computation unit calculate a kernel function.
The further design of the restructural multi-category support vector machines system is that the classification computing module will be counted
Resources re engineering is calculated into the multiply-add tree construction of multi input, the calculating of decision value is accelerated parallel.
Beneficial effects of the present invention:
Restructural multi-category support vector machines system of the invention makes full use of the concurrency of hardware, supports computing resource
Restructural, hardware can be cut, and supported parallel and water operation, the arithmetic speed of support vector cassification is improved, for difference
The sample of characteristic also has considerable flexibility.
Detailed description of the invention
Fig. 1 is support vector machines hardware implementing architecture figure.
Fig. 2 is kernel function computing module structure chart.
Fig. 3 is kernel function computing module computing unit schematic diagram all the way.
Fig. 4 is supporting vector data arrangement schematic diagram.
Fig. 5 is kernel function data arrangement schematic diagram.
Fig. 6 is coefficient data arrangement schematic diagram.
Fig. 7 is the multiply-add tree structure diagram of classification computing module.
Specific embodiment
The present invention is described in detail with reference to the accompanying drawing.
Multi-category support vector machines system provided in this embodiment inputs a sample and support vector machines to be sorted
The supporting vector of model, calculate test vector (testvector, TV) and all supporting vectors (support vector, SV) it
Between kernel function, kernel function calculates complete after carry out decision, carry out kernel function value and coefficient of correspondence to multiply accumulating calculatings, completion
Afterwards plus biasing, decision value is obtained.For multi-category support vector machines, multiple models can be trained, decision value has been calculated
Cheng Hou judges whether to be the last one support vector machines, and if yes then enter result comparison module, more each model is determined
Plan value size, using the maximum classification of decision value as the classification of final test vector.Otherwise next supporting vector machine model is carried out
Calculating, calculating process with it is consistent before, different models have different supporting vectors and its corresponding coefficient.
The restructural multi-category support vector machines system of the present embodiment, mainly by main control module, storage control module,
Kernel function computing module, classification computing module and result comparison module composition.
Wherein, main control module receives system enabling signal, starts kernel function computing module, is started based on data delay
Classification computing module, obtains classification, adds 1 to be stored in storage unit respective classes poll.Control module is sentenced after the completion of calculating
Whether disconnected be currently the last one support vector machines, if it is starts result comparison module, otherwise starts kernel function next time
Computing module.When all support vector machines all calculate classification, after obtaining classification poll, as a result comparison module is obtained to the end
Classification results, main control module provides reconfiguration information simultaneously, passes through the topological structure for changing basic processing unit and interconnection is closed
Computing resource dynamic restructuring is calculating structure needed for nonidentity operation module by system.
Storage control module controls the storage of data, including supporting vector memory module, kernel function memory module, coefficient
Memory module, test vector storage unit, biasing storage unit and classification poll storage unit.
Kernel function computing module be responsible for receive main control module enabling signal, carry out kernel function calculating, overall structure by
Multiplex arithmetric unit forms parallel.When running the module, computing resource is reconstructed into arithmetic element, each arithmetic element successively into
Row subtraction, square, it adds up, is added and asks index operation with biasing, kernel function is calculated.The arithmetic element has 16 tunnels altogether,
Parallel computation goes out the kernel function of test vector and 16 supporting vectors.Test vector is stored in a storage unit, support to
Amount data are stored in 16 storage units, and all characteristic values of the same supporting vector are stored in a storage unit, often
A address stores two characteristic values, and first supporting vector is stored in storage unit 1, and second supporting vector is stored in storage
In unit 2, after being filled with 16 supporting vectors, the 17th supporting vector continues to be stored in storage unit 1, and so on.It calculates
After the completion, kernel function result is stored in 16 storage units, and each address stores two kernel functions, test vector and the
One, the kernel function of two supporting vectors is stored in storage unit 1, and test vector and the kernel function of third and fourth supporting vector are deposited
It is stored in storage unit 2, after being filled with 32 kernel functions, test vector and the kernel function of the 33rd, 34 supporting vector store again
In storage unit 1, and so on.
Classification computing module starts to start after receiving the end signal of kernel function computing module.When running the module, meter
Resources re engineering is calculated into multiply-add tree construction, when calculating take out parallel 32 kernel function values in 16 kernel function storage units and
32 coefficient values in 16 coefficient memory units, coefficient 1 and coefficient 2 are stored in storage unit 1, and coefficient 3 and coefficient 4 store
In storage unit 2, after being filled with 16 storage units, coefficient 33 and coefficient 34 continue to be stored in storage unit 1, and so on.
Kernel function and coefficient are carried out by add tree the result of multiplier tired respectively as the source data 1 and source data 2 of multiplier
Adding, specific multiply-add tree construction has 7 grades altogether, there are 32 multipliers respectively, 16 adders, 8 adders, 4 adders, 2
A adder, 1 adder and 1 accumulator.Biasing is added after all kernel function and coefficient multiply accumulating calculating,
Decision value and classification are obtained, the poll of corresponding classification is added 1, is stored in classification poll storage unit.
As a result comparing unit receive control module to enabling signal after start, compare that be stored in decision value storage single
All decision values of member, using classification belonging to maximum decision value as final classification.
A specific example is given below, which is made of 16 road computing units.Classification operation
Module is made of the multiply-add tree of 32 inputs.Test vector and supporting vector characteristic are all 8, and one shares 1,2,3 classifications,
In 1 class and 2 classes, one support vector machines of training, referred to as support vector machines 1, supporting vector number be 66；2 classes and 3 classes training one
A support vector machines, referred to as support vector machines 2, supporting vector number are 66；1 class and 3 classes, one support vector machines of training, claim
For support vector machines 3, supporting vector number is 66.Kernel function uses gaussian kernel function
Step 1) inputs the supporting vector data of sample to be tested and support vector machines 1, and supporting vector is single using 16 storages
Member is stored, and test sample is stored using 1 storage unit, and storage mode is consistent with supporting vector storage unit.
Step 2) sample and each component of supporting vector subtract each other, cumulative after the result square of subtraction, after the completion of adding up
Index then is sought multiplied by nuclear parameter, obtains kernel function.
After the completion of step 3) kernel function calculates, classification computing module is carried out.The input of multiply-add tree is respectively core in the module
Function and coefficient multiply accumulating calculating after the completion plus biasing, obtain classification, be 1 at this time for specific example discussed above,
The support vector machines of 2 classes, it is 1 class that decision, which goes out classification, and the poll of 1 class adds 1.
Step 4) judges to be currently not the last one support vector machines, repeat the above steps 1), step 2), step 3),
Classification until completing all support vector machines, it should be noted that supporting vector and coefficient respectively correspond accordingly when repeating every time
Support vector machines.
Step 5) compares the poll value of each classification of current class poll storage unit, wherein 1 class, 2 ticket, 2 class, 1 ticket, 3
0 ticket of class, final classification are 1 class.
The restructural multi-category support vector machines system of the present embodiment makes full use of the concurrency of hardware, supports to calculate
Resource is restructural, and hardware can be cut, and supports parallel and water operation, improves the arithmetic speed of support vector cassification, for
The sample of different characteristic number also has considerable flexibility.
More than, it is merely preferred embodiments of the present invention, but scope of protection of the present invention is not limited thereto, appoints
In the technical scope disclosed by the present invention, any changes or substitutions that can be easily thought of, all by what those familiar with the art
It is covered by the protection scope of the present invention.Therefore, protection scope of the present invention should be with scope of protection of the claims
It is quasi-.
Claims (6)
1. a kind of restructural multi-category support vector machines system, characterized by comprising:
Main control module controls entire calculation process, provides the reconfiguration information of computing resource；
Storage control module controls the data storage of each computing module；
Kernel function computing module receives test data and supporting vector, and calculates the core letter between test data and supporting vector
Number, and obtain operation result；
Classification computing module, according to the operation result, calculating kernel function multiplies accumulating with corresponding Lagrange multiplier coefficient
As a result, result will be multiplied accumulating again plus biasing, decision value and classification are obtained, and corresponding classification poll is added 1；
As a result comparison module, the poll of more each classification obtain the final classification result of test data.
2. restructural multi-category support vector machines system according to claim 1, it is characterised in that: main control module receives
System enabling signal completes the sort operation of getting tickets of single vector machine, the operation specifically: starting kernel function computing module, core
Start classification computing module after the completion of functional operation, classification is calculated, and is controlled after respective classes poll is added 1 by storage
Module storage, whether then main control module judgement is currently the last one support vector machines, if it is starts result and compares
Module obtains classification results to the end；Otherwise the sort operation of getting tickets of single vector machine is repeated.
3. restructural multi-category support vector machines system according to claim 1, it is characterised in that: the kernel function operation
Module and classification computing module share computing resource, by the topological structure and the interconnection that change the basic processing unit of computing resource
Relationship is respectively formed kernel function computing module and classification computing module.
4. restructural multi-category support vector machines system according to claim 3, it is characterised in that: the basic operation list
Member includes adder, multiplier, accumulator and exponential operator.
5. restructural multi-category support vector machines system according to claim 3, it is characterised in that: the kernel function operation
In module, computing resource is reconstructed into multiple parallel computation units, and each parallel computation unit is by adder, multiplier, accumulator
And exponential operator composition, each parallel computation unit calculate a kernel function.
6. restructural multi-category support vector machines system according to claim 3, it is characterised in that: the classification operation mould
Computing resource is reconstructed into the multiply-add tree construction of multi input by block, is accelerated parallel to the calculating of decision value.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910528797.5A CN110427960B (en) | 2019-06-18 | 2019-06-18 | Reconfigurable multi-classification support vector machine system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910528797.5A CN110427960B (en) | 2019-06-18 | 2019-06-18 | Reconfigurable multi-classification support vector machine system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110427960A true CN110427960A (en) | 2019-11-08 |
| CN110427960B CN110427960B (en) | 2023-04-25 |
Family
ID=68408684
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910528797.5A Active CN110427960B (en) | 2019-06-18 | 2019-06-18 | Reconfigurable multi-classification support vector machine system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110427960B (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160189056A1 (en) * | 2014-12-30 | 2016-06-30 | Battelle Memorial Institute | Fast efficient evaluation of messages on automotive networks using look-up tables |
| CN106845401A (en) * | 2017-01-20 | 2017-06-13 | 中国科学院合肥物质科学研究院 | A kind of insect image-recognizing method based on many spatial convoluted neutral nets |
| CN106874958A (en) * | 2017-02-28 | 2017-06-20 | 中南大学 | A kind of supporting vector machine model approximation method and its application based on multinomial fitting |
-
2019
- 2019-06-18 CN CN201910528797.5A patent/CN110427960B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160189056A1 (en) * | 2014-12-30 | 2016-06-30 | Battelle Memorial Institute | Fast efficient evaluation of messages on automotive networks using look-up tables |
| CN106845401A (en) * | 2017-01-20 | 2017-06-13 | 中国科学院合肥物质科学研究院 | A kind of insect image-recognizing method based on many spatial convoluted neutral nets |
| CN106874958A (en) * | 2017-02-28 | 2017-06-20 | 中南大学 | A kind of supporting vector machine model approximation method and its application based on multinomial fitting |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110427960B (en) | 2023-04-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhang et al. | BoostGCN: A framework for optimizing GCN inference on FPGA | |
| CN106779060A (en) | A kind of computational methods of the depth convolutional neural networks for being suitable to hardware design realization | |
| CN107908536B (en) | Performance evaluation method and system for GPU application in CPU-GPU heterogeneous environment | |
| CN109543662A (en) | Object detection method, system, device and the storage medium proposed based on region | |
| CN106228183A (en) | A kind of semi-supervised learning sorting technique and device | |
| CN109740731A (en) | A kind of adaptive convolutional layer hardware accelerator design method | |
| CN111914728A (en) | Hyperspectral remote sensing image semi-supervised classification method and device and storage medium | |
| CN108776833B (en) | Data processing method, system and computer readable storage medium | |
| CN106250909A (en) | A kind of based on the image classification method improving visual word bag model | |
| CN110147808A (en) | A kind of novel battery screening technique in groups | |
| CN116596095A (en) | Training method and device of carbon emission prediction model based on machine learning | |
| CN102945553A (en) | Remote sensing image partition method based on automatic difference clustering algorithm | |
| Ying et al. | Processor free time forecasting based on convolutional neural network | |
| CN111815209A (en) | Data dimension reduction method and device applied to wind control model | |
| CN113971090A (en) | Layered federal learning method and device of distributed deep neural network | |
| CN105955896B (en) | A kind of restructural DBF hardware algorithm accelerator and control method | |
| CN117635418B (en) | Training method for generating countermeasure network, bidirectional image style conversion method and device | |
| CN115035334A (en) | Multi-classification change detection method and system for multi-scale fusion double-temporal remote sensing image | |
| Watanakeesuntorn et al. | Massively parallel causal inference of whole brain dynamics at single neuron resolution | |
| Wang et al. | Evaluating computing performance of deep neural network models with different backbones on IoT-based edge and cloud platforms | |
| CN110807159B (en) | Data marking method and device, storage medium and electronic equipment | |
| CN110427960A (en) | A kind of restructural multi-category support vector machines system | |
| CN104573331A (en) | K neighbor data prediction method based on MapReduce | |
| Dufrechou et al. | Machine learning for optimal selection of sparse triangular system solvers on GPUs | |
| Pashentsev et al. | Applying big data and machine learning approach to identify noised data |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |