CN110110863A

CN110110863A - A kind of distributed machines study tune ginseng system based on celery

Info

Publication number: CN110110863A
Application number: CN201910400730.3A
Authority: CN
Inventors: 刘宇为; 夏凡; 饶雪云
Original assignee: Chengdu Univeristy of Technology
Current assignee: Chengdu Univeristy of Technology
Priority date: 2019-05-15
Filing date: 2019-05-15
Publication date: 2019-08-09

Abstract

Automatic seeking is learnt as the distributed machines of core using Celery task queue frame join system the invention discloses a kind of, comprising the following steps: whole system includes a host node and several from node.Host node generates initial parameter task according to the hyper parameter of configuration and is put into message-oriented middleware, message-oriented middleware is monitored from node, when listening to task, task is obtained from node and is run, after end of run, host node is sent to from node by operation result, host node is generated parameter task using parameter generation algorithm according to the result run and places into message-oriented middleware, and so on, until reaching the number of iterations of setting, system stops, and eventually finds proper hyper parameter.

Description

A kind of distributed machines study tune ginseng system based on celery

Technical field

The invention belongs to machine learning fields, are related to automaton study aspect, design a kind of machine based on celery Study is distributed to adjust ginseng system.

Background technique

With the continuous development of IT application to our society, all trades and professions can all generate a large amount of basic data.Excavating number When according to value, machine learning modeling technique can be used.During modeling, in addition to preparing data, what is most taken time and effort is exactly Various super ginseng combinations are attempted, the process of model optimum efficiency is found.Even with experienced algorithm engineering teacher and data section Scholar is sometimes also difficult to hold rule therein, can only repeatedly attempt.Since data volume is big, one time cut-and-try process will consume Take for a long time, along with hyper parameter number of combinations is big, the number of trial is also very more, when can expend very long for single machine trial Between.The time is modeled to shorten, it would be desirable to be employed computer cluster, be joined using multiple host come distributed execution automatic seeking Journey.

Summary of the invention

Ginseng, which is sought, the technical problem to be solved by the present invention is to conventional machines study uses grid search and random mostly Search, parameter search require a great deal of time without directiveness to find suitable hyper parameter, this system uses distributed Mode search for hyper parameter, and the generation of hyper parameter is guidance with previous training result, greatly shorten parameter optimization Time.

In order to realize the target of foregoing invention, the present invention provides following technical schemes:

Machine learning distribution tune ginseng system based on celery, comprising the following steps:

S1, code and data is submitted to specified directory and to be distributed to distance host by web interface；

S2, page configuration hyper parameter type and range；

S3, system generate initial parameter task and are put into message-oriented middleware；

S4, task is obtained from node to message-oriented middleware and is executed；

S5, after having executed from node, result is returned into host node；

S6, the host node trained result before generate hyper parameter task using parameter generation algorithm and store to message Middleware supplies to consume from node.

S7, when frequency of training reaches specified value, whole system just stops.

By above step, suitable hyper parameter is finally found with the shorter time.

Main advantages of the present invention are using realizing distributed operation based on celery task queue frame, among message Part uses redis, and system coupling is good, and scalability is strong.After the completion of its subsystem ought be trained once when running, it can will transport Row result feeds back to host and the parameter of next round is instructed to generate, and generates parameter directional, shortens the parameter optimization time.In addition This system is operated using web page, and code configuration parameter need to be only submitted on the page, so that it may task is run, it is simple and convenient.

Detailed description of the invention

Fig. 1 system general frame.

Fig. 2 distributed structure/architecture figure.

Fig. 3 parameter task generation module.

Specific embodiment

Specific implementation design scheme of the invention is described in further detail below with reference to Fig. 1 Fig. 2 Fig. 3.

As shown in figure one, figure two, this system realizes distributed system using celery task queue frame, among message That part relies on is redis.Whole system includes a host node master and several from node work, and host node mainly manages Entire cluster and parameter task generate, and the parameter task of generation is put into message-oriented middleware, from node monitoring information middleware, When there is task, acquisition task is simultaneously executed, and thereby realizes distribution.

As shown in Figure 3, when having executed a task from node, implementing result can be sent to host node, host point root According to task of having completed as a result, using parameter optimization algorithm, the optimization algorithms such as Bayes, TPE can be used and generate hyper parameter Task realizes the directional guidance of generation of hyper parameter.

Claims

1. a kind of machine learning distribution tune ginseng system based on celery, which comprises the following steps:

S2, page configuration hyper parameter type and range；

S3, system generate initial parameter task according to configuration and are put into message-oriented middleware；

S5, after having executed task from node, result is returned into host node；

S6, the host node trained result before continue generation hyper parameter task using parameter generation algorithm and store to message Middleware supplies to consume from node；

2. the machine learning distribution tune ginseng system according to claim 1 based on celery, which is characterized in that step S1 The management of middle cluster uses socket communication protocol to complete the mutual access between host, and efficiency of transmission is high, stability It is good.

3. the machine learning distribution tune ginseng system according to claim 1 based on celery, which is characterized in that step S3 , in S4, S5, generation and the consumption of task use celery to dispatch, and using redis as message-oriented middleware, facilitate decoupling It closes.

4. the machine learning distribution tune ginseng system according to claim 1 based on celery, which is characterized in that hyper parameter Generating with housebroken result is foundation guide parameters generation, while the part is integrated with many kinds of parameters optimizing algorithm, including Bayes, TPE, Anneal etc., can also artificial spreading parameter optimizing algorithm, shorten the parameter optimization time.