CN110110861A - Determine method and apparatus, the storage medium of model hyper parameter and model training - Google Patents

Abstract

The embodiment of the present disclosure provides the technology and image processing techniques of a kind of determining model hyper parameter, is conducive to the image procossing performance of hoisting machine learning model, wherein the method for determining model hyper parameter comprises determining that the initial value of hyper parameter；According to the initial value of the hyper parameter and sample graph image set, M1 repetitive exercise is carried out to initial machine learning model by each path in parallel multiple paths, obtain each path first updates machine learning model, first based on each path in the multiple path updates the performance parameter of machine learning model, and the numerical value of the hyper parameter is updated to the first updated value；The first updated value and the sample graph image set based on the hyper parameter, the further numerical value update that machine learning model carries out M2 repetitive exercise and the hyper parameter is updated to the first of the multiple path, until reaching default cut-off condition, the final numerical value of the hyper parameter is obtained.

Description

Determine method and apparatus, the storage medium of model hyper parameter and model training

Technical field

This disclosure relates to machine learning techniques, and in particular to determine model hyper parameter and model training method and apparatus, Storage medium.

Background technique

In recent years, the machine learning models such as deep neural network various computer visions application in achieve significantly at Function.Under the driving of a large amount of flag datas, the performance of network has also reached surprising level.However, machine learning model is super Parameter is presently mainly to use engineer, after the good model hyper parameter of engineer, keeps the hyper parameter constant, to engineering It practises model to be trained, finally obtains the model parameter of machine learning model, it is finally obtained due to the limitation of engineer The performance of machine learning model need to be advanced optimized.

Summary of the invention

In view of this, the disclosure at least provides the technology and model training technology of a kind of determining model hyper parameter.

In a first aspect, providing a kind of method of determining model hyper parameter, which comprises determine the initial of hyper parameter Value；According to the initial value of the hyper parameter and sample graph image set, by each path in parallel multiple paths to initial machine Learning model carries out M1 repetitive exercise, and obtain each path first updates machine learning model, wherein the multiple The training parameter in different paths has the different numerical value sampled based on the hyper parameter in path, and M1 is greater than or equal to 1 and be less than or equal to the first numerical value；First based on each path in the multiple path updates the performance of machine learning model The numerical value of the hyper parameter is updated to the first updated value by parameter；The first updated value and the sample based on the hyper parameter Image set, to the first of the multiple path update machine learning model carry out M2 repetitive exercise and the hyper parameter into one Step Numerical updates, until reach default cut-off condition, obtains the final numerical value of the hyper parameter, wherein M2 be greater than or equal to 1 and Less than or equal to the first numerical value.

In one possible implementation, machine learning model is updated in first to the multiple path to carry out Before the further numerical value of M2 repetitive exercise and the hyper parameter updates, further includes: the first more new engine from multiple paths First object is chosen in learning model updates machine learning model；Machine learning model is updated by the first of the multiple path Model parameter is updated to the model parameter that the first object updates machine learning model.

It is in one possible implementation, described from the multiple road in conjunction with any embodiment that the disclosure provides The first of diameter, which updates, chooses first object update machine learning model in machine learning model, comprising: be based on the multiple path The first performance parameter for updating machine learning model, the is chosen from the first of the multiple path the update machine learning model One target update machine learning model.

It is in one possible implementation, described according to the super ginseng in conjunction with any embodiment that the disclosure provides Several initial values and sample graph image set carries out initial machine learning model M1 times by each path in parallel multiple paths Repetitive exercise, obtain each path first update machine learning model, comprising: initial value based on the hyper parameter and At least one first sample image in the sample image collection, by each path in multiple paths to the initial machine It practises model and carries out the first repetitive exercise, the first inner ring for obtaining each path updates machine learning model；Based on described super The second sample image of at least one of the initial value of parameter and the sample image collection, passes through each road in the multiple path Diameter updates machine learning model to first inner ring in each path and carries out secondary iteration training, obtains each path Second inner ring updates machine learning model；The second inner ring based on each path in the multiple path updates machine learning mould Type, obtain each path first update machine learning model.

It is in one possible implementation, described to be based on the super ginseng in conjunction with any embodiment that the disclosure provides At least one first sample image in several initial values and the sample image collection, by each path in multiple paths to institute It states initial machine learning model and carries out the first repetitive exercise, the first inner ring for obtaining each path updates machine learning mould Type, comprising: the initial value based on the hyper parameter carries out multiple repairing weld, obtains first instruction in each path in the multiple path Practice parameter；At least one of first training parameter and the sample image collection based on each path in the multiple path One sample image carries out the first repetitive exercise to the initial machine learning model, and the first inner ring for obtaining each path updates Machine learning model.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the of each path The training parameter used in one repetitive exercise and secondary iteration training is that the initial value based on the hyper parameter carries out different samplings It obtains.

It is in one possible implementation, described based on the multiple in conjunction with any embodiment that the disclosure provides The first of each path updates the performance parameter of machine learning model in path, and the numerical value of the hyper parameter is updated to first more New value, comprising: first based on each path in the multiple path updates the performance parameter of machine learning model, determine described in The model modification parameter in each path；The model modification parameter in the multiple path is averaging processing, average update is obtained Parameter；According to the average undated parameter, the numerical value of the hyper parameter is updated to the first updated value.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, described based on described more The first of each path updates the performance parameter of machine learning model in a path, and the numerical value of the hyper parameter is updated to first Before updated value, the method also includes: the property of machine learning model is updated to first of each path in the multiple path Energy parameter is normalized；First based on each path in the multiple path updates the property of machine learning model Energy parameter, is updated to the first updated value for the numerical value of the hyper parameter, comprising: described in obtaining after the normalized The first of each path updates the performance parameter of machine learning model in multiple paths, and the numerical value of the hyper parameter is updated to the One updated value.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the performance parameter includes Accuracy rate.

It is in one possible implementation, described to be based on the super ginseng in conjunction with any embodiment that the disclosure provides The first several updated value and the sample graph image set update machine learning model to the first of the multiple path and carry out M2 times repeatedly In generation, the trained further numerical value with the hyper parameter updated, comprising: the first updated value and the sample image based on hyper parameter Collection updates machine learning model to first of each path in the multiple path and carries out M2 repetitive exercise, obtains described every The second of a path updates machine learning model；Second based on each path in the multiple path updates machine learning model Performance parameter, the numerical value of the hyper parameter is updated to the second updated value.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the hyper parameter includes using In the enhancing distribution parameter for carrying out image enhancement processing to the sample graph image set；The initial value according to the hyper parameter and Sample graph image set carries out M1 repetitive exercise, packet to initial machine learning model by each path in parallel multiple paths It includes: according to the enhancing distribution parameter, determining enhancing probability distribution, include multiple images enhancing behaviour in the enhancing probability distribution The probability of work；Based on the enhancing probability distribution, by sampling parallel multiple paths in the multiple data enhancement operations In the target data in each path enhance operation, at least one sample image in each path is carried out at image enhancement Reason obtains at least one enhancing image；Based at least one enhancing image in each path in the multiple path, to described first Beginning machine learning model carries out M1 repetitive exercise.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the described first more new engine The acquisition of the performance parameter of learning model, including handling as follows: updating machine by first of each path in the multiple path Device learning model handles at least one test image in test image collection, obtains processing result image；Based on described The corresponding described image processing result in each path in multiple paths, obtain each path first update machine learning model Performance parameter.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the default cut-off condition Including in following at least one of: default update times are reached to the update times of the hyper parameter；Alternatively, the multiple path The performance of obtained update machine learning model reaches target capabilities.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the method also includes: from Target machine is chosen in the final updated machine learning model in the multiple paths obtained in the case where reaching the default cut-off condition Device learning model, wherein the target machine learning model is the trained machine learning model for image procossing.

Second aspect provides a kind of device of determining model hyper parameter, and described device includes: initialization module, for true Determine the initial value of hyper parameter；Model training module, for the initial value and sample graph image set according to the hyper parameter, by parallel Multiple paths in each path to initial machine learning model carry out M1 repetitive exercise, obtain each path first update Machine learning model, wherein the training parameter in different paths has in the multiple path is sampled based on the hyper parameter Obtained different numerical value, M1 are greater than or equal to 1 and are less than or equal to the first numerical value；Super ginseng update module, for based on described more The first of each path updates the performance parameter of machine learning model in a path, and the numerical value of the hyper parameter is updated to first Updated value；Super ginseng obtains module, for based on the hyper parameter the first updated value and the sample graph image set, to the multiple The first of path updates the further numerical value update that machine learning model carries out M2 repetitive exercise and the hyper parameter, until reaching To default cut-off condition, the final numerical value of the hyper parameter is obtained, wherein M2 is greater than or equal to 1 and is less than or equal to the first number Value.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the super ginseng obtains module, It is also used to: updating machine learning model in first to the multiple path and carry out M2 repetitive exercise and the hyper parameter Further numerical value update before, updated from the first of the multiple path and choose first object in machine learning model and update machine Device learning model；The model parameter of the first update machine learning model in the multiple path is updated to the first object more The model parameter of new engine learning model.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the super ginseng obtains module, When choosing first object update machine learning model in machine learning model for updating from the first of the multiple path, packet Include: first based on the multiple path updates the performance parameter of machine learning model, the first more new engine from multiple paths First object is chosen in learning model updates machine learning model.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the model training module, Be specifically used for: at least one first sample image in initial value and the sample image collection based on the hyper parameter passes through Each path carries out the first repetitive exercise to the initial machine learning model in multiple paths, obtains the of each path One inner ring updates machine learning model；At least one of initial value and the sample image collection based on the hyper parameter second Sample image, by each path in the multiple path to first inner ring in each path update machine learning model into The training of row secondary iteration, the second inner ring for obtaining each path update machine learning model；Based in the multiple path Second inner ring in each path updates machine learning model, and obtain each path first updates machine learning model.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the model training module, When the first inner ring for obtaining each path updates machine learning model, comprising: based on the initial of the hyper parameter Value carries out multiple repairing weld, obtains first training parameter in each path in the multiple path；Based on every in the multiple path At least one first sample image in first training parameter in a path and the sample image collection, to the initial machine It practises model and carries out the first repetitive exercise, the first inner ring for obtaining each path updates machine learning model.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the of each path The training parameter used in one repetitive exercise and secondary iteration training is that the initial value based on the hyper parameter carries out different samplings It obtains.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the super ginseng update module, Be specifically used for: based in the multiple path each path first update machine learning model performance parameter, determine described in The model modification parameter in each path；The network undated parameter in the multiple path is averaging processing, average update is obtained Parameter；According to the average undated parameter, the numerical value of the hyper parameter is updated to the first updated value.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the super ginseng update module, It is specifically used for: place is normalized to the performance parameter of the first update machine learning model in each path in the multiple path Reason；First based on each path in the multiple path obtained after normalized updates the performance ginseng of machine learning model Number, is updated to the first updated value for the numerical value of the hyper parameter.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the performance parameter includes Accuracy rate.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the super ginseng obtains module, It is specifically used for: the first updated value and the sample graph image set based on the hyper parameter, to each path in the multiple path First update machine learning model carry out M2 repetitive exercise, obtain each path second update machine learning mould Type；Second based on each path in multiple paths updates the performance parameter of machine learning model, by the numerical value of the hyper parameter It is updated to the second updated value.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the hyper parameter includes using In the enhancing distribution parameter for carrying out image enhancement processing to the sample graph image set；The model training module, is specifically used for: root According to the enhancing distribution parameter, enhancing probability distribution is determined, include multiple images enhancing operation in the enhancing probability distribution Probability；It is every in parallel multiple paths by being sampled in the multiple data enhancement operations based on the enhancing probability distribution The target data in a path enhances operation, carries out image enhancement processing at least one sample image in each path, obtains Enhance image at least one；Based at least one enhancing image in each path in the multiple path, to the initial machine Device learning model carries out M1 repetitive exercise.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the super ginseng update module, It is also used to obtain the acquisition of the performance parameter of the first update machine learning model, including handles as follows: by the multiple path In each path first update machine learning model at least one test image in test image collection is handled, obtain Processing result image；Based on the corresponding described image processing result in path each in the multiple path, each path is obtained First updates the performance parameter of machine learning model.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the default cut-off condition Including in following at least one of: default update times are reached to the update times of the hyper parameter；Alternatively, the multiple path The performance of obtained update machine learning model reaches target capabilities.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the super ginseng obtains module, The final updated machine learning model in the multiple path for being also used to obtain in the case where reach the default cut-off condition Middle selection target machine learning model, wherein the target machine learning model is the trained machine for image procossing Learning model.

In conjunction with any embodiment that the disclosure provides, in one possible implementation, the super ginseng obtains module, It is also used to after the final numerical value for obtaining hyper parameter, the final numerical value based on the hyper parameter, training initialization model parameter Initial machine learning model, obtain training completion target machine learning model.

The third aspect additionally provides a kind of training method of machine learning model, comprising: is based on above-mentioned any embodiment In the method for hyper parameter of determination machine learning model obtain the final numerical value of hyper parameter, and based on the final of the hyper parameter Numerical value, training have the initial machine learning model of original model parameter, obtain target machine learning model.

The device that the embodiment of the present disclosure additionally provides for realizing the training method.

Fourth aspect provides a kind of image processing method, which comprises obtains image to be processed；Utilize engineering It practises model to handle the image to be processed, obtains processing result image, wherein the hyper parameter of the machine learning model What the hyper parameter training that the method for being based upon cover half type hyper parameter really described in disclosure any embodiment determines obtained.

5th aspect provides a kind of electronic equipment, and the electronic equipment includes memory, processor, and the memory is used In the computer instruction that storage can be run on a processor, the processor is used to realize this when executing the computer instruction Really the training method of the method or machine learning model of cover half type hyper parameter described in open any embodiment.

6th aspect, provides a kind of computer readable storage medium, is stored thereon with computer program, described program is located Manage the instruction that the method or machine learning model of cover half type hyper parameter really described in disclosure any embodiment are realized when device executes Practice method.

7th aspect, provides a kind of training system, comprising: parameter management server and controller, wherein parameter management clothes Business device is used to manage and update the numerical value of hyper parameter.Controller be used for the machine learning model based on hyper parameter carry out circulation or Iteration updates, and the performance parameter for updating foundation as hyper parameter is fed back to parameter management server, so that parameter management Server carries out the update of hyper parameter accordingly.

The embodiment of the present disclosure provides the technology of cover half type hyper parameter really, by being based on multiple paths to machine learning model Be iterated training, update the hyper parameter after having carried out M1 repetitive exercise to model, and based on updated hyper parameter after Continuous be iterated by multiple paths to model is trained and the further update of hyper parameter numerical value, by it is this with numerical value more Newly-service check is the mode of cycling element come the hyper parameter for determining machine learning model, accelerates the optimization efficiency of hyper parameter, And hyper parameter is updated by the performance parameter based on machine learning model, enables the search of hyper parameter with machine learning mould The performance parameter of type is optimization direction, to promote the performance of the machine learning model based on determining hyper parameter.

Detailed description of the invention

Technical solution in order to illustrate more clearly of disclosure one or more embodiment or in the related technology, below will be right Attached drawing needed in embodiment or description of Related Art is briefly described, it should be apparent that, be described below in it is attached Figure is only some embodiments recorded in disclosure one or more embodiment, for those of ordinary skill in the art, Without any creative labor, it is also possible to obtain other drawings based on these drawings.

Fig. 1 shows a kind of method of determining model hyper parameter of at least one embodiment of disclosure offer；

Fig. 2 shows the methods that the another kind that at least one embodiment of the disclosure provides determines model hyper parameter；

Fig. 3, which shows at least one embodiment of the disclosure and provides the application scenarios of the method for cover half type hyper parameter really, to be shown Example；

Fig. 4 shows another method for determining model hyper parameter of at least one embodiment of disclosure offer；

Fig. 5 shows a kind of model training process of at least one embodiment of disclosure offer；

Fig. 6 shows a kind of device of determining model hyper parameter of at least one embodiment of disclosure offer；

Fig. 7 shows a kind of process of image processing method of at least one embodiment of disclosure offer；

Fig. 8 shows a kind of training method of machine learning model of at least one embodiment of disclosure offer；

Fig. 9 shows a kind of training device of machine learning model of at least one embodiment of disclosure offer.

Specific embodiment

In order to make those skilled in the art more fully understand the technical solution in disclosure one or more embodiment, under Face will combine disclosure one or more embodiment in attached drawing, to the technical solution in disclosure one or more embodiment into Row clearly and completely describes, it is clear that described embodiment is only disclosure a part of the embodiment, rather than whole realities Apply example.Based on disclosure one or more embodiment, those of ordinary skill in the art are without creative efforts The range of disclosure protection all should belong in every other embodiment obtained.

The embodiment of the present disclosure provides the method and model training method of cover half type hyper parameter really, can be applied to engineering Practise the training platform of model, such as cloud training platform or end training platform, wherein the training platform may include one or more Equipment, correspondingly, the above method can be executed by cloud device, the network equipment or terminal device etc., and the embodiment of the present disclosure is to this Without limitation.In order to make it easy to understand, below in this way by being described for training equipment.

In the embodiment of the present disclosure, based on designing reasonable search space and using efficient way of search, automation is searched The reasonable hyper parameter of rope, can be improved the performance of the machine learning model based on the hyper parameter.

In some embodiments, constant current journey includes repeatedly circulation to hyper parameter really, wherein in each cycle, training is set The standby current value based on hyper parameter is sampled, and the training parameter that each path uses in multiple paths is obtained, and being based on should Training parameter carries out the model parameter adjustment of the initial machine learning model of this circulation, obtains updated machine learning mould Type, then, based on the performance parameter of updated machine learning model in multiple paths, the numerical value for carrying out hyper parameter updates, more Determination of the hyper parameter for the training parameter in multiple paths in subsequent cycle after new.

In the embodiments of the present disclosure, machine learning model can be neural network or other are trained based on hyper parameter Model, the embodiment of the present disclosure do not limit this.

In some embodiments, hyper parameter can be the parameter for obtaining loss function, or including increasing for data Strong parameter, etc., the embodiment of the present disclosure do not limit this.

The embodiment of the present disclosure provides a kind of method of determining model hyper parameter, and this method is intended to quickly search performance Preferable hyper parameter, and the training process of the search process of hyper parameter and model is combined together and carries out simultaneously by this method. Referring to Figure 1, Fig. 1 shows a kind of method of determining model hyper parameter of at least one embodiment of disclosure offer.

In step 100, the initial value of hyper parameter is determined.

Optionally, step 100~104 can correspond to the first time circulation process in all circulations, correspondingly, in step In 100, hyper parameter can be initialized, such as assigns the hyper parameter initial value.Alternatively, step 100~104 can correspond to all follow Certain intercycle process in ring correspondingly recycles the numerical value for the hyper parameter that last circulation obtains as this The initial value of hyper parameter, the embodiment of the present disclosure do not limit this.

In a step 102, according to the initial value of the hyper parameter and sample graph image set, by every in parallel multiple paths A path carries out M1 repetitive exercise to initial machine learning model, and obtain each path first updates machine learning mould Type.

It is alternatively possible to by multiple parallel paths repetitive exercise machine learning model, such as uniform machinery is learnt Model is distributed to multiple processing units or thread, obtains multiple machine learning models in multiple processing units or thread, each The process object of machine learning model can be at least one sample image in sample image collection.

Training equipment can be sampled based on the hyper parameter, and the training parameter of model is obtained, wherein the sampling can be with It carries out repeatedly, to obtain the training parameter that each path uses in multiple paths.The training parameter in different paths in multiple paths It can be and carry out what different samplings obtained based on the hyper parameter, correspondingly, different paths can have different training parameter numbers Value.

After obtaining training parameter, machine learning model is carried out in conjunction with the training parameter and at least one sample image Repetitive exercise.

In order to make it easy to understand, the model in this being recycled before repetitive exercise is known as initial machine learning model, And the model obtained after repetitive exercise is known as the first update machine learning model.The repetitive exercise number of model can be M1 Secondary, which may be greater than or be equal to 1 and be less than or equal to the first numerical value.Specifically, the number of the repetitive exercise can be one Secondary or multiple, for example, the numerical value of M1 is 50 or 30 or other numerical value, which can preset or based on machine learning model Current performance determine that or determine by other means, correspondingly, repetitive exercise number in difference circulation can be identical or not Together, the embodiment of the present disclosure does not limit this.

In addition, in one cycle, multiple paths can carry out the repetitive exercise of same number to machine learning model, or Person, different paths can also carry out the repetitive exercise of different numbers to machine learning model, and the embodiment of the present disclosure does not limit this It is fixed.

In some embodiments, it is assumed that M1 is the integer greater than 1, in each path, processing unit or at least one line Journey can carry out an iteration to initial machine learning model based on training parameter, obtain network parameter machine learning adjusted Model, and using network parameter machine learning model adjusted as the input of following iteration, wherein in following iteration, It can be sampled to obtain another training parameter based on hyper parameter, M1 iteration is carried out back and forth with this, obtains the first of the path Update machine learning model.

In addition, parallel multiple paths, the embodiment of the present disclosure are not intended to limit the quantity in path.Empirical experimentation is adopted Different number of paths is taken, influence might have on the performance of model training, in certain quantitative range, the increasing of number of paths Add the performance of meeting lift scheme；After number of paths reaches certain numerical value, the influence to the performance of model training may also can drop It is low.It can integrate and be considered according to the consumption of model performance and computing resource, determine selected number of paths.

At step 104, first based on each path in the multiple path updates the performance ginseng of machine learning model Number, is updated to the first updated value for the numerical value of the hyper parameter.

The update of hyper parameter can be the performance parameter that machine learning model is updated based on first in multiple paths, one In a little embodiments, the mode based on intensified learning carries out the update of hyper parameter, for example, the purpose for updating hyper parameter is so that being based on The hyper parameter can train to obtain performance parameter more preferably model.

The updated numerical value of hyper parameter can be known as the first updated value by training equipment.In this way, just completing hyper parameter more New one cycle.

In step 106, the first updated value based on the hyper parameter and the sample graph image set, to the multiple path First update machine learning model and carry out the further numerical value of M2 repetitive exercise and the hyper parameter and update, up to reaching pre- If cut-off condition, the final numerical value of the hyper parameter is obtained.

It optionally, can be based on the first updated value and sample image of hyper parameter after step 104 updates hyper parameter Collection, continues to be iterated update to the machine learning model in multiple paths.For example, changing to the first update machine learning model It can be M2 times for frequency of training, which may be greater than or be equal to 1 and be less than or equal to the first numerical value, and M2 and M1 can phase Together, it can also be different.

With the continuation repetitive exercise of machine learning model, hyper parameter also will further update numerical value, likewise, for example, It can continue the numerical value for updating hyper parameter based on the performance parameter of the machine learning model after continuation repetitive exercise.Optionally, when When reaching preset cut-off condition, the final numerical value of hyper parameter can be obtained, which is the preferably super ginseng searched Number numerical value.For example, the preset cut-off condition, which can be, reaches default update times to the update times of the hyper parameter, or The performance of person, the update machine learning model that the multiple path obtains reach target capabilities.

Really the method for cover half type hyper parameter in some embodiments, by based on multiple paths to machine learning model into Row iteration training is updated the hyper parameter after having carried out M1 repetitive exercise to model, and is continued based on updated hyper parameter Trained and hyper parameter numerical value further update is iterated to model by multiple paths, by this with numerical value update- Service check is the mode of cycling element come the hyper parameter for determining machine learning model, accelerates the optimization efficiency of hyper parameter, and And hyper parameter is updated by the performance parameter based on machine learning model, enable the search of hyper parameter with machine learning model Performance parameter be optimization direction, thus promoted based on determining hyper parameter machine learning model performance.

Fig. 2 shows the method that the another kind that at least one embodiment of the disclosure provides determines model hyper parameter, the implementations Example describes the processing mode of more detailed search model hyper parameter.As shown in Fig. 2, this method may include:

In step 100, the initial value of hyper parameter is determined.

In a step 102, according to the initial value of the hyper parameter and sample graph image set, by every in parallel multiple paths A path carries out M1 repetitive exercise to initial machine learning model, and obtain each path first updates machine learning mould Type.

Optionally, M1 repetitive exercise is carried out to the initial machine learning model in each path and obtains the first update engineering Model is practised, may include following process:

For example, can be by least one first sample image in initial value and sample image collection based on hyper parameter to first Beginning machine learning model carries out the model obtained after the first repetitive exercise, and referred to as the first inner ring updates machine learning model.It is obtaining During obtaining first inner ring update machine learning model, multiple repairing weld can be carried out based on the initial value of hyper parameter, obtained First training parameter in each path in multiple paths is then based on first training parameter and sample graph image set to initial machine Learning model carries out the first repetitive exercise, and the first inner ring for obtaining each path updates machine learning model.

Then, can the second sample image of at least one of initial value based on hyper parameter and sample image collection, pass through Each path updates machine learning model progress secondary iteration instruction to first inner ring in each path in the multiple path Practice, obtained more new model is properly termed as the second inner ring and updates machine learning model.

The second inner ring based on each path in the multiple path updates machine learning model, continues iteration and updates, directly Machine learning model is updated to obtain each path first.Wherein, first repetitive exercise in each path and The training parameter that uses is that the initial value based on the hyper parameter carries out different samplings and obtains in secondary iteration training.

At step 104, first based on each path in the multiple path updates the performance ginseng of machine learning model Number, is updated to the first updated value for the numerical value of the hyper parameter.

In some embodiments, the performance parameter based on model updates hyper parameter, may include: based on the multiple path In each path first update machine learning model performance parameter, determine the model modification parameter in each path.It is right The model modification parameter in the multiple path is averaging processing, and obtains average undated parameter.And joined according to the average update Number, is updated to the first updated value for the numerical value of the hyper parameter.

Wherein, the acquisition of the performance parameter of model can update engineering by first of each path in multiple paths It practises model to handle at least one test image in test image collection, obtains processing result image；And based on described The processing result image in multiple paths obtains the first of each path performance parameter for updating machine learning model.For example, this property Energy parameter can be the accuracy rate of model.

In addition, before the numerical value of hyper parameter is updated to the first updated value, it can also be to each road in multiple paths The performance parameter of first update machine learning model of diameter is normalized.And it is obtained after being based on the normalized Performance parameter, the numerical value of hyper parameter is updated to the first updated value.

In step s 106, it is updated from the first of the multiple path and chooses first object update machine in machine learning model The model parameter of the first update machine learning model in the multiple path is updated to the first object more by device learning model The model parameter of new engine learning model.

Optionally, before carrying out the repetitive exercise of model next time, the machine learning model in multiple paths can be unified Model parameter.For example, the model parameter of machine learning model can be updated based on the first of multiple path, from multiple first It updates and chooses a model in machine learning model as first object update machine learning model, and by the multiple path The model parameter of model is updated to the model parameter that the first object updates machine learning model.Illustratively, the mould Type performance parameter includes but is not limited to accuracy rate of the model on verifying collection.

In step 106, the first updated value based on the hyper parameter and the sample graph image set, to the multiple path First update machine learning model and carry out the further numerical value of M2 repetitive exercise and the hyper parameter and update, up to reaching pre- If cut-off condition, the final numerical value of the hyper parameter is obtained.

For example, when updating the further repetitive exercise of machine learning model to first, can based on hyper parameter first more New value and the sample graph image set update machine learning model to first of each path in the multiple path and carry out M2 times repeatedly Generation training, obtain each path second update machine learning model.Also, likewise, based on each in multiple paths The second of path updates the performance parameter of machine learning model, and the numerical value of the hyper parameter is updated to the second updated value.

As above, the update of hyper parameter and the update of model parameter carry out simultaneously, in the training process of model, execute The performance parameter of interim feedback model training, and hyper parameter is updated based on the performance parameter.

In addition, may exist two kinds of situations after the hyper parameter after being optimized:

For example, trained model can be obtained simultaneously.The institute obtained in the case where reach the default cut-off condition It states in the final updated machine learning model in multiple paths, chooses target machine learning model, wherein the target machine learns mould Type is the trained machine learning model for image procossing.

In another example can use the hyper parameter of final optimization pass, re -training obtains model as finally trained engineering Practise model.I.e. based on the final numerical value of the hyper parameter, the initial machine learning model of training initialization model parameter is instructed Practice the target machine learning model completed.

Really the method for cover half type hyper parameter in some embodiments, by the training process to machine learning model In, while hyper parameter is updated, accelerate the search efficiency of hyper parameter；Also, this method is with engineering when updating hyper parameter The performance parameter of model is practised as optimization foundation, to ensure that the effect of hyper parameter, realize both fast and sound search compared with Excellent hyper parameter.

Really the method for cover half type hyper parameter as above, the chess game optimization of the hyper parameter suitable for any machine learning.It is as follows Will be by taking one of hyper parameter as an example, which is the enhancing point for carrying out image enhancement processing to the sample graph image set Cloth parameter illustrates the implementation procedure of this method with this, but it is understood that, this method is not limited to the excellent of enhancing distribution parameter Change.

Data enhancing strategy can be applied to the training process to network, by using data enhancing strategy to the defeated of network Enter data and carry out data enhancing, the overfitting problem in network training process can be improved.For example, can be to the picture number of input According to the data enhancing such as being rotated, shearing, move, new training data is obtained for training network, helps to improve network Generalization ability improves the accuracy rate of neural network forecast.

And the data enhancing strategy used is different, the training effect of network is also different.For example, being enhanced using a certain data The generalization ability of strategy, trained network is slightly worse, and accuracy rate of the network on verifying collection is relatively low.And another data is used to increase Strong strategy, the accuracy rate of trained network can get a promotion.Therefore, a kind of preferably data enhancing strategy is searched for, for instruction Get the network important role of better performances.

The embodiment of the present disclosure provides a kind of optimization method of data enhancing strategy, this method be a kind of automation search for compared with The method of excellent data enhancing strategy, this method described below:

Firstly, illustrating some basic contents in order to which the description to this method is clearer:

Search space (Search Space) and data enhancement operations: in network training, network inputs data are carried out Which data enhancement operations can be using preset some data enhancement operations, be used by selecting in these preset operations, Preset multiple operations can be known as " search space ".

For example, image can be rotated when the input data of network is image, color adjustment, shear, is mobile etc. The single processing such as " rotation ", " movement " can be known as data enhancing element (an augmentation by a variety of processing Element), the group of two kinds of data enhancing elements can be collectively referred to as " data enhancement operations " (an by the disclosure augmentation operation).Assuming that data enhancing element quantity have 36, it would be possible that combination of two quantity That is the quantity K=36 of data enhancement operations², i.e., include K data enhancement operations in search space.For each of network inputs A image can be applied to progress data enhancing in the image by selecting a data enhancement operations in the search space.

Wherein, in some embodiments, data enhancing element includes but is not limited to: horizontal shear (HorizontalShear), vertical shear (VerticalShear), move horizontally (HorizontalTranslate), be vertical Mobile (VerticalTranslate), rotation (Rotate), hue adjustment (ColorAdjust), tone separation (Posterize), exposure-processed (Solarize), comparison processing (Contrast), Edge contrast (Sharpness), at blast Manage (Brightness), automatic comparison (AutoContrast), hue balancing (Equalize), inversion processing (Invert) Deng.

Enhance distribution parameter (augmentation distribution parameter θ) and probability distribution (pθ): the increasing Strong distribution parameter can be a numerical value, each data enhancement operations can correspond to the numerical value of an enhancing distribution parameter. And probability distribution is converted to according to the enhancing distribution parameter, probability distribution is that each data enhancement operations are corresponding Enhancing distribution parameter numerical value be all converted to the numerical value between 0 to 1, that is, be converted to a probability.Also, the institute in search space The sum of corresponding probability of some data enhancement operations is 1.Illustratively, probability distribution can be { 0.1,0.08,0.32 ... }, K probability is shared, the sum of these probability are equal to 1, each of these probability can indicate that corresponding data enhancement operations are adopted The probability value that sample uses.

In embodiment of the disclosure, above-mentioned enhancing distribution parameter can be used as hyper parameter when network training (hyper-parameter), and can optimize simultaneously with the training process of network with network parameter.In addition, above-mentioned one kind Probability distribution can regard a kind of data enhancing strategy (an augmentation policy) as, because to the training data of network It carries out being to sample the data enhancement operations used based on the probability distribution when data enhancing, when probability distribution changes, sampling Data enhancement operations also change therewith.

Fig. 3, which shows at least one embodiment of the disclosure and provides the application scenarios of the method for cover half type hyper parameter really, to be shown Example.The application scenarios provide a kind of training system, which may include: parameter management server 11 and controller 12.Its In,

Parameter management server 11, for managing and updating the numerical value of hyper parameter.

Controller 12, including multiple machine learning models, the multiple machine learning model are based on hyper parameter and are recycled Or iteration updates, and obtains the performance parameter for the model that training obtains, which is fed back to parameter management by controller 12 Server, so that parameter management server carries out the update of hyper parameter accordingly.Controller 12 is also based on parameter management clothes The business updated hyper parameter of device 11 continues the training of model.

Fig. 3 is referred to, a kind of architectural form that the training system can use is illustrated, which exists using one kind The two-level architecture of line optimization can enable enhancing distribution parameter to carry out simultaneously with network parameter excellent by the two-level architecture Change.It wherein, include parameter management server 11 and controller 12 in the two-level architecture.Parameter management server 11 is responsible for depositing Storage and update enhancing distribution parameter, and the probability distribution of data enhancement operations is obtained based on the enhancing distribution parameterpθ,.Controller It may include one group of network to be trained (Network) in 12, it is assumed that N number of network is shared, for example, the N can be in 4 to 16 Any value.The network structure disclosure of the network does not limit, such as can be CNN (Convolutional Neural Networks, convolutional neural networks).

It can be using the entirety of parameter management server 11 and controller 12 as an outer ring (Outer Loop), the outer ring It is run in a manner of intensified learning.Outer ring can carry out the dimension of undated parameter management server 11 with T time step of iteration (time step) The enhancing distribution parameter of shieldθ, for the enhancing distribution parameter as a hyper parameter, updating the enhancing distribution parameter can be used as parameter Behavior (action) of the management server 11 in intensified learning training process.And controller 12 is based on enhancing distribution parameter instruction Accuracy rate of the network got on verifying collection as the return reward value (Reward) in intensified learning training process, according to The Reward carries out the update of enhancing distribution parameter, and the optimization of Reward, i.e. accuracy rate are realized after T time step of iteration most Height is the target direction of intensified learning.The T=1,2 ... ... T_max。

And N number of network to be trained in controller 12, it can be used as inner ring (Inner Loop), which can be with Parallel operation.The training of each network, which can be, is adopted based on parameter management server 11 in the enhancing distribution parameter that outer ring updates The data enhancement operations that sample uses, and network training is carried out based on the enhanced data of data.The training of network parameter can make With stochastic gradient descent method (SGD, Stochastic gradient descent).Each network can with iteration i times (i=1, 2 ... .I), and using accuracy rate of the iteration i times network on verifying collection as above-mentioned return reward value Reward, for more New enhancing distribution parameter.

Fig. 4 shows another method for determining model hyper parameter of at least one embodiment of disclosure offer, as follows Embodiment description in, parallel multiple paths may include parallel multiple networks, which can be machine learning model. Also, in this example, hyper parameter is the enhancing distribution parameter for carrying out image enhancement processing to sample graph image set.In conjunction with figure Shown in 4, this method may include following processing step:

In step 400, parallel enhancing distribution parameter of multiple networks based on initialization respectively carries out enhancing operation and adopts Sample obtains the data enhancement operations of the Web vector graphic.

Optionally, the multiple networks for including in controller 12 can be with parallel training.

Each network has the network parameter of initialization, and the enhancing that can be safeguarded based on parameter management server 11 Probability distributionpθ, sampling is applied to the data enhancement operations of the input data of network.The data of different network samples enhance behaviour It can be different.The data enhancement operations that sampling obtains can be known as to target data enhancing operation.

In step 402, each network in the multiple network is according to the data enhancement operations, to the network Input data carries out data enhancing, and carries out network training using enhanced data, obtains updated network parameter.

Optionally, each network carries out network training using enhanced data, can obtain after carrying out data enhancing Updated network parameter.The network training of this step can be with iteration preset times.

For example, data enhancement operations may include carrying out following processing to image: rotation, horizontal shear, is hung down at color adjustment Staight scissors cut, move horizontally, vertically moving, tone separation, exposure-processed etc..The enhancing operation of these target datas be can use to every At least one sample image in a path carries out image enhancement processing, obtains at least one enhancing image.And it is possible to based on institute At least one the enhancing image for stating each path in multiple paths, carries out M1 repetitive exercise to the initial network.

In step 404, the accuracy rate on each comfortable verifying collection of the multiple network that training obtains is obtained.

Optionally, the performance parameter of model is by taking accuracy rate as an example.

For example, can be by validation data set, the accuracy rate for the network that training obtains in verification step 404 (accuracy).In N number of network of controller 12, the data enhancement operations that heterogeneous networks sampling uses are different, correspondingly, training The effect of obtained network is also different, and the accuracy rate of each network may exist difference.

In a step 406, by nitrification enhancement, the accuracy rate based on the multiple network updates the enhancing distribution Parameter.

Optionally, the accuracy rate of the multiple networks obtained based on training is taken according to nitrification enhancement undated parameter management The enhancing distribution parameter that business device 11 is safeguarded.The optimization of accuracy rate can be used as the target direction of intensified learning, be based on accuracy rate Action (update for enhancing distribution parameter) is updated as Reward.

In a step 408, by the network parameter of the highest network of the accuracy rate, it is applied to the multiple network, is obtained The new network of next round iteration completes the update of a time step of intensified learning.

Collect the upper highest network of accuracy rate in verifying it is alternatively possible to determine in multiple networks of controller 12, and should The network parameter of the highest network of accuracy rate is synchronized to all networks, the i.e. network parameter of all-network in controller 12 all Using the network parameter of the synchronization.

Be synchronized the sum of network parameter, multiple networks in controller 12 are properly termed as new network, and can continue into Enter the update of next time step of intensified learning.

In step 410, multiple new networks are based on the updated enhancing distribution parameter, continue intensified learning The iteration of next time step updates.

The iteration of this step, which updates, repeats aforementioned step 400 to 408, for example, 11 base of parameter management server Updated probability distribution is obtained in updated enhancing distribution parameter, each network in controller 12 can be based on the update Probability distribution afterwards samples data enhancement operations to be used.And based on enhanced data training network, verifying accuracy rate, The operations such as enhancing distribution parameter are updated, are no longer described in detail.

In step 412, when reaching step of preset renewal time, verifying is collected into the upper highest network of accuracy rate and is determined as The network that final training obtains, and the enhancing distribution parameter of the last one time step update is obtained, to be distributed according to the enhancing Parameter determines the data enhancement operations taken.

Described reaches default renewal time step, is a kind of exemplary preset model cut-off condition.

The training of intensified learning can preset the quantity of renewal time step, for example, time step preferably at most updates T_max, then When reaching T_maxWhen, it can be using the highest network of accuracy rate as final network, and the enhancing distribution parameter of final updating is made To optimize obtained hyper parameter.Probability distribution can be obtained based on the hyper parameter, and can be instructed according to the probability distribution in network Sampled data enhancing operation is trained the enhancing of data during white silk.

Really the method for cover half type hyper parameter in some embodiments, by the way that distribution parameter will be enhanced as network training Hyper parameter, and in the training process of network, the interim feedback for taking intensified learning updates the mode of the hyper parameter, makes The optimization of the optimization and network parameter that obtain hyper parameter carries out simultaneously, has been obviously improved the search efficiency of hyper parameter；Furthermore this method The mode training hyper parameter for the intensified learning taken, using the verifying collection accuracy rate of network as optimization aim, the update of hyper parameter Optimal for training direction with accuracy rate, so that obtained hyper parameter is also more accurate, effect is more preferable.

In addition, it should be noted that, the training of network (Network) is primary in traditional hyperparameter optimization mode Property training on earth, be primary training from the Final Network of the network of initialization to the end similar in the frame diagram of Fig. 3 It obtains, then based on the trained preferable data enhancing strategy of web search, entire search process is very time-consuming, and needs to instruct The tactful search result of final data enhancing could be obtained by practicing a fairly large number of network, and time cost and calculating cost are very It is high.And the optimization method of the disclosure, it is equivalent to and interrupts the disposable training process of above-mentioned network, from " the network of initialization The feedback of multiple stages is increased during Final Network " to the end, the certain number of network iteration just feeds back one It is secondary, distribution parameter is enhanced based on interim feedback updated in the way of intensified learning, this mode by hyperparameter optimization and The optimization of network parameter carries out simultaneously, can greatly speed up the optimization efficiency of hyper parameter, and time cost significantly reduces, moreover, super The optimization of parameter is using the accuracy rate of network as optimization aim direction, and effect is preferable, can both fast and sound be searched preferably Hyper parameter, trained the number networks also greatly reduce compared to traditional approach.

There is provided the data enhancing of one embodiment tactful optimization method again as follows, in the embodiment, to input picture For the data enhancement operations such as being rotated, being sheared, translated, searching for enhances strategy to the preferably data that input picture is taken, So that according to this, preferably the enhanced data of strategy progress train the network performance of obtained network more preferable.

It firstly, initializing in frame shown in Fig. 4, include N number of network (Network) N number of network tool in controller 12 There is identical initialization network parameter ω ', and initializes the enhancing distribution ginseng of each data enhancement operations in search space Number θ.Parameter management server 11 is according to enhancing distribution parameter, the probability distribution p θ of available data enhancement operations.

Further, it is also possible to set, for each network in inner ring, in network training, extensive chemical may be set in Each time step (the at each time step) iteration practised updates I times, then carries out the enhancing distribution parameter of an outer ring Update.The iteration of the intensified learning of outer ring can set total iteration T_maxIt is secondary, i.e. iteration T_maxFinal optimization net is obtained after secondary Network and optimization hyper parameter.

Fig. 5 illustrates the training process of one of network, wherein when the network training, each time network parameter repeatedly Generation update is the input data by one group of B input picture for network, and iteration can be used as a grouping, i.e. network each time It is to be iterated training by multiple groupings.

In step 500, the enhancing probability distribution based on data enhancing strategy, is made by sampling in multiple data enhancement operations Target data enhancing operation.

For example, including B input picture in this group of data when using wherein one group of input picture is trained.For Each input picture can sample a data enhancement operations based on enhancing probability distribution, which is properly termed as number of targets It is operated according to enhancing.Also, obtained these data enhancement operations of sampling can be regarded as and sampled based on hyper parameter The training parameter of model, the data enhancement operations that the network samples in different paths obtain can be different.

Illustratively, for one of input picture, the data enhancement operations of sampling may include that two data increase Strong element " rotation " and " shearing ".The input picture can be enhanced using the data enhancement operations, increased Strong image.

In step 502, network training is carried out using enhanced data, obtains updated network parameter.

Optionally, enhanced data can be enhancing image.The network training of this step can be multiple with iteration, final To updated network parameter.

When network parameter updates, gradient descent method can be used, network parameter is optimized with the minimum principle of loss function.Such as Under formula (1) illustrate a kind of update mode of network parameter:

Formula (1) as above, network parameter is when updating, and following relating to parameters:

η_wIt is the learning rate of network parameter；

Loss function value when being the grouping iteration, the loss function value enhance according to the data that current iteration samples OperationCurrent network parameterWith the input data x of current iteration_B, y_B, determine.

In step 504, judge whether the number of iterations of network reaches default the number of iterations.

As previously mentioned, may be set in each time step (at each time step) of intensified learning, in controller Each network iteration update I time, that is, pass through I grouping progress network parameter update.

Optionally, if judging that the number of iterations i of network is less than presets the number of iterations I, in return step 300, under B input picture of one grouping continues trained and iterative network parameter.

If judging, the number of iterations i of network is equal to default the number of iterations I, continues step 306.

For example, the default the number of iterations of this step can be M1 times.In addition, in different times in step, the iteration of network Number may be the same or different.

In addition, what training obtained for the first time is properly termed as the first inner ring more new engine in the successive ignition training of network Learning model is properly termed as the second inner ring again and updates machine learning model after iteration, and so on.

In step 506, the network that training obtains is obtained.

Each network N etwork is trained all in accordance with process shown in Fig. 3, and training here is referred in intensified learning Each time step in training.These N number of networks can be parallel training, finally obtain following parameter: { ω_{T, n}}_N=1:N, Obtain the network parameter of each network of N number of network.Wherein, ω_{T, n}Indicate that the network parameter is the T in intensified learning The parameter of n-th of network in a time step.

Then, the network obtained based on N number of training can go the accuracy rate of verifying network by validation data set.It can be with Referring to the signal of Fig. 1, each network can obtain an accuracy rate Acc.In some embodiments, using accuracy rate as network Performance parameter for, certainly in other examples, expression of other parameters as performance parameter can also be used.

The acquisition of the accuracy rate can be the network obtained by training at least one test chart in test image collection As being handled, processing result image is obtained, and accuracy rate is obtained based on the processing result image.

Reward value reward of the above-mentioned accuracy rate as intensified learning can pass through reinforcing based on reward value reward Learning algorithm is updated enhancing distribution parameter.Intensified learning Policy-Gradient algorithm can be used in the embodiment of the present disclosure REINFORCE, the accuracy rate based on each network update enhancing distribution parameter.

Following formula (2) illustrates REINFORCE algorithm:

Formula (2) as above, acc (w_{T, n}) be n-th of network of the T time step of intensified learning verifying collection it is quasi- True rate,It is the data enhancement operations of n-th of network, which sampled according to probability distribution.Therefore, according to public affairs Formula (2) is the data enhancement operations of the accuracy rate and the network samples according to each network, determines the multiple network pair The average gradient of the enhancing distribution parameter.

In formula (2) as above, still by taking performance parameter is accuracy rate as an example, according to the available each path of accuracy rate Model modification parameter, the model modification parameter can be the gradient of the network in the path, and the gradient in multiple paths be carried out flat It handles, obtains above-mentioned average gradient (be averaged undated parameter).

In addition, place first can be normalized in the accuracy rate in each path in multiple paths before executing formula (2) Reason, and calculated based on the accuracy rate after normalization.

Based on above-mentioned average gradient, enhancing distribution parameter can be updated, following formula (3):

Wherein, η_θIt is the learning rate of probability distribution.First is properly termed as according to the updated hyper parameter numerical value of formula (3) Updated value, certainly, after iteration updates repeatedly, the second updated value of available hyper parameter, third updated value etc..

As above, in a time step, based on network in the accuracy rate of verifying collection, enhancing distribution parameter, parameter are had updated Management server 11 can be according to the updated enhancing distribution parameter, update probability distribution, with next time step according to Updated probability distribution continues sampled data enhancing operation.

Continuing with combining referring to Fig. 3, dotted arrow 13 indicates the accuracy rate of each network feeding back to parameter management service Device 11, parameter management server 11, can be with update probability distribution ps according to above-mentioned REINFORCE algorithm_θ.14 table of solid arrow Show the iteration that next time step is continued to execute based on updated probability distribution.In one embodiment, in next iteration Before beginning, the network parameter of the highest network of accuracy rate can be synchronized to the all-network of controller, as shown in figure 3, accurately The network parameter of the highest network of rate is parameter 15, indicates the parameter being applied to all-network by synchronous arrow 16.At it In his example, it is unitized that the network parameter before starting can also be iterated using other modes.

Updated network, which is based on updated probability distribution, to be continued to sample, and the iteration for starting next time step updates. In the iteration of next time step updates, obtained network is properly termed as the second update machine learning model, and equally can The performance parameter for updating machine learning model according to the second of path each in multiple path, more by the numerical value of the hyper parameter It is newly the second updated value.

Until the default renewal time for reaching intensified learning walks T_maxWhen, using the highest network of accuracy rate as Final The optimal network that Network, i.e. training are obtained, is referred to as target machine learning model.Meanwhile the last one time step makes The hyper parameter that the enhancing distribution parameter used is obtained as final optimization passThat is the final numerical value of hyper parameter.

The method flow of the above-mentioned Fig. 4 of the disclosure and determining model hyper parameter shown in fig. 5 can use the process such as the following table 1 It indicates, wherein obtaining hyper parameterWhile obtain Model Weight parameter.It is understood that other can also be used Process flow, as long as meeting the method for disclosure any embodiment cover half type hyper parameter really.

A kind of process of the method for determining model hyper parameter of following example:

For example, initialization hyper parameter is θ₀, and the network parameter for initializing each network is w₀。

For each time step (total T=T of intensified learning_maxA time step), it is carried out following processing:

Each network of controller carries out i=I iteration altogether and updates, and iteration updates (if 0≤i≤I) according to public affairs each time Formula (1) updates network parameterAfter I iteration, network parameter w is obtained_{T, n}, the w_{T, n}It indicates the of intensified learning The network parameter of n-th of network in T time step, wherein share N number of network.

When obtaining N number of respective network parameter w of network_{T, n}Afterwards, multiple networks are calculated to the increasing according to formula (2) The average gradient of strong distribution parameterAnd hyper parameter is updated according to formula (3), make in the T time step of intensified learning Hyper parameter can be expressed as θ_T。

Before the next time step for entering intensified learning carries out network training, verified by being selected in N number of network The network parameter w of the highest network of accuracy rate on collection_T, and by the network parameter synchronous applications to N number of network, so that this is N number of Network repetitive exercise starting point having the same.

When reaching preset intensified learning time step T_maxWhen, obtain final network parameter wT_maxAnd hyper parameter

In addition, being to reach the default renewal time of intensified learning and walk T in some embodiments_maxAs hyperparameter optimization Cut-off condition in other examples can also be using other cut-off conditions.Also, the example is with hyperparameter optimization For obtaining trained network simultaneously, it is also possible to that trained network is finally not yet received, but is obtained using final optimization pass The initial machine learning model of one initialization model parameter of the hyper parameter re -training arrived obtains the target machine of training completion Learning model.

After obtaining hyper parameter, when a network will train, if the search for the data enhancement operations that the network uses Space is identical as search space when hyperparameter optimization, and data when can directly be carried out the network training using the hyper parameter are increased It by force, will be so that network obtains preferable performance.If the search space of the corresponding data enhancement operations of network of training is become Change, the optimization method that can also reuse disclosure offer searches for new hyper parameter, that is, re-searches for preferably data enhancing Strategy.

Following table 1, it may be said that bright disclosure any embodiment provides the experiment effect of the method for cover half type hyper parameter really Fruit.Refer to table 1, ResNet-18, the WideResNet-28-10 in left side etc. indicate different network structures, above table Baseline, Cutout etc. indicate that different data enhance tactful preparation method, wherein OHL-Auto-Aug indicates to utilize this public affairs The method opened.The data enhancing strategy that a variety of different methods are obtained is applied to the training of multiple network structure, demonstrate,proves through test Bright, disclosed method can reduce error rate compared with various other ways, wherein disclosed method relative to Baseline method trains the lower error rate of obtained network about 30%.The hyper parameter obtained using method of disclosure Training machine learning model, the performances such as accuracy rate for the model that can be improved.

1 method of disclosure of table is compared with the effect of other methods

Fig. 6 provides a kind of device of determining model hyper parameter, as shown in fig. 6, the apparatus may include: initialization mould Block 61, model training module 62, super ginseng update module 63 and super ginseng obtain module 64.

Initialization module 61, for determining the initial value of hyper parameter；

Model training module 62, for the initial value and sample graph image set according to the hyper parameter, by parallel multiple Each path carries out M1 repetitive exercise to initial machine learning model in path, and obtain each path first updates engineering Practise model, wherein the training parameter in different paths has the difference sampled based on the hyper parameter in multiple paths Numerical value, M1 are greater than or equal to 1 and are less than or equal to the first numerical value；

Super ginseng update module 63 updates machine learning model for first based on each path in the multiple path The numerical value of the hyper parameter is updated to the first updated value by performance parameter；；

Super ginseng obtains module 64, for based on the hyper parameter the first updated value and the sample graph image set, to described The first of multiple paths updates the further numerical value update that machine learning model carries out M2 repetitive exercise and the hyper parameter, directly To default cut-off condition is reached, the final numerical value of the hyper parameter is obtained, wherein M2 is greater than or equal to 1 and is less than or equal to the One numerical value.

In some alternative embodiments, surpass ginseng and obtain module 64, be also used to: described to the first of the multiple path Before updating the further numerical value update that machine learning model carries out M2 repetitive exercise and the hyper parameter, from the multiple road The first of diameter, which updates, chooses first object update machine learning model in machine learning model；More by the first of the multiple path The model parameter of new engine learning model is updated to the model parameter that the first object updates machine learning model.

In some alternative embodiments, surpass ginseng and obtain module 64, for updating machine from the first of the multiple path When choosing first object update machine learning model in device learning model, comprising: first based on the multiple path updates machine The performance parameter of device learning model updates from the first of the multiple path and chooses first object update machine in machine learning model Device learning model.

In some alternative embodiments, the model training module 62, is specifically used for: based on the initial of the hyper parameter At least one first sample image in value and the sample image collection, by each path in multiple paths to the initial machine Device learning model carries out the first repetitive exercise, and the first inner ring for obtaining each path updates machine learning model；Based on institute The second sample image of initial value and at least one of the sample image collection for stating hyper parameter, by every in the multiple path A path updates machine learning model to first inner ring in each path and carries out secondary iteration training, obtains each road Second inner ring of diameter updates machine learning model；The second inner ring based on each path in the multiple path updates machine learning Model, obtain each path first update machine learning model.

In some alternative embodiments, model training module 62, in the first inner ring for obtaining each path When updating machine learning model, comprising: the initial value based on the hyper parameter carries out multiple repairing weld, obtains in the multiple path First training parameter in each path；The first training parameter and the sample image based on each path in the multiple path At least one first sample image in collection carries out the first repetitive exercise to initial machine learning model, obtains each road First inner ring of diameter updates machine learning model.

In some alternative embodiments, it is used in first repetitive exercise in each path and secondary iteration training Training parameter be the initial value based on the hyper parameter carry out it is different sampling obtain.

In some alternative embodiments, surpass ginseng update module 63, be specifically used for: based on each road in the multiple path The first of diameter updates the performance parameter of machine learning model, determines the model modification parameter in each path；To the multiple The network undated parameter in path is averaging processing, and obtains average undated parameter；It, will be described super according to the average undated parameter The numerical value of parameter is updated to the first updated value.

In some alternative embodiments, surpass ginseng update module 63, be specifically used for: to each path in the multiple path First update machine learning model performance parameter be normalized；Described in being obtained after the normalized The first of each path updates the performance parameter of machine learning model in multiple paths, and the numerical value of the hyper parameter is updated to the One updated value.

In some alternative embodiments, the performance parameter includes accuracy rate.

In some alternative embodiments, surpass ginseng and obtain module 64, be specifically used for: first based on the hyper parameter updates Value and the sample graph image set update machine learning model to first of each path in the multiple path and carry out M2 iteration Training, obtain each path second update machine learning model；Second based on each path in multiple paths updates The numerical value of the hyper parameter is updated to the second updated value by the performance parameter of machine learning model.

In some alternative embodiments, the hyper parameter includes for carrying out at image enhancement to the sample graph image set The enhancing distribution parameter of reason；Model training module 62, is specifically used for: according to the enhancing distribution parameter, determining enhancing probability point Cloth, described enhance in probability distribution includes that multiple images enhance the probability operated；Based on the enhancing probability distribution, by described more The target data enhancing operation that each path in parallel multiple paths is sampled in a data enhancement operations, to described each At least one sample image in path carries out image enhancement processing, obtains at least one enhancing image；Based on the multiple path In each path at least one enhancing image, to the initial machine learning model carry out M1 repetitive exercise.

In some alternative embodiments, surpass ginseng update module 63, be also used to obtain the first update machine learning model The acquisition of performance parameter, including handling as follows: updating machine learning model by first of each path in the multiple path At least one test image in test image collection is handled, processing result image is obtained；Based in the multiple path The corresponding described image processing result in each path obtains the first of each path performance parameter for updating machine learning model.

In some alternative embodiments, the default cut-off condition includes at least one in following: to the super ginseng Several update times reach default update times；Alternatively, the performance for the update machine learning model that the multiple path obtains reaches To target capabilities.

In some alternative embodiments, surpass ginseng and obtain module 64, be also used to from the feelings for reaching the default cut-off condition Target machine learning model is chosen in the final updated machine learning model in the multiple path obtained under condition, wherein described Target machine learning model is the trained machine learning model for image procossing.

In some alternative embodiments, surpass ginseng and obtain module 64, be also used in the final numerical value for obtaining the hyper parameter Later, based on the final numerical value of the hyper parameter, the initial machine learning model of training initialization model parameter obtains having trained At target machine learning model.

The disclosure additionally provides a kind of electronic equipment, and the equipment includes memory, processor, and the memory is for depositing The computer instruction that can be run on a processor is stored up, the processor is for realizing the disclosure when executing the computer instruction The training method of any embodiment method or machine learning model of cover half type hyper parameter really.

The disclosure additionally provides a kind of computer readable storage medium, is stored thereon with computer program, described program quilt The training of the disclosure any embodiment method or machine learning model of cover half type hyper parameter really is realized when processor executes Method.

Fig. 7 provides the flow diagram of the image processing method in an embodiment of the present disclosure, as shown in fig. 7, the party Method may include:

In step 700, image to be processed is obtained.

Which kind of image the image to be processed that this step does not limit input model is.

In a step 702, the image to be processed is handled using machine learning model, obtains image procossing knot Fruit, wherein the hyper parameter of the machine learning model is by cover half type hyper parameter really described in disclosure any embodiment What method determined.

Image processing method in some embodiments, since the hyper parameter of the machine learning model of processing image is to pass through The method of the above-mentioned type hyper parameter of cover half really of the disclosure determines that the hyper parameter has preferable effect, so at using the model Managing obtained processing result image also has good performance.

Fig. 8 shows a kind of training method of machine learning model of at least one embodiment of disclosure offer, such as Fig. 8 institute Show, this method may include:

In step 800, the final numerical value of the hyper parameter is obtained.

Optionally, the method for cover half type hyper parameter really can be provided by disclosure any embodiment, determines the super ginseng Several final numerical value.

In step 802, based on the final numerical value of the hyper parameter, training has the initial machine of original model parameter Model is practised, target machine learning model is obtained.

Since hyper parameter is determined by the method for the determination hyper parameter of disclosure any embodiment, the effect of hyper parameter compared with It is good, therefore also there is preferable performance using the machine learning model of hyper parameter training.

Fig. 9 shows a kind of training device of machine learning model of at least one embodiment of disclosure offer, such as Fig. 9 institute Show, which includes: that super ginseng obtains module 91 and model training module 92.

Super ginseng obtains module 91, for obtaining by the method for cover half type hyper parameter really described in disclosure any embodiment Obtain the final numerical value of the hyper parameter；

Model training module 92, for the final numerical value based on the hyper parameter, training has the first of original model parameter Beginning machine learning model obtains target machine learning model.

The disclosure additionally provides a kind of computer readable storage medium, is stored thereon with computer program, described program quilt The method or machine learning model of cover half type hyper parameter really are realized described in the disclosure any embodiment when processor executes Training method.

The disclosure additionally provides a kind of electronic equipment, and the electronic equipment includes memory, processor, and the memory is used In the computer instruction that storage can be run on a processor, the processor is used to realize this when executing the computer instruction Really the training method of the method or machine learning model of cover half type hyper parameter described in open any embodiment.

It will be understood by those skilled in the art that disclosure one or more embodiment can provide as method, system or computer Program product.Therefore, complete hardware embodiment, complete software embodiment or combination can be used in disclosure one or more embodiment The form of embodiment in terms of software and hardware.Moreover, disclosure one or more embodiment can be used it is one or more its In include computer usable program code computer-usable storage medium (including but not limited to magnetic disk storage, CD-ROM, Optical memory etc.) on the form of computer program product implemented.

The embodiment of the present disclosure also provides a kind of computer readable storage medium, can store computer on the storage medium Program realizes the neural network for Text region of the disclosure any embodiment description when described program is executed by processor The step of training method, and/or, realize the disclosure any embodiment description character recognition method the step of.Wherein, described "and/or" indicates at least with one of them in the two, for example, " more and/or B " includes three kinds of schemes: more, B and " more And B ".

Various embodiments are described in a progressive manner in the disclosure, same and similar part between each embodiment It may refer to each other, each embodiment focuses on the differences from other embodiments.Especially for data processing For apparatus embodiments, since it is substantially similar to the method embodiment, so being described relatively simple, related place is referring to method The part of embodiment illustrates.

It is above-mentioned that disclosure specific embodiment is described.Other embodiments are within the scope of the appended claims. In some cases, the behavior recorded in detail in the claims or step can be executed according to the sequence being different from embodiment And desired result still may be implemented.In addition, process depicted in the drawing not necessarily require the particular order shown or Person's consecutive order is just able to achieve desired result.In some embodiments, multitasking and parallel processing are also possible Or it may be advantageous.

The embodiment of theme and feature operation described in the disclosure can be realized in the following: Fundamental Digital Circuit has The existing computer software of body or firmware, the computer hardware including structure disclosed in the disclosure and its structural equivalents, Or the combination of one or more of which.The embodiment of theme described in the disclosure can be implemented as one or more meters Calculation machine program, i.e. coding are on tangible non-transitory program carrier to execute or control data processing equipment by data processing equipment Operation computer program instructions in one or more modules.Alternatively, or in addition, program instruction can be encoded On manually generated transmitting signal, such as electricity, light or electromagnetic signal that machine generates, the signal are generated to encode information onto And suitable receiver apparatus is transferred to be executed by data processing equipment.Computer storage medium can be machine readable storage Equipment, machine readable storage substrate, random or serial access memory equipment or one or more of which combination.

Processing and logic flow described in the disclosure can be by one or more of the one or more computer programs of execution A programmable calculator executes, to execute corresponding function by the way that output is operated and generated according to input data.It is described Processing and logic flow can also be by dedicated logic circuit-such as FPG more (field programmable gate arrays) or more SIC (dedicated collection At circuit) Lai Zhihang, and device also can be implemented as dedicated logic circuit.

The computer for being suitable for carrying out computer program includes, for example, general and/or special microprocessor or it is any its The central processing unit of his type.In general, central processing unit will refer to from read-only memory and/or random access memory reception Order and data.The basic module of computer includes central processing unit for being practiced or carried out instruction and for storing instruction With one or more memory devices of data.In general, computer will also be including one or more great Rong for storing data Amount storage equipment, such as disk, magneto-optic disk or CD etc. or computer will be coupled operationally with this mass-memory unit To receive from it data or have both at the same time to its transmission data or two kinds of situations.However, computer is not required to have in this way Equipment.In addition, computer can be embedded in another equipment, such as mobile phone, personal digital assistant (PD is more), mobile sound Frequency or video player, game console, global positioning system (GPS) receiver or such as universal serial bus (USB) flash memory The portable memory apparatus of driver, names just a few.

It is suitable for storing computer program instructions and the computer-readable medium of data including the non-volatile of form of ownership Memory, medium and memory devices, for example including semiconductor memory devices (such as EPROM, EEPROM and flash memory device), Disk (such as internal hard drive or removable disk), magneto-optic disk and CD ROM and DVD-ROM disk.Processor and memory can be by special It is supplemented or is incorporated in dedicated logic circuit with logic circuit.

Although the disclosure includes many specific implementation details, these are not necessarily to be construed as limiting any scope of disclosure Or range claimed, and be primarily used for describing the feature of specifically disclosed specific embodiment.Multiple in the disclosure Certain features described in embodiment can also be combined implementation in a single embodiment.On the other hand, in a single embodiment The various features of description can also be performed separately in various embodiments or be implemented with any suitable sub-portfolio.Though in addition, Right feature can work in certain combinations as described above and even initially so be claimed, but come from required guarantor One or more features in the combination of shield can be removed from the combination in some cases, and combination claimed The modification of sub-portfolio or sub-portfolio can be directed toward.

Similarly, although depicting operation in the accompanying drawings with particular order, this is understood not to require these behaviour Make the particular order shown in execute or sequentially carry out or require the operation of all illustrations to be performed, to realize desired knot Fruit.In some cases, multitask and parallel processing may be advantageous.In addition, the various system modules in above-described embodiment Separation with component is understood not to be required to such separation in all embodiments, and it is to be understood that described Program assembly and system can be usually integrated in together in single software product, or be packaged into multiple software product.

The specific embodiment of theme has been described as a result,.Other embodiments are within the scope of the appended claims.? In some cases, the movement recorded in claims can be executed in different order and still realize desired result.This Outside, the processing described in attached drawing and it is nonessential shown in particular order or sequential order, to realize desired result.In certain realities In existing, multitask and parallel processing be may be advantageous.

The foregoing is merely the preferred embodiments of disclosure one or more embodiment, not to limit the disclosure One or more embodiments, all any modifications within the spirit and principle of disclosure one or more embodiment, made, etc. With replacement, improvement etc., should be included within the scope of the protection of disclosure one or more embodiment.

Claims (10)

1. a kind of method of determining model hyper parameter, which is characterized in that the described method includes:

Determine the initial value of hyper parameter；

According to the initial value of the hyper parameter and sample graph image set, by each path in parallel multiple paths to initial machine Learning model carries out M1 repetitive exercise, and obtain each path first updates machine learning model, wherein the multiple The training parameter in different paths has the different numerical value sampled based on the hyper parameter in path, and M1 is greater than or equal to 1 and be less than or equal to the first numerical value；

First based on each path in the multiple path updates the performance parameter of machine learning model, by the hyper parameter Numerical value is updated to the first updated value；

The first updated value and the sample graph image set based on the hyper parameter update engineering to the first of the multiple path It practises model and carries out the further numerical value update of M2 repetitive exercise and the hyper parameter, until reach default cut-off condition, acquisition The final numerical value of the hyper parameter, wherein M2 is greater than or equal to 1 and is less than or equal to the first numerical value.

2. the method according to claim 1, wherein first based on each path in the multiple path The performance parameter for updating machine learning model, is updated to the first updated value for the numerical value of the hyper parameter, comprising:

First based on each path in the multiple path updates the performance parameter of machine learning model, determines each road The model modification parameter of diameter；

The model modification parameter in the multiple path is averaging processing, average undated parameter is obtained；

According to the average undated parameter, the numerical value of the hyper parameter is updated to the first updated value.

3. method according to claim 1 or 2, which is characterized in that described based on each path in the multiple path First update machine learning model performance parameter, it is described before the numerical value of the hyper parameter is updated to the first updated value Method further include:

The performance parameter of the first update machine learning model in each path in the multiple path is normalized；

First based on each path in the multiple path updates the performance parameter of machine learning model, by the super ginseng Several numerical value is updated to the first updated value, comprising:

First based on each path in the multiple path obtained after the normalized updates machine learning model The numerical value of the hyper parameter is updated to the first updated value by performance parameter.

4. method according to any one of claims 1 to 3, which is characterized in that the hyper parameter includes for the sample The enhancing distribution parameter of image set progress image enhancement processing；

The initial value and sample graph image set according to the hyper parameter, by each path in parallel multiple paths to initial Machine learning model carries out M1 repetitive exercise, comprising:

According to the enhancing distribution parameter, enhancing probability distribution is determined, include that multiple images enhance in the enhancing probability distribution The probability of operation；

It is each in parallel multiple paths by being sampled in the multiple data enhancement operations based on the enhancing probability distribution The target data in path enhances operation, carries out image enhancement processing at least one sample image in each path, obtains At least one enhancing image；

Based at least one enhancing image in each path in the multiple path, M1 is carried out to the initial machine learning model Secondary repetitive exercise.

5. a kind of training method of machine learning model characterized by comprising

The final numerical value of the hyper parameter is obtained by method described in any one of Claims 1-4；

Final numerical value based on the hyper parameter, training have the initial machine learning model of original model parameter, obtain target Machine learning model.

6. a kind of device of determining model hyper parameter, which is characterized in that described device includes:

Initialization module, for determining the initial value of hyper parameter；

Model training module, for the initial value and sample graph image set according to the hyper parameter, by parallel multiple paths Each path carries out M1 repetitive exercise to initial machine learning model, and obtain each path first updates machine learning mould Type, wherein the training parameter in different paths has the difference sampled based on the hyper parameter in the multiple path Numerical value, M1 are greater than or equal to 1 and are less than or equal to the first numerical value；

Super ginseng update module updates the performance ginseng of machine learning model for first based on each path in the multiple path Number, is updated to the first updated value for the numerical value of the hyper parameter；

Super ginseng obtains module, for based on the hyper parameter the first updated value and the sample graph image set, to the multiple road The first of diameter updates the further numerical value update that machine learning model carries out M2 repetitive exercise and the hyper parameter, until reaching Default cut-off condition, obtains the final numerical value of the hyper parameter, wherein M2 is greater than or equal to 1 and is less than or equal to the first numerical value.

7. device according to claim 6, which is characterized in that the default cut-off condition includes at least one in following :

Default update times are reached to the update times of the hyper parameter；

Alternatively, the performance for the update machine learning model that the multiple path obtains reaches target capabilities.

8. a kind of training device of machine learning model characterized by comprising

Super ginseng obtains module, for obtaining the final number of the hyper parameter by method described in any one of Claims 1-4 Value；

Model training module, for the final numerical value based on the hyper parameter, training has the initial machine of original model parameter Learning model obtains target machine learning model.

9. a kind of electronic equipment, which is characterized in that the equipment includes memory, processor, and the memory can for storing The computer instruction run on a processor, the processor are used to realize claim 1 when executing the computer instruction To method described in 4 one, or method described in realization claim 5.

10. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that described program is processed Claims 1-4 any method is realized when device executes, or realizes method described in claim 5.