The unified insertion of study

文档序号：1745755 发布日期：2019-11-26 浏览：13次中文

阅读说明：本技术 学习统一嵌入 (The unified insertion of study ) 是由宋扬李源武勃 C-Y.陈张晓� H.亚当于 2017-11-17 设计创作，主要内容包括：描述了一种用于在数据处理装置上使用神经网络生成统一机器学习模型的计算机实施的方法。该方法包括数据处理装置为多个对象顶点中的每一个确定相应学习目标。数据处理装置基于神经网络的两个或更多个嵌入输出来确定相应学习目标。该方法还包括数据处理装置训练神经网络以识别与多个对象顶点中的每一个相关联的数据。数据处理装置使用相应学习目标并基于第一损失函数来训练神经网络。数据处理装置使用经训练的神经网络来生成统一机器学习模型,其中该模型被配置为识别与多个对象顶点中的每一个相关联的特定数据项。(It describes a kind of for using neural network to generate the method implemented by computer of unified machine learning model on data processing equipment.This method includes that data processing equipment is that each of multiple object vertex determine corresponding learning objective.Data processing equipment two or more insertion outputs neural network based are to determine corresponding learning objective.This method further includes data processing equipment training neural network to identify data associated with each of multiple object vertex.Data processing equipment trains neural network using corresponding learning objective and based on first-loss function.Data processing equipment generates unified machine learning model using housebroken neural network, and wherein the model is configured as identification specific data item associated with each of multiple object vertex.)

1. a kind of computer for using neural network to generate unified machine learning computation model on data processing equipment is real The method applied, which comprises

The corresponding learning objective of each of multiple object vertex is determined by data processing equipment and for neural network, In, each object vertex defines the different classes of of the object for belonging to the vertex；

Neural network is trained by data processing equipment and based on first-loss function, with identification and the multiple object vertex Each of associated data, wherein the neural network is trained using corresponding learning objective；And

Unified machine learning model is generated by data processing equipment and using the neural network based on the training of first-loss function, The unified machine learning model is configured as identification and is included in number associated with each of the multiple object vertex Item in.

2. according to the method described in claim 1, wherein, determining corresponding learning objective for the neural network further include:

By the data processing equipment and based on at least one other neural network of the second loss function training, with identification and institute State the associated data of each of multiple object vertex；

In response to training, two or more insertions are generated by data processing equipment and are exported, wherein each insertion output instruction is special Determine the vector of learning objective and the parameter including corresponding to data associated with special object vertex；And

It is generated by data processing equipment and using at least one other neural network based on the training of the second loss function Corresponding machine learning model, each machine learning model are configured with specific insertion output.

3. according to the method described in claim 2, wherein, determining corresponding learning objective for the neural network further include:

The corresponding learning objective generated from corresponding independent model is provided, for training the neural network.

4. according to the method in claim 2 or 3, wherein each of the multiple object vertex corresponds to the spy of item Determine classification, and data associated with each of the multiple object vertex include the item in the particular category of item Image data.

5. according to the method described in claim 4, wherein, the particular category is dress ornament classification, and the particular category Item includes at least one of the following terms: handbag, shoes, one-piece dress, trousers or housing；And

Wherein, the image of described image data instruction at least one of the following terms: specific handbag, specific shoes, specific One-piece dress, specific trousers or specific housing.

6. according to the method described in claim 5, wherein:

Each of described corresponding machine learning model be configured as associated with the special object vertex data of identification and In first accuracy；And

The unified machine learning model is configured as identification data associated with each of the multiple object vertex And in the second accuracy for being more than first accuracy.

7. the method according to any one of claim 2 to 6, wherein determine each of the multiple object vertex Corresponding learning objective include:

The two or more insertion outputs are analyzed, it is specific right in the multiple object vertex that each insertion output corresponds to As vertex；And

Based on the analysis, the corresponding learning objective of each of the multiple object vertex is determined.

8. the method according to any one of claim 2 to 7, wherein the first-loss function is L2 loss function, and And it generates the unified machine learning model and includes:

It generates and minimizes the specific unified machine learning model for calculating output associated with the L2 loss function.

9. the method according to any one of claim 2 to 8, wherein the neural network includes receiving multiple layers of input Multiple neural net layers and wherein include: based on the first-loss function training neural network

Batch standardization is executed with the layer input to specific neural net layer that standardize；And

Standardize in response to executing batch, minimizes covariant offset.

10. the method according to any one of claim 2 to 9, wherein second loss function is triple loss letter Number, and generate the corresponding machine learning model and include:

Specific machine learning model is generated based on the association between anchor image, positive image and negative image.

11. method according to any of the preceding claims, wherein with each of the multiple object vertex phase Associated data include image data.

12. a kind of system for using neural network to generate unified machine learning computation model, the computing system include:

For the data processing equipment of neural network, the data processing equipment includes one or more processing equipments；And

One or more non-transitory machine readable storage devices can be run for storing by one or more of processing equipments So that execution includes the instruction according to the operation of the method for any preceding claims.

13. one or more non-transitory machine readable storage devices, can be run for storing by one or more processing equipments So that execute include according to claim 1 to any one of 10 method operation instruction.

14. a kind of computer for using neural network to generate unified machine learning computation model on data processing equipment is real The method applied, which comprises

15. according to the method for claim 14, wherein determine corresponding learning objective for the neural network further include:

16. according to the method for claim 15, wherein determine corresponding learning objective for the neural network further include:

The corresponding learning objective generated from corresponding independent model is provided, for training the neural network.

17. according to the method for claim 15, wherein each of the multiple object vertex corresponds to the specific of item Classification, and data associated with each of the multiple object vertex include the item in the particular category of item Image data.

18. according to the method for claim 17, wherein the particular category is dress ornament classification, and the particular category Item include at least one of the following terms: handbag, shoes, one-piece dress, trousers or housing；And

Wherein, the image of described image data instruction at least one of the following terms: specific handbag, specific shoes, specific One-piece dress, specific trousers or specific housing.

19. according to the method for claim 18, in which:

Each of described corresponding machine learning model be configured as associated with the special object vertex data of identification and In first accuracy；And

The unified machine learning model is configured as identification data associated with each of the multiple object vertex And in the second accuracy for being more than first accuracy.

20. according to the method for claim 15, wherein determine the corresponding study of each of the multiple object vertex Target includes:

The two or more insertion outputs are analyzed, it is specific right in the multiple object vertex that each insertion output corresponds to As vertex；And

Based on the analysis, the corresponding learning objective of each of the multiple object vertex is determined.

21. according to the method for claim 15, wherein the first-loss function is L2 loss function, and generates institute Stating unified machine learning model includes:

It generates and minimizes the specific unified machine learning model for calculating output associated with the L2 loss function.

22. according to the method for claim 15, wherein the neural network includes the multiple nerves for receiving multiple layers of input Network layer, and wherein include: based on the first-loss function training neural network

Batch standardization is executed with the layer input to specific neural net layer that standardize；And

Standardize in response to executing batch, minimizes covariant offset.

23. according to the method for claim 15, wherein second loss function is triple loss function, and raw Include: at the corresponding machine learning model

Specific machine learning model is generated based on the association between anchor image, positive image and negative image.

24. a kind of system for using neural network to generate unified machine learning computation model, the computing system include:

For the data processing equipment of neural network, the data processing equipment includes one or more processing equipments；And

One or more non-transitory machine readable storage devices can be run for storing by one or more of processing equipments So that executing the instruction of operation, the operation includes:

25. system according to claim 24, wherein determine corresponding learning objective for the neural network further include:

26. system according to claim 25, wherein each of the multiple object vertex corresponds to the specific of item Classification, and data associated with each of the multiple object vertex include the item in the particular category of item Image data.

27. system according to claim 26, wherein the particular category is dress ornament classification, and the particular category Item include at least one of the following terms: handbag, shoes, one-piece dress, trousers or housing；And

Wherein, the image of described image data instruction at least one of the following terms: specific handbag, specific shoes, specific One-piece dress, specific trousers or specific housing.

28. system according to claim 27, in which:

Each of described corresponding machine learning model be configured as associated with the special object vertex data of identification and In first accuracy；And

The unified machine learning model is configured as identification data associated with each of the multiple object vertex And in the second accuracy for being more than first accuracy.

29. system according to claim 25, wherein determine the corresponding study of each of the multiple object vertex Target includes:

The two or more insertion outputs are analyzed, it is specific right in the multiple object vertex that each insertion output corresponds to As vertex；And

Based on the analysis, the corresponding learning objective of each of the multiple object vertex is determined.

30. system according to claim 25, wherein the first-loss function is L2 loss function, and generates institute Stating unified machine learning model includes:

It generates and minimizes the specific unified machine learning model for calculating output associated with the L2 loss function.

31. system according to claim 25, wherein the neural network includes the multiple nerves for receiving multiple layers of input Network layer, and wherein include: based on the first-loss function training neural network

Batch standardization is executed with the layer input to specific neural net layer that standardize；And

Standardize in response to executing batch, minimizes covariant offset.

32. system according to claim 25, wherein second loss function is triple loss function, and raw Include: at the corresponding machine learning model

Specific machine learning model is generated based on the association between anchor image, positive image and negative image.

33. one or more non-transitory machine readable storage devices, can be run for storing by one or more processing equipments So that executing the instruction of operation, the operation includes:

Background technique

This specification is related to unified (unified) neural network model of training.

Neural network is machine learning model, one or more layers operation is used to generate output, example for received input Such as classification.Some neural networks further include one or more hidden layers other than output layer.The output of each hidden layer by with Make the input of next layer (i.e. the next hidden layer or output layer of network) in network.The some or all of layers of network are according to corresponding ginseng The current value of manifold is generated from received input and is exported.

Some neural networks include one or more convolutional neural networks layers.Each convolutional neural networks layer has one group of phase Associated kernel.Each kernel includes by the value of the Establishment of Neural Model of user's creation.In some embodiments, kernel Identify specific image profile, shape or color.Kernel can be expressed as the matrix structure of weight input.Each convolutional layer can also be with The set of processing activation input.The set of activation input can also be expressed as matrix structure.

Summary of the invention

Present specification describes for using neural network to generate unified machine learning model on data processing equipment System and method.According to described technology, data processing equipment is that each object vertex in a group objects vertex determines Practise target.Data processing equipment can two or more insertion (embedding) outputs neural network based it is every to determine A learning objective.Each insertion output can be by using triple loss function (triplet loss function) individually to instruct Experienced independent special purpose model generates.Each special purpose model is configured as identification data associated with special object vertex.

It is raw when data processing equipment training neural network is to identify data associated with each object vertex in group At unified machine learning model.Data processing equipment is instructed based on L2 loss function and using the corresponding learning objective of special purpose model Practice neural network.Data processing equipment generates unified machine learning model using housebroken neural network.Unified model can To be configured as the specific electron data item that the object that identification includes the item in each of object vertex indicates.

The one aspect of theme described in this specification can be embodied in for using nerve on data processing equipment Network generates in the method implemented by computer of unified machine learning computation model.This method comprises: simultaneously by data processing equipment And the corresponding learning objective of each of multiple object vertex is determined for neural network, wherein the definition of each object vertex belongs to The object on the vertex it is different classes of；Neural network is trained by data processing equipment and based on first-loss function, to know Data not associated with each of multiple object vertex, wherein neural network is trained using corresponding learning objective 's；And unified machine learning mould is generated by data processing equipment and using the neural network based on the training of first-loss function Type, this unify machine learning model be configured as identification be included in data associated with each of multiple object vertex Item.

These and other embodiment can respectively optionally include one or more of following characteristics.For example, one In a little embodiments, corresponding learning objective is determined for neural network further include: by data processing equipment and based on the second loss At least one other neural network of function training, to identify data associated with each of multiple object vertex；Response In training, two or more insertions are generated by data processing equipment and are exported, wherein each insertion output indicates specific study mesh The vector of mark and the parameter including corresponding to data associated with special object vertex；And by data processing equipment and Corresponding machine learning model, each engineering are generated using at least one other neural network based on the training of the second loss function It practises model and is configured with specific insertion output.

In some embodiments, corresponding learning objective is determined for neural network further include: provide from corresponding independent model The corresponding learning objective generated, for training neural network.In some embodiments, each of multiple object vertex pair Should be in the particular category of item, and data associated with each of multiple object vertex include in the particular category of item The image data of item.In some embodiments, particular category is dress ornament classification, and the item of particular category includes the following terms At least one of: handbag, shoes, one-piece dress, trousers or housing；And wherein in image data instruction the following terms extremely Few one image: specific handbag, specific shoes, specific one-piece dress, specific trousers or specific housing.

In some embodiments, each of corresponding machine learning model is configured as identification and special object vertex Associated data and in the first accuracy；And unified machine learning model is configured as in identification and multiple object vertex Each associated data and in the second accuracy more than the first accuracy.In some embodiments, determination is more The corresponding learning objective of each of a object vertex includes: two or more insertion outputs of analysis, each insertion output Corresponding to the special object vertex in multiple object vertex；And it is based on the analysis, determine each of multiple object vertex Corresponding learning objective.

In some embodiments, first-loss function is L2 loss function, and generates unified machine learning model packet It includes: generating and minimize the specific unified machine learning model for calculating output associated with L2 loss function.In some embodiment party In formula, neural network includes the multiple neural net layers for receiving multiple layers of input, and wherein based on the training of first-loss function Neural network includes: to execute batch standardization (batch normalization) with the layer for arriving specific neural net layer that standardizes Input；And standardize in response to executing batch, minimize covariant offset.In some embodiments, the second loss function It is triple loss function, and generating corresponding machine learning model includes: based between anchor image, positive image and negative image Association generates specific machine learning model.

The other embodiment of this aspect and other aspects includes correspondence system, device and computer program, is configured For the movement for executing method of the coding on computer memory device.The computing system of one or more computers or circuit can be with By the software, firmware, hardware or the their combination that are mounted in system come configured in this way, so that executing system in operation These movements.One or more computer programs can be by having instruction come configured in this way, which fills by data processing Device is made to execute these movements when setting operation.

Theme described in this specification can be implemented in a particular embodiment, to realize one or more following advantages. Object recognition receives more and more attention in vision research.In this background, described introduction includes using nerve Network to generate the process of unified machine learning model using L2 loss function, and wherein the unified model can be used for identifying or distinguishing The various objects (for example, one-piece dress, handbag, shoes) of not multiple object vertex.

For example, the image data of the given expression for describing dress, the unified model generated according to described introduction It can be used for positioning or retrieving the same or similar item.In some instances, the appearance of item may with illumination, viewpoint, block and Background condition and change.Different object vertex also can have different characteristics, so that the image from one-piece dress vertex can More deformations can be undergone than the image from shoes vertex.Therefore, because these are distinguished, independent model is trained to every to identify Item in a object vertex.

However, independent special purpose model requires the vast resources stored for model and increased calculating demand multiple to support The deployment of model.When using multiple models on a mobile platform, these resources burden may become more serious.Therefore, Unified model for the object recognition across different dress ornament vertex can reduce processor utilization, and provide increased example pair As the computational efficiency of discrimination system.In addition, object retrieval function can use more subtotal by being combined to multiple special purpose models It calculates in the single unified model of overlay capacity (footprint) and is effectively carried out.This causes model to generate and use technical field Technological improvement.

The details of one or more embodiments of theme described in this specification illustrates in the accompanying drawings and the description below. From specification, drawings and the claims, other potential features, aspects and advantages of theme be will become obvious.

Detailed description of the invention

Fig. 1 shows the neural network framework for generating machine learning model based on first-loss function.

Fig. 2 shows for generating the neural network framework of machine learning model based on the second loss function.

The example graph that Fig. 3 shows embedding data relevant to different object vertex indicates.

Fig. 4 is the exemplary flow for generating the process of unified machine learning model for multiple object vertex based on certain loss function Cheng Tu.

Fig. 5 shows the figured figure of the corresponding incorporation model of the object vertex including corresponding to specific dress ornament classification Show.

Fig. 6 is shown with the calculating function for obtaining the image data for the one or more machine learning models of training The diagram of energy.

Identical appended drawing reference and title indicate identical element in different attached drawings.

Specific embodiment

Deep neural network can be used to train, to distinguish certain kinds based on the reasoning learnt in machine learning system Other item.Deep neural network can generate reasoning based on the analysis to the received input data of machine learning system.Through instructing Experienced machine learning system can produce or generate one or more special purpose models, and the special purpose model uses the reasoning learnt Specific collection for particular category item identification or discrimination.

For example, special purpose model can be trained in the context that dress ornament distinguishes to distinguish and particular category or object vertex (object vertical) associated item, one-piece dress, trousers or portable such as into the image data of its application model Packet.In an alternative embodiment, classification or object vertex can correspond to various items or object, such as automobile, animal, the mankind Body and various physical objects, such as indicated in image data.In some embodiments, object vertex can correspond to sound Frequency signal data.

Generally, special purpose model may be significantly better than universal model, for example, being trained to distinguish the object top with wide scope The model of the associated item of point.Therefore, the item generated using deep neural network distinguishes model generally directed to different object tops Point is independently trained.Object vertex defines the different classes of of the object for belonging to the vertex, for example, for dress ornament, object top Point can be the item classification of cap, shoes, shirt, jacket etc..However, including the dedicated of the item of different object vertex for identification The item discrimination system deployment of model group is got up may be costly, and may not be able to sufficiently scalable (scalable).

In this context, theme described in this specification includes for using nerve net on data processing equipment The system and method that network generates unified machine learning model.Deep neural network can be used to generate in unified incorporation model, is somebody's turn to do Deep neural network utilizes the learning objective indicated by the insertion output generated by corresponding special purpose model.Example machine study Neural network (or deep neural network) can be used to learn for identification across multiple object vertex (for example, corresponding in system The item classification of various dress ornament types) various items reasoning.

E.g., including the phase of each object vertex in the machine learning system of neural network accessible group objects vertex Answer learning objective.System can it is neural network based two or more insertion output to determine corresponding learning objective.Each Insertion output can be generated by corresponding special purpose model, these special purpose models are individually trained using triple loss function, and Identify data (for example, image of luxury goods handbag) associated with special object vertex (for example, handbag).

The data processing equipment training neural network of system is associated with each vertex in the group objects vertex to identify Data.Neural network can be used the corresponding learning objective of special purpose model and be trained based on L2 loss function.Data processing Device generates unified machine learning model using housebroken neural network, this is unified machine learning model and is configured as identifying Associated with each object vertex (for example, shoes, handbag, jacket/shirt etc.) in group specific data item (for example, Branding campaign shoes, luxury goods wallet, luxury goods shirt etc.).

Fig. 1 shows the nerve network system framework for generating example machine learning model based on first-loss function 100 (" systems 100 ").Generating machine learning model may include that system 100 executes nerve associated with reasoning workload Network query function.Specifically, the calculating of reasoning workload may include handling neural network by the layer of neural network to input (example Such as, input activation).Each layer of neural network may include one group of parameter (for example, weight), and by neural net layer Reason input may include that input activation and parameter is used to calculate dot product as the operand calculated.

System 100 generally includes the exemplary neural network indicated by neural network framework 102.The neural network of framework 102 It may include basic network 103, pond layer 104, the first articulamentum collection 106, the second articulamentum collection 108 and insertion output 110.Base Plinth network 103 may include the subset of the neural net layer of framework 102.

For example, deep neural network may include the basic network 103 comprising multiple convolutional layers.These convolutional layers can be used for Execute complicated calculating, the calculating of the complexity is for including that the computer based of various items in various image datas is distinguished Not.In some embodiments, basic network 103 can be initial (inception) v2, initial v3, initial v4 or another phase Close neural network structure.Although being described in the context of image data, the process of this specification can be applied In the detection or discrimination of audio signal data.

Framework 102 may include executing various functions associated with the reasoning and calculation of training machine learning model is used for Various additional neural network layers.For example, pond layer 104 can be average pond layer or maximum pond layer, executes and be used for down Sampling operation, the relevant function of pondization output activation.Down-sampling operation can by modify it is relevant to input data set certain A little Spatial Dimensions reduce the size of output data set.

Articulamentum collection 106,108 can be the corresponding set for being fully connected layer comprising have all into preceding layer The artificial neuron of activation being fully connected.Insertion output 110 can correspond to include given output dimension (64-d, 256-d Deng) floating number/parameter vector one or more output characteristic sets.As described in more detail below, when system 100 When exemplary neural network is trained to execute the certain computing functions for distinguishing or identifying for object/item, insertion output 110 is by shape At, generate or generate.

System 100 may include the one or more processors to form one or more neural networks and other interlock circuits Component.Generally, various processor architectures can be used to implement in method and process described in this specification, such as centre Manage unit (Central Processing Unit, CPU), graphics processing unit (Graphics Processing Unit, GPU), digital signal processor (Digital Signal Processor, DSP) or other relevant processor architectures.

System 100 may include multiple computers, calculation server and other calculating equipment, each include processor and Storage can be by the memory for the calculating logic or software instruction that processor executes.In some embodiments, system 100 includes one A or multiple processors, memory and data storage device, one or more frameworks 102 are collectively formed in they.The place of framework 102 Reason device handles the instruction executed by system 100, including storing the instruction in equipment in memory or on storage devices.The instruction of storage Execution can to execute machine-learning process described herein.

Referring again to FIGS. 1, system 100 is configured as executing various calculating operations relevant to machine-learning process.For example, System 100 executes learning manipulation 112 and 114 and with training neural network to generate one or more special purpose machinery learning models Relevant various other operations.In some embodiments, system 100 executes programming code or software instruction to execute and learn Operate 112 and 114 associated calculating.As described in more detail below, learning manipulation 112 and 114 is executed by system 100, with Corresponding dedicated learning model is trained based on the triple loss function indicated by calculating logic 116.

Learning manipulation 112 includes that system 100 generates model training data using the neural network of framework 102.Model instruction Practicing data can correspond to be exported when system 100 is trained to generate specific special purpose model by the insertion that system 100 generates. In some embodiments, system 100 generates multiple and different special purpose models, and generates the set of individually insertion output, In be specifically embedded in output set correspond to specific special purpose model.

For example, independent special purpose model can be generated to distinguish and retrieve in the context that dress ornament distinguishes or dress ornament is retrieved The dress ornament item of different dress ornament classifications (for example, one-piece dress, jacket, handbag etc.).In some embodiments, independence can be used Sub-network learn for distinguishing image that image from various websites or other users generate (for example, setting using movement It is standby/smart phone capture " street " or " actual life " digital picture) incorporation model.

During model training, for the sub-network of each vertex (such as one-piece dress, handbag, glasses and trousers), all It can independently finely tune.For these sub-networks, the result of model training can make machine learning system generate up to 11 Independent special purpose model, each independent special purpose model correspond to one in 11 vertex.As it is used herein, " vertex (vertical) " or " object vertex " corresponds to object or item classification.Dress ornament is distinguished, object vertex can be dress ornament item class Not, one-piece dress, handbag, glasses, trousers etc..As discussed in more detail below, for the item in particular category/vertex The object recognition of (for example, dress ornament item) can produce basic accurately item discrimination results using independent model.

In some embodiments, system 100 trains the neural network of framework 102 using the image data of dress ornament item, The dress ornament item is respectively associated from different dress ornament classifications.For example, the handbag of multiple and different types can be used in system 100 Image data trains neural network.System 100 may then based on the insertion output generated in response to training neural network To generate identification or distinguish the special purpose model of certain types of handbag.

As described above, the set of specific insertion output can correspond to specific special purpose model.For example, being distinguished in dress ornament In, the first set for being embedded in output can correspond to for generating the first model for distinguishing specific shirt/shirt or jacket The reasoning learnt neural metwork training data (for example, with reference to operation 114).Similarly, it is embedded in the second set of output It can correspond to for generating the reasoning learnt for distinguishing the second model of specific jeans/trousers/skirt or lower dress Neural metwork training data.

Insertion output each set include insertion feature vector, these insertion feature vectors can be extracted and for pair As or item retrieval.As described in more detail below, the set of the insertion feature vector extracted can correspond to accordingly learn mesh Mark, and the insertion output of housebroken neural network model may include these insertion feature vectors.

One or more learning objectives can be used for training machine learning system (for example, system 100) to generate specific type Dedicated computing model.For example, multiple and different learning objectives can be used for instructing as the feature below with reference to Fig. 2 is discussed At least one of white silk discrimination item associated with multiple and different vertex or classification unifies machine learning model.

Referring again to FIGS. 1, system 100 executes learning manipulation 114, based on the model training determined in learning manipulation 112 Data generate dedicated learning model.In some instances, it is individual embedding corresponding to " study " to determine or generate model training data Enter the instantiation procedure of model.In some embodiments, for model training and characteristic vector pickup, system 100 is in training (example Such as, the first order) when and extract be used for object retrieval insertion feature vector (for example, second level) when use two-stage approach.

From the context for retrieving dress ornament in image data, the first order may include carrying out to the dress ornament item of image data It positions (localize) and classifies.In some instances, carrying out classification including system 100 to the dress ornament item of image data is image The dress ornament item of data determines object type label.For example, example object detector can be used to analyze picture number in system 100 According to.Then analysis data can be used to detect the image data including object properties associated with handbag in system 100 Object or dress ornament item.Based on the analysis and detection, then system 100 can determine that the object type label of dress ornament item is " portable Packet " class label.

In some embodiments, system 100 include object detection framework, be for basic network 103 single-shot it is more Frame (single-shot multi-box, SSD) detector, basic network 103 is initial V2 basic network.In other embodiment party In formula, system 100 can be configured as to be combined using or including various other object detection frameworks and basic network.SSD can be with It is the example calculations module of system 100, executes program code so that executing one or more object detection functions.

For example, the SSD detector module can provide the bounding box for defining the object of (bound) image data.SSD may be used also The dress ornament class label that object is handbag, glasses item or one-piece dress is defined to provide instruction.In some embodiments, scheme As the object pixel of data can be cut (crop), the specific incorporation model that system 100 then can be used is cutting image It is upper to extract various features.Subprocess step associated with the first order in two stage process can be used for based on various image datas To train dedicated incorporation model.

In response to determining object type label in the first order, system 100 may be advanced to the second level and the dedicated insertion of training Model calculates the similarity feature for object retrieval.For example, three indicated by calculating logic 116 can be used in system 100 Tuple loss function is trained to execute incorporation model.More specifically, system 100 uses triple sequence loss (triplet Ranking loss) it is embedded in learn each independent item/object vertex or class another characteristic.

For example, triple includes anchor image, positive image and negative image.During triple study, system 100 seeks to generate Insertion, so that positive image is close to anchor image in the feature space of neural network, and negative image pushed away from anchor image.From triple The insertion learnt in training be used to calculate image similarity.

For example, enablingFor triple, whereinIt respectively indicates anchor image, positive image and bears Image.The aim of learning of system 100 is to minimize calculating output associated with the following loss function as shown in equation (1),

(1)

Wherein: i) α is compulsory spacing (margin), ii between positive and negative couple) be f (I) image I feature insertion, and iii)D(fx,f_y) it is two feature insertion f_xAnd f_yThe distance between.

For being related to the embodiment of specific dress ornament object vertex training independent model, positive image is identical as anchor image Product (for example, how fragrant youngster's handbag), and negative image is another product, but on identical dress ornament vertex (for example, luxurious Product handbag) in.In some embodiments, system 100 is executed for negative excavation function (the semi-hard negative of semihard Mining function) calculating logic for obtaining negative image data.For example, system 100 it is accessible it is online/be based on net The resource of network, and strong negative object images are identified using the negative excavation of semihard.Then these object images can be used in system 100 Come enhance or improve specific special purpose model training validity.

As described above, the model training of the triple loss function of logic-based 116 generates such as including feature vector It is embedded in the training data of output.The set of the insertion feature vector extracted can correspond to corresponding learning objective.Therefore, it is learning Operation 118 is practised, system 100 can determine corresponding learning objective based on triple loss model training data.System 100 is then Two or more learning objectives can be used and come training machine learning system (for example, system 200 of Fig. 2), to generate at least One unified computation model, as described below.

Fig. 2 shows for generating example machine learning model based on the second loss function (for example, L2 loss function) Nerve network system framework 200 (" system 200 ").As shown, system 200 includes the spy essentially identical with above system 100 Sign.However, system 200 includes L2 standardization layer 212, this be will be described in greater detail below.In some embodiments, system 200 are the subsystem of system 100, and can be configured as the various computing functions for executing above system 100.

System 200 is configured as learning or generating the unified incorporation model trained based on the reasoning learnt.These are learned The reasoning practised, which makes it possible to be grouped various items, carries out object recognition, wherein each grouping correspond to different object vertex or Classification (for example, dress ornament classification).System 200 is by combination when system 100 when training corresponding special purpose model as described above for producing Raw training data learns one or more unified models.In some relevant model learning/Training scenes, work as triple When loss is for training unified model, training data of the combination from independent model will lead under performance with generating unified model Drop.

For example, using the triple loss function of logic 118 come the combined training data (example based on different object vertex Such as, 112) training unified model with accuracy for the model of each individually vertex training compared with may lead for insertion output It causes to distinguish being remarkably decreased for accuracy.As more object type (or vertex) accumulation is into single unified model, under these Drop may occur.However, as described in more detail below, when insertion of the combination from multiple special purpose models is to generate single system When one model, performance can be led to and distinguish that the essence of accuracy improves by reducing training difficulty and training complexity.

According to described introduction, unified incorporation model is can be generated in system 200, when compared with individual special purpose model When, this is unified incorporation model and realizes equivalent performance and distinguish accuracy.In addition, unified model can have with it is single individually Special purpose model is identical or even less model complexity.Therefore, present specification describes for mitigating or reducing the multiple tops of training The difficulty in model insertion put allows to generate the improved process of unified model and method.

For example, independent special purpose model can be trained first, the desired of the object for including in image data is distinguished to realize Threshold value level of accuracy.As set forth above, it is possible to be instructed using system 100 and based on the triple loss function of calculating logic 118 Practice independent model.Then the insertion output of the model of each stand-alone training is used as learning objective, with training example unified model.

In some embodiments, specific special purpose model can have 0.66 example accuracy metric, wherein The time of 66.1% (66.1), upper model accurately identified specific handbag.It can according to the unified model that described introduction generates To achieve over the accurate object recognition knot of the accuracy metric (for example, 66.1) of the object recognition result of specific special purpose model Fruit.For example, the unified model generated according to described introduction is for handbag dress ornament classification in the context that dress ornament distinguishes It can have the object retrieval of 0.723 accuracy metric or 72.3% accuracy or distinguish accuracy.

It in some embodiments, include the classification (example of determining object using unified model accurate discrimination/identification object Such as, " handbag "), determine the owner or designer's (for example, " how fragrant (Chanel) " or " Gucci (Gucci) ") of object, And/or determine type/style (for example, " fragrant how the classical flip lid of youngster 2.55 wraps ") of handbag.In some instances, pass through unification Model identification object may include the figured associated picture number that model index (for example, object retrieval) includes object According to.

Referring again to FIGS. 2, system 200 is configurable to generate unified model, which can be executed for accurate right As the multiple tasks for distinguishing and retrieving, in existing system, these tasks are to be executed by independent model, but accuracy is more It is low.In addition, described introduction includes improving emulation independent model insertion output by using L2 loss function (for example, learning Practise target) method and process.

For example, the instruction of combination two different object vertex (for example, handbag and shoes) is lost and passed through based on triple Practice data to train unified model, the unification for executing the object recognition of the item in those vertex with reasonable accuracy can be generated Model.However, when combining the training data of three or more different object vertex using triple loss may cause with The unified model that substantially poor object recognition accuracy executes.Triple loss function is being based on for several different vertex The difficulty and complicated calculating challenge that occur when training unified model cause the accuracy of difference.

In order to mitigate this trained difficulty, this description presents a kind of Learning Scheme, the Learning Scheme is using acting on one's own It uses the insertion of model to export as learning objective, allows to lose using L2 and be lost instead of triple.L2 loss function makes With the training difficulty for generating unified model is alleviated, and the more efficient use of the feature space of neural network is provided.Most terminate Fruit is a unified model, it can obtain with the retrieval accuracy of many independent special purpose models identical (or higher), have simultaneously There is the model complexity of single special purpose model.

For example, system 200 learns unified learning model using the corresponding learning objective of independent model, so that unified from this The insertion that model generates is identical as the insertion of independent special purpose model that system 100 generates (or very close).In some embodiment party In formula, system 200 determines the corresponding study of each object vertex in a group objects vertex using the neural network of framework 102 Target.Each of corresponding learning objective can be exported with specific insertion neural network based.

Fig. 2 shows for generating the calculating operation of unified machine learning model.For example, in the learning manipulation 214 of Fig. 2, System 200 accesses learning objective corresponding with the insertion of the feature of corresponding special purpose model.In learning manipulation 216, system 200 generate with Feature for detecting the object on various vertex is embedded in corresponding unified model training data.Feature insertion is based in unified model The ANN Reasoning occurred during study calculates.

For example, enablingWherein each V_iIt is that its data can be combined to train the vertex of incorporation model Set.It enablesFor the set of incorporation model, wherein each M_iIt is for vertex set V_iThe model of study.It enablesFor the set of N number of training image.If I_jVertex ∈ V_s, s=1...K, then its corresponding model M_sFor Generate image I_jInsertion feature.Enable f_sjIt is expressed as image I_jFrom M_sThe feature of generation is embedded in.

In learning manipulation 220, system 200 generates unified machine learning model, this is unified machine learning model and is configured as Identification includes the particular item in exemplary image data.Image data can be associated with each object vertex in group, and Unified model is generated using based on the neural network of certain loss function (for example, L2 loses) training.For example, 200 quilt of system It is configured to learning model U, so that the feature generated from model U and the spy generated from the independent special purpose model generated by system 100 It levies identical.

Specifically, f is enabled_ujIndicate the feature generated from model U insertion.The aim of learning of system 200 is that determination can be minimum Change the model U that calculates output associated with the following loss function as shown in equation (2).

(2)

With reference to equation (2), feature f_ujIt is calculated from model U, and feature f_sjIt can be calculated from above-mentioned different special purpose model. Model learning description above uses the L2 loss function indicated by calculating logic 218 and above equation (2), rather than by counting Calculate the triple loss function of logic 116 and above equation (1) instruction.

In some embodiments, system 200 is configurable to generate the unified model of the output dimension 107 with 256-d. It is greater than or the 256-d of substantially greater than unified model is defeated on the contrary, can have in the single special purpose model that learning manipulation 114 generates The output dimension 107 of dimension out, such as 4096-d.As described above, the use of L2 loss is provided than triple loss less Complicated and less difficult training process.

In addition, the use of L2 loss function is smaller multiple in providing in for learning art (such as batch standardizes) Miscellaneous degree and difficulty.For example, being standardized by batch, neural net layer input can be typically canonicalized, to allow higher study speed Rate.For example image disaggregated model, when compared with the learning art being used together with other loss functions, batch standardizes Less training step is used for realize desired threshold value accuracy metric (for example, 0.60 or higher).

In some embodiments, the batch standardization function applied via L2 standardization layer 212 is executed in response to system 200 Can, covariant offset (covariate shift) can be minimized.For example, deep neural network may include in sequence Multiple layers.Training deep neural network is typically due to following fact and is complicated: for example, as the parameter of previous layer in sequence changes Become, the possibility change during being distributed in model training of each layer of input.

This change can reduce the speed using deep neural network training pattern, thus need slower learning rate and Careful parameter initialization.It is inclined that covariant inside neural network can be described as to the parameter change that training speed adversely affects It moves.However, losing by using L2, the batch process of normalization for the layer input that standardize can be executed, with solution or minimum Change the adverse effect caused by covariant offset to training speed.

In addition, being allowed using the learning method that L2 loses generation unified model using relative to triple loss incrementss Unlabelled data.For example, using for training dress ornament to distinguish that the triple of model loses learning method, it may be required that product mark (for example, " fragrant how the classical flip lid of youngster 2.55 wraps ") is known to generate the embedding data of trained triple.However, the mould lost using L2 Type training and learning method only require vertex label, can be automatically generated by example position fixes/disaggregated model.Therefore, L2 is damaged The use of mistake can reduce processor utilization by the aforementioned calculating for determining product identification, and increase and be used for extra computation System bandwidth.

Alternatively, described introduction further includes that can be used for reducing special purpose model for selecting vertex data to be combined to produce Quantity particular model (for example, unified model or other relevant built-up pattern forms) method.This is specifically combined Model can successfully be learnt, and be can have and be similar to, is accurate equal to or more than the discrimination of each independent special purpose model The comparable object recognition accuracy of degree.

For example, the combination of selective or " intelligent " vertex can be used for determining built-up pattern (for example, example unified model). In some embodiments, system 200 may include calculating logic, which can combine for determination from die for special purpose Which embedding data on the different vertex of type is to generate example combination model.Specifically, since the first vertex, system 200 can Gradually to add the embedding data from other vertex.When adding embedding data, system 200, which can execute sample item and distinguish, appoints Business, to determine whether the model learnt from data splitting causes the accuracy observed to decline.

In some instances, system 200 can steadily add the embedding data from other vertex, until observing standard Exactness decline.In other examples, system 200 is the specific combination that multiple special purpose models determine vertex, wherein each die for special purpose Type is distinguished for the item across vertex subset.In addition, system 200 can determine the specific combination on vertex for multiple special purpose models, together When also keep threshold value level of accuracy.Then special purpose model corresponding with the certain vertex in subset can be used in system 200 Feature insertion, and built-up pattern is generated based on feature insertion.

Fig. 3 shows the graphical representation of the embedding data 300 of different object vertex in the feature space of exemplary neural network. Graphical representation, which is indicated generally at the unified model based on described introduction training (for example, study), can provide the spy of neural network Levy the more effective and wider use in space.For example, the described learning method using L2 loss can be by being utilized as The Feature Mapping (for example, learning objective) pre-established of independent special purpose model study effectively trains unified model.

Embedding data 300 includes the t distribution random neighborhood insertion (t- generated from the insertion of the feature of independent special purpose model Distributed stochastic neighbor embedding, t-SNE) visualization.Specifically, embedding data 300 includes 2,000 images from each vertex 302,304 and 306, wherein data are projected to downwards the space 2D for visualizing.

Fig. 3 indicates feature insertion f_sjIt is to be opened for 302,304,306 points across vertex in feature space.In other words, (come From model M_s) each vertex f_sjIncorporation model a part (for example, 64-d) of dimensional space, a therefore unification is used only The insertion on each dress ornament vertex (for example, 8 vertex in total) that model can be included in the insertion that learn to come data splitting 300 Output.

Fig. 4 is the exemplary flow for generating the process of unified machine learning model for multiple object vertex based on certain loss function Cheng Tu.Process 400 corresponds to the development for generating unified machine learning model, wherein model generated has at least The item of the accuracy metric of special purpose models different equal to two or more distinguishes accuracy metric.Process 400 can be used System 100 or 200 is stated to implement, wherein system 100 can execute be described function associated with subsystem 200.

Process 400 includes block 402, and in block 402, system 100 determines each object vertex in a group objects vertex Corresponding learning objective.In some embodiments, the neural network of system 100 two or more insertions neural network based Output is to determine corresponding learning objective.

For example, object vertex can be dress ornament classification, such as one-piece dress, shoes or handbag.In addition, each vertex can be with The insertion output generated when identifying or distinguish the dress ornament or clothes item in the vertex corresponding to being trained in particular model.Example Dress ornament item may include cocktail one-piece dress, basketball shoes or brand name monogram handbag.

In the block 404 of process 400, system 100 is based on first-loss function (for example, L2 loses) training neural network, with Identify data associated with object vertex each in group.In some embodiments, using for each object vertex determine Corresponding learning objective trains neural network.For example, given image file or image data, system 100 can train nerve net Network is come at least: i) identifying the one-piece dress item in image based on the analysis of the pixel data to image；Ii) based on the picture to image The analysis of prime number evidence is to identify the shoes item in image；Or iii) image identified based on the analysis of the pixel data to image In handbag item.

In block 406, system 100 generates unified machine learning model, this is unified machine learning model and is configured as identification packet Include the item in data associated with each object vertex in this group of vertex.For example, the data processing equipment of system 100 can One executed in object recognition function described herein is generated to use based on the neural network of first-loss function training Or multiple unified machine learning model.

In some instances, determine corresponding learning objective include: i) training neural network with identify in object vertex Each associated data, wherein based on the second loss function training neural network；And ii) generate at least two insertion it is defeated Out, wherein each insertion output indicates the specific learning objective of corresponding learning objective.Other than indicating specific learning objective, often It is a insertion output may include generally correspond to the attribute of image data associated with special object vertex floating number (or ginseng Number) vector.

In some embodiments, system 100 generate corresponding machine learning model, wherein each of model be using What the neural network based on the second loss function (for example, the triple loss) training for being different from first-loss function generated.This Outside, each of model can be used it is specific insertion output floating number vector come identify special object vertex dress ornament or Clothes item.In some instances, insertion is generated to be in response to occur in training neural network.

As discussed above referring briefly to Fig. 2, in some embodiments, generating unified machine learning model be can wrap Include combination and the different associated training datas in dress ornament vertex.In order to which calibration object identification and retrieval performance are (for example, determine study Target), learn the insertion on each vertex using triple loss first.The purpose of vertex combination can be using smaller amounts Individual special purpose model, the retrieval accuracy decline observed without any.

Table 1: the retrieval accuracy (percentage) of different vertex combinations compares

Table 1 shows (1) individually model, (2) one-piece dress-jacket joint or built-up pattern and (3) one-piece dress-jacket- The example of retrieval accuracy measurement of the housing conjunctive model on " one-piece dress ", " jacket " and " housing " vertex.With independent model It compares, on " one-piece dress " and " jacket ", one-piece dress-jacket conjunctive model performance is closely similar or slightly good, however, even clothing Skirt-jacket conjunctive model is performed poor in terms of the retrieval accuracy of dress ornament item in the classification of " housing " vertex.

In addition, one-piece dress-jacket-housing conjunctive model will lead under the significant accuracy on all three vertex Drop.The accuracy data of table 1 indicates that some vertex can be recombined to obtain the better accuracy of the model more trained than individually, But only to a certain extent, the model training difficulty of triple loss function causes accuracy to decline (as described above) later.

The instantiation procedure of system 100,200 may include the different vertex of combined training data.Specifically, nine dress ornament tops Point can be combined into four groups, wherein for every group of trained built-up pattern.In some instances, user demand can be based on Certain amount of group will be combined into more or less than nine dress ornament vertex.

Set of vertices
	One-piece dress, jacket
Footwear, handbag, glasses
	Housing
Skirt, shorts, trousers

Table 2: set of vertices is combined

Four groups are shown in table 2, and the individually trained comparable capability retrieval of model that can have with each group is quasi- Exactness.In some embodiments, " clean triple (clean triplets) " is used to finely tune each of four models, Wherein clean triple is (as described below) obtained from " picture search " triple.For example, system 200 can be configured as Model performance is finely tuned using clean data, with effectively improving for the retrieval accuracy of each of four models of realization.

In general, by combined training data four incorporation models can be obtained for nine different dress ornament vertex.Based on upper Introduction is stated, can be all nine vertex training or study unified model, then generate the model.In some embodiments, The model of generation is disposed operating with for multiple users.For example, unified model can receive the picture number of user's transmission According to wherein user seeks to obtain the identification data about the special object or dress ornament item that include in image.

In some embodiments, unified model is configured as receiving the image data of image, identifies or distinguishes in image Dress ornament item, and determine identification information about the dress ornament item in image.In some instances, unified model is configured as retrieving The associated picture of dress ornament item, and the identification information about dress ornament item and/or the associated picture of dress ornament item is provided, so as to via shifting Dynamic equipment is exported to user.

Fig. 5 shows the figured figure of the corresponding incorporation model of the object vertex including corresponding to specific dress ornament classification Show 500.In addition, the description of diagram 500 can correspond to be used to extract one or more features by the execution of system 100,200 The process of insertion.As described above, exemplary two-stage method can be used for extracting and item or dress ornament object in the context that dress ornament distinguishes Image data associated feature insertion.

Shown in as shown 500, given image/image data 502 can be detected in the picture first in block 504 With positioning clothing item.In block 506, it then is embedded in (for example, vector of floating number) from cutting image, to indicate clothes .System 100,200 is compared using insertion for item/object retrieval similarity image.In some embodiments, from sanction The insertion for cutting image acquisition can correspond to the learning objective on dress ornament vertex belonging to clothes item.

As shown in corresponding arrow 507, each incorporation model 508,510,512 and 514 corresponds to for identification and retrieves The object retrieval function of certain dress ornament items.Dress ornament item can correspond to the item described in the specific cutting part of image data.Phase Than under, arrow 515 indicate unified incorporation model 516 correspond to for identification with retrieval in block 506 each of image data Cut the object retrieval function of the dress ornament item for the object described in part.

Block 506 include it is corresponding cut image, each clothes for cutting the object vertex classification that image includes incorporation model or The expression of dress ornament item.Correspond to the figure with retrieval handbag for identification for example, describing the first of handbag and cutting image data As the incorporation model 510 of data.Similarly, describe trousers second cut image data correspond to for identification with retrieval trousers Image data incorporation model 512.However, each cutting image data at block 506 is corresponding to each with retrieval for identification The unified incorporation model 516 of the image data of the dress ornament item of seed type.

Fig. 6 is shown including for obtaining the calculating function for being used for the image data of the one or more machine learning models of training The example diagram 600 of energy.Diagram 600 can correspond to be executed by one or more computing modules of above system 100 or 200 Computing function.

In logical block 602, training data relevant to image is collected first from search inquiry.It can be from reception and storage Search inquiry is accessed in the example search system of large capacity image search query.It is looked into for example, system 100 can be configured as access Ask data storage device, such as storage using Google's picture search (Google Image Search) submit it is thousands of (for example, 200,000) the privately owned inquiry repository of a user query.Search inquiry may include specified product or dress ornament item title.

In block 604, system 100 executes the resolution logic for being used for sample query text resolver, is stored with obtaining from inquiry The dress ornament class label of each text query obtained in library.Text query can be different from nine vertex it is associated: i) even Clothing skirt, ii) jacket, iii) footwear, iv) handbag, v) glasses, vi) housing, vii) skirt, viii) shorts and ix) trousers. In block 606, system 100 is that each search inquiry selects certain amount of highest to grade image (for example, 30 images), wherein scheming It seem to be graded based on the degree of the dress ornament item of the image pixel quality and query text data accurate description image.

The image data of highest grading image can be used to form above-mentioned " picture search triple ", and wherein triple includes just Image, negative image and anchor image.System 100 can at least identify the subset of triple (for example, each object vertex 20,000 Triple), system or user rating and verifying are carried out for the correctness to image each in triple.In some embodiments In, grading and image authentication include whether anchor image in determining triple and positive image come from identical product/vertex classification. The subset for being rated and being verified as correct triple image can be used for being formed the second set of triple, referred to herein as " clean Net triple ".

It, can be from the image data (example for using one or more types before generating unified model in logical block 608 Such as, image net (ImageNet) data) pre-training model initialization basic network 103.In some embodiments, work as generation When unified incorporation model, trained number identical with the triple feature learning of two or more special purpose models is used for can be used According to learning unified incorporation model.For example, vertex label data, institute as above can be only required by generating unified insertion learning model It states, can be obtained via locator/classifier.Therefore, it is possible to use with the training that generates during triple insertion study The identical training image of image learns to generate unified insertion.

Theme described in this specification and the embodiment of operation can be implemented in Fundamental Digital Circuit, or include Implement in the computer software of structure and its equivalent structures disclosed in this specification, firmware or hardware, or in them One or more combinations in implement.The embodiment of theme described in this specification may be embodied as coding and deposit in computer One or more computer programs on storage media, i.e. one or more modules of computer program instructions, for by data Manage the operation that device executes or controls data processing equipment.

Alternatively or additionally, program instruction can be coded on manually generated transmitting signal, such as machine generates Electricity, light or electromagnetic signal, which is generated with encoded information, suitable acceptor device is used for transmission, by data Processing unit executes.Computer storage medium can be computer readable storage devices, computer-readable storage substrate, it is random or The combination of serial access memory array or equipment or one or more of which is included therein.Although in addition, calculating Machine storage medium is not transmitting signal, but computer storage medium can be the meter encoded in manually generated transmitting signal The source or destination of calculation machine program instruction.Computer storage medium is also possible to one or more independent physical assemblies or medium It (for example, multiple CD, disk or other storage equipment) or is included therein.

Operation described in this specification can be implemented for by data processing equipment to being stored in one or more computers The operation executed in readable storage device or from the received data in other sources.

Term " data processing equipment " covers device, equipment and the machine of all kinds for handling data, including example Such as programmable processor, computer, system on chip or among the above multiple or combination.Device may include dedicated logic circuit, Such as FPGA (Field Programmable Gate Array, field programmable gate array) or ASIC (Application- Specific Integrated Circuit, specific integrated circuit).In addition to hardware, device can also include to be discussed Computer program creation running environment code, for example, constitute processor firmware, protocol stack, data base management system, operation The combined code of system, cross-platform runtime environment, virtual machine or one or more of which.Device and running environment can To realize a variety of different computation model infrastructure, such as network service, distributed computing and grid computing infrastructure.

Computer program (also referred to as program, software, software application, script or code) can programming language in any form Speech is write, including compiling or interpretative code, statement or procedural language, and can dispose in any form, including as independence Program as module, component, subroutine, object or is suitble to other units for using in a computing environment.Computer program can With but not necessarily correspond to the file in file system.Program, which can store, is saving other programs or data (for example, being stored in One or more scripts in marking language document) file a part in, be stored in the list for being exclusively used in discussed program In a file, or multiple coordination files are stored in (for example, storing the part of one or more modules, subprogram or code File) in.Computer program can be deployed as being located at a website or being distributed across multiple websites and to pass through communication network mutual It is executed on multiple computers or a computer even.

Process and logic flow described in this specification can be executed by one or more programmable processors, the processing Device runs one or more computer programs to execute movement by being operated to input data and generating output.Process and Logic flow can also be executed by dedicated logic circuit, and device also may be embodied as dedicated logic circuit, for example, FPGA (field programmable gate array) or ASIC (specific integrated circuit).

For example, the processor suitable for running computer program includes both general and special microprocessors, Yi Jiren Any one or more processors of the digital computer of which kind of class.In general, processor will be from read-only memory or arbitrary access Memory or both receives instruction and data.The primary element of computer is the processor and use for acting according to instruction execution In one or more memory devices of store instruction and data.

In general, computer will also be including one or more mass-memory units for storing data, such as disk, magnetic CD or CD, or be operatively coupled to receive data from one or more mass-memory units or by data transmission To it or receive transmission the two.However, computer does not need have such equipment.In addition, computer can be embedded into it is another In equipment, such as mobile phone, personal digital assistant (Personal Digital Assistant, PDA), Mobile audio frequency or view Frequency player, game console, global positioning system (Global Positioning System, GPS) receiver or portable It stores equipment (such as universal serial bus (Universal Serial Bus, USB) flash drive), only lifts several examples.

Equipment suitable for storing computer program instructions and data include the nonvolatile memory of form of ownership, medium and Memory devices, including such as semiconductor memory devices, such as EPROM, EEPROM and flash memory device；Disk, for example, it is internal Hard disk or moveable magnetic disc；Magneto-optic disk；With CD-ROM and DVD-ROM disk.Processor and memory can be by special logic electricity Road supplements or is incorporated to dedicated logic circuit.

In order to provide the interaction with user, the embodiment of theme described in this specification can have for user Show the display apparatus of information (for example, CRT (Cathode Ray Tube, cathode-ray tube) or LCD (Liquid Crystal Display, liquid crystal display) monitor) and user can by its to computer provide input keyboard and Implement on the computer of pointing device (for example, mouse or trackball).

Also the equipment of other types can be used to provide the interaction with user；For example, the feedback for being supplied to user can be with It is any type of sense feedback, such as visual feedback, audio feedback or touch feedback；And it can receive and in any form From the input of user, including sound, voice or tactile input.In addition, computer can be sent by the equipment used to user Document and the equipment used from user receive document and interact with user；For example, by response to being connect from network (web) browser The request of receipts, the web browser on the user equipment of user send webpage.

The embodiment of theme described in this specification can be implemented in computing systems, which includes rear end group Part, such as data server, or including middleware component, such as application server, or including front end assemblies, such as The graphic user interface or net that can be interacted by the embodiment of itself and theme described in this specification with user Rear end as the subscriber computer or one or more of network browser, middleware or front end assemblies any combination.System Component can pass through any form or the digital data communications (for example, communication network) of medium and be connected with each other.Communication network Example includes local area network (Local Area Network, " LAN ") and wide area network (Wide Area Network, " WAN "), mutually Networking network (for example, internet) and peer-to-peer network (for example, self-organizing peer-to-peer network).

Computing system may include user and server.User and server are generally remote from each other, and usually by logical Communication network interacts.Relationship between user and server is used by running and having on corresponding computer each other Family-relationship server computer program and generate.In some embodiments, server sends data (example to user equipment Such as, html page) (for example, in order to show data to the user interacted with user equipment and be handed over from user equipment Mutual user receives the purpose of user's input).At server the number generated at user equipment can be received from user equipment According to (for example, result of user's interaction).

Although this specification includes many specific embodiment details, these be not necessarily to be construed as to any invention or It may require the limitation of the range of protection, but the description of the feature to the specific embodiment specific to specific invention.This explanation Certain features described in the context of separated embodiment can also combine implementation in a single embodiment in book.

On the contrary, the various features described in the context of single embodiment can also in various embodiments dividually or Implemented with any suitable sub-portfolio.In addition, although feature can be described as working with certain combinations above, even Initially so it is claimed, but in some cases, it can be from one deleted in combination claimed in combination Or multiple features, and combination claimed can be directed toward the modification of sub-portfolio or sub-portfolio.

Similarly, although depicting operation in the accompanying drawings with particular order, this should not be construed as requiring the spy shown in It is fixed sequentially or continuously sequentially to execute these operations, or require to execute it is all shown in operate, to obtain desired result.At certain In a little situations, multitasking and parallel processing be may be advantageous.In addition, point of the various system components in above-described embodiment From should not be construed as requiring this separation in all embodiments, and it is to be understood that described program assembly and system It usually can integrate in single software product or be encapsulated into multiple software product.

Therefore, it has been described that the specific embodiment of theme.Other embodiments are within the scope of the appended claims.One In a little situations, the movement enumerated in claim can be executed in different order, and still obtain desired result.This Outside, the process described in attached drawing not necessarily require shown in particular order or consecutive order obtain desired result.Certain In embodiment, multitasking and parallel processing be may be advantageous.

26页详细技术资料下载

上一篇：一种医用注射器针头装配设备

下一篇：RPU阵列的更新管理

The unified insertion of study

相关技术

网友询问留言