WO2021189960A1 - 对抗网络训练、医疗数据补充方法、装置、设备及介质 - Google Patents

对抗网络训练、医疗数据补充方法、装置、设备及介质 Download PDF

Info

Publication number
WO2021189960A1
WO2021189960A1 PCT/CN2020/135342 CN2020135342W WO2021189960A1 WO 2021189960 A1 WO2021189960 A1 WO 2021189960A1 CN 2020135342 W CN2020135342 W CN 2020135342W WO 2021189960 A1 WO2021189960 A1 WO 2021189960A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
loss value
sample
network
generated
Prior art date
Application number
PCT/CN2020/135342
Other languages
English (en)
French (fr)
Inventor
李彦轩
朱昭苇
孙行智
Original Assignee
平安科技(深圳)有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 平安科技(深圳)有限公司 filed Critical 平安科技(深圳)有限公司
Publication of WO2021189960A1 publication Critical patent/WO2021189960A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G16INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16HHEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/70ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214Generating training patterns; Bootstrap methods, e.g. bagging or boosting

Definitions

  • This application relates to the field of artificial intelligence technology, and in particular to a method, device, equipment, and medium for combating network training and medical data supplementation.
  • the embodiments of the present application provide a method, device, equipment, and medium for combating network training and medical data supplementation to solve the problem of lack of data and low model accuracy.
  • a countermeasure network training method including:
  • the initial confrontation network including a generator model containing initial parameters and a trained induction network model;
  • the initial parameters of the iterative generator model are updated, until the total loss value reaches the preset convergence condition, the initial confrontation after the convergence
  • the network is recorded as a confrontational network.
  • a countermeasure network training device including:
  • a confrontation network acquisition module for acquiring an initial confrontation network, the initial confrontation network including a generator model containing initial parameters and a trained induction network model;
  • a data generation module configured to input preset random noise into the initial countermeasure network, and generate generated data corresponding to the random noise through the generator model;
  • a loss value determining module configured to determine the total loss value of the generator model through the induction network model according to the generated data
  • the convergence judgment module is used to update the initial parameters of the generator model when the total loss value does not reach the preset convergence condition, and converge until the total loss value reaches the preset convergence condition
  • the subsequent initial confrontation network is recorded as a confrontation network.
  • a method for supplementing medical data including:
  • the full medical data set contains multiple samples of medical data and a first small sample of medical data; the first small sample of medical data is associated with a small sample label;
  • the second sample size is much smaller than the first sample size
  • the confrontation network completed through training generates a second small sample of medical data that is equal to the sample difference and is associated with the small sample label; wherein, the confrontation network is obtained according to the above-mentioned confrontation network training method;
  • the induction network model is obtained by training according to the medical full data set;
  • a computer device includes a memory, a processor, and computer-readable instructions that are stored in the memory and can run on the processor, and the processor implements the following steps when the processor executes the computer-readable instructions:
  • the initial confrontation network including a generator model containing initial parameters and a trained induction network model;
  • the initial parameters of the iterative generator model are updated, until the total loss value reaches the preset convergence condition, the initial confrontation after the convergence
  • the network is recorded as a confrontational network.
  • a computer device includes a memory, a processor, and computer-readable instructions that are stored in the memory and can run on the processor, and the processor implements the following steps when the processor executes the computer-readable instructions:
  • the full medical data set contains multiple samples of medical data and a first small sample of medical data; the first small sample of medical data is associated with a small sample label;
  • the confrontation network completed through training generates a second small sample of medical data that is equal to the sample difference and is associated with the small sample label; wherein, the confrontation network is obtained according to the above-mentioned confrontation network training method;
  • the induction network model is obtained by training according to the medical full data set;
  • One or more readable storage media storing computer readable instructions, when the computer readable instructions are executed by one or more processors, the one or more processors execute the following steps:
  • the initial confrontation network including a generator model containing initial parameters and a trained induction network model;
  • the initial parameters of the iterative generator model are updated, until the total loss value reaches the preset convergence condition, the initial confrontation after the convergence
  • the network is recorded as a confrontational network.
  • One or more readable storage media storing computer readable instructions, when the computer readable instructions are executed by one or more processors, the one or more processors execute the following steps:
  • the full medical data set contains multiple samples of medical data and a first small sample of medical data; the first small sample of medical data is associated with a small sample label;
  • the confrontation network completed through training generates a second small sample of medical data that is equal to the sample difference and is associated with the small sample label; wherein, the confrontation network is obtained according to the above-mentioned confrontation network training method;
  • the induction network model is obtained by training according to the medical full data set;
  • the above-mentioned countermeasure network training and medical data supplement methods, devices, equipment and media are obtained by obtaining an initial countermeasure network, which includes a generator model containing initial parameters and a trained induction network model; preset random noise is input to The initial confrontation network generates generated data corresponding to the random noise through the generator model; determines the total loss value of the generator model through the induction network model according to the generated data; When the loss value does not reach the preset convergence condition, the initial parameters of the iterative generator model are updated, until the total loss value reaches the preset convergence condition, the initial confrontation network after convergence is recorded as a confrontation The internet.
  • This application improves the structure of the GAN network in the prior art and uses an induction network model to replace the discriminator model, so that the trained adversarial network can judge whether the generated data conforms to the distribution of the full data set and the distribution of the sub-data sets of each category.
  • This application The function of the confrontation network is expanded, and the accuracy of the data generated in the confrontation network is improved; and the trained confrontation network can be applied to the supplementation of small sample data in different scenarios, so that the completed model can be trained through the supplemented small sample data.
  • the accuracy rate is higher, which provides convenience for intelligent research in various scenarios.
  • FIG. 1 is a schematic diagram of an application environment of the confrontation network training method and the medical data supplement method in an embodiment of the present application;
  • Fig. 2 is a flowchart of a countermeasure network training method in an embodiment of the present application
  • FIG. 3 is a flowchart of step S30 in the countermeasure network training method in an embodiment of the present application.
  • FIG. 4 is a flowchart of step S301 in the countermeasure network training method in an embodiment of the present application.
  • FIG. 5 is a flowchart of step S302 in the countermeasure network training method in an embodiment of the present application
  • FIG. 6 is a flowchart of a method for supplementing medical data in an embodiment of the present application.
  • Fig. 7 is a functional block diagram of a countermeasure network training device in an embodiment of the present application.
  • FIG. 8 is a functional block diagram of the loss value determining module in the countermeasure network training device in an embodiment of the present application.
  • FIG. 9 is a functional block diagram of the first loss value determining unit in the adversarial network training device in an embodiment of the present application.
  • FIG. 10 is a functional block diagram of the second loss value determining unit in the adversarial network training device in an embodiment of the present application.
  • FIG. 11 is a functional block diagram of a medical data supplement device in an embodiment of the present application.
  • Fig. 12 is a schematic diagram of a computer device in an embodiment of the present application.
  • the confrontation network training method provided by the embodiment of the present application can be applied to the application environment shown in FIG. 1.
  • the confrontation network training method is applied in a confrontation network training system.
  • the confrontation network training system includes a client and a server as shown in FIG.
  • the client is also called the client, which refers to the program that corresponds to the server and provides local services to the client.
  • the client can be installed on, but not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices.
  • the server can be implemented as an independent server or a server cluster composed of multiple servers.
  • a method for training a confrontation network is provided.
  • the method is applied to the server in FIG. 1 as an example for description, including the following steps:
  • S10 Obtain an initial confrontation network, where the initial confrontation network includes a generator model containing initial parameters and a trained induction network model.
  • the initial confrontation network is improved based on the GAN (Generative Adversarial Networks) network in the prior art.
  • the initial confrontation network retains the generator containing the initial parameters in the original GAN network, and the discrimination in the original GAN network Replace the sensor with the trained induction network model.
  • the induction network model includes the following three modules: encoder module, induction module and correlation module. The induction network model is obtained through training on the full data set.
  • step S10 before step S10, that is, before obtaining the initial confrontation network, the method further includes:
  • S11 Obtain a full data set.
  • the full data set includes several sub-data sets corresponding to several categories; one sub-data set is associated with one sub-data label.
  • the full data set can be a data set in any scenario.
  • the full data set can be all application data; it can be a full data set in the medical field; the full data set includes several sub-data sets corresponding to several categories.
  • the full data set is application data
  • the corresponding sub-data set can be classified according to specific applications (such as NetEase Cloud Music, Tencent Video, etc.), or according to different types (such as music, video, and games). ) Applications are classified, and a sub-data set is associated with a sub-data label (for example, a music application corresponds to a music label).
  • S12 Input each of the sub-data sets into the inductive network model, and perform encoding conversion on each of the sub-data sets through the encoder module in the inductive network model to obtain the corresponding data of each sub-data set. Sub data vector.
  • the encoder module is used to convert the data in each sub-data set into a low-dimensional embedding vector, which facilitates the identification and calculation of the subsequent steps.
  • each sub-data set of the full data set into the induction network model, and encode and convert the data in each sub-data set through the encoder module in the induction network model, so that the data It is transformed into the corresponding low-dimensional embedding vector, that is, the sub-data vector corresponding to each data in each sub-data set.
  • the induction network model uses the principle of dynamic routing to convert the sub-data vector corresponding to each sub-data set into its corresponding representation.
  • all sub-data vectors under each category need to be expressed as unified features, that is, all sub-data vectors in each sub-data set are converted into corresponding category vectors.
  • the correlation module is a module that provides correlation calculation methods.
  • the correlation module in the induction network model is used to iteratively determine each category. After iterating over the category vectors of the same category, iteratively determine the correlation between the category vectors of different categories, and then determine the correlation function corresponding to each category vector.
  • the same-dimensional transformation refers to the transformation of various correlation functions into the same-dimensional relationship.
  • each correlation function is converted in the same dimension to determine each sub-data set and its corresponding sub-data set.
  • the preset relational expression standard can be that when the relational coefficient between the sub-data set and the sub-data label in the relational expression changes little or no longer changes, the relational expression is determined to be the final relational expression, and then in all the relations After the equations are determined, the training of the representation induction network model is completed.
  • the trained induction network model can learn the distribution of the full data set, after the new data is input to the induction network model, the new data can be classified into the category closest to its distribution, that is The induction network model can determine whether the new data conforms to the distribution of the full data set, and it can also determine whether the new data conforms to the distribution of any category of sub-data set.
  • S20 Input preset random noise to the initial countermeasure network, and generate generated data corresponding to the preset random noise through the generator model.
  • the preset random noise can be generated by a random algorithm. Further, after the random algorithm generates the preset random noise, the preset random noise is received through the generator model, and generated data corresponding to the preset random noise is generated.
  • S30 Determine the total loss value of the generator model through the induction network model according to the generated data.
  • step S30 includes the following steps:
  • S301 Output a first loss value between the generated data and the small sample data through the induction network model.
  • the small sample data refers to the data corresponding to the category with a small sample size in the full data set.
  • the full data set is an application data set.
  • this full data set there are less data for book category applications.
  • the data corresponding to the book category can be called small sample data.
  • the first loss value is obtained by logarithmic calculation based on the matching degree between the generated data and the small sample data.
  • step S301 includes the following steps:
  • S3011 Obtain a generated label corresponding to the generated data and a sample label corresponding to the small sample data.
  • the generated tag represents the category of the generated data.
  • the sample label represents the category of the small sample data.
  • S3012 Determine a first relational expression corresponding to the small sample data in the induction network model according to the small sample data and the sample label.
  • the induction network model when the induction network model is trained through the full data set, for the trained induction network model, it learns the distribution of the full data set and at the same time identifies the distribution of the sub-data set corresponding to each category, that is, The first relationship between the small sample data and the sample label has been determined in the induction network model.
  • S3013 Determine the first loss value according to the generated data, the generated label, and the first relational expression.
  • the relationship between the generated data obtained according to the preset random noise and the corresponding generated tag for the first time is quite different from the first relationship.
  • the network model outputs a first loss value corresponding to the generator model, and the first loss value is determined according to the generated data, the generated label, and the first relational expression.
  • S302 Output a second loss value between the generated data and the full data set through the induction network model.
  • the second loss value is obtained by logarithmic calculation according to the matching degree between the generated data and the full data set.
  • step S302 includes the following steps:
  • S3022 Determine a second relational expression corresponding to the full data set in the induction network model according to the full data set and the full data set label.
  • the induction network model is trained through the full data set, for the trained induction network model, it has learned the distribution of the full data set, so the full data set and the full data have been determined in the induction network model The second relationship between tags.
  • S3023 Determine the second loss value according to the generated data, the generated label, and the second relational expression.
  • the generator model in order to make the data generated by the generator model conform to the distribution of the sub-data set of the corresponding category, it also conforms to the distribution of the full data set, so that the generated data can be supplemented to the full data set without destroying the full data Therefore, it is necessary to determine the second loss value of the generator model according to the generated data, the generated label, and the second relational expression.
  • S303 Determine the total loss value of the generator model through the induction network model according to the first loss value and the second loss value.
  • the total loss value of the generator model can be determined by the following loss function:
  • LOSS G is the total loss value
  • log (similarity part ) is the first loss value
  • log(similarity all ) is the second loss value; ⁇ is the weight corresponding to the second loss value; ⁇ is the weight corresponding to the first loss value.
  • the similarity part is the degree of matching between the generated data and the small sample data, that is, the judgment of whether the generated data conforms to the distribution of the small sample data; the similarity all is the difference between the generated data and the full data set The degree of matching between the two, that is, the judgment of whether the generated data conforms to the distribution of the full data set.
  • the convergence condition can be the condition that the total loss value is less than the set threshold, that is, when the total loss value is less than the set threshold, stop training; the convergence condition can also be that the total loss value is calculated after 10,000 times The condition that it is small and will not decrease, that is, when the total loss value is small and does not decrease after 10,000 calculations, stop training and record the initial confrontation network after convergence as a confrontation network.
  • the initial parameters of the generator model are adjusted according to the total loss value output by the induction network model, so that the generator model outputs
  • the generated data can continue to move closer to the full data set distribution and small sample data distribution, so that the matching degree between the generated data and the small sample data, and the matching degree between the generated data and the full data set are getting higher and higher, until the generator When the total loss value of the model reaches the preset convergence condition, the initial confrontation network after convergence is recorded as the confrontation network.
  • the induction network model is used instead of the discriminator model, so that the trained adversarial network can determine whether the generated data conforms to the distribution of the full data set and the distribution of the sub-data sets of each category. It is judged that this application expands the function of the confrontation network and improves the accuracy of the data generated in the confrontation network; and the trained confrontation network can be applied to small sample data supplementation in different scenarios, so that the supplemented small sample data can be used for training
  • the completed model has a higher accuracy rate, which provides convenience for intelligent research in various scenarios, so as to promote the construction of smart cities.
  • the full data set and the confrontation network may be stored in the blockchain.
  • the Blockchain is an encrypted and chained transaction storage structure formed by blocks.
  • the header of each block can not only include the hash value of all transactions in the block, but also the hash value of all transactions in the previous block, so as to achieve tamper-proof transactions in the block based on the hash value And anti-counterfeiting; newly generated transactions are filled in the block and after the consensus of the nodes in the block chain network, they will be appended to the end of the block chain to form chain growth.
  • a method for supplementing medical data which includes the following steps:
  • S50 Receive a data supplement instruction including a full medical data set; the full medical data set includes multi-sample medical data and a first small-sample medical data; the first small-sample medical data is associated with a small-sample label.
  • the medical full data set is a collection containing all medical data in a specific scenario (such as a specific hospital or a specific department).
  • Multi-sample medical data refers to data corresponding to a category with a larger sample size in the full data set.
  • Small-sample medical data refers to the data corresponding to the categories with a small sample size in the full data set.
  • S60 Acquire a first sample size of the multi-sample medical data and a second sample size of the first small sample medical data; the second sample size is smaller than the first sample size.
  • the medical full data set contains a total of 100,000 sets of data.
  • the number of data corresponding to the multi-sample medical data is tens of thousands, and the number of data corresponding to the small-sample medical data may be hundreds, that is, multi-sample medical treatment.
  • Data One type of medical data may have 50,000 sets of data, while there are only a few hundred sets of data in a small sample of medical data.
  • medical models such as triage models cannot obtain enough feature information from categories with small sample sizes, resulting in failure to identify small-sample medical data categories. The data is correctly classified.
  • S70 Record the difference between the first sample quantity and the second sample quantity as a sample difference.
  • the confrontation network generated through training generates second small-sample medical data that is equal to the sample difference and is associated with the small-sample label; wherein, the confrontation network is based on the confrontation network training method in the foregoing embodiment Obtained;
  • the induction network model is obtained by training according to the medical full data set.
  • the confrontation network is obtained according to the training method of the confrontation network in the above embodiment, and the full data set in the confrontation network is the medical full data set, that is, the induction network model in the confrontation network is obtained by training based on the medical full data set. of.
  • the generator model in the confrontation network completed through training, according to the random noise signal obtained by the random algorithm, generates the second small sample of medical data with the number and the sample difference and is associated with the small sample label, and then generates The second small sample of medical data is supplemented to the medical full data set, so that the number of multi-sample medical data is balanced with the sum of the first small sample of medical data and the second small sample of medical data, and the second small sample of medical data is generated After being supplemented to the full medical data set, the distribution of the full medical data set will not be destroyed. Therefore, when training related models of medical scenarios such as triage model based on the supplemented full data set, the model can overcome samples in certain categories The problem of too little data, and the small sample medical data in the medical full data can also achieve a higher model classification accuracy rate.
  • a confrontation network training device is provided, and the confrontation network training device corresponds to the confrontation network training method in the foregoing embodiment in a one-to-one correspondence.
  • the confrontation network training device includes a confrontation network acquisition module 10, a data generation module 20, a loss value determination module 30 and a convergence judgment module 40.
  • the detailed description of each functional module is as follows:
  • the confrontation network acquisition module 10 is configured to acquire an initial confrontation network, the initial confrontation network including a generator model containing initial parameters and a trained induction network model;
  • the data generation module 20 is configured to input preset random noise into the initial countermeasure network, and generate generated data corresponding to the random noise through the generator model;
  • the loss value determination module 30 is configured to determine the total loss value of the generator model through the induction network model according to the generated data
  • the convergence judgment module 40 is configured to update the initial parameters of the generator model iteratively when the total loss value does not reach the preset convergence condition, until the total loss value reaches the preset convergence condition, The initial confrontation network after convergence is recorded as a confrontation network.
  • the confrontation network training device further includes the following modules:
  • the full data set acquisition module 11 is configured to acquire a full data set, the full data set includes several sub-data sets corresponding to classifications; one sub-data set is associated with one sub-data label;
  • the code conversion module 12 is configured to input each of the sub-data sets into the induction network model, and perform code conversion on each of the sub-data sets through the encoder module in the induction network model to obtain The sub-data vector corresponding to the sub-data set;
  • the vector conversion module 13 is configured to convert each of the sub-data vectors into a category vector corresponding to each sub-data vector through the induction module in the induction network model;
  • the correlation calculation module 14 is used to determine the correlation function corresponding to each of the category vectors through the correlation module in the induction network model;
  • the dimension conversion module 15 is configured to perform the same-dimensional conversion of each correlation function, and determine the relationship between each sub-data set and the corresponding sub-data label;
  • the standard determination module 16 is configured to indicate that the training of the induction network model is completed after each of the relational expressions reaches a preset relational expression standard.
  • the loss value determining module 30 includes the following units:
  • the first loss value determining unit 301 is configured to output the first loss value between the generated data and the small sample data through the induction network model;
  • the second loss value determining unit 302 is configured to output a second loss value between the generated data and the full data set through the induction network model;
  • the total loss value determining unit 303 is configured to determine the total loss value of the generator model through the induction network model according to the first loss value and the second loss value.
  • the first loss value determining unit 301 includes the following subunits:
  • the first label obtaining subunit 3011 is configured to obtain a generated label corresponding to the generated data and a sample label corresponding to the small sample data.
  • the first relational expression determining subunit 3012 is configured to determine a first relational expression corresponding to the small sample data in the induction network model according to the small sample data and the sample label.
  • the first loss value determining subunit 3013 is configured to determine the first loss value according to the generated data, the generated label, and the first relational expression.
  • the second loss value determining unit 302 includes the following subunits:
  • the second label obtaining subunit 3021 is configured to obtain the full data label corresponding to the full data set
  • the second relational expression determining subunit 3022 is configured to determine a second relational expression corresponding to the full data set in the induction network model according to the full data set and the full data set label;
  • the second loss value determining subunit 3023 is configured to determine the second loss value according to the generated data, the generated label, and the second relational expression.
  • the total loss value determining unit 303 is further configured to determine the total loss value of the generator model by using the following loss function:
  • LOSS G is the total loss value
  • log (similarity part ) is the first loss value
  • log(similarity all ) is the second loss value; ⁇ is the weight corresponding to the second loss value; ⁇ is the weight corresponding to the first loss value.
  • the various modules in the above-mentioned confrontation network training device can be implemented in whole or in part by software, hardware and a combination thereof.
  • the above-mentioned modules may be embedded in the form of hardware or independent of the processor in the computer equipment, or may be stored in the memory of the computer equipment in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.
  • a medical data supplement device is provided, and the medical data supplement device corresponds to the medical data supplement method in the above-mentioned embodiment in a one-to-one correspondence.
  • the medical data supplement device includes a supplement instruction receiving module 50, a sample quantity acquisition module 60, a sample difference recording module 70, a data generation module 80 and a data supplement module 90.
  • the detailed description of each functional module is as follows:
  • the supplementary instruction receiving module 50 is configured to receive a data supplementary instruction containing a full medical data set; the full medical data set contains multi-sample medical data and a first small-sample medical data; the first small-sample medical data and a small-sample label Associated
  • the sample quantity acquisition module 60 is configured to acquire the first sample quantity of the multi-sample medical data and the second sample quantity of the first small-sample medical data; the second sample quantity is smaller than the first sample quantity ;
  • the sample difference recording module 70 is configured to record the difference between the first sample quantity and the second sample quantity as a sample difference
  • the data generation module 80 is configured to generate the second small sample medical data whose quantity is equal to the sample difference and is associated with the small sample label through the countermeasure network completed through training; wherein, the countermeasure network is according to the above-mentioned embodiment Obtained by a confrontation network training method; the induction network model is obtained by training according to the medical full data set;
  • the data supplement module 90 is configured to supplement the generated second small sample medical data to the medical full data set.
  • a computer device is provided.
  • the computer device may be a server, and its internal structure diagram may be as shown in FIG. 12.
  • the computer equipment includes a processor, a memory, a network interface, and a database connected through a system bus.
  • the processor of the computer device is used to provide calculation and control capabilities.
  • the memory of the computer device includes a readable storage medium and an internal memory.
  • the readable storage medium stores an operating system, computer readable instructions, and a database.
  • the internal memory provides an environment for the operation of the operating system and computer readable instructions in the readable storage medium.
  • the database of the computer device is used to store the data used in the countermeasure network training method or the medical data supplement method in the foregoing embodiment.
  • the network interface of the computer device is used to communicate with an external terminal through a network connection.
  • the computer-readable instruction is executed by the processor to implement a method for training against a network, or the computer-readable instruction is executed by the processor to implement a method for supplementing medical data.
  • the readable storage medium provided in this embodiment includes a non-volatile readable storage medium and a volatile readable storage medium.
  • a computer device including a memory, a processor, and computer readable instructions stored in the memory and capable of running on the processor, and the processor implements the following steps when the processor executes the computer readable instructions:
  • the initial confrontation network including a generator model containing initial parameters and a trained induction network model;
  • the initial parameters of the iterative generator model are updated, until the total loss value reaches the preset convergence condition, the initial confrontation after the convergence
  • the network is recorded as a confrontational network.
  • a computer device including a memory, a processor, and computer readable instructions stored in the memory and capable of running on the processor, and the processor implements the following steps when the processor executes the computer readable instructions:
  • the full medical data set contains multiple samples of medical data and a first small sample of medical data; the first small sample of medical data is associated with a small sample label;
  • the confrontation network completed through training generates a second small sample of medical data that is equal to the sample difference and is associated with the small sample label; wherein, the confrontation network is obtained according to the above-mentioned confrontation network training method;
  • the induction network model is obtained by training according to the medical full data set;
  • one or more readable storage media storing computer readable instructions are provided.
  • the readable storage media provided in this embodiment include non-volatile readable storage media and volatile readable storage. Medium; the readable storage medium stores computer readable instructions, and when the computer readable instructions are executed by one or more processors, the one or more processors implement the following steps:
  • the initial confrontation network including a generator model containing initial parameters and a trained induction network model;
  • the initial parameters of the iterative generator model are updated, until the total loss value reaches the preset convergence condition, the initial confrontation after the convergence
  • the network is recorded as a confrontational network.
  • one or more readable storage media storing computer readable instructions are provided.
  • the readable storage media provided in this embodiment include non-volatile readable storage media and volatile readable storage. Medium; the readable storage medium stores computer readable instructions, and when the computer readable instructions are executed by one or more processors, the one or more processors implement the following steps:
  • the full medical data set contains multiple samples of medical data and a first small sample of medical data; the first small sample of medical data is associated with a small sample label;
  • the confrontation network completed through training generates a second small sample of medical data that is equal to the sample difference and is associated with the small sample label; wherein, the confrontation network is obtained according to the above-mentioned confrontation network training method;
  • the induction network model is obtained by training according to the medical full data set;
  • Non-volatile memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • Volatile memory may include random access memory (RAM) or external cache memory.
  • RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Channel (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Public Health (AREA)
  • Medical Informatics (AREA)
  • Health & Medical Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Databases & Information Systems (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • Biomedical Technology (AREA)
  • General Physics & Mathematics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Pathology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Epidemiology (AREA)
  • General Health & Medical Sciences (AREA)
  • Primary Health Care (AREA)
  • Measuring And Recording Apparatus For Diagnosis (AREA)
  • Instructional Devices (AREA)

Abstract

本申请涉及人工智能技术领域,应用于智慧医疗领域中,揭露了一种对抗网络训练、医疗数据补充方法、装置、设备及介质。该对抗网络训练方法通过获取初始对抗网络,所述初始对抗网络包括含有初始参数的生成器模型以及训练完成的感应网络模型;将预设随机噪声输入至初始对抗网络,通过生成器模型生成与随机噪声对应的生成数据;根据生成数据,通过感应网络模型确定生成器模型的总损失值;在总损失值未达到预设的收敛条件时,更新迭代生成器模型的初始参数,直至总损失值达到所述预设的收敛条件时,将收敛之后的所述初始对抗网络记录为对抗网络。本申请通过改进GAN网络,扩展了训练得到的对抗网络功能,提高了对抗网络生成数据的准确性。

Description

对抗网络训练、医疗数据补充方法、装置、设备及介质
本申请要求于2020年10月22日提交中国专利局、申请号为202011140634.9,发明名称为“对抗网络训练、医疗数据补充方法、装置、设备及介质”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能技术领域,尤其涉及一种对抗网络训练、医疗数据补充方法、装置、设备及介质。
背景技术
随着科学技术的发展,人工智能技术被广泛应用在各个不同的领域中,如医疗领域、汽车领域。
例如,在医疗领域中,为了从医疗数据中进行人工智能学习,以智能化地完成诸如分诊和病例监控等,需要大量人力进行对应的工作。发明人意识到,由于人工智能学习的过程中,需要通过大量数据进行不断训练学习,进而将训练完成的模型替代人工进行智能化工作,但目前,由于医疗数据常常涉及用户隐私,因此医疗数据的获取途径有限会导致医疗数据的稀缺或医疗数据质量层次不齐,从而导致可用于人工智能训练(要求数据质量较好)的医疗数据缺乏,进而使得训练之后的模型应用在医疗领域中时,产生模型准确率低的问题。
申请内容
本申请实施例提供一种对抗网络训练、医疗数据补充方法、装置、设备及介质,以解决数据缺乏且模型准确率低的问题。
一种对抗网络训练方法,包括:
获取初始对抗网络,所述初始对抗网络包括含有初始参数的生成器模型以及训练完成的感应网络模型;
将预设随机噪声输入至所述初始对抗网络,通过所述生成器模型生成与所述随机噪声对应的生成数据;
根据所述生成数据,通过所述感应网络模型确定所述生成器模型的总损失值;
在所述总损失值未达到预设的收敛条件时,更新迭代所述生成器模型的初始参数,直至所述总损失值达到所述预设的收敛条件时,将收敛之后的所述初始对抗网络记录为对抗网络。
一种对抗网络训练装置,包括:
对抗网络获取模块,用于获取初始对抗网络,所述初始对抗网络包括含有初始参数的生成器模型以及训练完成的感应网络模型;
数据生成模块,用于将预设随机噪声输入至所述初始对抗网络,通过所述生成器模型生成与所述随机噪声对应的生成数据;
损失值确定模块,用于根据所述生成数据,通过所述感应网络模型确定所述生成器模型的总损失值;
收敛判断模块,用于在所述总损失值未达到预设的收敛条件时,更新迭代所述生成器模型的初始参数,直至所述总损失值达到所述预设的收敛条件时,将收敛之后的所述初始对抗网络记录为对抗网络。
一种医疗数据补充方法,包括:
接收包含医疗全数据集的数据补充指令;所述医疗全数据集中包含多样本医疗数据以及第一小样本医疗数据;所述第一小样本医疗数据与小样本标签关联;
获取所述多样本医疗数据的第一样本数量以及所述第一小样本医疗数据的第二样本数量;所述第二样本数量远小于所述第一样本数量;
将所述第一样本数量与所述第二样本数量之间的差值记录为样本差值;
通过训练完成的对抗网络生成数量与所述样本差值相等且与所述小样本标签关联的第二小样本医疗数据;其中,所述对抗网络是根据上述对抗网路训练方法得到的;所述感应网络模型根据所述医疗全数据集训练得到;
将生成的所述第二小样本医疗数据补充至所述医疗全数据集中。
一种计算机设备,包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令,所述处理器执行所述计算机可读指令时实现如下步骤:
获取初始对抗网络,所述初始对抗网络包括含有初始参数的生成器模型以及训练完成的感应网络模型;
将预设随机噪声输入至所述初始对抗网络,通过所述生成器模型生成与所述预设随机噪声对应的生成数据;
根据所述生成数据,通过所述感应网络模型确定所述生成器模型的总损失值;
在所述总损失值未达到预设的收敛条件时,更新迭代所述生成器模型的初始参数,直至所述总损失值达到所述预设的收敛条件时,将收敛之后的所述初始对抗网络记录为对抗网络。
一种计算机设备,包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令,所述处理器执行所述计算机可读指令时实现如下步骤:
接收包含医疗全数据集的数据补充指令;所述医疗全数据集中包含多样本医疗数据以及第一小样本医疗数据;所述第一小样本医疗数据与小样本标签关联;
获取所述多样本医疗数据的第一样本数量以及所述第一小样本医疗数据的第二样本数量;所述第二样本数量小于所述第一样本数量;
将所述第一样本数量与所述第二样本数量之间的差值记录为样本差值;
通过训练完成的对抗网络生成数量与所述样本差值相等且与所述小样本标签关联的第二小样本医疗数据;其中,所述对抗网络是根据上述对抗网路训练方法得到的;所述感应网络模型根据所述医疗全数据集训练得到;
将生成的所述第二小样本医疗数据补充至所述医疗全数据集中。
一个或多个存储有计算机可读指令的可读存储介质,所述计算机可读指令被一个或多个处理器执行时,使得所述一个或多个处理器执行如下步骤:
获取初始对抗网络,所述初始对抗网络包括含有初始参数的生成器模型以及训练完成的感应网络模型;
将预设随机噪声输入至所述初始对抗网络,通过所述生成器模型生成与所述预设随机噪声对应的生成数据;
根据所述生成数据,通过所述感应网络模型确定所述生成器模型的总损失值;
在所述总损失值未达到预设的收敛条件时,更新迭代所述生成器模型的初始参数,直至所述总损失值达到所述预设的收敛条件时,将收敛之后的所述初始对抗网络记录为对抗网络。
一个或多个存储有计算机可读指令的可读存储介质,所述计算机可读指令被一个或多个处理器执行时,使得所述一个或多个处理器执行如下步骤:
接收包含医疗全数据集的数据补充指令;所述医疗全数据集中包含多样本医疗数据以及第一小样本医疗数据;所述第一小样本医疗数据与小样本标签关联;
获取所述多样本医疗数据的第一样本数量以及所述第一小样本医疗数据的第二样本 数量;所述第二样本数量小于所述第一样本数量;
将所述第一样本数量与所述第二样本数量之间的差值记录为样本差值;
通过训练完成的对抗网络生成数量与所述样本差值相等且与所述小样本标签关联的第二小样本医疗数据;其中,所述对抗网络是根据上述对抗网路训练方法得到的;所述感应网络模型根据所述医疗全数据集训练得到;
将生成的所述第二小样本医疗数据补充至所述医疗全数据集中。
上述对抗网络训练、医疗数据补充方法、装置、设备及介质,通过获取初始对抗网络,所述初始对抗网络包括含有初始参数的生成器模型以及训练完成的感应网络模型;将预设随机噪声输入至所述初始对抗网络,通过所述生成器模型生成与所述随机噪声对应的生成数据;根据所述生成数据,通过所述感应网络模型确定所述生成器模型的总损失值;在所述总损失值未达到预设的收敛条件时,更新迭代所述生成器模型的初始参数,直至所述总损失值达到所述预设的收敛条件时,将收敛之后的所述初始对抗网络记录为对抗网络。
本申请通过改进现有技术中GAN网络的结构,采用感应网络模型代替判别器模型,使得训练得到的对抗网络可以对生成数据是否符合全数据集分布以及各类别子数据集分布进行判断,本申请扩展了对抗网络的功能,提高了对抗网络中生成数据的准确性;并且训练完成的对抗网络可以适用于不同场景下的小样本数据补充,使得通过补充之后的小样本数据进行训练完成的模型的准确率更高,为各场景下的智能化研究提供便利。
本申请的一个或多个实施例的细节在下面的附图和描述中提出,本申请的其他特征和优点将从说明书、附图以及权利要求变得明显。
附图说明
为了更清楚地说明本申请实施例的技术方案,下面将对本申请实施例的描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本申请的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动性的前提下,还可以根据这些附图获得其他的附图。
图1是本申请一实施例中对抗网络训练方法以及医疗数据补充方法的一应用环境示意图;
图2是本申请一实施例中对抗网络训练方法的一流程图;
图3是本申请一实施例中对抗网络训练方法中步骤S30的一流程图;
图4是本申请一实施例中对抗网络训练方法中步骤S301的一流程图;
图5是本申请一实施例中对抗网络训练方法中步骤S302的一流程图;
图6是本申请一实施例中医疗数据补充方法的一流程图;
图7是本申请一实施例中对抗网络训练装置的一原理框图;
图8是本申请一实施例中对抗网络训练装置中损失值确定模块的一原理框图;
图9是本申请一实施例中对抗网络训练装置中第一损失值确定单元的一原理框图;
图10是本申请一实施例中对抗网络训练装置中第二损失值确定单元的一原理框图;
图11是本申请一实施例中医疗数据补充装置的一原理框图;
图12是本申请一实施例中计算机设备的一示意图。
具体实施方式
下面将结合本申请实施例中的附图,对本申请实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例是本申请一部分实施例,而不是全部的实施例。基于本申请中的实施例,本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例,都属于本申请保护的范围。
本申请实施例提供的对抗网络训练方法,该对抗网络训练方法可应用如图1所示的应 用环境中。具体地,该对抗网络训练方法应用在对抗网络训练***中,该对抗网络训练***包括如图1所示的客户端和服务器,客户端与服务器通过网络进行通信,用于解决数据缺乏且模型准确率低的问题。其中,客户端又称为用户端,是指与服务器相对应,为客户提供本地服务的程序。客户端可安装在但不限于各种个人计算机、笔记本电脑、智能手机、平板电脑和便携式可穿戴设备上。服务器可以用独立的服务器或者是多个服务器组成的服务器集群来实现。
在一实施例中,如图2所示,提供一种对抗网络训练方法,以该方法应用在图1中的服务器为例进行说明,包括如下步骤:
S10:获取初始对抗网络,所述初始对抗网络包括含有初始参数的生成器模型以及训练完成的感应网络模型。
其中,初始对抗网络是基于现有技术中GAN(Generative Adversarial Networks,生成对抗网络)网络进行改进得到,该初始对抗网络保留原有GAN网络中含有初始参数的生成器,将原GAN网络中的判别器替换成训练完成的感应网络模型。感应网络模型中包含如下三个模块:编码器模块、感应模块以及相关性模块,该感应网络模型通过全数据集进行训练得到。
在一具体实施方式中,步骤S10之前,也即获取初始对抗网络之前,还包括:
S11:获取全数据集,所述全数据集中包括若干分类对应的子数据集;一个所述子数据集关联一个子数据标签。
其中,全数据集可以为任意场景下的数据集,示例性地,该全数据集可以为所有应用程序数据;可以为医疗领域全数据集;该全数据集中包括若干分类对应的子数据集,假设该全数据集为应用程序数据,则对应的子数据集可以根据具体的应用程序(如网易云音乐、腾讯视频等)进行分类,也可以根据不同种类(如音乐类、视频类、游戏类)的应用程序进行分类,一个子数据集关联一个子数据标签(如音乐类应用程序对应于音乐标签)。
S12:将各所述子数据集输入至所述感应网络模型中,通过所述感应网络模型中的编码器模块对各所述子数据集进行编码转换,得到与各所述子数据集对应的子数据向量。
其中,编码器模块用于将各子数据集中的数据转化为低维度的嵌入向量,利于后续步骤对其进行识别计算。
具体地,在获取全数据集之后,将全数据集中各子数据集输入至感应网络模型中,通过感应网路模型中的编码器模块对各子数据集中的数据进行编码转换,以将该数据转化为与其对应的低维度的嵌入向量,也即各子数据集中各数据对应的子数据向量。
S13:通过所述感应网络模型中的感应模块将各所述子数据向量转化为与各子数据向量对应的类别向量。
其中,感应网络模型利用了动态路由的原理,将每一子数据集对应的子数据向量转化为与其对应的表征。
可以理解地,在感应网络模型的感应模块中,需要将各分类下的所有子数据向量表示为统一特征,也即将各子数据集中的所有子数据向量转化为与其对应的类别向量。
S14:通过所述感应网络模型中的相关性模块,确定与各所述类别向量对应的相关性函数。
其中,相关性模块是提供相关性计算方法的模块。
具体地,在通过所述感应网络模型中的感应模块将各所述子数据向量转化为与各子数据向量对应的类别向量之后,通过感应网络模型中的相关性模块,迭代确定每一类别下的类别向量之间的相关性,并在迭代完相同类别下的类别向量之后,迭代确定不同类别之间的类别向量之间的相关性,进而确定与各类别向量对应的相关性函数。假设类别向量对应为x 1、x 2、x 3,而相关性模块中存在如下关系式:y=λ 1x 12x 23x 3,进而根 据各类别向量得到的相关性函数可以为:y 1=k 1x 1,y 2=k 2x 2,y 3=k 3x 3
S15:将各所述相关性函数进行同维转换,确定各所述子数据集和与其对应的所述子数据标签之间的关系式。
其中,同维转换指的是将各相关性函数转化成同维度关系式。
具体地,在通过所述感应网络模型中的相关性模块,确定与各所述类别向量对应的相关性函数之后,将各相关性函数进行同维转换,确定各子数据集和与其对应的子数据标签之间的关系式。示例性地,假设一子数据集对应的特征为x,子数据标签为z,该关系式可以为,z=x+1。可以理解地,该关系式是感应网路模型中隐藏的关系式,只有通过感应网路模型学习识别得到,以令感应网络模型可以根据各关系式确定新输入数据是否符合任意一个子数据集分类。
S16:在各所述关系式达到预设关系式标准之后,表征所述感应网络模型训练完成。
可以理解地,预设关系式标准可以为当关系式中子数据集与子数据标签之间的关系系数变化很小或者不再改变时,确定该关系式为最终的关系式,进而在所有关系式均确定完成之后,表征感应网络模型训练完成。示例性地,假设一子数据集对应的特征为x,子数据标签为z,该关系式可以为,z=x+1,1则为该关系式中的关系系数,若在后续迭代训练过程中,该关系系数变化小于0.00001,则认为该子数据集与子数据标签之前的关系式确定成功。
进一步地,训练完成的感应网络模型可以学习到全数据集的分布的同时,在新的数据输入至感应网络模型之后,可以将该新的数据归类到与其分布最接近的类别中,也即感应网络模型可以判定新的数据是否符合全数据集的分布,还可以判定新的数据是否符合任意一个类别的子数据集的分布。
S20:将预设随机噪声输入至所述初始对抗网络,通过所述生成器模型生成与所述预设随机噪声对应的生成数据。
其中,预设随机噪声可以通过随机算法产生。进一步地,在随机算法产生预设随机噪声之后,通过生成器模型接收该预设随机噪声,并生成与该预设随机噪声对应的生成数据。
S30:根据所述生成数据,通过所述感应网络模型确定所述生成器模型的总损失值。
具体地,如图3所示,步骤S30中包括以下步骤:
S301:通过所述感应网络模型输出所述生成数据与小样本数据之间的第一损失值。
其中,小样本数据指的是全数据集中样本量较少的类别对应的数据,示例性地,假设全数据集为应用程序数据集,在该全数据集中,书籍类别应用程序的数据较少,则可以将书籍类别对应的数据称为小样本数据。第一损失值是根据生成数据与小样本数据之间的匹配度进行对数计算得到的。
进一步地,如图4所示,步骤S301中包括如下步骤:
S3011:获取与所述生成数据对应的生成标签,以及与所述小样本数据对应的样本标签。其中,生成标签表征生成数据的类别。样本标签表征小样本数据的类别。
S3012:根据所述小样本数据以及所述样本标签,确定所述感应网络模型中与所述小样本数据对应的第一关系式。
可以理解地,在感应网络模型通过全数据集进行训练的时候,对于训练完成的感应网路模型,其学习了全数据集的分布,同时识别了各类别对应的子数据集的分布,也即该感应网络模型中已经确定出小样本数据与样本标签之间的第一关系式。
S3013:根据所述生成数据、生成标签以及所述第一关系式,确定所述第一损失值。
可以理解地,针对含有初始参数的生成器模型,在第一次根据预设随机噪声得到的生成数据和与其对应的生成标签之间的关系式,与第一关系式之间差别较大,感应网络模型会输出一个与生成器模型对应的第一损失值,该第一损失值是根据生成数据、生成标签以 及第一关系式确定的。
S302:通过所述感应网络模型输出所述生成数据与全数据集之间的第二损失值。
其中,第二损失值是根据生成数据与全数据集之间的匹配度进行对数计算得到的。
进一步地,如图5所示,步骤S302中包括如下步骤:
S3021:获取与所述全数据集对应的全数据标签。
S3022:根据所述全数据集以及所述全数据集标签,确定所述感应网络模型中与所述全数据集对应的第二关系式。
可以理解地,在感应网络模型通过全数据集进行训练的时候,对于训练完成的感应网路模型,其学习了全数据集的分布,因此该感应网络模型中已经确定出全数据集与全数据标签之间的第二关系式。
S3023:根据所述生成数据、生成标签以及所述第二关系式,确定所述第二损失值。
可以理解地,为了使得生成器模型生成的数据,在符合对应的类别的子数据集的分布的同时,符合全数据集的分布,进而使得生成数据补充至全数据集之后,可以不破坏全数据集的分布,因此需要根据生成数据、生成标签以及所述第二关系式,确定生成器模型的第二损失值。
S303:根据所述第一损失值以及所述第二损失值,通过所述感应网络模型确定所述生成器模型的总损失值。
具体地,可以通过以下损失函数确定生成器模型的总损失值:
LOSS G=-α*log(similarity all)-βlog(similarity part)
其中,LOSS G为所述总损失值;log(similarity part)为所述第一损失值;
log(similarity all)为所述第二损失值;α为所述第二损失值对应的权重;β为所述第一损失值对应的权重。
进一步地,similarity part为所述生成数据与所述小样本数据之间的匹配度,也即对生成数据是否符合小样本数据分布的判断;similarity all为所述生成数据与所述全数据集之间的匹配度,也即对生成数据是否符合全数据集分布的判断。
S40:在所述总损失值未达到预设的收敛条件时,更新迭代所述生成器模型的初始参数,直至所述总损失值达到所述预设的收敛条件时,将收敛之后的所述初始对抗网络记录为对抗网络。
可以理解地,该收敛条件可以为总损失值小于设定阈值的条件,也即在总损失值小于设定阈值时,停止训练;收敛条件还可以为总损失值经过了10000次计算后值为很小且不会再下降的条件,也即总损失值经过10000次计算后值很小且不会下降时,停止训练,并将收敛之后的初始对抗网络记录为对抗网络。
如此,在通过不同的预设随机噪声输入至初始对抗网络,通过生成器模型生成对应的生成数据之后,根据感应网络模型输出的总损失值,调整生成器模型的初始参数,使得生成器模型输出的生成数据可以不断向全数据集分布以及小样本数据分布靠拢,让生成数据与小样本数据之间的匹配度,以及生成数据与全数据集之间的匹配度越来越高,直至生成器模型的总损失值达到预设的收敛条件时,将收敛之后的初始对抗网络记录为对抗网络。
在本实施例中,通过改进现有技术中GAN网络的结构,采用感应网络模型代替判别器模型,使得训练得到的对抗网络可以对生成数据是否符合全数据集分布以及各类别子数据集分布进行判断,本申请扩展了对抗网络的功能,提高了对抗网络中生成数据的准确性; 并且训练完成的对抗网络可以适用于不同场景下的小样本数据补充,使得通过补充之后的小样本数据进行训练完成的模型的准确率更高,为各场景下的智能化研究提供便利,以便推动智慧城市的建设。
在另一具体实施例中,为了保证上述实施例中的全数据集以及对抗网络的私密以及安全性,可以将全数据集以及对抗网络存储在区块链中。其中,区块链(Blockchain),是由区块(Block)形成的加密的、链式的交易的存储结构。
例如,每个区块的头部既可以包括区块中所有交易的哈希值,同时也包含前一个区块中所有交易的哈希值,从而基于哈希值实现区块中交易的防篡改和防伪造;新产生的交易被填充到区块并经过区块链网络中节点的共识后,会被追加到区块链的尾部从而形成链式的增长。
在一实施例中,如图6所示,提供一种医疗数据补充方法,包括如下步骤:
S50:接收包含医疗全数据集的数据补充指令;所述医疗全数据集中包含多样本医疗数据以及第一小样本医疗数据;所述第一小样本医疗数据与小样本标签关联。
其中,医疗全数据集为包含某一具体场景下(如某一具体医院、或者某一具体科室)的所有医疗数据的集合。多样本医疗数据指的是全数据集中样本量较多的类别对应的数据。小样本医疗数据指的是全数据集中样本量较少的类别对应的数据。
S60:获取所述多样本医疗数据的第一样本数量以及所述第一小样本医疗数据的第二样本数量;所述第二样本数量小于所述第一样本数量。
示例性地,假设医疗全数据集中一共包含十万组数据,其中多样本医疗数据对应的数据数量是万级别的,而小样本医疗数据对应的数据数量可能是百级别的,也即多样本医疗数据其中一类医疗数据可能有五万组数据,而小样本医疗数据中仅只有几百组数据。而在医疗领域中,常常会因为小样本医疗数据的不充足,导致诸如分诊模型等医疗模型无法从样本量少的类别中获得足够多的特征信息,从而导致无法对属于小样本医疗数据类别的数据进行正确分类。
S70:将所述第一样本数量与所述第二样本数量之间的差值记录为样本差值。
S80:通过训练完成的对抗网络生成数量与所述样本差值相等且与所述小样本标签关联的第二小样本医疗数据;其中,所述对抗网络是根据上述实施例中对抗网路训练方法得到的;所述感应网络模型根据所述医疗全数据集训练得到。
S90:将生成的所述第二小样本医疗数据补充至所述医疗全数据集中。
其中,对抗网络是根据上述实施例中对抗网络训练方法得到的,并且该对抗网络中的全数据集为医疗全数据集,也即该对抗网络中的感应网络模型是根据医疗全数据集训练得到的。
在获取所述多样本医疗数据的第一样本数量以及所述第一小样本医疗数据的第二样本数量之后,将所述第一样本数量与所述第二样本数量之间的差值记录为样本差值;通过训练完成的对抗网络中的生成器模型,根据随机算法得到的随机噪声信号生成数量与样本差值且与小样本标签关联的第二小样本医疗数据,进而将生成的第二小样本医疗数据补充至医疗全数据集中,以令多样本医疗数据的数量,与第一小样本医疗数据以及第二小样本医疗数据之和达到均衡,并且生成的第二小样本医疗数据补充至医疗全数据集之后,不会破坏医疗全数据集的分布,从而在根据补充之后的全数据集训练例如分诊模型等医疗场景的相关模型时,可以令该模型克服某些类别下样本数据过少的难题,进而在医疗全数据里的小样本医疗数据上同样达到较高的模型分类准确率。
应理解,上述实施例中各步骤的序号的大小并不意味着执行顺序的先后,各过程的执行顺序应以其功能和内在逻辑确定,而不应对本申请实施例的实施过程构成任何限定。
在一实施例中,提供一种对抗网络训练装置,该对抗网络训练装置与上述实施例中对抗网络训练方法一一对应。如图7所示,该对抗网络训练装置包括对抗网络获取模块10、 数据生成模块20、损失值确定模块30和收敛判断模块40。各功能模块详细说明如下:
对抗网络获取模块10,用于获取初始对抗网络,所述初始对抗网络包括含有初始参数的生成器模型以及训练完成的感应网络模型;
数据生成模块20,用于将预设随机噪声输入至所述初始对抗网络,通过所述生成器模型生成与所述随机噪声对应的生成数据;
损失值确定模块30,用于根据所述生成数据,通过所述感应网络模型确定所述生成器模型的总损失值;
收敛判断模块40,用于在所述总损失值未达到预设的收敛条件时,更新迭代所述生成器模型的初始参数,直至所述总损失值达到所述预设的收敛条件时,将收敛之后的所述初始对抗网络记录为对抗网络。
优选地,对抗网络训练装置还包括如下模块:
全数据集获取模块11,用于获取全数据集,所述全数据集中包括若干分类对应的子数据集;一个所述子数据集关联一个子数据标签;
编码转换模块12,用于将各所述子数据集输入至所述感应网络模型中,通过所述感应网络模型中的编码器模块对各所述子数据集进行编码转换,得到与各所述子数据集对应的子数据向量;
向量转化模块13,用于通过所述感应网络模型中的感应模块将各所述子数据向量转化为与各子数据向量对应的类别向量;
相关性计算模块14,用于通过所述感应网络模型中的相关性模块,确定与各所述类别向量对应的相关性函数;
维度转换模块15,用于将各所述相关性函数进行同维转换,确定各所述子数据集和与其对应的所述子数据标签之间的关系式;
标准判定模块16,用于在各所述关系式达到预设关系式标准之后,表征所述感应网络模型训练完成。
优选地,如图8所示,损失值确定模块30包括如下单元:
第一损失值确定单元301,用于通过所述感应网络模型输出所述生成数据与小样本数据之间的第一损失值;
第二损失值确定单元302,用于通过所述感应网络模型输出所述生成数据与全数据集之间的第二损失值;
总损失值确定单元303,用于根据所述第一损失值以及所述第二损失值,通过所述感应网络模型确定所述生成器模型的总损失值。
优选地,如图9所示,第一损失值确定单元301包括如下子单元:
第一标签获取子单元3011,用于获取与所述生成数据对应的生成标签,以及与所述小样本数据对应的样本标签。
第一关系式确定子单元3012,用于根据所述小样本数据以及所述样本标签,确定所述感应网络模型中与所述小样本数据对应的第一关系式。
第一损失值确定子单元3013,用于根据所述生成数据、生成标签以及所述第一关系式,确定所述第一损失值。
优选地,如图10所示,第二损失值确定单元302包括如下子单元:
第二标签获取子单元3021,用于获取与所述全数据集对应的全数据标签;
第二关系式确定子单元3022,用于根据所述全数据集以及所述全数据集标签,确定所述感应网络模型中与所述全数据集对应的第二关系式;
第二损失值确定子单元3023,用于根据所述生成数据、生成标签以及所述第二关系式,确定所述第二损失值。
优选地,总损失值确定单元303还用于通过如下损失函数确定所述生成器模型的总损 失值:
LOSS G=-α*log(similarity all)-βlog(similarity part)
其中,LOSS G为所述总损失值;log(similarity part)为所述第一损失值;
log(similarity all)为所述第二损失值;α为所述第二损失值对应的权重;β为所述第一损失值对应的权重。
关于对抗网络训练装置的具体限定可以参见上文中对于对抗网络训练方法的限定,在此不再赘述。上述对抗网络训练装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中,也可以以软件形式存储于计算机设备中的存储器中,以便于处理器调用执行以上各个模块对应的操作。
在一实施例中,提供一种医疗数据补充装置,该医疗数据补充装置与上述实施例中医疗数据补充方法一一对应。如图11所示,该医疗数据补充装置包括补充指令接收模块50、样本数量获取模块60、样本差值记录模块70、数据生成模块80和数据补充模块90。各功能模块详细说明如下:
补充指令接收模块50,用于接收包含医疗全数据集的数据补充指令;所述医疗全数据集中包含多样本医疗数据以及第一小样本医疗数据;所述第一小样本医疗数据与小样本标签关联;
样本数量获取模块60,用于获取所述多样本医疗数据的第一样本数量以及所述第一小样本医疗数据的第二样本数量;所述第二样本数量小于所述第一样本数量;
样本差值记录模块70,用于将所述第一样本数量与所述第二样本数量之间的差值记录为样本差值;
数据生成模块80,用于通过训练完成的对抗网络生成数量与所述样本差值相等且与所述小样本标签关联的第二小样本医疗数据;其中,所述对抗网络是根据上述实施例中对抗网路训练方法得到的;所述感应网络模型根据所述医疗全数据集训练得到;
数据补充模块90,用于将生成的所述第二小样本医疗数据补充至所述医疗全数据集中。
在一个实施例中,提供了一种计算机设备,该计算机设备可以是服务器,其内部结构图可以如图12所示。该计算机设备包括通过***总线连接的处理器、存储器、网络接口和数据库。其中,该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括可读存储介质、内存储器。该可读存储介质存储有操作***、计算机可读指令和数据库。该内存储器为可读存储介质中的操作***和计算机可读指令的运行提供环境。该计算机设备的数据库用于存储上述实施例中对抗网络训练方法或者医疗数据补充方法所使用到的数据。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读指令被处理器执行时以实现一种对抗网络训练方法,或者该计算机可读指令被处理器执行时以实现一种医疗数据补充方法。本实施例所提供的可读存储介质包括非易失性可读存储介质和易失性可读存储介质。
在一个实施例中,提供了一种计算机设备,包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机可读指令,处理器执行计算机可读指令时实现如下步骤:
获取初始对抗网络,所述初始对抗网络包括含有初始参数的生成器模型以及训练完成的感应网络模型;
将预设随机噪声输入至所述初始对抗网络,通过所述生成器模型生成与所述预设随机噪声对应的生成数据;
根据所述生成数据,通过所述感应网络模型确定所述生成器模型的总损失值;
在所述总损失值未达到预设的收敛条件时,更新迭代所述生成器模型的初始参数,直至所述总损失值达到所述预设的收敛条件时,将收敛之后的所述初始对抗网络记录为对抗网络。
在一个实施例中,提供了一种计算机设备,包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机可读指令,处理器执行计算机可读指令时实现如下步骤:
接收包含医疗全数据集的数据补充指令;所述医疗全数据集中包含多样本医疗数据以及第一小样本医疗数据;所述第一小样本医疗数据与小样本标签关联;
获取所述多样本医疗数据的第一样本数量以及所述第一小样本医疗数据的第二样本数量;所述第二样本数量小于所述第一样本数量;
将所述第一样本数量与所述第二样本数量之间的差值记录为样本差值;
通过训练完成的对抗网络生成数量与所述样本差值相等且与所述小样本标签关联的第二小样本医疗数据;其中,所述对抗网络是根据上述对抗网路训练方法得到的;所述感应网络模型根据所述医疗全数据集训练得到;
将生成的所述第二小样本医疗数据补充至所述医疗全数据集中。在一个实施例中,提供了一个或多个存储有计算机可读指令的可读存储介质,本实施例所提供的可读存储介质包括非易失性可读存储介质和易失性可读存储介质;该可读存储介质上存储有计算机可读指令,该计算机可读指令被一个或多个处理器执行时,使得一个或多个处理器实现如下步骤:
获取初始对抗网络,所述初始对抗网络包括含有初始参数的生成器模型以及训练完成的感应网络模型;
将预设随机噪声输入至所述初始对抗网络,通过所述生成器模型生成与所述预设随机噪声对应的生成数据;
根据所述生成数据,通过所述感应网络模型确定所述生成器模型的总损失值;
在所述总损失值未达到预设的收敛条件时,更新迭代所述生成器模型的初始参数,直至所述总损失值达到所述预设的收敛条件时,将收敛之后的所述初始对抗网络记录为对抗网络。
在一个实施例中,提供了一个或多个存储有计算机可读指令的可读存储介质,本实施例所提供的可读存储介质包括非易失性可读存储介质和易失性可读存储介质;该可读存储介质上存储有计算机可读指令,该计算机可读指令被一个或多个处理器执行时,使得一个或多个处理器实现如下步骤:
接收包含医疗全数据集的数据补充指令;所述医疗全数据集中包含多样本医疗数据以及第一小样本医疗数据;所述第一小样本医疗数据与小样本标签关联;
获取所述多样本医疗数据的第一样本数量以及所述第一小样本医疗数据的第二样本数量;所述第二样本数量小于所述第一样本数量;
将所述第一样本数量与所述第二样本数量之间的差值记录为样本差值;
通过训练完成的对抗网络生成数量与所述样本差值相等且与所述小样本标签关联的第二小样本医疗数据;其中,所述对抗网络是根据上述对抗网路训练方法得到的;所述感应网络模型根据所述医疗全数据集训练得到;
将生成的所述第二小样本医疗数据补充至所述医疗全数据集中。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程,是可以通过计算机可读指令来指令相关的硬件来完成,所述的计算机可读指令可存储于一非易失性计算机可读取存储介质中,该计算机可读指令在执行时,可包括如上述各方法的实施例的流程。其中,本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用,均可包括非易失性和/或易失性存储器。非易失性存储器可包括只读存储器 (ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限,RAM以多种形式可得,诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
所属领域的技术人员可以清楚地了解到,为了描述的方便和简洁,仅以上述各功能单元、模块的划分进行举例说明,实际应用中,可以根据需要而将上述功能分配由不同的功能单元、模块完成,即将所述装置的内部结构划分成不同的功能单元或模块,以完成以上描述的全部或者部分功能。
以上所述实施例仅用以说明本申请的技术方案,而非对其限制;尽管参照前述实施例对本申请进行了详细的说明,本领域的普通技术人员应当理解:其依然可以对前述各实施例所记载的技术方案进行修改,或者对其中部分技术特征进行等同替换;而这些修改或者替换,并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围,均应包含在本申请的保护范围之内。

Claims (20)

  1. 一种对抗网络训练方法,其中,包括:
    获取初始对抗网络,所述初始对抗网络包括含有初始参数的生成器模型以及训练完成的感应网络模型;
    将预设随机噪声输入至所述初始对抗网络,通过所述生成器模型生成与所述预设随机噪声对应的生成数据;
    根据所述生成数据,通过所述感应网络模型确定所述生成器模型的总损失值;
    在所述总损失值未达到预设的收敛条件时,更新迭代所述生成器模型的初始参数,直至所述总损失值达到所述预设的收敛条件时,将收敛之后的所述初始对抗网络记录为对抗网络。
  2. 如权利要求1所述的对抗网络训练方法,其中,所述获取初始对抗网络之前,还包括:
    获取全数据集,所述全数据集中包括若干分类对应的子数据集;一个所述子数据集关联一个子数据标签;
    将各所述子数据集输入至所述感应网络模型中,通过所述感应网络模型中的编码器模块对各所述子数据集进行编码转换,得到与各所述子数据集对应的子数据向量;
    通过所述感应网络模型中的感应模块将各所述子数据向量转化为与各子数据向量对应的类别向量;
    通过所述感应网络模型中的相关性模块,确定与各所述类别向量对应的相关性函数;
    将各所述相关性函数进行同维转换,确定各所述子数据集和与其对应的所述子数据标签之间的关系式;
    在各所述关系式达到预设关系式标准之后,表征所述感应网络模型训练完成。
  3. 如权利要求1所述的对抗网络训练方法,其中,所述根据所述生成数据,通过所述感应网络模型确定所述生成器模型的总损失值,包括:
    通过所述感应网络模型输出所述生成数据与小样本数据之间的第一损失值;
    通过所述感应网络模型输出所述生成数据与全数据集之间的第二损失值;
    根据所述第一损失值以及所述第二损失值,通过所述感应网络模型确定所述生成器模型的总损失值。
  4. 如权利要求3所述的对抗网络训练方法,其中,所述通过所述感应网络输出所述生成数据与所述小样本数据之间的第一损失值,包括:
    获取与所述生成数据对应的生成标签,以及与所述小样本数据对应的样本标签;
    根据所述小样本数据以及所述样本标签,确定所述感应网络模型中与所述小样本数据对应的第一关系式;
    根据所述生成数据、生成标签以及所述第一关系式,确定所述第一损失值。
  5. 如权利要求4所述的对抗网络训练方法,其中,所述通过所述感应网络模型输出所述生成数据与所述全数据集之间的第二损失值,包括:
    获取与所述全数据集对应的全数据标签;
    根据所述全数据集以及所述全数据集标签,确定所述感应网络模型中与所述全数据集对应的第二关系式;
    根据所述生成数据、生成标签以及所述第二关系式,确定所述第二损失值。
  6. 如权利要求3所述的对抗网络训练方法,其中,所述根据所述第一损失值以及所述第二损失值,通过所述感应网络模型确定所述生成器模型的总损失值,包括:
    通过如下损失函数确定所述生成器模型的总损失值:
    LOSS G=-α*log(similarity all)-βlog(similarity part)
    其中,LOSSG为所述总损失值;
    log(similarity part)为所述第一损失值;
    log(similarity all)为所述第二损失值;
    α为所述第二损失值对应的权重;
    β为所述第一损失值对应的权重。
  7. 一种医疗数据补充方法,其中,包括:
    接收包含医疗全数据集的数据补充指令;所述医疗全数据集中包含多样本医疗数据以及第一小样本医疗数据;所述第一小样本医疗数据与小样本标签关联;
    获取所述多样本医疗数据的第一样本数量以及所述第一小样本医疗数据的第二样本数量;所述第二样本数量小于所述第一样本数量;
    将所述第一样本数量与所述第二样本数量之间的差值记录为样本差值;
    通过训练完成的对抗网络生成数量与所述样本差值相等且与所述小样本标签关联的第二小样本医疗数据;其中,所述对抗网络是根据如权利要求1至6任一项所述对抗网路训练方法得到的;所述感应网络模型根据所述医疗全数据集训练得到;
    将生成的所述第二小样本医疗数据补充至所述医疗全数据集中。
  8. 一种对抗网络训练装置,其中,包括:
    对抗网络获取模块,用于获取初始对抗网络,所述初始对抗网络包括含有初始参数的生成器模型以及训练完成的感应网络模型;
    数据生成模块,用于将预设随机噪声输入至所述初始对抗网络,通过所述生成器模型生成与所述随机噪声对应的生成数据;
    损失值确定模块,用于根据所述生成数据,通过所述感应网络模型确定所述生成器模型的总损失值;
    收敛判断模块,用于在所述总损失值未达到预设的收敛条件时,更新迭代所述生成器模型的初始参数,直至所述总损失值达到所述预设的收敛条件时,将收敛之后的所述初始对抗网络记录为对抗网络。
  9. 一种计算机设备,包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令,其中,所述处理器执行所述计算机可读指令时实现如下步骤:
    获取初始对抗网络,所述初始对抗网络包括含有初始参数的生成器模型以及训练完成的感应网络模型;
    将预设随机噪声输入至所述初始对抗网络,通过所述生成器模型生成与所述预设随机噪声对应的生成数据;
    根据所述生成数据,通过所述感应网络模型确定所述生成器模型的总损失值;
    在所述总损失值未达到预设的收敛条件时,更新迭代所述生成器模型的初始参数,直至所述总损失值达到所述预设的收敛条件时,将收敛之后的所述初始对抗网络记录为对抗网络。
  10. 如权利要求9所述的计算机设备,其中,所述获取初始对抗网络之前,所述处理器执行所述计算机可读指令时还实现如下步骤:
    获取全数据集,所述全数据集中包括若干分类对应的子数据集;一个所述子数据集关联一个子数据标签;
    将各所述子数据集输入至所述感应网络模型中,通过所述感应网络模型中的编码器模块对各所述子数据集进行编码转换,得到与各所述子数据集对应的子数据向量;
    通过所述感应网络模型中的感应模块将各所述子数据向量转化为与各子数据向量对应的类别向量;
    通过所述感应网络模型中的相关性模块,确定与各所述类别向量对应的相关性函数;
    将各所述相关性函数进行同维转换,确定各所述子数据集和与其对应的所述子数据标签之间的关系式;
    在各所述关系式达到预设关系式标准之后,表征所述感应网络模型训练完成。
  11. 如权利要求9所述的计算机设备,其中,所述根据所述生成数据,通过所述感应网络模型确定所述生成器模型的总损失值,包括:
    通过所述感应网络模型输出所述生成数据与小样本数据之间的第一损失值;
    通过所述感应网络模型输出所述生成数据与全数据集之间的第二损失值;
    根据所述第一损失值以及所述第二损失值,通过所述感应网络模型确定所述生成器模型的总损失值。
  12. 如权利要求11所述的计算机设备,其中,所述通过所述感应网络输出所述生成数据与所述小样本数据之间的第一损失值,包括:
    获取与所述生成数据对应的生成标签,以及与所述小样本数据对应的样本标签;
    根据所述小样本数据以及所述样本标签,确定所述感应网络模型中与所述小样本数据对应的第一关系式;
    根据所述生成数据、生成标签以及所述第一关系式,确定所述第一损失值。
  13. 如权利要求12所述的计算机设备,其中,所述通过所述感应网络模型输出所述生成数据与所述全数据集之间的第二损失值,包括:
    获取与所述全数据集对应的全数据标签;
    根据所述全数据集以及所述全数据集标签,确定所述感应网络模型中与所述全数据集对应的第二关系式;
    根据所述生成数据、生成标签以及所述第二关系式,确定所述第二损失值。
  14. 一种计算机设备,包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机可读指令,其中,所述处理器执行所述计算机可读指令时实现如下步骤:
    接收包含医疗全数据集的数据补充指令;所述医疗全数据集中包含多样本医疗数据以及第一小样本医疗数据;所述第一小样本医疗数据与小样本标签关联;
    获取所述多样本医疗数据的第一样本数量以及所述第一小样本医疗数据的第二样本数量;所述第二样本数量小于所述第一样本数量;
    将所述第一样本数量与所述第二样本数量之间的差值记录为样本差值;
    通过训练完成的对抗网络生成数量与所述样本差值相等且与所述小样本标签关联的第二小样本医疗数据;其中,所述对抗网络是根据如权利要求1至6任一项所述对抗网路训练方法得到的;所述感应网络模型根据所述医疗全数据集训练得到;
    将生成的所述第二小样本医疗数据补充至所述医疗全数据集中。
  15. 一个或多个存储有计算机可读指令的可读存储介质,其中,所述计算机可读指令被一个或多个处理器执行时,使得所述一个或多个处理器执行如下步骤:
    获取初始对抗网络,所述初始对抗网络包括含有初始参数的生成器模型以及训练完成的感应网络模型;
    将预设随机噪声输入至所述初始对抗网络,通过所述生成器模型生成与所述预设随机噪声对应的生成数据;
    根据所述生成数据,通过所述感应网络模型确定所述生成器模型的总损失值;
    在所述总损失值未达到预设的收敛条件时,更新迭代所述生成器模型的初始参数,直至所述总损失值达到所述预设的收敛条件时,将收敛之后的所述初始对抗网络记录为对抗网络。
  16. 如权利要求15所述的可读存储介质,其中,所述获取初始对抗网络之前,所述计算机可读指令被一个或多个处理器执行时,使得所述一个或多个处理器还执行如下步骤:
    获取全数据集,所述全数据集中包括若干分类对应的子数据集;一个所述子数据集关联一个子数据标签;
    将各所述子数据集输入至所述感应网络模型中,通过所述感应网络模型中的编码器模块对各所述子数据集进行编码转换,得到与各所述子数据集对应的子数据向量;
    通过所述感应网络模型中的感应模块将各所述子数据向量转化为与各子数据向量对应的类别向量;
    通过所述感应网络模型中的相关性模块,确定与各所述类别向量对应的相关性函数;
    将各所述相关性函数进行同维转换,确定各所述子数据集和与其对应的所述子数据标签之间的关系式;
    在各所述关系式达到预设关系式标准之后,表征所述感应网络模型训练完成。
  17. 如权利要求15所述的可读存储介质,其中,所述根据所述生成数据,通过所述感应网络模型确定所述生成器模型的总损失值,包括:
    通过所述感应网络模型输出所述生成数据与小样本数据之间的第一损失值;
    通过所述感应网络模型输出所述生成数据与全数据集之间的第二损失值;
    根据所述第一损失值以及所述第二损失值,通过所述感应网络模型确定所述生成器模型的总损失值。
  18. 如权利要求17所述的可读存储介质,其中,所述通过所述感应网络输出所述生成数据与所述小样本数据之间的第一损失值,包括:
    获取与所述生成数据对应的生成标签,以及与所述小样本数据对应的样本标签;
    根据所述小样本数据以及所述样本标签,确定所述感应网络模型中与所述小样本数据对应的第一关系式;
    根据所述生成数据、生成标签以及所述第一关系式,确定所述第一损失值。
  19. 如权利要求18所述的可读存储介质,其中,所述通过所述感应网络模型输出所述生成数据与所述全数据集之间的第二损失值,包括:获取与所述全数据集对应的全数据标签;
    根据所述全数据集以及所述全数据集标签,确定所述感应网络模型中与所述全数据集对应的第二关系式;
    根据所述生成数据、生成标签以及所述第二关系式,确定所述第二损失值。
  20. 一个或多个存储有计算机可读指令的可读存储介质,其中,所述计算机可读指令被一个或多个处理器执行时,使得所述一个或多个处理器执行如下步骤:
    接收包含医疗全数据集的数据补充指令;所述医疗全数据集中包含多样本医疗数据以及第一小样本医疗数据;所述第一小样本医疗数据与小样本标签关联;
    获取所述多样本医疗数据的第一样本数量以及所述第一小样本医疗数据的第二样本数量;所述第二样本数量小于所述第一样本数量;
    将所述第一样本数量与所述第二样本数量之间的差值记录为样本差值;
    通过训练完成的对抗网络生成数量与所述样本差值相等且与所述小样本标签关联的第二小样本医疗数据;其中,所述对抗网络是根据如权利要求1至6任一项所述对抗网路训练方法得到的;所述感应网络模型根据所述医疗全数据集训练得到;
    将生成的所述第二小样本医疗数据补充至所述医疗全数据集中。
PCT/CN2020/135342 2020-10-22 2020-12-10 对抗网络训练、医疗数据补充方法、装置、设备及介质 WO2021189960A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011140634.9A CN112259247B (zh) 2020-10-22 2020-10-22 对抗网络训练、医疗数据补充方法、装置、设备及介质
CN202011140634.9 2020-10-22

Publications (1)

Publication Number Publication Date
WO2021189960A1 true WO2021189960A1 (zh) 2021-09-30

Family

ID=74264222

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/135342 WO2021189960A1 (zh) 2020-10-22 2020-12-10 对抗网络训练、医疗数据补充方法、装置、设备及介质

Country Status (2)

Country Link
CN (1) CN112259247B (zh)
WO (1) WO2021189960A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114254739A (zh) * 2021-12-21 2022-03-29 南方电网数字电网研究院有限公司 多模态电力传感器的数据处理方法、装置和计算机设备
CN114548367A (zh) * 2022-01-17 2022-05-27 中国人民解放军国防科技大学 基于对抗网络的多模态数据的重构方法及装置
CN117291252A (zh) * 2023-11-27 2023-12-26 浙江华创视讯科技有限公司 稳定视频生成模型训练方法、生成方法、设备及存储介质
CN117933250A (zh) * 2024-03-22 2024-04-26 南京泛美利机器人科技有限公司 一种基于改进生成对抗网络的新菜谱生成方法

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113239022B (zh) * 2021-04-19 2023-04-07 浙江大学 医疗诊断缺失数据补全方法及补全装置、电子设备、介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109948717A (zh) * 2019-03-26 2019-06-28 江南大学 一种生成对抗网络的自生长训练方法
CN110070174A (zh) * 2019-04-10 2019-07-30 厦门美图之家科技有限公司 一种生成对抗网络的稳定训练方法
US20200193269A1 (en) * 2018-12-18 2020-06-18 Samsung Electronics Co., Ltd. Recognizer, object recognition method, learning apparatus, and learning method for domain adaptation
CN111797078A (zh) * 2019-04-09 2020-10-20 Oppo广东移动通信有限公司 数据清洗方法、模型训练方法、装置、存储介质及设备

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107220600B (zh) * 2017-05-17 2019-09-10 清华大学深圳研究生院 一种基于深度学习的图片生成方法及生成对抗网络
CN108805418B (zh) * 2018-05-22 2021-08-31 福州大学 一种基于生成式对抗网络的交通数据填充方法
EP3830793A4 (en) * 2018-07-30 2022-05-11 Memorial Sloan Kettering Cancer Center MULTIMODE, MULTI-RESOLUTION DEEP LEARNING NETWORKS FOR SEGMENTATION, OUTCOME PREDICTION, AND MONITORING LONGITUDINAL RESPONSES TO IMMUNOTHERAPY AND RADIATION THERAPY
CN109522973A (zh) * 2019-01-17 2019-03-26 云南大学 基于生成式对抗网络与半监督学习的医疗大数据分类方法及***
CN110503187B (zh) * 2019-07-26 2024-01-16 深圳万知达科技有限公司 一种用于功能核磁共振成像数据生成的生成对抗网络模型的实现方法
CN111275686B (zh) * 2020-01-20 2023-05-26 中山大学 用于人工神经网络训练的医学图像数据的生成方法及装置
CN111738346A (zh) * 2020-06-28 2020-10-02 辽宁大学 一种生成式对抗网络估值的不完整数据聚类方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200193269A1 (en) * 2018-12-18 2020-06-18 Samsung Electronics Co., Ltd. Recognizer, object recognition method, learning apparatus, and learning method for domain adaptation
CN109948717A (zh) * 2019-03-26 2019-06-28 江南大学 一种生成对抗网络的自生长训练方法
CN111797078A (zh) * 2019-04-09 2020-10-20 Oppo广东移动通信有限公司 数据清洗方法、模型训练方法、装置、存储介质及设备
CN110070174A (zh) * 2019-04-10 2019-07-30 厦门美图之家科技有限公司 一种生成对抗网络的稳定训练方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
LIU QIKAI , JIANG DAIHONG , LI WENJI: "Generative Adversarial Network Based on Piecewise Loss", COMPUTER ENGINEERING, vol. 45, no. 5, 31 May 2019 (2019-05-31), pages 155 - 160+168, XP055853386, ISSN: 1000-3428, DOI: 10.19678/j.issn.1000-3428.0050529 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114254739A (zh) * 2021-12-21 2022-03-29 南方电网数字电网研究院有限公司 多模态电力传感器的数据处理方法、装置和计算机设备
CN114548367A (zh) * 2022-01-17 2022-05-27 中国人民解放军国防科技大学 基于对抗网络的多模态数据的重构方法及装置
CN114548367B (zh) * 2022-01-17 2024-02-20 中国人民解放军国防科技大学 基于对抗网络的多模态数据的重构方法及装置
CN117291252A (zh) * 2023-11-27 2023-12-26 浙江华创视讯科技有限公司 稳定视频生成模型训练方法、生成方法、设备及存储介质
CN117291252B (zh) * 2023-11-27 2024-02-20 浙江华创视讯科技有限公司 稳定视频生成模型训练方法、生成方法、设备及存储介质
CN117933250A (zh) * 2024-03-22 2024-04-26 南京泛美利机器人科技有限公司 一种基于改进生成对抗网络的新菜谱生成方法

Also Published As

Publication number Publication date
CN112259247A (zh) 2021-01-22
CN112259247B (zh) 2022-08-23

Similar Documents

Publication Publication Date Title
WO2021189960A1 (zh) 对抗网络训练、医疗数据补充方法、装置、设备及介质
US11645833B2 (en) Generative adversarial network medical image generation for training of a classifier
WO2021258348A1 (zh) 异常流量检测方法和***、及计算机存储介质
WO2021121129A1 (zh) 雷同病例检测方法、装置、设备及存储介质
WO2021121127A1 (zh) 样本类别识别方法、装置、计算机设备及存储介质
WO2022142613A1 (zh) 训练语料扩充方法及装置、意图识别模型训练方法及装置
CN108563782B (zh) 商品信息格式处理方法、装置、计算机设备和存储介质
AU2020385264B2 (en) Fusing multimodal data using recurrent neural networks
US20210182660A1 (en) Distributed training of neural network models
CN113157863B (zh) 问答数据处理方法、装置、计算机设备及存储介质
KR20200068050A (ko) 인공지능 수행을 위한 학습 데이터 생성장치 및 방법
JP6979661B2 (ja) 量子分類器としての浅い回路のためのシステム、コンピュータ実装方法およびコンピュータ・プログラム
CN112016318B (zh) 基于解释模型的分诊信息推荐方法、装置、设备及介质
US20170185913A1 (en) System and method for comparing training data with test data
US11042710B2 (en) User-friendly explanation production using generative adversarial networks
WO2022141864A1 (zh) 对话意图识别模型训练方法、装置、计算机设备及介质
US20210182661A1 (en) Neural Network Training From Private Data
WO2023065635A1 (zh) 命名实体识别方法、装置、存储介质及终端设备
WO2021068563A1 (zh) 样本数据处理方法、装置、计算机设备及存储介质
WO2022103682A1 (en) Face recognition from unseen domains via learning of semantic features
KR20210015531A (ko) 뉴럴 네트워크 모델을 업데이트하는 방법 및 시스템
US20220405529A1 (en) Learning Mahalanobis Distance Metrics from Data
US20230082014A1 (en) Artificial Intelligence Based Technologies for Improving Patient Intake
WO2022257468A1 (zh) 对话管理***更新方法、装置、计算机设备及存储介质
CN111091198B (zh) 一种数据处理方法及装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20927519

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20927519

Country of ref document: EP

Kind code of ref document: A1