WO2023123926A1 - 人工智能任务处理方法、装置、电子设备及可读存储介质 - Google Patents

人工智能任务处理方法、装置、电子设备及可读存储介质 Download PDF

Info

Publication number
WO2023123926A1
WO2023123926A1 PCT/CN2022/100481 CN2022100481W WO2023123926A1 WO 2023123926 A1 WO2023123926 A1 WO 2023123926A1 CN 2022100481 W CN2022100481 W CN 2022100481W WO 2023123926 A1 WO2023123926 A1 WO 2023123926A1
Authority
WO
WIPO (PCT)
Prior art keywords
task
training
loss
information
accuracy
Prior art date
Application number
PCT/CN2022/100481
Other languages
English (en)
French (fr)
Inventor
张潇澜
李峰
周镇镇
Original Assignee
苏州浪潮智能科技有限公司
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 苏州浪潮智能科技有限公司 filed Critical 苏州浪潮智能科技有限公司
Publication of WO2023123926A1 publication Critical patent/WO2023123926A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00Arrangements for program control, e.g. control units
    • G06F9/06Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46Multiprogramming arrangements
    • G06F9/50Allocation of resources, e.g. of the central processing unit [CPU]
    • G06F9/5005Allocation of resources, e.g. of the central processing unit [CPU] to service a request
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00Machine learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods

Definitions

  • the present application relates to the technical field of artificial intelligence, in particular to an artificial intelligence task processing method, device, electronic equipment and readable storage medium.
  • machine vision is to use machines instead of human eyes for measurement and judgment. It uses machine vision products, namely image capture devices such as CMOS (Complementary Metal Oxide Semiconductor, Complementary Metal Oxide Semiconductor) and CCD (Charge-coupled Device, charge-coupled component), etc., convert the subject into an image signal, and then send it to a special image processing system to obtain the shape information of the subject, and convert it into a digital signal according to pixel distribution, brightness, color and other information; the image processing system processes these signals Perform various calculations to extract the characteristics of the target, and then control the on-site equipment actions according to the results of the discrimination.
  • image capture devices such as CMOS (Complementary Metal Oxide Semiconductor, Complementary Metal Oxide Semiconductor) and CCD (Charge-coupled Device, charge-coupled component), etc.
  • Natural language processing is a variety of theories and methods to realize effective communication between humans and computers using natural language. By processing text information or voice signals, effective signals that can be understood by computers are obtained. The recognition accuracy and recognition efficiency of speech information will greatly affect the performance of natural language processing.
  • Deep learning uses pre-labeled data sets, such as sample image data sets or natural language sample sets, to train the parameter information of relevant neurons in the model to better complete prediction tasks in specific scenarios, such as image classification, target detection, Image segmentation, speech recognition, and more.
  • the training process of the model is a process of continuous optimization of parameters, and the training process directly affects the performance of the model. The training of the model depends on the settings of the initial parameters and hyperparameters.
  • the early stopping strategy of the model is usually determined by a trade-off between the training time of the model and the generalization error (Validation Error) on the verification set.
  • the generalization error (Validation Error) on the verification set.
  • N Epoch a specified period such as N Epoch
  • one Epoch is the process of training all training samples once, if the error of the model on the verification set is higher than the training error of the last model on the training set, stop the model training .
  • the specific implementation is to set the appropriate stop standard. Stopping criteria in the prior art include:
  • the verification error curve is not a smooth and monotonous curve, and it may continue to improve after several bad trainings. Therefore, only considering the verification error to set the early stopping strategy, its accuracy will be limited to a certain extent, and eventually lead to The resulting models performed poorly and did not perform well on AI processing tasks.
  • Embodiments of the present invention provide an artificial intelligence task processing method on the one hand, comprising:
  • the AI data set includes an AI training set and an AI verification set; both the AI training set and the AI verification set include multiple consecutive subsets corresponding to a sliding window, and each subset corresponds to a window of the sliding window;
  • the task loss information is generated;
  • the AI task execution model based on the expected value of the positive performance index of the AI verification set in each sliding window training process, generate task accuracy expectation information
  • an instruction to continue training the AI task execution model is output.
  • the period model accuracy representation information is determined according to the task loss information and the task accuracy expectation information, including:
  • the call cycle result represents the calculation relational expression, and the calculation cycle model accuracy represents information;
  • the cycle result represents the calculation relational expression as:
  • S is the periodic model accuracy representation information
  • i is the i-th training cycle
  • is the training tolerance
  • M is the task loss information
  • N is the task accuracy expectation information
  • S (i+ ⁇ ) is the cycle result of the i+ ⁇ th training AI task execution model
  • M (i+ ⁇ ) is the task loss information of the i+ ⁇ th training AI task execution model
  • N (i+ ⁇ ) is the i+ ⁇ th time Task accuracy expectation information for training the AI task execution model
  • represents a logical AND operator.
  • judging whether the period model accuracy representation information matches the task accuracy requirement information includes:
  • the periodic result of training the AI task execution model at least once is not 0, and it is judged whether each element of the periodic model accuracy representation information is 0; correspondingly, in response to the periodic model accuracy representation information and the task accuracy
  • the process of outputting an instruction to stop training the AI task execution model after matching the demand information includes:
  • the process of outputting instructions to continue training the AI task execution model includes:
  • the task loss information is generated based on the loss value of the AI training set in each sliding window training process, including:
  • the task loss information of the current training cycle is determined according to the current standard deviation, the forward standard deviation and the loss change threshold, including:
  • the periodic task loss information calculation relationship is:
  • M i is the task loss information of the i-th training cycle
  • loss ik is the loss value of the k-th sliding window in the i-th training cycle
  • loss ik is the k-th sliding window of the i-1-th training cycle
  • ⁇ (lossi1, lossi2,..., lossik) is the current standard deviation of all loss values in the i-th training cycle
  • ⁇ (loss i1 , loss i2 ,...,loss ik ) is the i-1th
  • is the loss change threshold.
  • task accuracy expectation information is generated, including:
  • the task accuracy expectation information of the current training cycle is determined.
  • the task accuracy expectation information of the current training period is determined according to the current expectation value, the forward expectation value and the performance change threshold, including:
  • the periodic task accuracy calculation relational expression is:
  • N i is the expected task accuracy information of the i-th training cycle
  • perf ik is the expected value of the k-th sliding window in the i-th training cycle
  • perf (i-1)k is the expected value of the i-1-th training cycle
  • the expected value of the k-th sliding window, E(perf i1 ,perf i2 ,....,perf ik ) is the current expected value of all positive performance indicators in the i-th training cycle
  • E(perf (i-1)1 , perf (i-1)2 ,....,perf (i-1)k ) is the forward expected value of all forward performance indicators in the i-1th training cycle
  • is the performance change threshold.
  • obtaining the AI data set and AI task execution model corresponding to the artificial intelligence task to be processed includes:
  • an information processing interface When an information input instruction is received, an information processing interface is displayed; the information processing interface includes an information input area and a result display area; and
  • the result display area is used to display the trained AI task execution model and/or the task execution result of the artificial intelligence task to be processed.
  • an artificial intelligence task processing device including:
  • the information acquisition module is used to obtain the AI data set and AI task execution model corresponding to the artificial intelligence task to be processed;
  • the AI data set includes an AI training set and an AI verification set; both the AI training set and the AI verification set include multiple consecutive subsets, each subset corresponds to a window of the sliding window;
  • the loss calculation module is used to generate task loss information based on the loss value of the AI training set in each sliding window training process according to the AI task execution model;
  • the expectation calculation module is used to generate task accuracy expectation information based on the AI task execution model and based on the expected value of the positive performance index of the AI verification set in each sliding window training process;
  • the model training end determination module is used to determine whether to stop the training of the AI task execution model based on the task loss information, the task accuracy expectation information, and the task accuracy requirement information of the artificial intelligence task to be processed, and execute the pending AI task execution model based on the trained AI task execution model. artificial intelligence tasks.
  • the embodiment of the present application also provides an electronic device, including a memory and one or more processors, where computer-readable instructions are stored in the memory, and when the computer-readable instructions are executed by one or more processors, one or more The processor executes the steps of the artificial intelligence task processing method in any of the foregoing embodiments.
  • the embodiment of the present application also provides one or more non-volatile computer-readable storage media storing computer-readable instructions.
  • the computer-readable instructions are executed by one or more processors, the one or more processors Execute the steps of the artificial intelligence task processing method in any of the above embodiments.
  • FIG. 1 is a schematic flowchart of an artificial intelligence task processing method provided in one or more embodiments of the present application
  • FIG. 2 is a schematic flowchart of an artificial intelligence task processing method in a schematic example provided in one or more embodiments of the present application;
  • FIG. 3 is a schematic structural diagram of an information processing system provided in one or more embodiments of the present application.
  • FIG. 4 is a structural diagram of a specific implementation of an artificial intelligence task processing device provided in one or more embodiments of the present application;
  • Fig. 5 is a structural diagram of a specific implementation manner of an electronic device provided in one or more embodiments of the present application.
  • FIG. 1 is a schematic flowchart of an artificial intelligence task processing method provided by an embodiment of the present application.
  • the embodiment of the present application may include the following:
  • S101 Obtain an AI data set and an AI task execution model corresponding to an artificial intelligence task to be processed.
  • the artificial intelligence task to be processed may be any task that needs to be performed in the field of artificial intelligence, such as image classification and recognition tasks, object detection tasks, image segmentation tasks, speech recognition tasks, and so on.
  • the AI task execution model is a network model based on a deep learning model, such as an artificial neural network model, a convolutional neural network model, a recursive tensor neural network, a generative confrontation network, a long-term short-term memory network, and so on.
  • the AI task execution model is used to execute the artificial intelligence task to be processed.
  • the artificial intelligence task to be processed is an image segmentation task
  • the AI task execution model is used to perform image segmentation on the input image.
  • the AI data set refers to the sample data set used in the training process of the AI task execution model.
  • the AI data set can be, for example, an image data set, a voice signal data set, or a text information data set.
  • the AI dataset includes an AI training set and an AI verification set; the AI training set is used to train the AI task execution model, and the AI verification set is used to verify the performance of the AI task execution model.
  • the AI training set in this embodiment also includes an AI test set. Both the AI training set and the AI verification set are divided into multiple consecutive subsets, the AI training set corresponds to a sliding window, the AI verification set corresponds to a sliding window, and each subset is a window of the sliding window.
  • the AI training set and AI verification set slide from the first window corresponding to the first subset to the last window corresponding to the last subset to complete a sliding window model training.
  • the AI training set includes 1000 images
  • the first 100 images can be used as the first subset
  • the 100th image to the 500th image can be used as the second subset
  • the 500th image to the 700th image can be used as the first subset.
  • the four subsets and the fifth subset are both a window of the sliding window of the AI training set.
  • the sliding window will slide from the first subset to the fifth subset in order to read the sample images in the current subset.
  • train k times on the training data set that is, the AI training set and the AI verification set, that is, k epochs, also called a training cycle
  • each training is called a sliding window, that is, a training cycle includes k epochs or k sliding windows.
  • k ⁇ 1 one epoch means that the model is trained once on the training set.
  • a training cycle means that the AI task execution model continuously completes k times of training on the specified training data set.
  • the loss value refers to the gap between the forward calculation result of each iteration of the AI task execution model and the real value.
  • Each training cycle will train multiple times on the AI training set, and each time is a sliding window. Any loss function in the prior art can be used to calculate the performance of training the AI task execution model based on the AI training set under each sliding window.
  • Loss value according to the loss value under each sliding window, determine the task loss information of the AI task execution model in one training cycle or multiple training cycles. The task loss information is used to reflect the current training cycle or the AI task in the actual training period Error in executing the model.
  • the positive performance index is used to evaluate the performance of the AI task execution model obtained after one training cycle or multiple training cycles.
  • the performance of the AI task execution model can be verified after each training cycle using the AI verification set.
  • Each training cycle will be verified multiple times in the AI verification set, and each time is a sliding window, which can be Use any expected calculation method in the prior art to calculate the expected value of the forward performance index based on the AI verification set to verify the AI task execution model under each sliding window, and determine the AI according to the expected value of the forward performance index under each sliding window
  • the task accuracy expectation information of the task execution model in one training cycle or multiple training cycles is used to reflect the performance of the AI task execution model in the current training cycle or the actual training period.
  • S104 Determine whether to stop the training of the AI task execution model according to the task loss information, the task accuracy expectation information, and the task accuracy requirement information of the AI task to be processed, and execute the AI task to be processed based on the trained AI task execution model.
  • the task accuracy requirement information is based on the accuracy of the artificial intelligence task to be processed required by the actual application scenario. For example, for the execution of image recognition and segmentation tasks, for lesion identification and lesion cutting , the accuracy of identifying and cutting the lesion needs to be greater than 99%, and for the vehicle identification task, the accuracy of identifying the vehicle only needs to be 90%. Judging whether the AI task execution model trained in the current training cycle meets the requirements according to the task loss information, task accuracy expectation information, and task accuracy requirement information of the AI task to be processed. If not, the AI task execution model needs to be continuously trained. If it is satisfied, the training of the AI task execution model is stopped, and the AI task execution model obtained at this time is the trained model, and the AI task execution model can be used to execute the artificial intelligence task to be processed.
  • a method for judging whether the AI task execution model stops early may include the following steps:
  • the cycle model accuracy representation information is used to represent the performance of the AI task execution model or the task execution accuracy after the current training cycle ends.
  • the periodic model accuracy representation information can be calculated by calling the periodic result representation calculation relational expression, and the periodic result representation calculation relational expression can be expressed as:
  • S is the periodic model accuracy representation information, i is the ith training cycle, ⁇ is the training tolerance; M is the task loss information, N is the task accuracy expectation information, S (i+ ⁇ ) is the i+ ⁇ training AI task
  • M is the task loss information
  • N is the task accuracy expectation information
  • S (i+ ⁇ ) is the i+ ⁇ training AI task
  • the periodic results of the execution model represent that M (i+ ⁇ ) is the task loss information of the i+ ⁇ -th training AI task execution model, and N (i+ ⁇ ) is the task accuracy expectation of the i+ ⁇ -th training AI task execution model information, ⁇ represents the logical AND operator.
  • "true” is replaced by the number 1
  • "false” is replaced by 0.
  • the training tolerance ⁇ indicates that the early stopping strategy of the AI task execution model is based on the results of ⁇ training cycles.
  • the early stopping strategy refers to the criteria for judging whether the training of the AI task execution model is over.
  • the periodic model accuracy representation information is a set, and the value of each element in the set is 0 or 1, that is, the periodic model accuracy representation information is a set composed of 0 and 1. If the element is 0, it means that at least one of the task loss information and the task accuracy expectation information is 0, and the task loss information is 0, which means that the model convergence performance of the AI task execution model in the i-th training cycle on the AI training set is not satisfied Requirements, the task accuracy expectation information is 0, which means that the model generalization ability of the AI task execution model in the i-th training cycle on the AI verification set does not meet the requirements.
  • the process of judging whether the periodic model accuracy representation information matches the task accuracy requirement information may include:
  • the task accuracy requirement information is that the cycle result of training the AI task execution model at least once is not 0, and it is judged whether each element of the cycle model accuracy representation information is 0;
  • This embodiment does not limit how to execute steps S102 and S103.
  • This embodiment also provides a calculation method of task loss information, that is, according to the AI task execution model based on the AI training set in each sliding window training process
  • the process of generating task loss information by loss value may include the following:
  • the task loss information can be calculated by calling the periodic task loss information calculation relational expression, and the periodic task loss information calculation relational expression can be expressed as:
  • M i is the task loss information of the i-th training cycle
  • loss ik is the loss value of the k-th sliding window in the i-th training cycle
  • loss ik is the k-th sliding window of the i-1-th training cycle
  • ⁇ (loss i1 ,loss i2 ,...,loss ik ) is the current standard deviation of all loss values in the i-th training cycle
  • ⁇ (loss i1 ,loss i2 ,...,loss ik ) is The forward standard deviation of all loss values for the i-1th training epoch
  • is the loss change threshold.
  • This embodiment also provides a calculation method for the expected information of task accuracy, that is, the process of generating expected information of task accuracy based on the expected value of the forward performance index of the AI verification set in each sliding window training process according to the AI task execution model can be Include the following:
  • the task accuracy expectation information of the current training cycle is determined.
  • the task accuracy expectation information can be calculated by calling the periodic task accuracy calculation relational expression, and the periodic task accuracy calculation relational expression can be expressed as:
  • N i is the expected task accuracy information of the i-th training cycle
  • perf ik is the expected value of the k-th sliding window in the i-th training cycle
  • perf (i-1)k is the expected value of the i-1-th training cycle
  • the expected value of the k-th sliding window, E(perf i1 ,perf i2 ,....,perf ik ) is the current expected value of all positive performance indicators in the i-th training cycle
  • E(perf (i-1)1 , perf (i-1)2 ,....,perf (i-1)k ) is the forward expected value of all forward performance indicators in the i-1th training cycle
  • is the performance change threshold.
  • b is a performance change degree factor, and its value range may be (0, + ⁇ ), and the value of b may describe the acceptable performance improvement degree of the early stop strategy of this embodiment. The larger the value of b, the higher the acceptable performance improvement degree, and the smaller the b value, the lower the acceptable performance improvement degree.
  • the output of the AI task execution model in the i-th training cycle is denoted as S i
  • the training cycle includes performing k training processes on the training set.
  • ⁇ (loss i1 , loss i2 ,...,loss ik ) is the current standard deviation of all loss values loss in the i-th training cycle, which includes k training rounds, that is, the number of epochs, which can be abbreviated as ⁇ (loss i[1,k] ), representing 1 to k epochs of the i-th training cycle.
  • E(perf i1 ,perf i2 ,....,perf ik ) represents the current expected value of the forward performance indicator perf of the AI task execution model on the AI verification set in the i-th training cycle, which can be abbreviated as E(perf i [1,k] ).
  • ⁇ >1 is the training tolerance of the artificial intelligence task to be processed, which is an integer greater than or equal to 1, which means continuous training for ⁇ cycles.
  • f ⁇ is defined as a mapping relationship: ⁇ S i+1 ,S i+2 ,...,S i+ ⁇ ⁇ e 1 ,e 2 , whil,e ⁇ ⁇ e i ⁇ 0,1 ⁇ .
  • the output of S is a set (S i+1 , S i+2 ,....,S i+ ⁇ ) consisting of the output of continuous ⁇ training cycles.
  • each Each element has a value of 0 or 1.
  • the evaluation standard of the early stopping strategy of the embodiment of the present application can be obtained: given the tolerance parameter ⁇ , if the elements in the output set of the periodic model accuracy representation information S are all 0, then the training task needs to be stopped.
  • S i M i ⁇ N i, it can be known that the result of S i corresponds to the following four situations, and the specific explanation is given below:
  • this embodiment can help the user make a judgment on whether to stop early for the training task according to his actual needs.
  • display an information processing interactive interface that is, display an information processing interactive interface in response to receiving an information input instruction
  • the information processing interactive interface includes an information input area and a result display area; the result display area is used to display the trained AI task execution model and/or the task execution results of the artificial intelligence tasks to be processed.
  • the information input instruction is an instruction issued by the user.
  • the user wants to execute the artificial intelligence task to be processed, the user sends the information input instruction to the system, and the AI task execution model and The corresponding AI data set is input, and the relevant parameters of the AI task execution model and the relevant parameters of the task accuracy requirement information are initialized, including two types of model parameters and hyperparameters of the AI task execution model.
  • the model parameters include the weight and offset matrix of each neuron, the initial convolution kernel of each level of convolutional layer, the weight matrix and offset vector of the fully connected layer.
  • Hyperparameters include learning rate, momentum value, optimizer, training times (epoch number), batch training data size (batch-size) and other related parameters.
  • the relevant parameters of the task accuracy requirement information may include the definition of the system tolerance ⁇ and the sliding window size (k) parameters.
  • the present application also provides a schematic example in conjunction with FIG. 2 and FIG. It is defined as an information processing system S, which can be applied to the training process of deep learning models in any field.
  • the system S will first obtain the model M to be trained, that is, the AI task execution model and the AI data set (including the AI training set and the AI verification set); secondly, under the specified sliding window, the evaluation model is The output of ⁇ training epochs. Finally, comprehensively consider the error value of the AI training set and the positive performance index of the AI verification set to decide whether to stop the training task.
  • the AI task execution model can be a network model based on Resnet50, and the artificial intelligence task to be processed is an image recognition task.
  • the corresponding positive performance index can be the performance accuracy acc, which can include the following:
  • the AI task execution model is M (Resnet50), the AI training set is train_set, and the AI verification set is valid_set.
  • the model parameters include the weight and offset matrix of each neuron, the initial convolution kernel of each level of convolutional layer, the weight matrix and offset vector of the fully connected layer.
  • Hyperparameters include learning rate, momentum value, optimizer, training times (epoch number), batch training data size (batch-size) and other related parameters.
  • Initialize the relevant parameters of the information system S including the definition of system tolerance ⁇ and sliding window size (k) parameters.
  • the coefficient a 0.1 in the training set loss change threshold
  • the coefficient b 0.2 in the verification set performance change threshold
  • the training tolerance Parameter ⁇ 2.
  • the tolerance parameter ⁇ repeat the following process ⁇ times to obtain the output result of the information processing system S: that is, the output set with a length ⁇ .
  • A1 Calculate the standard deviation of the loss of the training process of the first sliding window (that is, epoch from 1 to 10) and the expected value of the accuracy on the AI verification set:
  • ⁇ 1> is the expected value of 10 loss values for the first sliding window.
  • A2 Calculate the standard deviation of the loss of the second sliding window (epoch from 11 to 20) training process and the expected value of the accuracy on the AI verification set:
  • A3 Calculate the change threshold of loss on the AI training set according to the following calculation relationship:
  • A4 Calculate the change threshold of accuracy on the AI verification set according to the following calculation relationship:
  • A5 Calculate the difference ( ⁇ (loss 1[2,10] )- ⁇ (loss 1[1,10] ) and ⁇ to obtain the value of M 2 , M 2 ⁇ ⁇ 0, 1 ⁇ .
  • A6 Calculate the difference (E(acc 2[1,10] )-E(acc 1[1,10] ) and ⁇ to obtain the value of N 2 , N 2 ⁇ 0,1 ⁇ .
  • the output set of information processing system S can be obtained: ⁇ S 2 , S 3 ⁇ . If the elements in the set are all 0, the training task can be stopped. Otherwise, continue training, and judge the results of ⁇ S 3, S 4 ⁇ ... in the same steps to decide whether to stop early.
  • a judgment is given whether the training task is early stopping. If the elements in the output set ⁇ e 1 , e 2 , e 3 ,...,e ⁇ ⁇ are all 0, perform early stop operation; otherwise, continue training, and judge again in the next training cycle until the AI task execution model The training process is over.
  • the information processing system S is used to control the training process of the AI task execution model. Its sliding-window-based early-stop technology during model training can not only be applied to the traditional learning model training process, but also can be applied to the training process of automatic hyperparameter tuning, which can terminate the training process of a certain group of hyperparameters whose performance tends to decline in advance. Free up hardware resources and provide more opportunities for better hyperparameter combination searches.
  • some specific problems such as classification, image recognition, target detection, natural language processing, etc.
  • the ideas of the above embodiments can also be adopted during model training, and the early stopping strategy is used in the training process to make the final training model performance Better, better generalization.
  • the user can limit the relevant parameter factors of the early stop strategy in this embodiment according to his own needs. If the amount of data is large enough, these parameter factors can also be trained as hyperparameters to obtain appropriate early stopping policy criteria.
  • the embodiment of the present application also provides a corresponding device for the artificial intelligence task processing method, which further makes the method more practical.
  • the device can be described separately from the perspective of functional modules and hardware.
  • the following is an introduction to the artificial intelligence task processing device provided by the embodiment of the present application.
  • the artificial intelligence task processing device described below and the artificial intelligence task processing method described above can be referred to in correspondence.
  • FIG. 4 is a structural diagram of an artificial intelligence task processing device provided in an embodiment of the present application in a specific implementation manner.
  • the device may include:
  • the information acquisition module 401 is used to obtain the AI data set and the AI task execution model corresponding to the artificial intelligence task to be processed; the AI data set includes an AI training set and an AI verification set; both the AI training set and the AI verification set include a corresponding sliding window Multiple consecutive subsets, each corresponding to a window of the sliding window.
  • the loss calculation module 402 is configured to generate task loss information based on the loss value of the AI training set during each sliding window training process according to the AI task execution model.
  • the expectation calculation module 403 is configured to generate task accuracy expectation information based on the AI task execution model and based on the expected value of the positive performance index of the AI verification set in each sliding window training process.
  • the model training end determination module 404 is used to determine whether to stop the training of the AI task execution model based on the task loss information, the task accuracy expectation information, and the task accuracy requirement information of the artificial intelligence task to be processed, and execute the pending AI task execution model based on the trained AI task execution model. Handle artificial intelligence tasks.
  • the above-mentioned model training completion determination module 404 can be further used to:
  • an instruction to stop training the AI task execution model is output; in response to a mismatch between the periodic model accuracy representation information and the task accuracy requirement information, an instruction to continue training the AI task execution model is output.
  • the above-mentioned model training completion determination module 404 can be further used to: call the cycle result to represent the calculation relational expression, and calculate the cycle model accuracy to represent information;
  • S is the periodic model accuracy representation information
  • i is the i-th training cycle
  • is the training tolerance
  • M is the task loss information
  • N is the task accuracy expectation information
  • S (i+ ⁇ ) is the cycle result of the i+ ⁇ th training AI task execution model
  • M (i+ ⁇ ) is the task loss information of the i+ ⁇ th training AI task execution model
  • N (i+ ⁇ ) is the i+ ⁇ th time Task accuracy expectation information for training the AI task execution model
  • represents a logical AND operator.
  • the above-mentioned model training completion determination module 404 may be further used to:
  • the task accuracy requirement information means that at least one cycle result of training the AI task execution model is not 0;
  • the loss calculation module 402 may be further used to:
  • the above-mentioned loss calculation module 402 may further be used for:
  • the periodic task loss information calculation relationship is:
  • M i is the task loss information of the i-th training cycle
  • loss ik is the loss value of the k-th sliding window in the i-th training cycle
  • loss ik is the k-th sliding window of the i-1-th training cycle
  • ⁇ (lossi1, lossi2,..., lossik) is the current standard deviation of all loss values in the i-th training cycle
  • ⁇ (loss i1 , loss i2 ,...,loss ik ) is the i-1th
  • is the loss change threshold.
  • the above-mentioned expected calculation module 403 may be further used to:
  • the task accuracy expectation information of the current training cycle is determined.
  • the above-mentioned expected calculation module 403 may further be used for:
  • the periodic task accuracy calculation relational expression is:
  • N i is the expected task accuracy information of the i-th training cycle
  • perf ik is the expected value of the k-th sliding window in the i-th training cycle
  • perf (i-1)k is the expected value of the i-1-th training cycle
  • the expected value of the k-th sliding window, E(perf i1 ,perf i2 ,....,perf ik ) is the current expected value of all positive performance indicators in the i-th training cycle
  • E(perf (i-1)1 , perf (i-1)2 ,....,perf (i-1)k ) is the forward expected value of all forward performance indicators in the i-1th training cycle
  • is the performance change threshold.
  • the above information acquisition module 401 may be further used to:
  • the information processing interactive interface includes an information input area and a result display area
  • the result display area is used to display the trained AI task execution model and/or the task execution results of the artificial intelligence tasks to be processed.
  • each functional module of the artificial intelligence task processing device in the embodiment of the present application can be specifically realized according to the method in the above method embodiment, and the specific implementation process can refer to the relevant description of the above method embodiment, and will not be repeated here.
  • the embodiments of the present application can effectively improve the performance of artificial intelligence task processing and reduce the computing resources consumed in the process of artificial intelligence task processing.
  • FIG. 5 is a schematic structural diagram of an electronic device provided in an embodiment of the present application in an implementation manner. As shown in Figure 5, the electronic device includes a memory 50 for storing computer-readable instructions; a processor 51 for implementing the steps of the artificial intelligence task processing method mentioned in any of the above-mentioned embodiments when executing the computer-readable instructions .
  • the processor 51 may include one or more processing cores, such as a 4-core processor or an 8-core processor, and the processor 51 may also be a controller, a microcontroller, a microprocessor, or other data processing chips.
  • the processor 51 can adopt at least one hardware form in DSP (Digital Signal Processing, digital signal processing), FPGA (Field-Programmable Gate Array, field programmable gate array), PLA (Programmable Logic Array, programmable logic array) accomplish.
  • Processor 51 may also include a main processor and a coprocessor, the main processor is a processor for processing data in a wake-up state, also called CPU (Central Processing Unit, central processing unit); the coprocessor is Low-power processor for processing data in standby state.
  • the processor 51 may be integrated with a GPU (Graphics Processing Unit, image processor), and the GPU is used for rendering and drawing the content that needs to be displayed on the display screen.
  • the processor 51 may also include an AI (Artificial Intelligence, artificial intelligence) processor, where the AI processor is used to process computing operations related to machine learning.
  • AI Artificial Intelligence, artificial intelligence
  • Memory 50 may include one or more computer-readable storage media, which may be non-transitory.
  • the memory 50 may also include high-speed random access memory and non-volatile memory, such as one or more magnetic disk storage devices and flash memory storage devices.
  • the memory 50 may be an internal storage unit of an electronic device, such as a hard disk of a server.
  • Memory 50 can also be the external storage device of electronic equipment in other embodiments, such as plug-in hard disk equipped on the server, smart memory card (Smart Media Card, SMC), secure digital (Secure Digital, SD) card, flash memory Card (Flash Card), etc.
  • the memory 50 may also include both an internal storage unit of the electronic device and an external storage device.
  • the memory 50 can not only be used to store application software and various data installed in the electronic device, such as: program codes for executing the vulnerability handling method, etc., but also can be used to temporarily store data that has been output or will be output.
  • the memory 50 is at least used to store the following computer-readable instructions 501, wherein, after the computer-readable instructions are loaded and executed by the processor 51, it is possible to realize the relevant aspects of the artificial intelligence task processing method disclosed in any of the foregoing embodiments. step.
  • the resources stored in the memory 50 may also include an operating system 502 and data 503, etc., and the storage method may be temporary storage or permanent storage.
  • the operating system 502 may include Windows, Unix, Linux and so on.
  • Data 503 may include, but is not limited to, data corresponding to artificial intelligence task processing results and the like.
  • the above-mentioned electronic device may further include a display screen 52 , an input/output interface 53 , a communication interface 54 or network interface, a power supply 55 and a communication bus 56 .
  • the display screen 52 and the input/output interface 53 such as a keyboard are user interfaces, and optional user interfaces may also include standard wired interfaces, wireless interfaces, and the like.
  • the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode, organic light-emitting diode) touch device, and the like.
  • a display may also be properly called a display screen or a display unit, and is used for displaying information processed in an electronic device and for displaying a visualized user interface.
  • the communication interface 54 may optionally include a wired interface and/or a wireless interface, such as a WI-FI interface, a Bluetooth interface, etc., and is generally used to establish a communication connection between an electronic device and other electronic devices.
  • the communication bus 56 may be a peripheral component interconnect standard (PCI for short) bus or an extended industry standard architecture (EISA for short) bus or the like.
  • PCI peripheral component interconnect standard
  • EISA extended industry standard architecture
  • FIG. 5 does not constitute a limitation to the electronic device, and may include more or less components than shown in the figure, for example, may also include a sensor 57 for implementing various functions.
  • each functional module of the electronic device in the embodiment of the present application can be specifically implemented according to the method in the above method embodiment, and the specific implementation process can refer to the relevant description of the above method embodiment, and will not be repeated here.
  • the embodiments of the present application can effectively improve the performance of artificial intelligence task processing and reduce the computing resources consumed in the process of artificial intelligence task processing.
  • the artificial intelligence task processing method in the above embodiments is implemented in the form of a software function unit and sold or used as an independent product, it can be stored in a computer-readable storage medium.
  • the technical solution of the present application is essentially or part of the contribution to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium , executing all or part of the steps of the methods in the various embodiments of the present application.
  • the aforementioned storage media include: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), electrically erasable programmable ROM, registers, hard disk, multimedia Cards, card-type memories (such as SD or DX memories, etc.), magnetic memories, removable disks, CD-ROMs, magnetic disks, or optical disks, and other media that can store program codes.
  • the embodiment of the present application also provides one or more non-volatile computer-readable storage media storing computer-readable instructions, which is characterized in that, when the computer-readable instructions are executed by one or more processors, such that One or more processors execute the steps of the artificial intelligence task processing method in any of the above embodiments.
  • each embodiment in this specification is described in a progressive manner, each embodiment focuses on the difference from other embodiments, and the same or similar parts of each embodiment can be referred to each other.
  • the hardware including devices and electronic equipment disclosed in the embodiments since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and for related details, please refer to the description of the methods.

Abstract

一种人工智能任务处理方法、装置、电子设备及可读存储介质。其中,方法包括:获取待处理人工智能任务对应的AI数据集和AI任务执行模型;AI数据集包括AI训练集和AI验证集(S101);根据AI任务执行模型,基于AI训练集在每个滑窗训练过程中的损失值,生成任务损失信息(S102);根据AI任务执行模型,基于AI验证集在每个滑窗训练过程中的正向性能指标的期望值,生成任务精度期望信息(S103);根据任务损失信息、任务精度期望信息和待处理人工智能任务的任务精度需求信息确定是否停止AI任务执行模型的训练,基于训练好的AI任务执行模型执行待处理人工智能任务(S104)。

Description

人工智能任务处理方法、装置、电子设备及可读存储介质
相关申请的交叉引用
本申请要求于2021年12月28日提交中国专利局,申请号为202111616394.X,申请名称为“人工智能任务处理方法、装置、电子设备及可读存储介质”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,特别是涉及一种人工智能任务处理方法、装置、电子设备及可读存储介质。
背景技术
随着人工智能技术的快速发展,机器视觉、自然语言处理作为人工智能的一个分支,也得到相应的发展。简单来说,机器视觉就是用机器代替人眼进行测量和判断,其通过机器视觉产品即图像摄取装置如CMOS(Complementary Metal Oxide Semiconductor,互补金属氧化物半导体)和CCD(Charge-coupled Device,电荷耦合元件)等,将被摄目标转换成图像信号,然后传送给专用图像处理系统,得到被摄目标的形态信息,根据像素分布、亮度、颜色等信息,转变成数字化信号;图像处理系统对这些信号进行各种运算来抽取目标的特征,进而根据判别的结果来控制现场的设备动作。可见,机器视觉在实现过程中很大一部分工作就是图像处理,对图像摄取装置所采集图像的识别准确程度和识别效率,会影响机器视觉性能。自然语言处理是实现人与计算机之间用自然语言进行有效通信的各种理论和方法,通过对文本信息或者是语音信号进行处理得到可使计算机理解的有效信号,对输入的这些文本信息或者是语音信息的识别准确程度和识别效率,很大程度会影响自然语言处理性能。
除此之外,随着大数据+深度学习+超大算力范式的人工智能研究已经深入到社会生活的方方面面,目前的处理系统通常采用深度学习训练神经网络模型来执行人工智能数据的处理任务。深度学习通过预先标注好的数据集,如样本图像数据集或者是自然语言样本集,训练模型中相关神经元的参数信息,来更好地完成特定场景的预测任务,比如图像分类、目标检测、图像分割、语音识别等等。可以理解的是,模型的训练过程是参数不断优化的过程,训练过程直接影响模型性能。模型的训练依赖于初始参数及超参数的设置,不合适的参数会导致整个训练过程出现局部最优、梯度消失或者梯度爆炸、训练缓慢、损失值不收敛等问题,最终导致人工智能任务处理性能较差。在相关技术中,通常通过模型的训练时间和在验证集上泛化错误(Validation Error)之间的权衡来确定模型的早停策略。具体来说,在指定周期内如N个Epoch,一个Epoch就是将所有训练样本训练一次的过程,如果模型在验证集上的误差高于上一次模型在训练集上的训练误差,则停止模型训练。具体实施的时候是设定合适的停止标准。现有技术中的停止标准包括:
1、泛化损失超过一定阈值时停止训练,以避免过拟合;
2、在1的基础上加入度量进展:也即如果泛化损失和度量进展的商高于指定阈值的时候,停止训练;
3、当泛化误差也即验证集上的误差在连续多个周期增长,则停止训练。
发明人意识到,由于相关技术是围绕训练集和验证集的误差也即损失值loss的变化趋势来设定早停策略,会导致模型在新的数据集上的泛化性能较低的问题。此外,验证误差曲线并非光滑单调的曲线,可能会在几次变差的训练之后又持续变好,所以只考虑验证误差来设定早停策略,其准确度会受到一定程度的限制,最终导致所得到的模型性能不佳,并不能很好地执行人工智能处理任务。
发明内容
本发明实施例一方面提供了一种人工智能任务处理方法,包括:
获取待处理人工智能任务对应的AI数据集和AI任务执行模型;
AI数据集包括AI训练集和AI验证集;AI训练集和AI验证集均包括对应一个滑窗的多个连续子集,每个子集对应滑窗的一个窗口;
根据AI任务执行模型,基于AI训练集在每个滑窗训练过程中的损失值,生成任务损失信息;
根据AI任务执行模型,基于AI验证集在每个滑窗训练过程中的正向性能指标的期望值,生成任务精度期望信息;及
根据任务损失信息、任务精度期望信息和待处理人工智能任务的任务精度需求信息确定是否停止AI任务执行模型的训练,并基于训练好的AI任务执行模型执行待处理人工智能任务。
在一个实施例中,根据任务损失信息、任务精度期望信息和待处理人工智能任务的任务精度需求信息确定是否停止AI任务执行模型的训练,包括:
根据任务损失信息和任务精度期望信息确定周期模型精度表示信息;
判断周期模型精度表示信息是否与任务精度需求信息相匹配;
响应于周期模型精度表示信息与任务精度需求信息相匹配,输出停止训练AI任务执行模型的指令;或
响应于周期模型精度表示信息与任务精度需求信息不匹配,输出继续训练AI任务执行模型的指令。
在一个实施例中,根据任务损失信息和任务精度期望信息确定周期模型精度表示信息,包括:
调用周期结果表示计算关系式,计算周期模型精度表示信息;周期结果表示计算关系式为:
S=f ε(S (i+1),S (i+2),....,S (i+ε))
其中,S (i+ε)=M (i+ε)∧N (i+ε),f ε表示映射关系:
{S i+1,S i+2,......,S i+ε}→{e 1,e 2,.....,e ε}
e i∈{0,1};S为周期模型精度表示信息,i为第i个训练周期,ε为训练容忍度;M为任务损失信息,N为任务精度期望信息,S (i+ε)为第i+ε次训练AI任务执行模型的周期结果表示,M (i+ε)为第i+ε次训练AI任务执行模型的任务损失信息,N (i+ε)为第i+ε次训练AI任务执行模型的任务精度期望信息,∧表示逻辑与运算符。
在一个实施例中,判断周期模型精度表示信息是否与任务精度需求信息相匹配,包括:
响应于任务精度需求信息为至少有一次训练AI任务执行模型的周期结果表示不为0,判断周期模型精度表示信息的各元素是否均为0;相应的,响应于周期模型精度表示信息与任务精度需求信息相匹配,输出停止训练AI任务执行模型的指令的过程包括:
响应于周期模型精度表示信息的各元素均为0,输出停止训练AI任务执行模型的指令;及
相应的,响应于周期模型精度表示信息与任务精度需求信息不匹配,输出继续训练AI任务执行模型的指令的过程包括:
响应于周期模型精度表示信息的各元素不均为0,输出继续训练AI任务执行模型的指令。
在一个实施例中,根据AI任务执行模型,基于AI训练集在每个滑窗训练过程中的损失值,生成任务损失信息,包括:
基于AI训练集,对每个训练周期,根据每个滑窗的AI任务执行模型的损失值,计算当前训练周期的所有损失值的当前标准差;
根据损失变化程度因子和当前训练周期的前一个训练周期的所有损失值的前向标准差,确定损失变化阈值;及
根据当前标准差、前向标准差和损失变化阈值确定当前训练周期的任务损失信息。
在一个实施例中,根据当前标准差、前向标准差和损失变化阈值确定当前训练周期的任务损失信息,包括:
调用周期任务损失信息计算关系式计算当前训练周期的任务损失信息,周期任务损失信息计算关系式为:
响应于σ(loss i1,loss i2,...,loss ik)-σ(loss (i-1)1,loss (i-2)2,...,loss (i-1)k)≤η,确定M i=1;响应于σ(loss i1,loss i2,...,loss ik)-σ(loss (i-1)1,loss (i-2)2,...,loss (i-1)k)>η,确定M i=0
式中,M i为第i个训练周期的任务损失信息,loss ik为第i个训练周期的第k个滑窗的损失值,loss ik为第i-1个训练周期的第k个滑窗的损失值,σ(lossi1,lossi2,…,lossik)为第i个训练周期的所有损失值的当前标准差,σ(loss i1,loss i2,...,loss ik)为第i-1个训练周期的所有损失值的前向标准差,η为损失变化阈值。
在一个实施例中,根据AI任务执行模型,基于AI验证集在每个滑窗训练过程中的正向性能指标的期望值,生成任务精度期望信息,包括:
基于AI验证集,对每个训练周期,根据每个滑窗的AI任务执行模型的正向性能指标,计算当前训练周期的所有正向性能指标的当前期望值;
根据性能变化程度因子和当前训练周期的前一个训练周期的所有正向性能指标的前向期望值,确定性能变化阈值;及
根据当前期望值、前向期望值和性能变化阈值确定当前训练周期的任务精度期望信息。
在一个实施例中,根据当前期望值、前向期望值和性能变化阈值确定当前训练周期的任务精度期望信息,包括:
调用周期任务精度计算关系式计算当前训练周期的任务精度期望信息,周期任务精度计算关系式为:
响应于E(perf i1,perf i2,....,perf ik)-E(perf (i-1)1,perf (i-1)2,....,perf (i-1)k)>μ,确定N i=1;响应于E(perf i1,perf i2,....,perf ik)-E(perf (i-1)1,perf (i-1)2,....,perf (i-1)k)≤μ,确定N i=0;
响应于σ(loss i1,loss i2,...,loss ik)-σ(loss (i-1)1,loss (i-2)2,...,loss (i-1)k)≤η,确定M i=1;响应于σ(loss i1,loss i2,...,loss ik)-σ(loss (i-1)1,loss (i-2)2,...,loss (i-1)k)>η 确定M i=0
式中,N i为第i个训练周期的任务精度期望信息,perf ik为第i个训练周期的第k个滑窗的期望值,perf (i-1)k为第i-1个训练周期的第k个滑窗的期望值,E(perf i1,perf i2,....,perf ik)为第i个训练周期 的所有正向性能指标的当前期望值,E(perf (i-1)1,perf (i-1)2,....,perf (i-1)k)为第i-1个训练周期的所有正向性能指标的前向期望值,μ为性能变化阈值。
在一个实施例中,获取待处理人工智能任务对应的AI数据集和AI任务执行模型,包括:
当接收到信息输入指令,展示信息处理交互界面;信息处理交互界面包括信息输入区域和结果展示区域;及
响应用户通过信息输入区域下发的信息输入指令,从信息输入指令中获取待处理人工智能任务对应的AI数据集、AI任务执行模型、待处理人工智能任务的任务精度需求信息和滑窗参数值;其中,结果展示区域用于展示训练好的AI任务执行模型和/或待处理人工智能任务的任务执行结果。
本申请实施例另一方面提供了一种人工智能任务处理装置,包括:
信息获取模块,用于获取待处理人工智能任务对应的AI数据集和AI任务执行模型;AI数据集包括AI训练集和AI验证集;AI训练集和AI验证集均包括对应一个滑窗的多个连续子集,每个子集对应滑窗的一个窗口;
损失计算模块,用于根据AI任务执行模型,基于AI训练集在每个滑窗训练过程中的损失值,生成任务损失信息;
期望计算模块,用于根据AI任务执行模型,基于AI验证集在每个滑窗训练过程中的正向性能指标的期望值,生成任务精度期望信息;及
模型训练结束确定模块,用于根据任务损失信息、任务精度期望信息和待处理人工智能任务的任务精度需求信息确定是否停止AI任务执行模型的训练,并基于训练好的AI任务执行模型执行待处理人工智能任务。
本申请实施例还提供了一种电子设备,包括存储器及一个或多个处理器,存储器中储存有计算机可读指令,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执行前述任一实施例中人工智能任务处理方法的步骤。
本申请实施例最后还提供了一个或多个存储有计算机可读指令的非易失性计算机可读存储介质,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执行如上述任一实施例中人工智能任务处理方法的步骤。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。本申请的其它特征和优点将从说明书、附图以及权利要求书变得明显。
附图说明
为了更清楚的说明本申请实施例或相关技术的技术方案,下面将对实施例或相关技术描述中所需要使用的附图作简单的介绍,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的前提下,还可以根据这些附图获得其他的附图。
图1为本申请一个或多个实施例中提供的一种人工智能任务处理方法的流程示意图;
图2为本申请一个或多个实施例中提供的一个示意性例子中的人工智能任务处理方法的流程示意图;
图3为本申请一个或多个实施例中提供的信息处理系统的架构示意图;
图4为本申请一个或多个实施例中提供的人工智能任务处理装置的一种具体实施方式结构图;
图5为本申请一个或多个实施例中提供的电子设备的一种具体实施方式结构图。
具体实施方式
为了使本技术领域的人员更好地理解本申请方案,下面结合附图和具体实施方式对本申请作进一步的详 细说明。显然,所描述的实施例仅仅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做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。
本申请的说明书和权利要求书及上述附图中的术语“第一”、“第二”、“第三”、“第四”等是用于区别不同的对象,而不是用于描述特定的顺序。此外术语“包括”和“具有”以及他们任何变形,意图在于覆盖不排他的包含。例如包含了一系列步骤或单元的过程、方法、系统、产品或设备没有限定于已列出的步骤或单元,而是可包括没有列出的步骤或单元。
在介绍了本申请实施例的技术方案后,下面详细的说明本申请的各种非限制性实施方式。
首先参见图1,图1为本申请实施例提供的一种人工智能任务处理方法的流程示意图,本申请实施例可包括以下内容:
S101:获取待处理人工智能任务对应的AI数据集和AI任务执行模型。
在本实施例中,待处理人工智能任务可为任何一种人工智能领域中需要执行的任务,例如图像分类识别任务、目标检测任务、图像分割任务、语音识别任务等等。AI任务执行模型为基于深度学习模型构建的网络模型,例如人工神经网络模型、卷积神经网络模型、递归张量神经网络、生成对抗网络、长短期记忆网络等等。AI任务执行模型是用于执行待处理人工智能任务,举例来说,待处理人工智能任务为图像分割任务,AI任务执行模型则是用于对输入图像进行图像分割。AI数据集是指AI任务执行模型训练过程中所使用的样本数据集,AI数据集例如可为图像数据集、语音信号数据集、文本信息数据集。AI数据集包括AI训练集和AI验证集;AI训练集用于训练AI任务执行模型,AI验证集用于验证AI任务执行模型的性能。当然,本实施例的AI训练集还包括AI测试集。AI训练集和AI验证集均被划分为多个连续子集,AI训练集对应一个滑窗,AI验证集对应一个滑窗,每个子集均为滑窗的一个窗口。在AI训练集和AI验证集时,从第一个子集对应的第一窗口一直滑动至最后一个子集对应的最后一个窗口,完成一次滑窗的模型训练。举例来说,AI训练集包括1000张图像,可将前100张图像作为第一子集,第100张图像至第500张图像作为第二子集,第500张图像至第700张图像作为第三子集,第700张图像至第900张图像作为第四子集,第900张图像至第1000张图像作为第五子集,第一子集、第二子集、第三子集、第四子集和第五子集均为AI训练集的滑窗的一个窗口,训练过程中,滑窗会依次从第一子集滑到第五子集来读取当前子集中的样本图像。在本步骤中,在训练数据集即AI训练集和AI验证集上训练k次,也即k个epoch,也叫一个训练周期,每一次训练称为一个滑窗,也即一个训练周期中包括k个epoch或者是k个滑窗。其中k≥1,一个epoch表示模型在训练集上训练一次。一个训练周期指AI任务执行模型连续完成k次指定训练数据集上的训练。
S102:根据AI任务执行模型,基于AI训练集在每个滑窗训练过程中的损失值,生成任务损失信息。
在本步骤中,损失值是指AI任务执行模型每次迭代的前向计算结果与真实值的差距,本实施例在使用AI训练集对AI任务执行模型进行多个训练周期的训练过程中,每个训练周期会在AI训练集训练多次,而每一次即为一个滑窗,可使用现有技术中的任何一种损失函数计算每个滑窗下基于AI训练集训练AI任务执行模型的损失值,根据每个滑窗下的损失值确定AI任务执行模型在一个训练周期或多个训练周期下的任务损失信息,任务损失信息用于反映当前训练周期或者是当前训练实际段内AI任务执行模型的误差。
S103:根据AI任务执行模型,基于AI验证集在每个滑窗训练过程中的正向性能指标的期望值,生成任务精度期望信息。
在本实施例中,正向性能指标用于评价在一个训练周期或多个训练周期结束后所得的AI任务执行模型的性能,正向性能指标的数值越大,表示AI任务执行模型的性能越好。比如在分类和检测任务中,可以分别使用“精度(precision)”和“各类别平均精度的均值(map,mean average precision)”值来衡量AI任务执行模型的分类性能或目标检测性能。本实施例在使用AI验证集对AI任务执行模型在各训练周期结束之后,可对其进行性能验证,每个训练周期会在AI验证集验证多次,而每一次即为一个滑窗,可使用现有技术中的任何一种期望计算方法计算每个滑窗下基于AI验证集验证AI任务执行模型的正向性能指标的期望 值,根据每个滑窗下的正向性能指标的期望值确定AI任务执行模型在一个训练周期或多个训练周期下的任务精度期望信息,任务精度期望信息用于反映当前训练周期或者是当前训练实际段内AI任务执行模型的性能的优劣。
S104:根据任务损失信息、任务精度期望信息和待处理人工智能任务的任务精度需求信息确定是否停止AI任务执行模型的训练,并基于训练好的AI任务执行模型执行待处理人工智能任务。
在本实施例中,任务精度需求信息是基于实际应用场景所需求的待处理人工智能任务执行的精准度,举例来说,对于执行图像识别和分割任务来说,对于病灶识别和病灶切割来说,需要识别病灶并切割病灶的精准度需大于99%,而对于车辆识别任务来说,需要识别车辆的精准度只需90%即可。根据任务损失信息、任务精度期望信息和待处理人工智能任务的任务精度需求信息判断当前训练周期所训练的AI任务执行模型是否满足要求,如果不满足,则需要对AI任务执行模型继续进行训练。如果满足,则停止对AI任务执行模型的训练,此时所得的AI任务执行模型即为训练好的模型,可使用该AI任务执行模型去执行待处理人工智能任务。
在本申请实施例中提供的技术方案中,综合考虑AI任务执行模型在AI训练集的损失值变化趋势以及在AI验证集上的正向性能指标在多个训练周期的性能变化趋势,能够提前反映AI任务执行模型是否过拟合、模型的泛化能力以及数据集的分布情况,根据模型在训练过程中的性能趋势来控制AI任务执行模型的训练是继续还是结束,不仅保证AI任务执行模型更好的学习数据特征,提高找到更优模型的概率,而且在新的数据集上有更高的泛化性能,可以有效地提升人工智能任务处理性能。在合适的时候停止训练任务,避免无效的训练过程,可释放出更多的硬件资源,有效降低人工智能任务处理中消耗的计算资源。
在上述实施例中,对于如何执行步骤S104并不做限定,本实施例中给出AI任务执行模型是否早停的一种判断方式,可包括如下步骤:
根据任务损失信息和任务精度期望信息确定周期模型精度表示信息;判断周期模型精度表示信息是否与任务精度需求信息相匹配;响应于周期模型精度表示信息与任务精度需求信息相匹配,输出停止训练AI任务执行模型的指令;响应于周期模型精度表示信息与任务精度需求信息不匹配,输出继续训练AI任务执行模型的指令。
在本实施例中,周期模型精度表示信息用于表示当前训练周期结束之后,AI任务执行模型的性能或者是在任务执行精度。作为一种可选的实施方式,周期模型精度表示信息可通过调用周期结果表示计算关系式计算得到,周期结果表示计算关系式可表示为:
S=f ε(S (i+1),S (i+2),....,S (i+ε))
其中,S (i+ε)=M (i+ε)∧N (i+ε),f ε表示映射关系:
{S i+1,S i+2,......,S i+ε}→{e 1,e 2,.....,e ε}e i∈{0,1};
S为周期模型精度表示信息,i为第i个训练周期,ε为训练容忍度;M为任务损失信息,N为任务精度期望信息,S (i+ε)为第i+ε次训练AI任务执行模型的周期结果表示,M (i+ε)为第i+ε次训练AI任务执行模型的任务损失信息,N (i+ε)为第i+ε次训练AI任务执行模型的任务精度期望信息,∧表示逻辑与运算符。
其中,逻辑与运算符通常表示为P=A 1∧A 2∧...∧A n,表示当所有的A i(1≤i≤n)都为真时,逻辑表达式P才为真,否则为假。在本实施例中,“真”用数字1来代替,“假”使用0来代替。训练容忍 度ε表示AI任务执行模型的早停策略基于ε个训练周期的结果来做的决策。早停策略是指判断AI任务执行模型的训练是否结束的标准。基于周期结果表示计算关系式可知,周期模型精度表示信息为一个集合,集合中的各元素的取值为0或1,也就是说,周期模型精度表示信息为0和1组成的集合。如果元素为0,则表示任务损失信息和任务精度期望信息至少有一个为0,而任务损失信息为0,表示AI任务执行模型在AI训练集上的第i个训练周期的模型收敛性能不满足要求,任务精度期望信息为0,表示AI任务执行模型在AI验证集上的第i个训练周期的模型泛化能力不满足要求。
为了获取最优模型参数,得到性能较佳的AI任务执行模型,作为一种可选的实施方式,判断周期模型精度表示信息是否与任务精度需求信息相匹配的过程可包括:
任务精度需求信息为至少有一次训练AI任务执行模型的周期结果表示不为0,判断周期模型精度表示信息的各元素是否均为0;
响应于周期模型精度表示信息的各元素均为0,输出停止训练AI任务执行模型的指令;
响应于周期模型精度表示信息的各元素不均为0,输出继续训练AI任务执行模型的指令。
上述实施例对于如何执行步骤S102和S103并不做限定,本实施例还给出任务损失信息的一种计算方式,也即根据AI任务执行模型基于AI训练集在每个滑窗训练过程中的损失值生成任务损失信息的过程可包括下述内容:
基于AI训练集,对每个训练周期,根据每个滑窗的AI任务执行模型的损失值,计算当前训练周期的所有损失值的当前标准差;
根据损失变化程度因子和当前训练周期的前一个训练周期的所有损失值的前向标准差,确定损失变化阈值;
根据当前标准差、前向标准差和损失变化阈值确定当前训练周期的任务损失信息。
其中,任务损失信息可通过调用周期任务损失信息计算关系式计算得到,周期任务损失信息计算关系式可表示为:
响应于σ(loss i1,loss i2,...,loss ik)-σ(loss (i-1)1,loss (i-2)2,...,loss (i-1)k)≤η,确定M i=1;响应于σ(loss i1,loss i2,...,loss ik)-σ(loss (i-1)1,loss (i-2)2,...,loss (i-1)k)>η,确定M i=0;
式中,M i为第i个训练周期的任务损失信息,loss ik为第i个训练周期的第k个滑窗的损失值,loss ik为第i-1个训练周期的第k个滑窗的损失值,σ(loss i1,loss i2,...,loss ik)为第i个训练周期的所有损失值的当前标准差,σ(loss i1,loss i2,...,loss ik)为第i-1个训练周期的所有损失值的前向标准差,η为损失变化阈值。η表示的是损失值的标准差的变化阈值,可通过计算关系式η=a*σ(loss (i-1)1,loss (i-1)2,....,loss (i-1)k)计算得到,该计算 x系式可缩写记为η=a*σ(loss (i-1)[1,k]);a为损失变化程度因子,其取值范围可为(0,+∞),a的大小可以刻画早停策略可以接受的损失值loss的变化程度的高低。a值越大,可接受的loss的变化程度越高,a值越小,可接受的loss的变化程度越低。
本实施例还给出任务精度期望信息的一种计算方式,也即根据AI任务执行模型基于AI验证集在每个滑窗训练过程中的正向性能指标的期望值生成任务精度期望信息的过程可包括下述内容:
基于AI验证集,对每个训练周期,根据每个滑窗的AI任务执行模型的正向性能指标,计算当前训练周期的所有正向性能指标的当前期望值;
根据性能变化程度因子和当前训练周期的前一个训练周期的所有正向性能指标的前向期望值,确定性能变化阈值;
根据当前期望值、前向期望值和性能变化阈值确定当前训练周期的任务精度期望信息。
其中,任务精度期望信息可通过调用周期任务精度计算关系式计算得到,周期任务精度计算关系式可表示为:
响应于E(perf i1,perf i2,....,perf ik)-E(perf (i-1)1,perf (i-1)2,....,perf (i-1)k)>μ,确定N i=1;响应于E(perf i1,perf i2,....,perf ik)-E(perf (i-1)1,perf (i-1)2,....,perf (i-1)k)≤μ,确定N i=0;
式中,N i为第i个训练周期的任务精度期望信息,perf ik为第i个训练周期的第k个滑窗的期望值,perf (i-1)k为第i-1个训练周期的第k个滑窗的期望值,E(perf i1,perf i2,....,perf ik)为第i个训练周期的所有正向性能指标的当前期望值,E(perf (i-1)1,perf (i-1)2,....,perf (i-1)k)为第i-1个训练周期的所有正向性能指标的前向期望值,μ为性能变化阈值。μ表示性能的变化阈值,该值可通过关系式μ=b*E(perf (i-1)1,perf (i-1)2,...,perf (i-1)k)计算得到,其可缩写为μ=b*E(perf (i-1)[1,k])。b为性能变化程度因子,其取值范围可为(0,+∞),b的大小可以刻画本实施例的早停策略可以接受的性能的提升程度的高低。b的值越大,可接受的性能提升程度越高,b的值越小,可接受的性能提升程度越低。
在上述实施例中,对于AI训练集和AI验证集,AI任务执行模型在第i个训练周期的输出记为S i,训练周期包含在训练集上执行k次训练过程。具体定义为如下逻辑表达式:
S i=M i∧N i,i≥2;
响应于σ(loss i1,loss i2,...,loss ik)-σ(loss (i-1)1,loss (i-2)2,...,loss (i-1)k)≤η,确定M i=1,响应于σ(loss i1,loss i2,...,loss ik)-σ(loss (i-1)1,loss (i-2)2,...,loss (i-1)k)>η,确定M i=0;
响应于E(perf i1,perf i2,....,perf ik)-E(perf (i-1)1,perf (i-1)2,....,perf (i-1)k)>μ,确定N i=1,响应于E(perf i1,perf i2,....,perf ik)-E(perf (i-1)1,perf (i-1)2,....,perf (i-1)k)≤μ,确定N i=0;
其中,σ(loss i1,loss i2,...,loss ik)为第i个训练周期的所有损失值loss的当前标准差,其中包含k个训练轮数,即epoch数量,这里可以缩写为σ(loss i[1,k]),表示第i个训练周期的1到k个epoch。E(perf i1,perf i2,....,perf ik)表示在第i个训练周期内AI任务执行模型在AI验证集上正向性能指标perf的当前期望值,这里可以缩写为E(perf i[1,k])。其中包含k个训练轮数,即epoch数量。perf具体可以是精度acc(accuracy,准确率)、map(mean average precision,均值平均精度)值,或者其他正向性能指标。根据计算关系式(1)、(2)和(3)可知:S i的取值为0或者1。根据S i的定义可得到周期模型精度表示信息S的形式化定义:
S=f ε(S (i+1),S (i+2),....,S (i+ε))
其中,ε>1为待处理人工智能任务的训练容忍度,为一个大于等于1的整数,表示连续训练ε个周期。f ε定义为一个映射关系:{S i+1,S i+2,......,S i+ε}→{e 1,e 2,.....,e ε}e i∈{0,1}。S的输出是一个由连续ε个训练周期的输出组成的集合(S i+1,S i+2,....,S i+ε)根据计算关系式(1)得知集合中的每个元素取值为0或者1。基于上述内容可得出本申请实施例的早停策略的评估标准:给定容忍度参数ε,如果周期模型精度表示信息S的输出集合中元素均为0,那么就需要停止该训练任务。根据公式S i=M i∧N i可知S i的结果对应以下四种情况,下面给出具体的解释:
(a):(S i=0|M i=0,N i=0):M i和N i同时为0,表示AI任务执行模型在AI训练集上第i个训练周期的损失值和第(i-1)个训练周期的损失波动差值已经超过了阈值η,说明AI任务执行模型在AI训练集的训练损失值已经出现了震荡,收敛的概率很低。同时AI任务执行模型在AI验证集上的精度出现了下降,这也说明AI任务执行模型在损失值震荡的情况下,在AI验证集上已经无法取得很好的泛化能力,因此需要考虑停止训练任务。
(b):(S i=0|M i=0,N i=1):M i=0同理(a)中的分析,N i=1的时候表示AI任务执行模型在AI验证集上的精度上升幅度超过阈值μ,一般这种情况出现的概率比较低。出现的原因之一很可能是AI训练集和AI验证集上的数据分布不一致,或者是训练的数据规模太小,AI任务执行模型无法学到合适的特征。精度的上升具有偶然性,往往在后续训练中精度急速下降。因此需要考虑停止训练任务,进一步分析数据的情况。
(c):(S i=0|M i=1,N i=0):M i=1表明loss震荡的频率降低,AI任务执行模型训练趋势良好。N i=0表示AI任务执行模型在对应AI验证集上的精度下降幅度超过阈值μ。这种情况往往是AI任务执 行模型在训练过程出现过拟合,AI任务执行模型在AI验证集上无法取得和训练集上一样的好的表现,泛化性能降低。因此也需要考虑停止任务。
(d):(S i=1|M i=1,N i=1):M i、N i同时为1的时候,表示AI任务执行模型训练过程的趋势良好,同时在AI验证集上的泛化性能表现良好,这个时候不需要停止任务。
由上可知,结合上述四种情况和训练容忍度参数ε,本实施例可以帮助用户根据自己的实际需要对训练任务做出是否早停的判断。
为了使得本实施例所提供的技术方案实用性更强,提高用户使用体验,基于上述实施例,还可包括下述内容:
当接收到信息输入指令,展示信息处理交互界面(即响应于接收到信息输入指令,展示信息处理交互界面);
响应用户通过信息输入区域下发的信息输入指令,从信息输入指令中获取待处理人工智能任务对应的AI数据集、AI任务执行模型、待处理人工智能任务的任务精度需求信息和滑窗参数值;
其中,信息处理交互界面包括信息输入区域和结果展示区域;结果展示区域用于展示训练好的AI任务执行模型和/或待处理人工智能任务的任务执行结果。
在本实施例中,信息输入指令为用户下发的指令,用户想要执行待处理人工智能任务时,向系统下发信息输入指令,将执行待处理人工智能任务所需的AI任务执行模型和对应的AI数据集进行输入,初始化AI任务执行模型的相关参数和任务精度需求信息的相关参数,其中包括AI任务执行模型的模型参数和超参数两种类型。模型参数包括各个神经元的权重和偏置矩阵、各级卷积层的初始卷积核、全连接层的权重矩阵和偏置向量。超参数包括学习率、动量值、优化器、训练次数(epoch数)、批训练数据大小(batch-size)等等相关参数。任务精度需求信息的相关参数可包括系统容忍度ε、滑窗大小(k)参数的定义。初始化参数之后,调用上述任何任意一个实施例人工智能任务处理方法的步骤实现对待处理人工智能任务的执行,最后将训练好的AI任务执行模型和任务执行结果输出。
为了使所属领域技术人员更加清楚明白本申请的技术方案,本申请还结合图2和图3提供了一个示意性例子,本实施例提供了基于滑窗的早停策略,并将这个早停策略定义为一个信息处理系统S,该系统可以应用在任何领域的深度学习模型训练过程。对于一个给定场景的训练任务,系统S首先会获取要训练的模型M也即AI任务执行模型、AI数据集(包括AI训练集、AI验证集);其次在指定滑窗下,评估模型在ε个训练周期的输出。最后综合考虑AI训练集的误差值和AI验证集的正向性能指标,来决定本次训练任务是否该停止,AI任务执行模型可为基于Resnet50的网络模型,待处理人工智能任务为图像识别任务,正向性能指标相应的可为性能精度acc,可包括下述内容:
假设AI任务执行模型为M(Resnet50),AI训练集为train_set,AI验证集为valid_set。初始化AI任务执行模型的相关参数,其中包括AI任务执行模型的模型参数和超参数两种类型。模型参数包括各个神经元的权重和偏置矩阵、各级卷积层的初始卷积核、全连接层的权重矩阵和偏置向量。超参数包括学习率、动量值、优化器、训练次数(epoch数)、批训练数据大小(batch-size)等等相关参数。初始化信息系统S的相关参数,包括系统容忍度ε、滑窗大小(k)参数的定义。例如,滑窗大小记为k=10,训练总的epoch数记为N=100,训练集loss变化阈值中的系数a=0.1,验证集的性能变化阈值中的系数b=0.2,训练容忍度参数ε=2。基于容忍度参数ε,重复下述流程ε次,获得信息处理系统S的输出结果:也即长度为ε的输出集合。根据信息处理系统S定义的早停策略评价标准,评估训练任务是否早停;重复下述过程,直至训练过程结束,获得性能好、泛化能力高的模型。
A1:计算第一个滑窗(也即epoch从1到10)训练过程的loss的标准差及在AI验证集上精度的期望值:
<1>
Figure PCTCN2022100481-appb-000001
为第一个滑窗10个loss 值的期望值。
<2>
Figure PCTCN2022100481-appb-000002
为第一个滑窗10个loss值的标准差。
<3>
Figure PCTCN2022100481-appb-000003
为第一个滑窗10个精度值acc的期望值。
A2:计算第二个滑窗(epoch从11到20)训练过程的loss的标准差及在AI验证集上精度的期望值:
<1>
Figure PCTCN2022100481-appb-000004
为第二个滑窗10个loss值的期望值。
<2>
Figure PCTCN2022100481-appb-000005
为第一个滑窗10个loss值的标准差。
<3>
Figure PCTCN2022100481-appb-000006
为第二个滑窗10个精度值acc期望值。
A3:根据下述计算关系式计算AI训练集上loss的变化阈值:
η=a*σ(loss 1[1,10])=0.1*σ(loss 1[1,10])
A4:根据下述计算关系式计算AI验证集上精度的变化阈值:
μ=b*E(acc 1[1,10])=0.2*E(acc 1[1,10])
A5:计算差值(σ(loss 1[2,10])-σ(loss 1[1,10])与η的大小,得到M 2的值,M 2∈{0,1}。
A6:计算差值(E(acc 2[1,10])-E(acc 1[1,10])与μ的大小,得到N 2的值,N 2∈{0,1}。
A7:根据M 2∈{0,1}和N 2的值得到S 2的值,S 2=M 2∧N 2
A8:按照A 2-A 8的步骤,可以同理求出第三个滑窗与第二个滑窗对应S 3的值。
A9:由S 2和S 3,可以得到信息处理系统S的输出集合:{S 2,S 3}。如果集合中元素均为0,则可以停止这个训练任务。否则继续训练,并且同样的步骤判断{S 3,S 4}……的结果,来决定是否需要早停。重复以上计算过程,基于训练容忍度参数ε,可以得到信息处理系统S的ε个输入(S 1,S 2,S 3,……, S ε)及对应的长度为ε的输出集合{e 1,e 2,e 3,……,e ε}。基于信息处理系统S提出的早停策略,给出这个训练任务是否早停的判断。如果输出集合{e 1,e 2,e 3,……,e ε}中元素均为0,则执行早停操作;否则,继续训练,并且在下一个训练周期进行再次判断,直至AI任务执行模型训练过程结束。
本技术方案提供的信息处理系统S,用来控制AI任务执行模型的训练过程。其基于滑窗的模型训练过程早停技术,不仅可以应用在传统学习模型训练过程中,而且可以应用于自动超参调优的训练过程,可以提前终止性能趋势下降的某组超参数训练过程,释放硬件资源,为更优的超参数组合搜索提供更多机会。对于一些特定问题的分析,比如分类,图像识别,目标检测,自然语言处理等,在进行模型训练时也可以采取上述实施例的思路,在训练过程中使用早停策略,使得最终训练的模型性能更优,泛化性更好。当然,用户可以根据自己的需求,对本实施例的早停策略相关参数因子进行限制。如果数据量足够大,也可以将这些参数因子作为超参数进行训练,从而获得合适的早停策略标准。
需要说明的是,本申请中各步骤之间没有严格的先后执行顺序,只要符合逻辑上的顺序,则这些步骤可以同时执行,也可按照某种预设顺序执行,图1-图2只是一种示意方式,并不代表只能是这样的执行顺序。
本申请实施例还针对人工智能任务处理方法提供了相应的装置,进一步使得方法更具有实用性。其中,装置可从功能模块的角度和硬件的角度分别说明。下面对本申请实施例提供的人工智能任务处理装置进行介绍,下文描述的人工智能任务处理装置与上文描述的人工智能任务处理方法可相互对应参照。
基于功能模块的角度,参见图4,图4为本申请实施例提供的人工智能任务处理装置在一种具体实施方式下的结构图,该装置可包括:
信息获取模块401,用于获取待处理人工智能任务对应的AI数据集和AI任务执行模型;AI数据集包括AI训练集和AI验证集;AI训练集和AI验证集均包括对应一个滑窗的多个连续子集,每个子集对应滑窗的一个窗口。
损失计算模块402,用于根据AI任务执行模型,基于AI训练集在每个滑窗训练过程中的损失值,生成任务损失信息。
期望计算模块403,用于根据AI任务执行模型,基于AI验证集在每个滑窗训练过程中的正向性能指标的期望值,生成任务精度期望信息。
模型训练结束确定模块404,用于根据任务损失信息、任务精度期望信息和待处理人工智能任务的任务精度需求信息确定是否停止AI任务执行模型的训练,并基于训练好的AI任务执行模型执行待处理人工智能任务。
可选的,在本实施例的一些实施方式中,上述模型训练结束确定模块404可进一步用于:
根据任务损失信息和任务精度期望信息确定周期模型精度表示信息;判断周期模型精度表示信息是否与任务精度需求信息相匹配;
响应于周期模型精度表示信息与任务精度需求信息相匹配,输出停止训练AI任务执行模型的指令;响应于周期模型精度表示信息与任务精度需求信息不匹配,输出继续训练AI任务执行模型的指令。
作为上述实施例的一种可选的实施方式,上述模型训练结束确定模块404还可进一步用于:调用周期结果表示计算关系式,计算周期模型精度表示信息;及
周期结果表示计算关系式为:
S=f ε(S (i+1),S (i+2),....,S (i+ε))
其中,S (i+ε)=M (i+ε)∧N (i+ε),f ε表示映射关系:
{S i+1,S i+2,......,S i+ε}→{e 1,e 2,.....,e ε}
e i∈{0,1};S为周期模型精度表示信息,i为第i个训练周期,ε为训练容忍度;M为任务损失信息,N为任务精度期望信息,S (i+ε)为第i+ε次训练AI任务执行模型的周期结果表示,M (i+ε)为第i+ε次训练AI任务执行模型的任务损失信息,N (i+ε)为第i+ε次训练AI任务执行模型的任务精度期望信息,∧表示逻辑与运算符。
作为上述实施例的另一种可选的实施方式,上述模型训练结束确定模块404还可进一步用于:
任务精度需求信息为至少有一次训练AI任务执行模型的周期结果表示不为0;
判断周期模型精度表示信息的各元素是否均为0;
响应于周期模型精度表示信息的各元素均为0,输出停止训练AI任务执行模型的指令;
响应于周期模型精度表示信息的各元素不均为0,输出继续训练AI任务执行模型的指令。
可选的,在本实施例的另一些实施方式中,上述损失计算模块402可进一步用于:
基于AI训练集,对每个训练周期,根据每个滑窗的AI任务执行模型的损失值,计算当前训练周期的所有损失值的当前标准差;
根据损失变化程度因子和当前训练周期的前一个训练周期的所有损失值的前向标准差,确定损失变化阈值;
根据当前标准差、前向标准差和损失变化阈值确定当前训练周期的任务损失信息。
作为上述本实施例的一种可选的实施方式中,上述损失计算模块402还可进一步用于:
调用周期任务损失信息计算关系式计算当前训练周期的任务损失信息,周期任务损失信息计算关系式为:
响应于σ(loss i1,loss i2,...,loss ik)-σ(loss (i-1)1,loss (i-2)2,...,loss (i-1)k)≤η,确定M i=1,响应于σ(loss i1,loss i2,...,loss ik)-σ(loss (i-1)1,loss (i-2)2,...,loss (i-1)k)>η,确定M i=0;
式中,M i为第i个训练周期的任务损失信息,loss ik为第i个训练周期的第k个滑窗的损失值,loss ik为第i-1个训练周期的第k个滑窗的损失值,σ(lossi1,lossi2,…,lossik)为第i个训练周期的所有损失值的当前标准差,σ(loss i1,loss i2,...,loss ik)为第i-1个训练周期的所有损失值的前向标准差,η为损失变化阈值。
可选的,在本实施例的其他一些实施方式中,上述期望计算模块403可进一步用于:
基于AI验证集,对每个训练周期,根据每个滑窗的AI任务执行模型的正向性能指标,计算当前训练周期的所有正向性能指标的当前期望值;
根据性能变化程度因子和当前训练周期的前一个训练周期的所有正向性能指标的前向期望值,确定性能变化阈值;
根据当前期望值、前向期望值和性能变化阈值确定当前训练周期的任务精度期望信息。
作为上述实施例的一种可选的实施方式,上述期望计算模块403还可进一步用于:
调用周期任务精度计算关系式计算当前训练周期的任务精度期望信息,周期任务精度计算关系式为:
响应于E(perf i1,perf i2,....,perf ik)-E(perf (i-1)1,perf (i-1)2,....,perf (i-1)k)>μ,确定N i=1;响应于E(perf i1,perf i2,....,perf ik)-E(perf (i-1)1,perf (i-1)2,....,perf (i-1)k)≤μ,确定N i=0;
式中,N i为第i个训练周期的任务精度期望信息,perf ik为第i个训练周期的第k个滑窗的期望值,perf (i-1)k为第i-1个训练周期的第k个滑窗的期望值,E(perf i1,perf i2,....,perf ik)为第i个训练周期的所有正向性能指标的当前期望值,E(perf (i-1)1,perf (i-1)2,....,perf (i-1)k)为第i-1个训练周期的所有正向性能指标的前向期望值,μ为性能变化阈值。
可选的,在本实施例的其他一些实施方式中,上述信息获取模块401还可进一步用于:
当接收到信息输入指令,展示信息处理交互界面;
信息处理交互界面包括信息输入区域和结果展示区域;
响应用户通过信息输入区域下发的信息输入指令,从信息输入指令中获取待处理人工智能任务对应的AI数据集、AI任务执行模型、待处理人工智能任务的任务精度需求信息和滑窗参数值;
其中,结果展示区域用于展示训练好的AI任务执行模型和/或待处理人工智能任务的任务执行结果。
本申请实施例人工智能任务处理装置的各功能模块的功能可根据上述方法实施例中的方法具体实现,其具体实现过程可以参照上述方法实施例的相关描述,此处不再赘述。
由上可知,本申请实施例可有效提升人工智能任务处理性能,降低人工智能任务处理过程中消耗的计算资源。
上文中提到的人工智能任务处理装置是从功能模块的角度描述,进一步的,本申请还提供一种电子设备,是从硬件角度描述。图5为本申请实施例提供的电子设备在一种实施方式下的结构示意图。如图5所示,该电子设备包括存储器50,用于存储计算机可读指令;处理器51,用于执行计算机可读指令时实现如上述任一实施例提到的人工智能任务处理方法的步骤。
其中,处理器51可以包括一个或多个处理核心,比如4核心处理器、8核心处理器,处理器51还可为控制器、微控制器、微处理器或其他数据处理芯片等。处理器51可以采用DSP(Digital Signal Processing,数字信号处理)、FPGA(Field-Programmable Gate Array,现场可编程门阵列)、PLA(Programmable Logic Array,可编程逻辑阵列)中的至少一种硬件形式来实现。处理器51也可以包括主处理器和协处理器,主处理器是用于对在唤醒状态下的数据进行处理的处理器,也称CPU(Central Processing Unit,中央处理器);协处理器是用于对在待机状态下的数据进行处理的低功耗处理器。在一些实施例中,处理器51可以集成有GPU(Graphics Processing Unit,图像处理器),GPU用于负责显示屏所需要显示的内容的渲染和绘制。一些实施例中,处理器51还可以包括AI(Artificial Intelligence,人工智能)处理器,该AI处理器用于处理有关机器学习的计算操作。
存储器50可以包括一个或多个计算机可读存储介质,该计算机可读存储介质可以是非暂态的。存储器50还可包括高速随机存取存储器以及非易失性存储器,比如一个或多个磁盘存储设备、闪存存储设备。存储器50在一些实施例中可以是电子设备的内部存储单元,例如服务器的硬盘。存储器50在另一些实施例中也可以是电子设备的外部存储设备,例如服务器上配备的插接式硬盘,智能存储卡(Smart Media Card,SMC), 安全数字(Secure Digital,SD)卡,闪存卡(Flash Card)等。进一步地,存储器50还可以既包括电子设备的内部存储单元也包括外部存储设备。存储器50不仅可以用于存储安装于电子设备的应用软件及各类数据,例如:执行漏洞处理方法的程序的代码等,还可以用于暂时地存储已经输出或者将要输出的数据。本实施例中,存储器50至少用于存储以下计算机可读指令501,其中,该计算机可读指令被处理器51加载并执行之后,能够实现前述任一实施例公开的人工智能任务处理方法的相关步骤。另外,存储器50所存储的资源还可以包括操作系统502和数据503等,存储方式可以是短暂存储或者永久存储。其中,操作系统502可以包括Windows、Unix、Linux等。数据503可以包括但不限于人工智能任务处理结果对应的数据等。
在一些实施例中,上述电子设备还可包括有显示屏52、输入输出接口53、通信接口54或者称为网络接口、电源55以及通信总线56。其中,显示屏52、输入输出接口53比如键盘(Keyboard)属于用户接口,可选的用户接口还可以包括标准的有线接口、无线接口等。可选地,在一些实施例中,显示器可以是LED显示器、液晶显示器、触控式液晶显示器以及OLED(Organic Light-Emitting Diode,有机发光二极管)触摸器等。显示器也可以适当的称为显示屏或显示单元,用于显示在电子设备中处理的信息以及用于显示可视化的用户界面。通信接口54可选的可以包括有线接口和/或无线接口,如WI-FI接口、蓝牙接口等,通常用于在电子设备与其他电子设备之间建立通信连接。通信总线56可以是外设部件互连标准(peripheral component interconnect,简称PCI)总线或扩展工业标准结构(extended industry standard architecture,简称EISA)总线等。该总线可以分为地址总线、数据总线、控制总线等。为便于表示,图5中仅用一条粗线表示,但并不表示仅有一根总线或一种类型的总线。
本领域技术人员可以理解,图5中示出的结构并不构成对该电子设备的限定,可以包括比图示更多或更少的组件,例如还可包括实现各类功能的传感器57。
本申请实施例中的电子设备的各功能模块的功能可根据上述方法实施例中的方法具体实现,其具体实现过程可以参照上述方法实施例的相关描述,此处不再赘述。
由上可知,本申请实施例可有效提升人工智能任务处理性能,降低人工智能任务处理过程中消耗的计算资源。
可以理解的是,如果上述实施例中的人工智能任务处理方法以软件功能单元的形式实现并作为独立的产品销售或使用时,可以存储在一个计算机可读取存储介质中。基于这样的理解,本申请的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的全部或部分可以以软件产品的形式体现出来,该计算机软件产品存储在一个存储介质中,执行本申请各个实施例方法的全部或部分步骤。而前述的存储介质包括:U盘、移动硬盘、只读存储器(Read-Only Memory,ROM)、随机存取存储器(Random Access Memory,RAM)、电可擦除可编程ROM、寄存器、硬盘、多媒体卡、卡型存储器(例如SD或DX存储器等)、磁性存储器、可移动磁盘、CD-ROM、磁碟或者光盘等各种可以存储程序代码的介质。
基于此,本申请实施例还提供了一个或多个存储有计算机可读指令的非易失性计算机可读存储介质,其特征在于,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执行如上述任一实施例中人工智能任务处理方法的步骤。
本说明书中各个实施例采用递进的方式描述,每个实施例重点说明的都是与其它实施例的不同之处,各个实施例之间相同或相似部分互相参见即可。对于实施例公开的硬件包括装置及电子设备而言,由于其与实施例公开的方法相对应,所以描述的比较简单,相关之处参见方法部分说明即可。
专业人员还可以进一步意识到,结合本文中所公开的实施例描述的各示例的单元及算法步骤,能够以电子硬件、计算机软件或者二者的结合来实现,为了清楚地说明硬件和软件的可互换性,在上述说明中已经按照功能一般性地描述了各示例的组成及步骤。这些功能究竟以硬件还是软件方式来执行,取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能,但是这种实现不应认为超出本申请的范围。
以上对本申请所提供的一种人工智能任务处理方法、装置、电子设备及可读存储介质进行了详细介绍。本文中应用了具体个例对本申请的原理及实施方式进行了阐述,以上实施例的说明只是用于帮助理解本申请的方法及其核心思想。应当指出,对于本技术领域的普通技术人员来说,在不脱离本申请原理的前提下,还可以对本申请进行若干改进和修饰,这些改进和修饰也落入本申请权利要求的保护范围内。

Claims (12)

  1. 一种人工智能任务处理方法,其特征在于,包括:
    获取待处理人工智能任务对应的AI数据集和AI任务执行模型;所述AI数据集包括AI训练集和AI验证集;所述AI训练集和所述AI验证集均包括对应一个滑窗的多个连续子集,每个子集对应所述滑窗的一个窗口;
    根据所述AI任务执行模型,基于所述AI训练集在每个滑窗训练过程中的损失值,生成任务损失信息;
    根据所述AI任务执行模型,基于所述AI验证集在每个滑窗训练过程中的正向性能指标的期望值,生成任务精度期望信息;及
    根据所述任务损失信息、所述任务精度期望信息和所述待处理人工智能任务的任务精度需求信息确定是否停止所述AI任务执行模型的训练,并基于训练好的AI任务执行模型执行所述待处理人工智能任务。
  2. 根据权利要求1所述的人工智能任务处理方法,其特征在于,所述根据所述任务损失信息、所述任务精度期望信息和所述待处理人工智能任务的任务精度需求信息确定是否停止所述AI任务执行模型的训练,包括:
    根据所述任务损失信息和所述任务精度期望信息确定周期模型精度表示信息;
    判断所述周期模型精度表示信息是否与所述任务精度需求信息相匹配;及
    响应于所述周期模型精度表示信息与所述任务精度需求信息相匹配,输出停止训练所述AI任务执行模型的指令;或,响应于所述周期模型精度表示信息与所述任务精度需求信息不匹配,输出继续训练所述AI任务执行模型的指令。
  3. 根据权利要求2所述的人工智能任务处理方法,其特征在于,所述根据所述任务损失信息和所述任务精度期望信息确定周期模型精度表示信息,包括:
    调用周期结果表示计算关系式,计算所述周期模型精度表示信息;所述周期结果表示计算关系式为:
    S=f ε(S (i+1),S (i+2),....,S (i+ε))
    其中,S (i+ε)=M (i+ε)∧N (i+ε),f ε表示映射关系:
    {S i+1,S i+2,......,S i+ε}→{e 1,e 2,.....,e ε}
    e i∈{0,1};S为所述周期模型精度表示信息,i为第i个训练周期,ε为训练容忍度;M为所述任务损失信息,N为所述任务精度期望信息,S (i+ε)为第i+ε次训练所述AI任务执行模型的周期结果表示,M (i+ε)为第i+ε次训练所述AI任务执行模型的任务损失信息,N (i+ε)为第i+ε次训练所述AI任务执行模型的任务精度期望信息,∧表示逻辑与运算符。
  4. 根据权利要求3所述的人工智能任务处理方法,其特征在于,所述判断所述周期模型精度表示信息是否与所述任务精度需求信息相匹配,包括:
    响应于所述任务精度需求信息为至少有一次训练所述AI任务执行模型的周期结果表示不为0,判断所述周期模型精度表示信息的各元素是否均为0;
    相应的,所述响应于所述周期模型精度表示信息与所述任务精度需求信息相匹配,输出停止训练所述AI任务执行模型的指令的过程包括:
    响应于所述周期模型精度表示信息的各元素均为0,输出停止训练所述AI任务执行模型的指令;
    相应的,所述响应于所述周期模型精度表示信息与所述任务精度需求信息不匹配,输出继续训练所述AI任务执行模型的指令的过程包括:
    响应于所述周期模型精度表示信息的各元素不均为0,输出继续训练所述AI任务执行模型的指令。
  5. 根据权利要求1所述的人工智能任务处理方法,其特征在于,所述根据所述AI任务执行模型,基于所述AI训练集在每个滑窗训练过程中的损失值,生成任务损失信息,包括:
    基于所述AI训练集,对每个训练周期,根据每个滑窗的所述AI任务执行模型的损失值,计算当前训练周期的所有损失值的当前标准差;
    根据损失变化程度因子和所述当前训练周期的前一个训练周期的所有损失值的前向标准差,确定损失变化阈值;及
    根据所述当前标准差、所述前向标准差和所述损失变化阈值确定所述当前训练周期的任务损失信息。
  6. 根据权利要求5所述的人工智能任务处理方法,其特征在于,所述根据所述当前标准差、所述前向标准差和所述损失变化阈值确定所述当前训练周期的任务损失信息,包括:
    调用周期任务损失信息计算关系式计算所述当前训练周期的任务损失信息,所述周期任务损失信息计算关系式为:
    响应于σ(loss i1,loss i2,...,loss ik)-σ(loss (i-1)1,loss (i-2)2,...,loss (i-1)k)≤η,确定M i=1;响应于σ(loss i1,loss i2,...,loss ik)-σ(loss (i-1)1,loss (i-2)2,...,loss (i-1)k)>η,确定M i=0;式中,M i为第i个训练周期的任务损失信息,loss ik为第i个训练周期的第k个滑窗的损失值,loss ik为第i-1个训练周期的第k个滑窗的损失值,σ(lossi1,lossi2,…,lossik)为第i个训练周期的所有损失值的当前标准差,σ(loss i1,loss i2,...,loss ik)为第i-1个训练周期的所有损失值的前向标准差,η为所述损失变化阈值。
  7. 根据权利要求1所述的人工智能任务处理方法,其特征在于,所述根据所述AI任务执行模型,基于所述AI验证集在每个滑窗训练过程中的正向性能指标的期望值,生成任务精度期望信息,包括:
    基于所述AI验证集,对每个训练周期,根据每个滑窗的所述AI任务执行模型的正向性能指标,计算当前训练周期的所有正向性能指标的当前期望值;
    根据性能变化程度因子和所述当前训练周期的前一个训练周期的所有正向性能指标的前向期望值,确定性能变化阈值;及
    根据所述当前期望值、所述前向期望值和所述性能变化阈值确定所述当前训练周期的任务精度期望信息。
  8. 根据权利要求7所述的人工智能任务处理方法,其特征在于,所述根据所述当前期望值、所述前向期望值和所述性能变化阈值确定所述当前训练周期的任务精度期望信息,包括:
    调用周期任务精度计算关系式计算所述当前训练周期的任务精度期望信息,所述周期任务精度计算关系式为:
    响应于E(perf i1,perf i2,....,perf ik)-E(perf (i-1)1,perf (i-1)2,....,perf (i-1)k)>μ,确定N i=1;响应于E(perf i1,perf i2,....,perf ik)-E(perf (i-1)1,perf (i-1)2,....,perf (i-1)k)≤μ,确定N i=0;
    式中,N i为第i个训练周期的任务精度期望信息,perf ik为第i个训练周期的第k个滑窗的期望值,perf (i-1)k为第i-1个训练周期的第k个滑窗的期望值,E(perf i1,perf i2,....,perf ik)为第i个训练周期的所有正向性能指标的当前期望值,E(perf (i-1)1,perf (i-1)2,....,perf (i-1)k)为第i-1个训练周期的所有正向性能指标的前向期望值,μ为所述性能变化阈值。
  9. 根据权利要求1至8任意一项所述的人工智能任务处理方法,其特征在于,所述获取待处理人工智能任务对应的AI数据集和AI任务执行模型,包括:
    当接收到信息输入指令,展示信息处理交互界面;所述信息处理交互界面包括信息输入区域和结果展示区域;及
    响应用户通过所述信息输入区域下发的信息输入指令,从所述信息输入指令中获取待处理人工智能任务对应的AI数据集、AI任务执行模型、所述待处理人工智能任务的任务精度需求信息和滑窗参数值;
    其中,所述结果展示区域用于展示训练好的AI任务执行模型和/或所述待处理人工智能任务的任务执行结果。
  10. 一种人工智能任务处理装置,其特征在于,包括:
    信息获取模块,用于获取待处理人工智能任务对应的AI数据集和AI任务执行模型;所述AI数据集包括AI训练集和AI验证集;所述AI训练集和所述AI验证集均包括对应一个滑窗的多个连续子集,每个子集对应所述滑窗的一个窗口;
    损失计算模块,用于根据所述AI任务执行模型,基于所述AI训练集在每个滑窗训练过程中的损失值,生成任务损失信息;
    期望计算模块,用于根据所述AI任务执行模型,基于所述AI验证集在每个滑窗训练过程中的正向性能指标的期望值,生成任务精度期望信息;及
    模型训练结束确定模块,用于根据所述任务损失信息、所述任务精度期望信息和所述待处理人工智能任务的任务精度需求信息确定是否停止所述AI任务执行模型的训练,并基于训练好的AI任务执行模型执行所述待处理人工智能任务。
  11. 一种电子设备,其特征在于,包括存储器及一个或多个处理器,存储器中储存有计算机可读指令,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执行如权利要求1至9中任一项所述的方法的步骤。
  12. 一个或多个存储有计算机可读指令的非易失性计算机可读存储介质,其特征在于,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权利要求1至9中任一项所述的方法的步骤。
PCT/CN2022/100481 2021-12-28 2022-06-22 人工智能任务处理方法、装置、电子设备及可读存储介质 WO2023123926A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111616394.X 2021-12-28
CN202111616394.XA CN113986561B (zh) 2021-12-28 2021-12-28 人工智能任务处理方法、装置、电子设备及可读存储介质

Publications (1)

Publication Number Publication Date
WO2023123926A1 true WO2023123926A1 (zh) 2023-07-06

Family

ID=79734627

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/100481 WO2023123926A1 (zh) 2021-12-28 2022-06-22 人工智能任务处理方法、装置、电子设备及可读存储介质

Country Status (2)

Country Link
CN (1) CN113986561B (zh)
WO (1) WO2023123926A1 (zh)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113986561B (zh) * 2021-12-28 2022-04-22 苏州浪潮智能科技有限公司 人工智能任务处理方法、装置、电子设备及可读存储介质
CN114741269B (zh) * 2022-04-14 2022-09-23 网思科技股份有限公司 一种推理系统业务性能评估的方法

Citations (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109002790A (zh) * 2018-07-11 2018-12-14 广州视源电子科技股份有限公司 一种人脸识别的方法、装置、设备和存储介质
CN109977871A (zh) * 2019-03-27 2019-07-05 中国人民解放军战略支援部队航天工程大学 一种基于宽带雷达数据和gru神经网络的卫星目标识别方法
CN110991346A (zh) * 2019-12-04 2020-04-10 厦门市美亚柏科信息股份有限公司 一种疑似吸毒人员识别的方法、装置及存储介质
CN111860833A (zh) * 2020-07-09 2020-10-30 苏州浪潮智能科技有限公司 一种模型的训练方法、设备以及介质
CN112115876A (zh) * 2020-09-21 2020-12-22 四川弘和通讯有限公司 一种基于3d卷积神经网络的水溶法实验过程识别方法
CN112132092A (zh) * 2020-09-30 2020-12-25 四川弘和通讯有限公司 一种基于卷积神经网络的灭火器和灭火毯的识别方法
CN112163528A (zh) * 2020-09-30 2021-01-01 四川弘和通讯有限公司 一种基于卷积神经网络的油管拖地的识别方法
CN112766549A (zh) * 2021-01-07 2021-05-07 清华大学 一种空气污染物浓度预报方法、装置及存储介质
US20210319314A1 (en) * 2020-04-09 2021-10-14 Naver Corporation End-To-End Graph Convolution Network
CN113762625A (zh) * 2021-09-09 2021-12-07 国网山东省电力公司经济技术研究院 一种基于图卷积网络的配电网状态评估方法和系统
CN113792783A (zh) * 2021-09-13 2021-12-14 陕西师范大学 一种基于深度学习的和面阶段自动识别方法及系统
US20210390449A1 (en) * 2020-06-12 2021-12-16 Beijing Xiaomi Pinecone Electronics Co., Ltd. Method and device for data processing, and storage medium
CN113822428A (zh) * 2021-08-06 2021-12-21 中国工商银行股份有限公司 神经网络训练方法及装置、图像分割方法
CN113986561A (zh) * 2021-12-28 2022-01-28 苏州浪潮智能科技有限公司 人工智能任务处理方法、装置、电子设备及可读存储介质

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110232334A (zh) * 2019-05-26 2019-09-13 天津大学 一种基于卷积神经网络的钢结构锈蚀识别方法
CN111260594B (zh) * 2019-12-22 2023-10-31 天津大学 一种无监督的多模态图像融合方法
CN111277606B (zh) * 2020-02-10 2022-04-15 北京邮电大学 检测模型训练方法、检测方法及装置、存储介质
CN111461345B (zh) * 2020-03-31 2023-08-11 北京百度网讯科技有限公司 深度学习模型训练方法及装置
CN115699209A (zh) * 2020-04-03 2023-02-03 普雷萨根私人有限公司 用于人工智能(ai)模型选择的方法
CN113762514B (zh) * 2020-06-05 2024-04-16 京东科技控股股份有限公司 数据处理方法、装置、设备及计算机可读存储介质
CN113222942A (zh) * 2021-05-17 2021-08-06 北京百度网讯科技有限公司 多标签分类模型的训练方法和预测标签的方法

Patent Citations (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109002790A (zh) * 2018-07-11 2018-12-14 广州视源电子科技股份有限公司 一种人脸识别的方法、装置、设备和存储介质
CN109977871A (zh) * 2019-03-27 2019-07-05 中国人民解放军战略支援部队航天工程大学 一种基于宽带雷达数据和gru神经网络的卫星目标识别方法
CN110991346A (zh) * 2019-12-04 2020-04-10 厦门市美亚柏科信息股份有限公司 一种疑似吸毒人员识别的方法、装置及存储介质
US20210319314A1 (en) * 2020-04-09 2021-10-14 Naver Corporation End-To-End Graph Convolution Network
US20210390449A1 (en) * 2020-06-12 2021-12-16 Beijing Xiaomi Pinecone Electronics Co., Ltd. Method and device for data processing, and storage medium
CN111860833A (zh) * 2020-07-09 2020-10-30 苏州浪潮智能科技有限公司 一种模型的训练方法、设备以及介质
CN112115876A (zh) * 2020-09-21 2020-12-22 四川弘和通讯有限公司 一种基于3d卷积神经网络的水溶法实验过程识别方法
CN112163528A (zh) * 2020-09-30 2021-01-01 四川弘和通讯有限公司 一种基于卷积神经网络的油管拖地的识别方法
CN112132092A (zh) * 2020-09-30 2020-12-25 四川弘和通讯有限公司 一种基于卷积神经网络的灭火器和灭火毯的识别方法
CN112766549A (zh) * 2021-01-07 2021-05-07 清华大学 一种空气污染物浓度预报方法、装置及存储介质
CN113822428A (zh) * 2021-08-06 2021-12-21 中国工商银行股份有限公司 神经网络训练方法及装置、图像分割方法
CN113762625A (zh) * 2021-09-09 2021-12-07 国网山东省电力公司经济技术研究院 一种基于图卷积网络的配电网状态评估方法和系统
CN113792783A (zh) * 2021-09-13 2021-12-14 陕西师范大学 一种基于深度学习的和面阶段自动识别方法及系统
CN113986561A (zh) * 2021-12-28 2022-01-28 苏州浪潮智能科技有限公司 人工智能任务处理方法、装置、电子设备及可读存储介质

Also Published As

Publication number Publication date
CN113986561A (zh) 2022-01-28
CN113986561B (zh) 2022-04-22

Similar Documents

Publication Publication Date Title
WO2023123926A1 (zh) 人工智能任务处理方法、装置、电子设备及可读存储介质
CN113159147B (zh) 基于神经网络的图像识别方法、装置、电子设备
WO2022007823A1 (zh) 一种文本数据处理方法及装置
WO2022141861A1 (zh) 情感分类方法、装置、电子设备及存储介质
CN109033068A (zh) 基于注意力机制的用于阅读理解的方法、装置和电子设备
CN111352965B (zh) 序列挖掘模型的训练方法、序列数据的处理方法及设备
WO2020082734A1 (zh) 文本情感识别方法、装置、电子设备及计算机非易失性可读存储介质
JP2022141931A (ja) 生体検出モデルのトレーニング方法及び装置、生体検出の方法及び装置、電子機器、記憶媒体、並びにコンピュータプログラム
CN109345553A (zh) 一种手掌及其关键点检测方法、装置和终端设备
CN111784595B (zh) 一种基于历史记录的动态标签平滑加权损失方法及装置
CN111563158B (zh) 文本排序方法、排序装置、服务器和计算机可读存储介质
CN113128671B (zh) 一种基于多模态机器学习的服务需求动态预测方法及系统
CN107622413A (zh) 一种价格敏感度计算方法、装置及其设备
CN111160049B (zh) 文本翻译方法、装置、机器翻译系统和存储介质
TW202119297A (zh) 訓練方法、特徵提取方法、裝置及電子設備
CN112115714B (zh) 深度学习序列标注方法、装置及计算机可读存储介质
CN110033165A (zh) 逾期借款催收方式的推荐方法、装置、介质、电子设备
CN113902260A (zh) 信息预测方法、装置、电子设备和介质
WO2020167156A1 (ru) Способ отладки обученной рекуррентной нейронной сети
CN111177493B (zh) 数据处理方法、装置、服务器和存储介质
CN114864043A (zh) 基于vr设备的认知训练方法、装置及介质
CN111178630A (zh) 一种负荷预测方法及装置
CN111091420A (zh) 一种电力价格的预测方法及装置
CN115146596B (zh) 召回文本的生成方法、装置、电子设备及存储介质
US20240144373A1 (en) Financial investment predictions and recommendations using neural networks

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: 22913215

Country of ref document: EP

Kind code of ref document: A1