Migration capability acquisition method and device, display equipment and storage medium

Abstract

The invention provides a migration capability acquisition method and device, display equipment and a storage medium, and the method comprises the steps: responding to the detected operation on a task set and an algorithm set, constructing a task migration graph, and enabling each vertex of the task migration graph to be a task in the task set, edges of the task migration graph refer to different mobility values of tasks of two vertexes of the edges under different algorithms; the mobility value is determined by performance index data of each vertex task processed by each algorithm; in response to the detected operation of selecting a clustering algorithm, clustering the task migration graph to obtain at least one task family; and in response to the detected selection operation for obtaining the migration capability, obtaining the migration capability of each algorithm in the algorithm set, the migration capability representing an average value of mobility values of each algorithm for migrating one task to another task.

Description

Migration capability acquisition method, device, display device, and storage medium Technical Field

[0001] The present disclosure relates to the field of data processing technology, and in particular to a migration capability acquisition method, apparatus, display device, and storage medium. Background Art

[0002] Transfer learning involves transferring previously learned model parameters to a new model to aid in its training. Given that most data or tasks are related, transfer learning allows previously learned model parameters to be shared with the new model in some way, accelerating and optimizing the new model's learning efficiency.

[0003] However, in practical applications, the effectiveness of different transfer learning methods or the difficulty of transferring to different fields require strong quantitative evaluation, which increases the difficulty of transfer learning.

[0004] Summary of the Invention

[0005] The present disclosure provides a migration capability acquisition method, apparatus, display device, and storage medium to solve the above technical problems.

[0006] According to a first aspect of the present disclosure, a method for acquiring migration capability is provided, the method comprising:

[0007] In response to detecting operations on a task set and an algorithm set, a task migration graph is constructed, wherein each vertex of the task migration graph is a task in the task set, and an edge of the task migration graph refers to different transferability values ​​of the tasks at two vertices of the edge under different algorithms; the transferability value is determined by accuracy performance indicator data of each algorithm migrating tasks from one vertex to another for processing each vertex task;

[0008] In response to detecting an operation of selecting a clustering algorithm, clustering the task transition graph to obtain at least one task family;

[0009] In response to detecting a selection operation of obtaining transfer capability, the transfer capability of each algorithm in the algorithm set is obtained, where the transfer capability represents an average value of transferability values ​​of each algorithm in migrating one task to another task.

[0010] Optionally, build a task migration graph, including:

[0011] Obtaining a task set and an algorithm set; the task set includes a plurality of tasks, and the algorithm set includes a plurality of algorithms for implementing transfer learning between any two tasks;

[0012] Using any one algorithm in the algorithm set to perform transfer learning on any two tasks in the task set, obtaining performance indicator data corresponding to the task-algorithm combination; the performance indicator data includes absolute indicator data and / or relative indicator data; the absolute indicator data refers to the accuracy of each algorithm when migrating from one vertex task to another; the relative indicator data refers to the difference in accuracy of each algorithm when processing two vertex tasks on the same edge;

[0013] Obtaining transferability values ​​of any two tasks according to the performance indicator data;

[0014] The task migration graph is constructed with each task in the task set as a vertex and the transferability value between any two tasks as an edge.

[0015] Optionally, using any one algorithm in the algorithm set to perform transfer learning on any two tasks in the task set to obtain performance indicator data corresponding to the task-algorithm combination includes:

[0016] Using a first portion of samples of the source task in the two tasks to train the algorithm to obtain a trained first algorithm;

[0017] Verifying the first algorithm using a second portion of samples of the source task to obtain a first accuracy rate;

[0018] Using a first portion of samples of a target task in the two tasks to train the first algorithm to obtain a trained second algorithm;

[0019] Verifying the second algorithm using a second portion of samples of the target task to obtain a second accuracy rate;

[0020] The performance indicator data is obtained according to the first accuracy rate and the second accuracy rate.

[0021] Optionally, obtaining the performance indicator data according to the first accuracy rate and the second accuracy rate includes:

[0022] Determining the second accuracy rate as absolute indicator data;

[0023] Obtaining a difference between the second accuracy rate and the first accuracy rate as relative indicator data;

[0024] The absolute indicator data and / or the relative indicator data are determined as the performance indicator data.

[0025] Optionally, the transferability value includes a forward transferability value, and obtaining the transferability values ​​of any two tasks according to the performance indicator data includes:

[0026] Obtaining performance indicator data of each algorithm in the algorithm set during transfer learning from a source task to a target task in any two tasks;

[0027] Obtain the average value of the performance indicator data corresponding to each algorithm as the positive transferability value of the two tasks;

[0028] Alternatively, the transferability value includes a reverse transferability value, and obtaining the transferability values ​​of any two tasks according to the performance indicator data includes:

[0029] Obtaining performance indicator data of each algorithm in the algorithm set when transferring learning from the target task to the source task in the arbitrary two tasks;

[0030] The average value of the performance indicator data corresponding to each algorithm is obtained as the reverse transferability value of the two tasks.

[0031] Optionally, the transferability value includes a bidirectional transferability value, and obtaining the transferability values ​​of any two tasks according to the performance indicator data includes:

[0032] An average value of the forward transferability value and the reverse transferability value is obtained as the bidirectional transferability value of the arbitrary two tasks.

[0033] Optionally, the task transition graph includes a directed transition graph and an undirected transition graph;

[0034] When the edge of the task transition graph is a forward transferability value or a reverse transferability value, the task transition graph is a directed transition graph;

[0035] When the edges of the task transition graph have bidirectional transferability values, the task transition graph is an undirected transition graph.

[0036] Optionally, clustering the task transition graph to obtain at least one task family includes:

[0037] Randomly select the initial center points of a preset number of categories;

[0038] Obtaining the distance between each task and each initial center point in the task migration graph, and classifying each task into the category of the initial center point with the smallest distance;

[0039] Update the center point of each category, and update the eigenvalue of the new center point to the average eigenvalue of all samples in the category;

[0040] When it is determined that the preset conditions are met, tasks of each category are determined to form a task family, and at least one task family is obtained;

[0041] When it is determined that the preset condition is not satisfied, the step of obtaining the distance between each task and each initial center point in the task transition graph is performed again.

[0042] Optionally, the preset condition means that the number of iterations is greater than a preset iteration threshold, or the tasks in each category remain unchanged.

[0043] Optionally, obtain the migration capability of each algorithm in the algorithm set, including:

[0044] Obtain the intra-family transfer capability of each algorithm in the algorithm set within each task family and the inter-family transfer capability between any two task families;

[0045] The migration capabilities of the respective algorithms are obtained according to the intra-cluster migration capabilities and the inter-cluster migration capabilities.

[0046] Optionally, obtain the intra-family transfer capability of each algorithm in the algorithm set within each task family, including:

[0047] For each algorithm in the algorithm set, obtaining performance indicator data of each task in each task family under the algorithm;

[0048] Obtaining an average value of performance indicator data of the algorithm in each task family as a first transfer capability of the algorithm in each task family;

[0049] An average value of the first transfer capabilities of the algorithm within each task family is obtained as the intra-family transfer capability of the algorithm within each task family.

[0050] Optionally, obtain the inter-family transfer capability of each algorithm in the algorithm set between any two task families, including:

[0051] Get any two task families as the source task family and the target task family;

[0052] For each algorithm in the algorithm set, obtaining performance indicator data of each source task in the source task family migrated to each target task in the target task family under the algorithm;

[0053] Obtaining an average value of performance indicator data corresponding to the source task family and the target task family as a second migration capability;

[0054] An average value of the second transfer capabilities corresponding to all task families is obtained as the inter-family transfer capability of each algorithm between any two task families.

[0055] Optionally, obtaining the migration capability of each algorithm according to the intra-cluster migration capability and the inter-cluster migration capability includes:

[0056] Obtaining weight values ​​of the intra-clan migration capability and the inter-clan migration capability respectively;

[0057] A weighted sum of the intra-cluster migration capability and the inter-cluster migration capability is obtained as the migration capability of each algorithm.

[0058] Optionally, when two tasks are in the same task family, the weight value of the intra-family migration capability is greater than or equal to the weight value of the inter-family migration capability;

[0059] When two tasks are in different task families, the weight value of the intra-family migration capability is smaller than the weight value of the inter-family migration capability.

[0060] According to a second aspect of the present disclosure, a display control method is provided, the method comprising:

[0061] In response to detecting the start operation, displaying the candidate tasks and the candidate algorithms in the display interface;

[0062] In response to detecting an operation of selecting a candidate task, displaying the selected task set in a display interface;

[0063] In response to detecting an operation of selecting at least one candidate algorithm, displaying the selected algorithm set in a display interface;

[0064] In response to detecting an operation of determining the task set and the algorithm set, a task migration graph constructed based on the task set and the algorithm set is displayed in a display interface; each vertex of the task migration graph is a task in the task set, and an edge of the task migration graph refers to different transferability values ​​of the tasks at two vertices of the edge under different algorithms; the transferability value is determined by accuracy performance indicator data of each algorithm migrating a task from one vertex to another to process each vertex task;

[0065] In response to detecting a selection operation of a clustering algorithm for clustering tasks in the task transition graph, at least one task family after clustering the task transition graph according to the clustering algorithm is displayed, each task family including a plurality of tasks.

[0066] Optionally, the method further includes:

[0067] In response to detecting a selection operation of obtaining migration capability, the migration capability of each algorithm in the algorithm set is displayed in the display interface, where the migration capability represents an average value of the migration capability values ​​of each algorithm for migrating one task to another task.

[0068] Optionally, the method further includes:

[0069] In response to detecting an operation of sorting the migration capabilities, various algorithms sorted according to the size relationship of the migration capabilities are displayed in a display interface.

[0070] According to a third aspect of the present disclosure, a migration capability acquisition device is provided, the device comprising:

[0071] A migration graph acquisition module is configured to construct a task migration graph in response to detecting operations on a task set and an algorithm set, wherein each vertex of the task migration graph is a task in the task set, and an edge of the task migration graph is a task with different transferability values ​​under different algorithms between the tasks at two vertices of the edge; the transferability value is determined by accuracy performance indicator data of each algorithm migrating tasks from one vertex to another for processing each vertex task;

[0072] a task family acquisition module, configured to cluster the task transition graph in response to detecting an operation of selecting a clustering algorithm, to obtain at least one task family;

[0073] The migration capability acquisition module is configured to acquire the migration capability of each algorithm in the algorithm set in response to detecting a selection operation for acquiring migration capability, wherein the migration capability represents an average value of the transferability values ​​of each algorithm in migrating one task to another.

[0074] Optionally, the migration graph acquisition module includes:

[0075] A task set acquisition submodule is used to acquire a task set; the task set includes a plurality of tasks;

[0076] An algorithm set acquisition submodule is used to acquire an algorithm set; the algorithm set includes several algorithms that implement transfer learning between any two tasks;

[0077] a performance index acquisition submodule, configured to perform transfer learning on any two tasks in the task set using any one algorithm in the algorithm set, and obtain performance index data corresponding to the task-algorithm combination; the performance index data includes absolute index data and / or relative index data; the absolute index data refers to the accuracy of each algorithm when migrating from one vertex task to another; the relative index data refers to the difference in accuracy of each algorithm when processing two vertex tasks on the same edge;

[0078] A transferability value acquisition submodule, configured to obtain transferability values ​​of any two tasks based on the performance indicator data;

[0079] The migration graph construction submodule is used to construct the task migration graph with each task in the task set as a vertex and with the transferability value between any two tasks as an edge.

[0080] Optionally, the performance index acquisition submodule includes:

[0081] A first algorithm acquisition unit is configured to train the algorithm using a first portion of samples of the source task in the two tasks to obtain a trained first algorithm;

[0082] A first accuracy rate acquisition unit, configured to verify the first algorithm using a second portion of samples of the source task to obtain a first accuracy rate;

[0083] a second algorithm acquiring unit, configured to train the first algorithm using a first portion of samples of a target task among the two tasks to obtain a trained second algorithm;

[0084] A second accuracy rate obtaining unit, configured to verify the second algorithm using a second portion of samples of the target task to obtain a second accuracy rate;

[0085] A performance indicator acquisition unit is used to acquire the performance indicator data according to the first accuracy rate and the second accuracy rate.

[0086] Optionally, the performance indicator acquisition unit includes:

[0087] an absolute indicator acquisition subunit, configured to determine the second accuracy rate as absolute indicator data;

[0088] a relative indicator acquisition subunit, configured to acquire a difference between the second accuracy rate and the first accuracy rate as relative indicator data;

[0089] The performance indicator determination subunit is configured to determine the absolute indicator data and / or the relative indicator data as the performance indicator data.

[0090] Optionally, the mobility value includes a forward mobility value, and the mobility value acquisition submodule includes:

[0091] A performance indicator acquisition unit, configured to acquire performance indicator data of each algorithm in the algorithm set during transfer learning from a source task to a target task in any two tasks;

[0092] The positive value acquisition unit is used to obtain the average value of the performance indicator data corresponding to each algorithm as the positive transferability value of the two tasks.

[0093] Optionally, the portability value includes a reverse portability value, and the portability value acquisition submodule includes:

[0094] A performance indicator acquisition unit, configured to acquire performance indicator data of each algorithm in the algorithm set during transfer learning from a target task to a source task in any two tasks;

[0095] The reverse value acquisition unit is used to obtain the average value of the performance indicator data corresponding to each algorithm as the reverse transferability value of the two tasks.

[0096] Optionally, the portability value includes a bidirectional portability value, and the portability value acquisition submodule includes:

[0097] The bidirectional value acquisition unit is configured to acquire an average value of the forward transferability value and the reverse transferability value as the bidirectional transferability value of the arbitrary two tasks.

[0098] Optionally, the task transition graph includes a directed transition graph and an undirected transition graph;

[0099] When the edge of the task transition graph is a forward transferability value or a reverse transferability value, the task transition graph is a directed transition graph;

[0100] When the edges of the task transition graph have bidirectional transferability values, the task transition graph is an undirected transition graph.

[0101] Optionally, the task family acquisition module includes:

[0102] The initial point selection submodule is used to randomly select the initial center points of a preset number of categories;

[0103] A task division submodule is used to obtain the distance between each task and each initial center point in the task migration graph, and divide each task into the category of the initial center point with the smallest distance;

[0104] The center point update submodule is used to update the center point of each category and update the feature value of the new center point to the average feature value of all samples in the category;

[0105] The judgment submodule is used to determine that tasks of each category constitute a task family and obtain at least one task family when it is determined that the preset conditions are met; when it is determined that the preset conditions are not met, re-execute the step of obtaining the distance between each task and each initial center point in the task migration diagram.

[0106] Optionally, the preset condition means that the number of iterations is greater than a preset iteration threshold, or the tasks in each category remain unchanged.

[0107] Optionally, the migration capability acquisition module includes:

[0108] The intra-family capability acquisition submodule is used to obtain the intra-family transfer capability of each algorithm in the algorithm set within each task family and the inter-family transfer capability between any two task families;

[0109] The inter-family capability acquisition submodule is used to obtain the inter-family transfer capability of each algorithm in the algorithm set between any two task families;

[0110] The migration capability acquisition submodule is configured to acquire the migration capability of each algorithm according to the intra-cluster migration capability and the inter-cluster migration capability.

[0111] Optionally, the intra-family capability acquisition submodule includes:

[0112] A performance indicator acquisition unit, configured to acquire, for each algorithm in the algorithm set, performance indicator data of each task in each task family under the algorithm;

[0113] a first capability acquisition unit, configured to acquire an average value of performance indicator data of the algorithm in each task family as a first transfer capability of the algorithm in each task family;

[0114] The intra-family capability acquisition unit is configured to acquire an average value of the first transfer capabilities of the algorithm within each task family as the intra-family transfer capability of the algorithm within each task family.

[0115] Optionally, the inter-clan capability acquisition submodule includes:

[0116] A task family acquisition unit is used to acquire any two task families as a source task family and a target task family;

[0117] a performance indicator acquisition unit, configured to acquire, for each algorithm in the algorithm set, performance indicator data of each source task in the source task family migrated to each target task in the target task family under the algorithm;

[0118] A second capability acquisition unit, configured to acquire an average value of performance indicator data corresponding to the source task family and the target task family as a second migration capability;

[0119] The inter-family capability acquisition unit is configured to acquire an average value of the second transfer capabilities corresponding to all task families as the inter-family transfer capability of each algorithm between any two task families.

[0120] Optionally, the migration capability acquisition submodule includes:

[0121] a weight value obtaining unit, configured to obtain weight values ​​of the intra-clan migration capability and the inter-clan migration capability respectively;

[0122] The migration capability acquisition unit is configured to acquire a weighted sum of the intra-cluster migration capability and the inter-cluster migration capability as the migration capability of each algorithm.

[0123] Optionally, when two tasks are in the same task family, the weight value of the intra-family migration capability is greater than or equal to the weight value of the inter-family migration capability;

[0124] When two tasks are in different task families, the weight value of the intra-family migration capability is smaller than the weight value of the inter-family migration capability.

[0125] According to a fourth aspect of the present disclosure, a display control device is provided, the device comprising:

[0126] a task algorithm display module, configured to display candidate tasks and candidate algorithms in a display interface in response to detecting a start operation;

[0127] a task set selection module, configured to display the selected task set in the display interface in response to detecting an operation of selecting a candidate task;

[0128] an algorithm set selection module, configured to display the selected algorithm set in a display interface in response to detecting an operation of selecting at least one candidate algorithm;

[0129] a migration graph display module, configured to, in response to detecting an operation of determining the task set and the algorithm set, display a task migration graph constructed based on the task set and the algorithm set within a display interface; wherein each vertex of the task migration graph is a task in the task set, and an edge of the task migration graph is a task having different migration values ​​under different algorithms at two vertices of the edge; wherein the migration value is determined by accuracy performance indicator data of each algorithm migrating a task from one vertex to another for processing each vertex task;

[0130] The task family display module is used to display at least one task family after clustering the task transition graph according to the clustering algorithm in response to detecting the selection operation of the clustering algorithm for clustering the tasks of the task transition graph, each task family including multiple tasks.

[0131] Optionally, the device further comprises:

[0132] The migration capability display module is used to display the migration capability of each algorithm in the algorithm set in a display interface in response to detecting a selection operation for obtaining migration capability, wherein the migration capability represents the accuracy of each algorithm in migrating one task to another task.

[0133] Optionally, the device further comprises:

[0134] The migration capability ranking module is configured to, in response to detecting an operation of ranking the migration capabilities, display various algorithms ranked according to the size relationship of the migration capabilities in a display interface.

[0135] According to a fifth aspect of the present disclosure, there is provided a display device, including:

[0136] processor and memory;

[0137] The memory is used to store a computer program executable by the processor;

[0138] The processor is configured to execute the computer program in the memory to implement the method as described in any one of the first aspects.

[0139] According to a sixth aspect of the present disclosure, a non-transitory computer-readable storage medium is provided, which, when an executable computer program in the storage medium is executed by a processor, can implement the method described in any one of the first aspects.

[0140] The technical solutions provided by the embodiments of the present disclosure may have the following beneficial effects:

[0141] The solution provided by the solution of this embodiment can construct a task migration graph in response to detecting an operation on a task set and an algorithm set, wherein each vertex of the task migration graph is a task in the task set, and the edge of the task migration graph refers to the different transferability values ​​of the tasks of the two vertices of the edge under different algorithms; the transferability value is determined by the accuracy performance index data of each algorithm migrating from one vertex task to another to process each vertex task; then, in response to detecting an operation of selecting a clustering algorithm, the task migration graph is clustered to obtain at least one task family; thereafter, in response to detecting a selection operation of obtaining transferability, the transferability of each algorithm in the algorithm set is obtained, and the transferability represents the average value of the transferability value of each algorithm migrating one task to another. In this way, by quantifying the transferability of the algorithm, this embodiment can better evaluate the difficulty of transfer learning, which is conducive to improving the efficiency of transfer learning.

[0142] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0143] FIG1 is a flowchart of a method for acquiring migration capability according to an embodiment of the present disclosure.

[0144] FIG2 is a schematic diagram of a selection task and algorithm according to an embodiment of the present disclosure.

[0145] FIG3 is a flowchart of constructing a task migration graph according to an embodiment of the present disclosure.

[0146] FIG4 is a schematic diagram of displaying performance indicator data according to an embodiment of the present disclosure.

[0147] FIG5 is a flow chart of obtaining a forward transferability value according to an embodiment of the present disclosure.

[0148] FIG6 is a flow chart of obtaining a reverse transferability value according to an embodiment of the present disclosure.

[0149] FIG7 is a schematic diagram of a task family according to an embodiment of the present disclosure.

[0150] FIG8 is a flowchart of obtaining the migration capability of each algorithm according to an embodiment of the present disclosure.

[0151] FIG9 is a flowchart of obtaining the intra-family migration capability of each algorithm according to an embodiment of the present disclosure.

[0152] FIG10 is a flowchart of obtaining the inter-family migration capability of each algorithm according to an embodiment of the present disclosure.

[0153] FIG11 is a schematic diagram of obtaining a ranking of migration capabilities of algorithms according to an embodiment of the present disclosure.

[0154] FIG12 is a block diagram of a migration capability acquisition device according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0155] Exemplary embodiments will be described in detail herein, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with the present disclosure. Rather, they are merely examples of devices consistent with certain aspects of the present disclosure, as detailed in the appended claims.

[0156] To address the above technical issues, embodiments of the present disclosure provide a migration capability acquisition method, apparatus, display device, and storage medium. The migration capability acquisition method can be applied to display devices, including but not limited to mobile phones, computers, tablets, e-books, 3D display screens, all-in-one conference machines, electronic whiteboards, servers, or server clusters, and other devices or combinations thereof that have both data processing and display capabilities.

[0157] An embodiment of the present disclosure provides a migration capability acquisition method, as shown in FIG1 , including steps 11 to 13 .

[0158] In step 11, in response to detecting operations on a task set and an algorithm set, a task migration graph is constructed, wherein each vertex of the task migration graph is a task in the task set, and an edge of the task migration graph refers to different transferability values ​​of the tasks at the two vertices of the edge under different algorithms; the transferability value is determined by the accuracy performance index data of each algorithm migrating from one vertex task to another to process each vertex task.

[0159] In this step, the display device can display candidate tasks and candidate algorithms within the display interface. The candidate tasks may include, but are not limited to, classification tasks, target detection, target tracking, and the like. Each candidate task is a sample set for a training algorithm. The sample set can be an open-source sample set, such as the COCO2017 target detection dataset, the DAVIS video segmentation dataset, the LVIS image segmentation dataset, or the Kaggle garbage classification image dataset. Of course, the sample set can also be a private sample set, and the corresponding sample set can be selected based on the specific scenario. The candidate algorithms may include, but are not limited to, ResNet, MobileNet, DenseNet, NASNetLarge, EfficientNet, Xception, and the like, and the corresponding algorithm can be selected based on the specific scenario.

[0160] In this step, the candidate tasks and candidate algorithms can be pre-stored by the user in the display device. When the user needs to obtain migration capabilities, they can click the start operation button in the display interface. In response to detecting the start operation, the display device can display the candidate tasks and candidate algorithms in the display interface. It is understood that the candidate tasks and candidate algorithms can be displayed separately, for example, displaying the candidate tasks first and then the candidate algorithms, or displaying the candidate algorithms first and then the candidate tasks, etc.

[0161] Taking the simultaneous display of candidate tasks and candidate algorithms as an example, see Figure 2. Candidate tasks can be displayed in multiple rows and columns, where each candidate task can serve as either a source task or a target task. Candidate tasks can also be arranged in two columns, with the left column representing source tasks and the right column representing target tasks. The arrangement of candidate algorithms is not limited. Continuing with Figure 2, the user can select task D3, then select task D7, and then select algorithm A6; then, press Ctrl+A to query the unidirectional transferability value from task D3 to task D7. In one example, the user can select task D3, then select task D7, then select task D3, then select task D7, and then select algorithm A6; then, press Ctrl+A to query the bidirectional transferability value from task D3 to task D7. Alternatively, in one example, the above select operation can be modified to a double-click operation to query the bidirectional transferability value.

[0162] The user can select at least two candidate tasks, or select at least one candidate task as a source task and at least one candidate task as a target task. In this way, the display device can display the selected task set within the display interface in response to detecting the operation of selecting a candidate task. The user can select at least one candidate algorithm to obtain an algorithm set. In this way, the display device can display the selected algorithm set within the display interface in response to detecting the operation of selecting a candidate algorithm.

[0163] It should be noted that the algorithm set disclosed in the present invention is not limited to transfer learning, but can also be applied to scenarios such as small sample learning and meta-learning. Taking classification task transfer as an example, the indicators (absolute and relative indicators) for evaluating inter-task transferability and algorithm transfer performance are all based on accuracy. Specifically, the sample size of the target task data set for transfer learning is relatively large, such as 500 samples per class or even more, and the risk of overfitting is relatively small, so the inter-task transferability is relatively high, and thus the algorithm transfer performance is also relatively high; while the sample size of small samples is relatively small, such as 10 samples per class, which is prone to overfitting, so the task transferability and the corresponding algorithm transfer performance are relatively low; meta-learning is learning how to learn, the training unit is the task, and the indicator for evaluating the learning effect is still accuracy. Usually, the task transferability and the corresponding algorithm transfer performance of meta-learning are also relatively low.

[0164] When displaying the task set and algorithm set, the user can confirm whether the displayed task set and algorithm set match the selected task and algorithm. If a match is confirmed, the user can select a confirmation operation. In response to detecting the operation of confirming the task set and algorithm set, the display device can display a task transition diagram constructed based on the task set and algorithm set within the display interface.

[0165] 3 , a task migration diagram constructed by a display device according to a task set and an algorithm set is shown, including steps 31 to 34 .

[0166] In step 31 , the display device may obtain a task set and an algorithm set; the task set includes a plurality of tasks, and the algorithm set includes a plurality of algorithms for implementing transfer learning between any two tasks.

[0167] In step 32, the display device may use any algorithm in the algorithm set to perform transfer learning on any two tasks in the task set, thereby obtaining performance indicator data corresponding to the task-algorithm combination. The performance indicator data may include absolute indicator data and / or relative indicator data; the absolute indicator data refers to the accuracy of each algorithm transferring from one vertex task to another; the relative indicator data refers to the difference in accuracy between each algorithm processing two vertex tasks on the same edge.

[0168] In this step, the display device can select one task from the task set as the source task and another task as the target task, and the source task and the target task are different tasks. Then, the display device can use any algorithm in the algorithm set to perform transfer learning on the source task and the target task to obtain performance indicator data corresponding to the task-algorithm combination, including:

[0169] The display device can divide the samples in the source task into a first part and a second part, and then use the first part of the samples of the source task to train the algorithm to obtain a trained algorithm, which will be referred to as the first algorithm for distinction; then use the second part of the samples of the source task to verify the first algorithm to obtain a first accuracy rate.

[0170] The display device can divide the samples in the target task into a first part and a second part, and then use the first part of the samples of the target task to train the first algorithm to obtain a trained algorithm, which will be referred to as the second algorithm for distinction; then use the second part of the samples of the target task to verify the second algorithm to obtain a second accuracy rate.

[0171] Finally, the display device may obtain the performance indicator data based on the first accuracy rate and the second accuracy rate. For example, the display device may determine the second accuracy rate as absolute indicator data. Alternatively, the display device may obtain the difference between the second accuracy rate and the first accuracy rate as relative indicator data. Ultimately, the display device may determine absolute indicator data and / or relative indicator data as performance indicator data. In other words, performance indicator data may select to use absolute indicator data and / or relative indicator data based on the specific scenario, and the corresponding solution falls within the scope of protection of this disclosure.

[0172] Taking the first accuracy rate as 95% and the second accuracy rate as 93% as an example, the absolute index data of the performance index data of the task algorithm combination is determined to be 93%, and the relative index data is -2% (=93%-95%). In one example, when a user queries the performance index data of the task algorithm combination, the display device can display the performance index data, as shown in Figure 4.

[0173] It should be noted that the above accuracy refers to the ratio of the number of correct results obtained by the algorithm when processing a task to the total number of images processed. For example, the algorithm classifies 100 images in a task, of which the classification results of 93 images are the same as the labeled classification, while the classification results of 7 images are different from the labeled classification. In this case, 93 images are classified correctly and 7 images are classified incorrectly. In this case, the accuracy of the algorithm is 93 / 100*100%=93%.

[0174] In step 33, the display device may obtain transferability values ​​of any two tasks according to the performance indicator data.

[0175] In one example, the transferability value includes a forward transferability value, that is, a unidirectional transferability value from a source task to a target task. The display device can obtain the transferability values ​​of any two tasks based on the performance indicator data, as shown in FIG5 , including steps 51 and 52 .

[0176] In step 51 , the display device may obtain performance indicator data of each algorithm in the algorithm set during transfer learning from a source task to a target task in any two tasks.

[0177] In step 52 , the display device may obtain an average value of the performance indicator data corresponding to each algorithm as the forward transferability value of the two tasks.

[0178] For example, the source task is Di, the target task is Dj, and the positive transferability value of the algorithm from the source task to the target task is shown in formula (1).

[0179] In formula (1), i represents the source task Di, j represents the target task Dj, k represents the algorithm Ak, and J represents the number of algorithms in the algorithm set.

[0180] In another example, the transferability value includes a reverse transferability value, i.e., a unidirectional transferability value from the target task to the source task in the example solution of Figure 5. The display device can obtain the transferability values ​​of any two tasks based on the performance indicator data, as shown in Figure 6, including steps 61 and 62.

[0181] In step 61 , the display device may obtain performance indicator data of each algorithm in the algorithm set during transfer learning from the target task to the source task in the arbitrary two tasks.

[0182] In step 62 , the display device may obtain an average value of the performance indicator data corresponding to each algorithm as the reverse transferability value of the two tasks.

[0183] It is understandable that the solution illustrated in FIG6 is similar to the solution illustrated in FIG5 , the difference being that the source task and the target task are exchanged between the two solutions.

[0184] For example, the source task is Di, the target task is Dj, and the reverse transferability value of the algorithm from the target task to the source task is shown in formula (2).

[0185] In formula (2), i represents the source task Di (which is actually used as the target task in the example scheme shown in FIG6 ), j represents the target task Dj (which is actually used as the source task in the example scheme shown in FIG6 ), k represents the algorithm Ak, and J represents the number of algorithms in the algorithm set.

[0186] In another example, the transferability value includes a bidirectional transferability value. In this case, the display device can obtain the average value of the forward transferability value of the example scheme shown in Figure 5 and the reverse transferability value of the example scheme shown in Figure 6 as the bidirectional transferability value of the any two tasks, as shown in formula (3).

[0187] In formula (3), represents the bidirectional transferability value, represents the positive transferability value, Indicates the reverse transferability value.

[0188] In step 34 , the display device may construct the task migration graph using each task in the task set as a vertex and using the transferability value between any two tasks as an edge.

[0189] Considering that the transferability value includes a unidirectional transferability value and a bidirectional transferability value, the task transition graph can include a directed transition graph and an undirected transition graph. In one example, when the edges of the task transition graph are positive transferability values ​​or negative transferability values, the task transition graph is a directed transition graph. In another example, when the edges of the task transition graph are bidirectional transferability values, the task transition graph is an undirected transition graph.

[0190] It should be noted that the choice of whether to use an undirected or directed task migration graph depends on the specific scenario and is not limited here. For example, in one example, Figure 7 illustrates a schematic diagram including edges with bidirectional and unidirectional migration values, including tasks D1 to D5. The edges between any two tasks are migration values ​​from one task to another, including unidirectional and bidirectional migration values.

[0191] Based on the above content, it can be seen that the edges of the task migration graph are the different transferability values ​​of the tasks at the two vertices of this edge under different algorithms, and the transferability value is the accuracy of each algorithm in migrating from one vertex task to another vertex task. The accuracy rate refers to the ratio of the number of correct results of the processing results when each algorithm processes another vertex task to the total number of results, that is, the above-mentioned performance indicator data (absolute indicator data and / or relative indicator data).

[0192] In step 12, in response to detecting an operation of selecting a clustering algorithm, clustering is performed on the task transition graph to obtain at least one task family.

[0193] In this step, the display device can display the task transition diagram, either graphically or in a table format, without limitation. Furthermore, the display interface can display a clustering algorithm for the task transition diagram, including but not limited to a K-nearest neighbor algorithm or a K-means algorithm. The number and type of clustering algorithms can be determined based on the specific scenario, without limitation. The user can select one of the clustering algorithms within the display interface to cluster the task transition diagram.

[0194] In response to detecting an operation of selecting a clustering algorithm, the display device may cluster the task transition graph to obtain at least one task family.

[0195] For example, the display device can randomly select initial center points from a preset number of categories. The device then obtains the distance between each task in the task transition graph and each initial center point, and assigns each task to the category with the smallest distance to the initial center point. The center point of each category is then updated, and the feature value of the new center point is updated to the average feature value of all samples within the category. If the preset conditions are determined to be met, the tasks in each category are determined to constitute a task family, resulting in at least one task family. Continuing with Figure 7, the first task family includes tasks D1, D2, and D3, the second task family includes task D4, and the third task family includes task D5. It will be appreciated that the distances between different tasks in Figure 7 represent different transferability, with shorter distances indicating greater transferability. Furthermore, different task families can be marked with different colored frames to facilitate locating tasks within each task family. If the preset conditions are determined not to be met, the step of obtaining the distances between each task in the task transition graph and each initial center point is repeated, i.e., the last two steps of the aforementioned steps are repeated to update the center points until the preset conditions are met.

[0196] Among them, the above-mentioned preset conditions mean that the number of iterations is greater than the preset iteration threshold, or the tasks within each category remain unchanged. The preset conditions can be selected according to the specific scenario. If the tasks can be clustered, the corresponding solution falls within the protection scope of this disclosure.

[0197] In step 13 , in response to detecting a selection operation of obtaining transfer capability, the transfer capability of each algorithm in the algorithm set is obtained, where the transfer capability represents an average value of transferability values ​​of each algorithm for migrating one task to another.

[0198] In this step, the display device may display an option to select migration capabilities when displaying the task family. If the user desires to obtain the migration capabilities of each algorithm, they may select the migration capability option, for example, by checking the migration capability box. In response to detecting the selection operation to obtain migration capabilities, the display device may obtain the migration capabilities of each algorithm in the algorithm set, as shown in Figure 8, including steps 81 and 82.

[0199] In step 81 , the display device may obtain the intra-family migration capability of each algorithm in the algorithm set within each task family and the inter-family migration capability between any two task families.

[0200] In this step, intra-family transfer capability refers to the algorithm's transfer capability when the source and target tasks belong to the same task family; inter-family transfer capability refers to the algorithm's transfer capability when the source and target tasks belong to different task families. It is understood that users can choose to use intra-family transfer capability, inter-family transfer capability, or both, depending on the degree of similarity between the two tasks. Such solutions fall within the scope of protection of this disclosure.

[0201] In this step, the display device obtains the intra-family migration capability of each algorithm in the algorithm set within each task family, as shown in FIG. 9 , which includes steps 91 to 93 .

[0202] In step 91 , the display device may obtain, for each algorithm in the algorithm set, performance indicator data of each task in each task family under the algorithm.

[0203] The method for obtaining the performance indicator data in step 91 is the same as that in step 32 . The performance indicator data in step 32 can be directly read and will not be described in detail here.

[0204] In step 92 , the display device may obtain an average value of the performance indicator data of the algorithm in each task family as the first migration capability of the algorithm in each task family.

[0205] In step 93 , the display device may obtain an average value of the first transfer capabilities of the algorithm within each task family as the intra-family transfer capability of the algorithm within each task family.

[0206] For example, for any task family δe containing Ne tasks {De0, De1, …, DeNe-1}, the display device can traverse the performance indicator data migrated under algorithm Ak; the average value of the performance indicator data is calculated as the migration capability of algorithm Ak on task family δe, which is subsequently referred to as the first migration capability Sk,e. Then, after traversing all task families, the average value of the first migration capability is obtained to obtain the migration capability within the family, as shown in Equation (4).

[0207] In formula (4), P k内 represents the migration capability within the family, and E represents the number of task families.

[0208] In this step, the display device obtains the intra-family migration capability of each algorithm in the algorithm set within each task family. Referring to FIG. 10 , the step includes steps 101 to 104 .

[0209] In step 101 , the display device may obtain any two task families as a source task family and a target task family.

[0210] In step 102 , the display device may obtain, for each algorithm in the algorithm set, performance indicator data of each source task in the source task family migrated to each target task in the target task family under the algorithm.

[0211] In step 103 , the display device may obtain an average value of performance indicator data corresponding to the source task family and the target task family as a second migration capability.

[0212] In step 104 , the display device may obtain an average value of the second transfer capabilities corresponding to all task families as the inter-family transfer capability of each algorithm between any two task families.

[0213] For example, for any two task families δe1 and δe2, take task family δe1 as the source task family and task family δe2 as the target task family, select a task Di from δe1 and a task Dj from δe2, obtain the performance index data P(Di, Dj, Ak) corresponding to the algorithm Ak, traverse all possible Di and Dj; calculate the average value of all performance index data and obtain the second migration capability P(δ e1 ,δ e2 , A k ); obtain the second migration capabilities of all possible combinations of δe1 and δe2, and calculate the average value of the second migration capabilities to obtain the inter-clan migration capability, as shown in formula (5).

[0214] In formula (5), P k间 It represents the inter-family migration capability; F represents the number of source task families and target task families. When the migration task graph is a bidirectional task graph, F = E*(E-1) / 2; when the migration task graph is a unidirectional task graph, F = E*(E-1).

[0215] In step 82 , the display device may obtain the migration capabilities of the respective algorithms according to the intra-cluster migration capabilities and the inter-cluster migration capabilities.

[0216] In this step, the display device may obtain weight values ​​of the intra-clan migration capability and the inter-clan migration capability respectively; and then obtain a weighted sum of the intra-clan migration capability and the inter-clan migration capability as the migration capability of each algorithm.

[0217] It should be noted that the above-mentioned weight values ​​can be set according to specific scenarios. When two tasks are in the same task family, the weight value of the intra-family migration capability is greater than or equal to the weight value of the inter-family migration capability; when two tasks are in different task families, the weight value of the intra-family migration capability is less than the weight value of the inter-family migration capability. Appropriate weight values ​​can be selected according to the specific scenario, and the corresponding scheme falls within the protection scope of this disclosure.

[0218] In one embodiment, after obtaining the migration capability of each algorithm, the display device can display the migration capability option. When the user desires to obtain the migration capability, the user can check and confirm the above option. In response to detecting the selection operation to obtain the migration capability, the display device can display the migration capability of each algorithm in the algorithm set in the display interface. The migration capability represents the accuracy of each algorithm in migrating one task to another. In one example, in response to detecting the operation of sorting the migration capability, the display device can display the algorithms sorted by the size of the migration capability in the display interface, as shown in Figure 11.

[0219] It should be noted that the algorithm set disclosed in the present invention is not limited to transfer learning, but can also be applied to scenarios such as small sample learning and meta-learning. Taking classification task transfer as an example, the indicators (absolute and relative indicators) for evaluating inter-task transferability and algorithm transfer performance are all based on accuracy. Specifically, the sample size of the target task data set for transfer learning is relatively large, such as 500 samples per class or even more, and the risk of overfitting is relatively small, so the inter-task transferability is relatively high, and thus the algorithm transfer performance is also relatively high; while the sample size of small samples is relatively small, such as 10 samples per class, which is prone to overfitting, so the task transferability and the corresponding algorithm transfer performance are relatively low; meta-learning is learning how to learn, the training unit is the task, and the indicator for evaluating the learning effect is still accuracy. Usually, the task transferability and the corresponding algorithm transfer performance of meta-learning are also relatively low.

[0220] Thus, the solution provided by this embodiment can, in response to detecting an operation on a task set and an algorithm set, construct a task migration graph, where each vertex of the task migration graph represents a task in the task set; then, in response to detecting an operation to select a clustering algorithm, cluster the task migration graph to obtain at least one task family; and then, in response to detecting an operation to select a migration capability, obtain the migration capability of each algorithm in the algorithm set. In this way, by quantifying the migration capability of an algorithm, this embodiment can better assess the difficulty of transfer learning, which is conducive to improving the efficiency of transfer learning.

[0221] Based on a migration capability acquisition method provided in an embodiment of the present disclosure, an embodiment of the present disclosure further provides a display control method, including:

[0222] In response to detecting the start operation, displaying the candidate tasks and the candidate algorithms in the display interface;

[0223] In response to detecting an operation of selecting a candidate task, displaying the selected task set in a display interface;

[0224] In response to detecting an operation of selecting at least one candidate algorithm, displaying the selected algorithm set in a display interface;

[0225] In response to detecting an operation of determining the task set and the algorithm set, a task migration graph constructed based on the task set and the algorithm set is displayed in a display interface; each vertex of the task migration graph is a task in the task set, and an edge of the task migration graph refers to different transferability values ​​of the tasks at two vertices of the edge under different algorithms; the transferability value is determined by accuracy performance indicator data of each algorithm migrating a task from one vertex to another to process each vertex task;

[0226] In response to detecting a selection operation of a clustering algorithm for clustering tasks in the task transition graph, at least one task family after clustering the task transition graph according to the clustering algorithm is displayed, each task family including a plurality of tasks.

[0227] In one embodiment, the method further comprises:

[0228] In response to detecting a selection operation of obtaining migration capability, the migration capability of each algorithm in the algorithm set is displayed in the display interface, where the migration capability represents an average value of the migration capability values ​​of each algorithm for migrating one task to another task.

[0229] In one embodiment, the method further comprises:

[0230] In response to detecting an operation of sorting the migration capabilities, various algorithms sorted according to the size relationship of the migration capabilities are displayed in a display interface.

[0231] It should be noted that the display control method shown in this embodiment is described in the process of describing a migration capability acquisition method. Please refer to the content of the above method embodiment and will not be repeated here.

[0232] Based on the migration capability acquisition method provided in the embodiment of the present disclosure, the embodiment of the present disclosure further provides a migration capability acquisition device. Referring to FIG12 , the device includes:

[0233] A migration graph acquisition module 121 is configured to construct a task migration graph in response to detecting operations on a task set and an algorithm set, wherein each vertex of the task migration graph is a task in the task set, and an edge of the task migration graph is a task with different transferability values ​​under different algorithms between the tasks at two vertices of the edge; the transferability value is determined by the accuracy performance indicator data of each algorithm migrating tasks from one vertex to another for processing each vertex task;

[0234] A task family acquisition module 122 is configured to cluster the task transition graph in response to detecting an operation of selecting a clustering algorithm to obtain at least one task family;

[0235] The migration capability acquisition module 123 is configured to acquire the migration capability of each algorithm in the algorithm set in response to detecting a selection operation for acquiring migration capability, wherein the migration capability represents an average value of the transferability values ​​of each algorithm for migrating one task to another.

[0236] In one embodiment, the transition graph acquisition module includes:

[0237] A task set acquisition submodule is used to acquire a task set; the task set includes a plurality of tasks;

[0238] An algorithm set acquisition submodule is used to acquire an algorithm set; the algorithm set includes several algorithms that implement transfer learning between any two tasks;

[0239] a performance index acquisition submodule, configured to perform transfer learning on any two tasks in the task set using any one algorithm in the algorithm set, and obtain performance index data corresponding to the task-algorithm combination; the performance index data includes absolute index data and / or relative index data; the absolute index data refers to the accuracy of each algorithm when migrating from one vertex task to another; the relative index data refers to the difference in accuracy of each algorithm when processing two vertex tasks on the same edge;

[0240] A transferability value acquisition submodule, configured to obtain transferability values ​​of any two tasks based on the performance indicator data;

[0241] The migration graph construction submodule is used to construct the task migration graph with each task in the task set as a vertex and with the transferability value between any two tasks as an edge.

[0242] In one embodiment, the performance index acquisition submodule includes:

[0243] A first algorithm acquisition unit is configured to train the algorithm using a first portion of samples of the source task in the two tasks to obtain a trained first algorithm;

[0244] A first accuracy rate acquisition unit, configured to verify the first algorithm using a second portion of samples of the source task to obtain a first accuracy rate;

[0245] a second algorithm acquiring unit, configured to train the first algorithm using a first portion of samples of a target task among the two tasks to obtain a trained second algorithm;

[0246] A second accuracy rate obtaining unit, configured to verify the second algorithm using a second portion of samples of the target task to obtain a second accuracy rate;

[0247] A performance indicator acquisition unit is used to acquire the performance indicator data according to the first accuracy rate and the second accuracy rate.

[0248] In one embodiment, the performance indicator acquisition unit includes:

[0249] an absolute indicator acquisition subunit, configured to determine the second accuracy rate as absolute indicator data;

[0250] a relative indicator acquisition subunit, configured to acquire a difference between the second accuracy rate and the first accuracy rate as relative indicator data;

[0251] The performance indicator determination subunit is configured to determine the absolute indicator data and / or the relative indicator data as the performance indicator data.

[0252] In one embodiment, the mobility value includes a forward mobility value, and the mobility value acquisition submodule includes:

[0253] A performance indicator acquisition unit, configured to acquire performance indicator data of each algorithm in the algorithm set during transfer learning from a source task to a target task in any two tasks;

[0254] a positive value obtaining unit, configured to obtain an average value of the performance indicator data corresponding to each algorithm as a positive transferability value of the two tasks;

[0255] Alternatively, the portability value includes a reverse portability value, and the portability value acquisition submodule includes:

[0256] A performance indicator acquisition unit, configured to acquire performance indicator data of each algorithm in the algorithm set during transfer learning from a target task to a source task in any two tasks;

[0257] The reverse value acquisition unit is used to obtain the average value of the performance indicator data corresponding to each algorithm as the reverse transferability value of the two tasks.

[0258] In one embodiment, the portability value includes a bidirectional portability value, and the portability value acquisition submodule includes:

[0259] The bidirectional value acquisition unit is configured to acquire an average value of the forward transferability value and the reverse transferability value as the bidirectional transferability value of the arbitrary two tasks.

[0260] In one embodiment, the task transition graph includes a directed transition graph and an undirected transition graph;

[0261] When the edge of the task transition graph is a forward transferability value or a reverse transferability value, the task transition graph is a directed transition graph;

[0262] When the edges of the task transition graph have bidirectional transferability values, the task transition graph is an undirected transition graph.

[0263] In one embodiment, the task family acquisition module includes:

[0264] The initial point selection submodule is used to randomly select the initial center points of a preset number of categories;

[0265] A task division submodule is used to obtain the distance between each task and each initial center point in the task migration graph, and divide each task into the category of the initial center point with the smallest distance;

[0266] The center point update submodule is used to update the center point of each category and update the feature value of the new center point to the average feature value of all samples in the category;

[0267] The judgment submodule is used to determine that tasks of each category constitute a task family and obtain at least one task family when it is determined that the preset conditions are met; when it is determined that the preset conditions are not met, re-execute the step of obtaining the distance between each task and each initial center point in the task migration diagram.

[0268] In one embodiment, the preset condition means that the number of iterations is greater than a preset iteration threshold, or the tasks in each category remain unchanged.

[0269] In one embodiment, the migration capability acquisition module includes:

[0270] The intra-family capability acquisition submodule is used to obtain the intra-family transfer capability of each algorithm in the algorithm set within each task family and the inter-family transfer capability between any two task families;

[0271] The inter-family capability acquisition submodule is used to obtain the inter-family transfer capability of each algorithm in the algorithm set between any two task families;

[0272] The migration capability acquisition submodule is configured to acquire the migration capability of each algorithm according to the intra-cluster migration capability and the inter-cluster migration capability.

[0273] In one embodiment, the intra-family capability acquisition submodule includes:

[0274] A performance indicator acquisition unit, configured to acquire, for each algorithm in the algorithm set, performance indicator data of each task in each task family under the algorithm;

[0275] a first capability acquisition unit, configured to acquire an average value of performance indicator data of the algorithm in each task family as a first transfer capability of the algorithm in each task family;

[0276] The intra-family capability acquisition unit is configured to acquire an average value of the first transfer capabilities of the algorithm within each task family as the intra-family transfer capability of the algorithm within each task family.

[0277] In one embodiment, the inter-family capability acquisition submodule includes:

[0278] A task family acquisition unit is used to acquire any two task families as a source task family and a target task family;

[0279] a performance indicator acquisition unit, configured to acquire, for each algorithm in the algorithm set, performance indicator data of each source task in the source task family migrated to each target task in the target task family under the algorithm;

[0280] A second capability acquisition unit, configured to acquire an average value of performance indicator data corresponding to the source task family and the target task family as a second migration capability;

[0281] The inter-family capability acquisition unit is configured to acquire an average value of the second transfer capabilities corresponding to all task families as the inter-family transfer capability of each algorithm between any two task families.

[0282] In one embodiment, the migration capability acquisition submodule includes:

[0283] a weight value obtaining unit, configured to obtain weight values ​​of the intra-clan migration capability and the inter-clan migration capability respectively;

[0284] The migration capability acquisition unit is configured to acquire a weighted sum of the intra-cluster migration capability and the inter-cluster migration capability as the migration capability of each algorithm.

[0285] In one embodiment, when two tasks are in the same task family, the weight value of the intra-family migration capability is greater than or equal to the weight value of the inter-family migration capability;

[0286] When two tasks are in different task families, the weight value of the intra-family migration capability is smaller than the weight value of the inter-family migration capability.

[0287] It should be noted that the apparatus shown in this embodiment matches the contents of the method embodiment, and reference may be made to the contents of the above method embodiment, which will not be repeated here.

[0288] Based on a display control method provided in an embodiment of the present disclosure, an embodiment of the present disclosure further provides a display control device, the device comprising:

[0289] a task algorithm display module, configured to display candidate tasks and candidate algorithms in a display interface in response to detecting a start operation;

[0290] a task set selection module, configured to display the selected task set in the display interface in response to detecting an operation of selecting a candidate task;

[0291] an algorithm set selection module, configured to display the selected algorithm set in a display interface in response to detecting an operation of selecting at least one candidate algorithm;

[0292] a migration graph display module, configured to, in response to detecting an operation of determining the task set and the algorithm set, display a task migration graph constructed based on the task set and the algorithm set within a display interface; wherein each vertex of the task migration graph is a task in the task set, and an edge of the task migration graph is a task having different migration values ​​under different algorithms at two vertices of the edge; wherein the migration value is determined by accuracy performance indicator data of each algorithm migrating a task from one vertex to another for processing each vertex task;

[0293] The task family display module is used to display at least one task family after clustering the task transition graph according to the clustering algorithm in response to detecting the selection operation of the clustering algorithm for clustering the tasks of the task transition graph, each task family including multiple tasks.

[0294] In one embodiment, the apparatus further comprises:

[0295] The migration capability display module is used to display the migration capability of each algorithm in the algorithm set in a display interface in response to detecting a selection operation for obtaining migration capability, wherein the migration capability represents the accuracy of each algorithm in migrating one task to another task.

[0296] In one embodiment, the apparatus further comprises:

[0297] The migration capability ranking module is configured to, in response to detecting an operation of ranking the migration capabilities, display various algorithms ranked according to the size relationship of the migration capabilities in a display interface.

[0298] It should be noted that the apparatus shown in this embodiment matches the contents of the method embodiment, and reference may be made to the contents of the above method embodiment, which will not be repeated here.

[0299] In some possible embodiments, a display device is provided, including:

[0300] processor and memory;

[0301] The memory is used to store a computer program executable by the processor;

[0302] The processor is configured to execute the computer program in the memory to implement the above method.

[0303] In some possible embodiments, a non-transitory computer-readable storage medium is provided. When an executable computer program in the storage medium is executed by a processor, the above-mentioned method can be implemented.

[0304] The terms used in this disclosure are for the purpose of describing specific embodiments only and are not intended to limit the disclosure. Unless otherwise defined, technical or scientific terms used in this disclosure should have the ordinary meaning understood by a person of ordinary skill in the art to which this disclosure belongs. The use of "a" or "an," and similar terms in this disclosure and the claims does not indicate a limitation of quantity, but rather indicates the presence of at least one. "Multiple" means at least two. "Include" or "comprising," and similar terms mean that the elements or objects preceding "include" or "comprise" include the elements or objects listed after "include" or "comprise," and their equivalents, and do not exclude other elements or objects. "Connected" or "connected," and similar terms are not limited to physical or mechanical connections and may include electrical connections, whether direct or indirect. As used in this disclosure and the appended claims, the singular forms "a," "the," and "the" are intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0305] As for the method embodiment, since it basically corresponds to the apparatus embodiment, the relevant parts can be referred to the partial description of the apparatus embodiment. The method embodiment and the apparatus embodiment complement each other.

[0306] The above description is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present disclosure should be included in the scope of protection of the present disclosure.

Claims

1. A migration capability acquisition method, characterized in that: The method comprises: In response to detecting the operation on the task set and the algorithm set, a task migration graph is constructed, wherein each vertex of the task migration graph is a task in the task set, and an edge of the task migration graph refers to different migration values ​​of the tasks at two vertices of the edge under different algorithms; the migration value is determined by the performance indicator data of each algorithm processing the tasks at each vertex; In response to detecting an operation of selecting a clustering algorithm, clustering the task transition graph to obtain at least one task family; In response to detecting a selection operation of obtaining transfer capability, the transfer capability of each algorithm in the algorithm set is obtained, where the transfer capability represents an average value of transferability values ​​of each algorithm for migrating one task to another task.

2. The method according to claim 1, characterized in that Build a task migration graph, including: Obtaining a task set and an algorithm set; the task set includes a plurality of tasks, and the algorithm set includes a plurality of algorithms for implementing transfer learning between any two tasks; Using any one of the algorithms in the algorithm set to perform transfer learning on any two tasks in the task set, obtain performance indicator data corresponding to the task algorithm combination; the performance indicator data includes absolute indicator data and / or relative indicator data; the absolute indicator data refers to the accuracy of each algorithm migrating from one vertex task to another vertex task; the relative indicator data refers to the difference in accuracy of each algorithm processing two vertex tasks of the same edge; Obtain transferability values ​​of any two tasks according to the performance indicator data; The task migration graph is constructed with each task in the task set as a vertex and with the migration value between any two tasks as an edge.

3. The method according to claim 2, characterized in that Using any one algorithm in the algorithm set to perform transfer learning on any two tasks in the task set, the performance indicator data corresponding to the task algorithm combination is obtained, including: Using a first part of samples of the source task in the two tasks to train the algorithm to obtain a trained first algorithm; Using a second portion of samples of the source task to verify the first algorithm, to obtain a first accuracy rate; Using a first part of samples of a target task in the two tasks to train the first algorithm to obtain a trained second algorithm; Using a second part of samples of the target task to verify the second algorithm, and obtain a second accuracy rate; The performance indicator data is acquired according to the first accuracy rate and the second accuracy rate.

4. The method according to claim 3, characterized in that Acquiring the performance indicator data according to the first accuracy rate and the second accuracy rate includes: Determining the second accuracy rate as absolute indicator data; Obtaining a difference between the second accuracy rate and the first accuracy rate as relative indicator data; The absolute indicator data and / or the relative indicator data are determined as the performance indicator data.

5. The method according to claim 2, characterized in that: The transferability value includes a positive transferability value, and the transferability values ​​of any two tasks are obtained according to the performance indicator data, including: Obtaining performance indicator data of each algorithm in the algorithm set during transfer learning from a source task to a target task in any two tasks; Obtaining the average value of the performance indicator data corresponding to each algorithm as the positive transferability value of the two tasks; or, The transferability value includes a reverse transferability value, and the transferability values ​​of any two tasks are obtained according to the performance indicator data, including: Obtaining performance indicator data of each algorithm in the algorithm set during transfer learning from a target task to a source task in any two tasks; The average value of the performance indicator data corresponding to each algorithm is obtained as the reverse transferability value of the two tasks.

6. The method according to claim 5, characterized in that The transferability value includes a bidirectional transferability value, and the transferability values ​​of any two tasks are obtained according to the performance indicator data, including: An average value of the forward transferability value and the reverse transferability value is obtained as the bidirectional transferability value of the arbitrary two tasks.

7. The method according to claim 6, characterized in that The task migration graph includes a directed migration graph and an undirected migration graph; When the edge of the task migration graph is a positive transferability value or a negative transferability value, the task migration graph is a directed migration graph; When the edge of the task transition graph has a bidirectional transferability value, the task transition graph is an undirected transition graph.

8. The method according to claim 1, characterized in that: Clustering the task migration graph to obtain at least one task family, including: Randomly select the initial center points of a preset number of categories; Obtaining the distance between each task and each initial center point in the task migration graph, and classifying each task into the category of the initial center point with the smallest distance; Update the center point of each category, and update the feature value of the new center point to the average feature value of all samples in the category; When it is determined that the preset conditions are met, it is determined that the tasks of each category constitute a task family, and at least one task family is obtained; When it is determined that the preset condition is not met, the step of obtaining the distance between each task and each initial center point in the task transition graph is re-executed.

9. The method according to claim 8, characterized in that The preset condition means that the number of iterations is greater than a preset iteration number threshold, or the tasks in each category remain unchanged.

10. The method according to claim 1, characterized in that Get the migration capability of each algorithm in the algorithm set, including: Obtain the intra-family transfer capability of each algorithm in the algorithm set within each task family and the inter-family transfer capability between any two task families; The migration capabilities of the respective algorithms are obtained according to the intra-cluster migration capabilities and the inter-cluster migration capabilities.

11. The method according to claim 10, characterized in that Obtain the intra-family migration capability of each algorithm in the algorithm set within each task family, including: For each algorithm in the algorithm set, obtaining performance indicator data of each task in each task family under the algorithm; Obtaining an average value of the performance indicator data of the algorithm in each task family as the first migration capability of the algorithm in each task family; An average value of the first transfer capability of the algorithm in each task family is obtained as the intra-family transfer capability of the algorithm in each task family.

12. The method according to claim 10, characterized in that Obtain the inter-family migration capability of each algorithm in the algorithm set between any two task families, including: Get any two task families as the source task family and the target task family; For each algorithm in the algorithm set, obtaining performance indicator data of each source task in the source task family migrated to each target task in the target task family under the algorithm; Obtaining an average value of performance indicator data corresponding to the source task family and the target task family as a second migration capability; An average value of the second migration capabilities corresponding to all task families is obtained as the inter-family migration capability of each algorithm between any two task families.

13. The method according to claim 10, characterized in that Acquiring the migration capability of each algorithm according to the intra-cluster migration capability and the inter-cluster migration capability includes: respectively obtaining weight values ​​of the intra-clan migration capability and the inter-clan migration capability; A weighted sum of the intra-cluster migration capability and the inter-cluster migration capability is obtained as the migration capability of each algorithm.

14. The method according to claim 13, characterized in that When two tasks are in the same task family, the weight value of the intra-family migration capability is greater than or equal to the weight value of the inter-family migration capability; When two tasks are in different task families, the weight value of the intra-family migration capability is smaller than the weight value of the inter-family migration capability.

15. A display control method, characterized in that: The method comprises: In response to detecting a start operation, displaying candidate tasks and candidate algorithms in a display interface; In response to detecting an operation of selecting a candidate task, displaying the selected task set in a display interface; In response to detecting an operation of selecting at least one candidate algorithm, displaying the selected algorithm set in a display interface; In response to detecting the operation of determining the task set and the algorithm set, a task migration graph constructed according to the task set and the algorithm set is displayed in a display interface; each vertex of the task migration graph is a task in the task set, and an edge of the task migration graph refers to different transferability values ​​of the tasks at two vertices of the edge under different algorithms; the transferability value is determined by the accuracy performance indicator data of each algorithm migrating from one vertex task to another to process each vertex task; In response to detecting a selection operation of a clustering algorithm for clustering the tasks of the task transition graph, at least one task family after the task transition graph is clustered according to the clustering algorithm is displayed, each task family including a plurality of tasks.

16. The method according to claim 15, characterized in that The method further comprises: In response to detecting a selection operation of obtaining migration capability, the migration capability of each algorithm in the algorithm set is displayed in the display interface, where the migration capability represents an average value of the migration capability values ​​of each algorithm for migrating one task to another task.

17. The method according to claim 15, characterized in that The method further comprises: In response to detecting an operation of sorting the migration capabilities, various algorithms sorted according to the size relationship of the migration capabilities are displayed in a display interface.

18. A migration capability acquisition device, characterized in that: The device comprises: A migration graph acquisition module is used to construct a task migration graph in response to detecting an operation on a task set and an algorithm set, wherein each vertex of the task migration graph is a task in the task set, and an edge of the task migration graph refers to different transferability values ​​of the tasks at two vertices of the edge under different algorithms; the transferability value is determined by the accuracy performance indicator data of each algorithm migrating from one vertex task to another to process each vertex task; A task family acquisition module, configured to cluster the task transition graph to obtain at least one task family in response to detecting an operation of selecting a clustering algorithm; The migration capability acquisition module is used to acquire the migration capability of each algorithm in the algorithm set in response to detecting a selection operation of acquiring the migration capability, wherein the migration capability represents an average value of the transferability values ​​of each algorithm in migrating one task to another task.

19. A display control device, characterized in that: The device comprises: A task algorithm display module, for displaying candidate tasks and candidate algorithms in a display interface in response to detecting a start operation; A task set selection module, configured to display the selected task set in the display interface in response to detecting an operation of selecting a candidate task; an algorithm set selection module, configured to display the selected algorithm set in a display interface in response to detecting an operation of selecting at least one candidate algorithm; A migration graph display module, for displaying a task migration graph constructed according to the task set and the algorithm set in a display interface in response to detecting an operation of determining the task set and the algorithm set; each vertex of the task migration graph is a task in the task set, and an edge of the task migration graph refers to different migration values ​​of the tasks at two vertices of the edge under different algorithms; the migration value is determined by the accuracy performance index data of each algorithm migrating from one vertex task to another to process each vertex task; The task family display module is used to display at least one task family after clustering the task transition graph according to the clustering algorithm in response to detecting the selection operation of the clustering algorithm for clustering the tasks of the task transition graph, each task family including multiple tasks.

20. A display device, characterized in that: include: Processor and memory; The memory is used to store a computer program executable by the processor; The processor is configured to execute the computer program in the memory to implement the method according to any one of claims 1 to 17.

21. A non-transitory computer-readable storage medium, characterized in that: When the executable computer program in the storage medium is executed by a processor, the method according to any one of claims 1 to 17 can be implemented.

Images