WO2023221371A1

WO2023221371A1 - Task search method and apparatus, server and storage medium

Info

Publication number: WO2023221371A1
Application number: PCT/CN2022/123598
Authority: WO
Inventors: 郑辉煌; 陈特峰; 陈浩泽; 王悦; 王震; 刘益群; 孙黎; 姜程; 石晓伟; 蓝翔
Original assignee: 北京百度网讯科技有限公司
Priority date: 2022-05-19
Filing date: 2022-09-30
Publication date: 2023-11-23
Also published as: CN114968520A; CN114968520B

Abstract

Disclosed are a task search method and apparatus, a server and a storage medium. A specific implementation scheme comprises: acquiring a task scheduling strategy corresponding to an initial task, wherein the task scheduling strategy comprises a reasoning duration task scheduling strategy and a training task scheduling strategy, the reasoning duration task scheduling strategy is used for adjusting a reasoning duration corresponding to the initial task in a reasoning scenario, and the training task scheduling strategy is used for adjusting a search duration corresponding to the initial task in a training scenario; acquiring a training operation mode corresponding to the initial task; and searching the initial task by using the reasoning duration task scheduling strategy, the training task scheduling strategy and the training operation mode to obtain a target task.

Description

Task search method and device, server and storage medium

Cross-references to related applications

This application is filed based on a Chinese patent application with application number 2022105481337 and a filing date of May 19, 2022, and claims the priority of the Chinese patent application. The entire content of the Chinese patent application is hereby incorporated by reference into this application.

Technical field

The present disclosure relates to the field of computer technology, and specifically to a task search method and device, a server and a storage medium.

Background technique

Deep learning is a research direction in the field of machine learning. It can learn the inherent laws and representation levels of sample data, so that the machine can imitate human activities such as audio-visual and thinking. In related technologies, a deep learning compiler can be used to train and infer a deep learning model. However, in related technologies, using a deep learning compiler to automatically optimize a deep learning model requires a long search time and is not suitable for training scenarios.

Contents of the invention

The present disclosure provides a task search method and device, a server and a storage medium. The main purpose is to reduce the search time and improve the applicability of the task search solution.

According to an aspect of the present disclosure, a task search method is provided, including:

Obtain the task scheduling policy corresponding to the initial task, where the task scheduling policy includes an inference duration task scheduling policy and a training task scheduling policy. The inference duration task scheduling policy is used to adjust the inference duration corresponding to the initial task in the inference scenario. , the training task scheduling policy is used to adjust the search duration corresponding to the initial task in the training scenario;

Obtain the training operation mode corresponding to the initial task;

The initial task is searched using the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode to obtain the target task.

According to another aspect of the present disclosure, a task search device is provided, including:

A policy acquisition unit is used to obtain a task scheduling policy corresponding to the initial task, wherein the task scheduling policy includes an inference duration task scheduling policy and a training task scheduling policy, and the inference duration task scheduling policy is used to adjust the inference duration task scheduling policy in the inference scenario. The inference duration corresponding to the initial task, and the training task scheduling policy is used to adjust the search duration corresponding to the initial task in the training scenario;

A mode acquisition unit, used to acquire the training operation mode corresponding to the initial task;

A task acquisition unit is used to search the initial task using the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode to obtain the target task.

According to another aspect of the present disclosure, a server is provided, including:

at least one processor; and

a memory communicatively connected to the at least one processor; wherein,

The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of the preceding aspects. .

According to another aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause the computer to perform the method according to any one of the preceding aspects.

According to another aspect of the present disclosure, there is provided a computer program product comprising a computer program that, when executed by a processor, implements the method of any one of the preceding aspects.

According to another aspect of the present disclosure, a computer program is provided, the computer program including computer program code, when the computer program code is run on a computer, causing the computer to perform the method according to any one of the preceding aspects. method.

According to another aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the above-mentioned steps are implemented. The method of any one of the aspects.

In one or more embodiments of the present disclosure, by obtaining a task scheduling policy corresponding to the initial task, where the task scheduling policy includes an inference duration task scheduling policy and a training task scheduling policy, the inference duration task scheduling policy is used Adjust the inference duration corresponding to the initial task in the inference scenario, and the training task scheduling policy is used to adjust the search duration corresponding to the initial task in the training scenario; obtain the training operation mode corresponding to the initial task; use the inference duration The task scheduling strategy, the training task scheduling strategy and the training operation mode search the initial task to obtain the target task. Therefore, the search time can be reduced and the applicability of the task search solution can be improved.

It should be understood that what is described in this section is not intended to identify key or important features of the embodiments of the disclosure, nor is it intended to limit the scope of the disclosure. Other features of the present disclosure will become readily understood from the following description.

Description of the drawings

The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure. in:

Figure 1 is a schematic flowchart of a task search method according to the first embodiment of the present disclosure;

Figure 2 is a schematic flowchart of a task search method according to a second embodiment of the present disclosure;

Figure 3 is a flowchart of selecting a scheduling strategy for inference duration tasks provided according to an embodiment of the present disclosure;

Figure 4a is a schematic structural diagram of a first task search device used to implement the task search method according to an embodiment of the present disclosure;

Figure 4b is a schematic structural diagram of a second task search device used to implement the task search method according to the embodiment of the present disclosure;

Figure 4c is a schematic structural diagram of a third task search device used to implement the task search method according to the embodiment of the present disclosure;

Figure 4d is a schematic structural diagram of a fourth task search device used to implement the task search method according to the embodiment of the present disclosure;

Figure 4e is a schematic structural diagram of a fifth task search device used to implement the task search method according to the embodiment of the present disclosure;

Figure 4f is a schematic structural diagram of a sixth task search device used to implement the task search method according to the embodiment of the present disclosure;

Figure 5 is a block diagram of a server used to implement the task search method of an embodiment of the present disclosure.

Detailed ways

Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings, in which various details of the embodiments of the present disclosure are included to facilitate understanding and should be considered to be exemplary only. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the disclosure. Also, descriptions of well-known functions and constructions are omitted from the following description for clarity and conciseness.

With the development of science and technology, server technology has become increasingly mature, which has improved the convenience of users' production and life. In the server application scenario, users can train and infer deep learning models through the server.

According to some embodiments, the deep learning compiler is a compiler software used to solve multiple hardware platforms and deep learning docking problems. The deep learning compiler can be composed of multiple layers of intermediate representation (Intermediate Representation, IR) and corresponding operating methods. Among them, the high-level intermediate expression is used to express the deep learning calculation graph structure, which includes the representation of deep learning variables (Variable) and operators (Operator). The low-level intermediate representation is the specific calculation of the operator, such as the matrix multiplication operator of the high-level intermediate representation. The low-level intermediate representation is more specific loop, multiplication, and summation operations. These operations are also closer to the low-level instructions of the hardware.

In some embodiments, a deep learning compiler may be provided in the server. Furthermore, when the server trains and infers the deep learning model, it can use the deep learning compiler to search for a configuration that matches the deep learning model.

In some embodiments, when the server uses the deep learning compiler to search for a configuration that matches the deep learning model, it needs to input the calculation graph optimized by deep learning to the deep learning compiler for the deep learning compiler to search.

However, when the deep learning compiler searches the calculation graph, due to the need to try and search for different optimization configurations, the search volume is large, and the entire calculation graph needs to be searched. Therefore, the server takes longer to search.

It is easy to understand that when inferring a deep learning model, you can spend time searching offline and then infer the deep learning model online. However, when training a deep learning model, users need to understand the training effect of the deep learning model through training iterations of the deep learning model. If the search time is too long, the training time of the deep learning model will be increased. Since automatic optimization technology requires a long optimization time, often dozens of hours, it is only suitable for deep learning inference scenarios, but its applicability in training scenarios is poor.

The present disclosure will be described in detail below with reference to specific embodiments.

In the first embodiment, as shown in Figure 1, Figure 1 is a schematic flowchart of a task search method according to the first embodiment of the present disclosure. This method can be implemented relying on a computer program and can be run on a device that performs task search. , can be a server with task search function. The computer program can be integrated into an application or run as a stand-alone utility application.

Specifically, the task search method includes: S101-S103.

S101. Obtain the task scheduling policy corresponding to the initial task.

According to some embodiments, the task Task refers to the computational graph subgraph used when the deep learning neural network model performs training and inference. This task is not specific to a fixed task. For example, the task can change when the computational graph changes. When the computational graph subgraph changes, the task can also change.

In some embodiments, the initial task refers to a task that requires tuning search. This initial task is not specific to a fixed task. For example, this initial task can change when the computational graph changes. When the computational graph subgraph changes, this initial task can also change.

According to some embodiments, the task scheduling policy refers to the policy adopted by the server when searching for initial tasks. The task scheduling strategy includes but is not limited to inference duration task scheduling strategy, training task scheduling strategy, and so on. This task scheduling strategy does not specifically refer to a fixed strategy. For example, when the initial task changes, the task scheduling policy can change. When the computation graph changes, the task scheduling policy can also change.

In some embodiments, the inference duration task scheduling policy refers to a strategy for adjusting the inference duration corresponding to the initial task in the inference scenario. The inference duration task scheduling strategy does not specifically refer to a fixed strategy. For example, when the task scheduling policy changes, the inference duration task scheduling policy can change. When the initial task changes, the inference duration task scheduling policy can also change.

In some embodiments, the training task scheduling policy refers to a policy used to adjust the search duration corresponding to the initial task in the training scenario. The training task scheduling strategy does not specifically refer to a fixed strategy. For example, when the task scheduling policy changes, the training task scheduling policy can change. When the initial task changes, the training task scheduling policy can also change.

It is easy to understand that when the server performs task search, the server can obtain the task scheduling policy corresponding to the initial task.

S102: Obtain the training operation mode corresponding to the initial task.

According to some embodiments, the training operation mode refers to the training operation mode adopted by the server when training the deep learning neural network model. The way the training is run is not specific to a fixed way. For example, when the deep learning neural network model changes, the way the training is run can change. When the initial task changes, the way the training is run can also change.

It is easy to understand that when the server obtains the task scheduling policy corresponding to the initial task, the server can obtain the training operation mode corresponding to the initial task.

S103. Use the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode to search for the initial task and obtain the target task.

According to some embodiments, the target task refers to a task obtained after performing a tuning search on the initial task. The target task does not refer to a fixed task. For example, the target task can change when the initial task changes. When the task scheduling policy changes, the target task can also change.

It is easy to understand that when the server obtains the training operation mode corresponding to the initial task, the server can use the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode to search for the initial task and obtain the target task.

In the embodiment of the present disclosure, by obtaining the task scheduling strategy corresponding to the initial task; obtaining the training operation mode corresponding to the initial task; and using the inference duration task scheduling strategy, training task scheduling strategy and training operation mode to search for the initial task to obtain the target Task. Therefore, by adopting the inference time task scheduling strategy, the task search time when inferring the deep learning model can be reduced, and the inference time of the deep learning model can be reduced. At the same time, by using the training task scheduling strategy and training operation mode, the task search time of the deep learning model can be reduced. The task search time when the model is trained can reduce the training time of the deep learning model. In turn, the task search time can be reduced while improving the applicability of the task search solution, making the task search method applicable to inference scenarios and training scenarios.

Please refer to Figure 2, which is a schematic flowchart of a task search method according to a second embodiment of the present disclosure. Specifically, the method includes: S201-S208.

S201: When it is determined that the second target configuration information corresponding to the initial task exists in the cache, obtain the hardware code based on the second target configuration information.

According to some embodiments, the target configuration information refers to tuning configuration information corresponding to the task. The target configuration information does not refer to certain fixed information. The target configuration information includes but is not limited to loop block size, vectorization, loop unrolling, calculation position adjustment, thread parallelism, graphics processing unit (GPU) parallelism, etc.

In some embodiments, the second target configuration information refers to tuning configuration information corresponding to the initial task. The second target configuration information does not specifically refer to certain fixed information. For example, when the initial task changes, the second target configuration information may change. When the server obtains the information modification instruction for the second target configuration information, the second target configuration information may also change.

In some embodiments, hardware code refers to code generated by running on hardware. The hardware code is not specific to a fixed code. For example, when the second target configuration information changes, the hardware code may change. When the initial task changes, this hardware code can also change.

In some embodiments, when the server obtains hardware code based on the second target configuration information, the server may convert the second target configuration information into the underlying IR. In turn, the server can generate hardware code through the underlying IR.

It is easy to understand that when the server performs a task search, if the server determines that the second target configuration information corresponding to the initial task exists in the cache, the server can obtain the hardware code based on the second target configuration information.

According to some embodiments, the server may partition the computational graph to obtain at least one task.

S202: Control the hardware to run the hardware code and obtain the operation information corresponding to the initial task.

According to some embodiments, the running information refers to the running information of the initial task in hardware. This operating information does not specifically refer to a certain fixed operating information. The operating information includes operating speed. For example, this running information can change when the initial task changes. When the hardware code changes, this operating information can also change.

In some embodiments, when the server performs a task search, the server needs to arrange and combine the target configuration information and search for the arrangement and combination that results in the lowest running speed corresponding to the task.

In some embodiments, when the server needs to arrange and combine the target configuration information and search for the permutation and combination with the lowest running speed corresponding to the task, the search algorithms used include but are not limited to genetic search algorithms, exhaustive search algorithms, and grid search. Algorithms and more.

In some embodiments, the server can also directly search from the search space for the permutation and combination that results in the lowest running speed corresponding to the task. The search space refers to the space including all executable permutations and combinations corresponding to the target configuration information.

According to some embodiments, when the server controls the hardware to run the hardware code and obtains the running information corresponding to the initial task, if the server determines that the running information meets the running information conditions, the server can end the search for the initial task and save the initial task. Set as target task.

In some embodiments, the running information conditions refer to the conditions used by the server to determine whether the initial task needs to perform a tuning search. This operating information condition does not specifically refer to a fixed condition. For example, when the server obtains a condition modification instruction for a running information condition, the running information condition may change.

It is easy to understand that when the server obtains the hardware code based on the second target configuration information, the server can control the hardware to run the hardware code. Furthermore, the server can obtain the running information corresponding to the initial task.

S203. Display the inference duration task scheduling policy set corresponding to the initial task on the display interface.

According to some embodiments, the display interface refers to the display interface used when the server interacts with the user. The display interface does not specifically refer to a fixed interface. For example, when the server changes, the presentation interface can change.

In some embodiments, the inference duration task scheduling policy set refers to a set aggregated from at least one inference duration task scheduling policy. The set of inference duration task scheduling strategies does not specifically refer to a fixed set. For example, when the thrust duration corresponding to the inference duration task scheduling policy changes, the inference duration task scheduling policy set may change. When the number of inference duration task scheduling policies changes, the set of inference duration task scheduling policies may change.

In some embodiments, different inference duration task scheduling strategies correspond to different inference durations and tasks corresponding to different speed improvement values. For example, the inference duration task scheduling policy may be, for example, 10% of the search duration corresponds to a 90% speed improvement value; for example, the inference duration task scheduling policy may be, for example, 20% of the search duration, corresponding to a 92% speed improvement value; for example, the inference duration The task scheduling policy may be, for example, that 100% of the search time corresponds to 100% of the speed improvement value.

In some embodiments, the set of inference duration task scheduling policies includes at least one inference duration task scheduling policy, which includes but is not limited to long-duration task scheduling policies, short-duration task scheduling policies, and the like. Among them, the inference time of the long-term task scheduling strategy is longer than the inference time of the short-term task scheduling strategy. The inference performance of the long-duration task scheduling strategy is higher than that of the short-duration task scheduling strategy.

It is easy to understand that when the server performs task search, the server can display a set of inference duration task scheduling strategies corresponding to the initial task on the display interface.

S204: Obtain the selection instructions input for the inference duration task scheduling policy set.

According to some embodiments, the selection instruction refers to the instruction obtained by the terminal and entered by the user when selecting the inference duration task scheduling strategy. This selection instruction does not refer to a fixed instruction. The selection instructions include but are not limited to voice selection instructions, click selection instructions, and so on. For example, when the server detects that the user speaks voice information corresponding to any inference duration task scheduling policy, the server can obtain the selection instruction corresponding to the inference duration task scheduling policy. When the server detects that the user clicks the selection button corresponding to any inference duration task scheduling policy, the server can also obtain the selection instruction corresponding to the inference duration task scheduling policy.

It is easy to understand that when the server displays the inference duration task scheduling policy set corresponding to the initial task on the display interface, the server can obtain the selection instruction input for the inference duration task scheduling policy set.

S205: Obtain the inference duration task scheduling policy corresponding to the selection instruction and obtain the training task scheduling policy.

According to some embodiments, FIG. 3 is a flow chart of selecting a scheduling strategy for inference duration tasks provided according to an embodiment of the present disclosure. As shown in Figure 3. The server displays the inference duration task scheduling policy set on the display interface. Among them, the set of inference-duration task scheduling strategies includes long-duration task scheduling strategies and short-duration task scheduling strategies. When the server detects that the user clicks on the short-term task scheduling policy, the server can obtain the selection instruction entered for the short-term task scheduling policy. Furthermore, the server can set the inference duration task scheduling policy to a short-duration task scheduling policy.

It is easy to understand that when the server obtains the selection instruction entered for the inference duration task scheduling policy set, the server can obtain the inference duration task scheduling policy corresponding to the selection instruction.

S206: Obtain the training operation mode corresponding to the initial task.

The specific process is as mentioned above and will not be described again here.

According to some embodiments, the training operation mode corresponding to the initial task includes but is not limited to the overall training operation mode, the cross-training operation mode, and so on.

In some embodiments, when the server adopts the overall training operation mode, the server can first perform a tuning search on all tasks and then train the model.

In some embodiments, when the server adopts cross-training operation mode, the server can train the model once and then perform task tuning once.

S207: Use the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode to search for the initial task and obtain the target task.

According to some embodiments, when searching for initial tasks using the inference duration task scheduling strategy, training task scheduling strategy, and training operation mode, the server can obtain the optimization potential value corresponding to the initial task in the training scenario. Furthermore, the server can perform task search based on the optimization potential value.

In some embodiments, the optimization potential value is used to indicate the optimization potential of the task. The optimization potential value does not refer to a fixed value. For example, when the task changes, the optimization potential value can change. The optimization potential value can be obtained based on derivatives or Bayesian models.

In some embodiments, when the server obtains the optimization potential value corresponding to the initial task, and the server determines that the optimization potential value is less than the potential threshold, the server can stop searching for the initial task, that is, early search and early stop task scheduling for training. Strategy. Therefore, the search for tasks whose optimization potential value is smaller than the potential threshold can be stopped, thereby reducing the task search time.

In some embodiments, the potential threshold refers to a threshold used by the server to evaluate whether a task has optimization potential. The potential threshold is not specific to a fixed threshold. For example, when the terminal obtains a threshold modification instruction for the potential threshold, the potential threshold may change.

In some embodiments, when the server obtains the optimization potential value corresponding to the initial task, the server may also obtain time resource information corresponding to the optimization potential value. Furthermore, the server can allocate a search duration corresponding to the time resource information to the initial task. Therefore, the search time can be allocated according to the optimization potential value corresponding to the task, thereby improving the efficiency of task search and reducing the total task search time.

According to some embodiments, when the server uses the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode to search for the initial task and obtain the target task, the server can obtain the first running time to control the iterative running of the training sample data of the initial task, and obtaining a second running time for controlling the hardware to run the training sample data.

In some embodiments, the first run time and the second run time overlap. That is to say, the time when the initial task iteratively runs the training sample data overlaps with the time when the control hardware runs the training sample data, which can be a complete overlap or a partial overlap. Therefore, task search time can be reduced.

According to some embodiments, the server may determine a search algorithm for permuting and combining different running configurations, including but not limited to loop block sizing, vectorization, loop unrolling, calculation position adjustment, etc. The server can use the machine learning cost model to predict the running speed of the optimized configuration, select the faster one to run on real hardware, and measure the real faster configuration as the optimization result. Depending on the system, there are different search algorithms, such as genetic search algorithm, exhaustive search algorithm, grid search algorithm, etc.

According to some embodiments, the cost data database Database refers to the DataBase of the data of the actual operating speed of the hardware used to train the cost model Cost Model, and is used to train a more accurate Cost Model. The Cost Model can be used to determine the search algorithm. The server can use the machine learning Cost Model to predict the running speed of the optimized configuration, which can speed up the search speed of the search algorithm and reduce automatic tuning time. At the same time, the real speed of task running on the hardware is fed back to the Cost Model for machine learning training to optimize the Cost Model.

In some embodiments, the server may use optimization methods such as ring block size, vectorization, loop unrolling, calculation position adjustment, thread parallelism, GPU parallelism, etc. in the matrix loop. These basic optimization methods are called schedule primitives in the automatic tuning system, and all runnable combinations composed of permutations and combinations of basic optimization methods are called search spaces. The server can use a search algorithm to search for a fast task running method in the search space.

S208: Obtain and store the first target configuration information corresponding to the target task.

According to some embodiments, the first target configuration information refers to tuning configuration information corresponding to the target task. The first target configuration information does not specifically refer to certain fixed information. For example, when the target task changes, the first target configuration information may change.

In some embodiments, when the server performs model training and the server obtains the first target configuration information corresponding to the target task, the server may store the first target configuration information. Furthermore, when the server performs inference on the deep learning model, the first target configuration information can be reused, thereby reducing the task search time during model inference.

It is easy to understand that when the server obtains the target task, the server can obtain and store the first target configuration information corresponding to the target task.

In the embodiment of the present disclosure, first, by determining that the second target configuration information corresponding to the initial task exists in the cache, the hardware code is obtained based on the second target configuration information; the hardware is controlled to run the hardware code, and the hardware code corresponding to the initial task is obtained. Running information; therefore, if it is determined based on the cached configuration information that there is no need to continue searching for the initial task, the time required to search for the initial task can be reduced, thereby reducing the task search time. Secondly, by displaying the inference duration task scheduling policy set corresponding to the initial task on the display interface; obtaining the selection instructions entered for the inference duration task scheduling policy set; obtaining the inference duration task scheduling policy corresponding to the selection instruction; therefore, you can choose according to your needs The required inference duration task scheduling strategy can improve the flexibility of task search. Then, by obtaining the training operation mode corresponding to the initial task; using the inference duration task scheduling strategy, training task scheduling strategy and training operation mode to search the initial task to obtain the target task; therefore, by using the inference duration task scheduling strategy, the depth of the task can be reduced The task search time when the learning model performs inference, thereby reducing the inference time of the deep learning model. By adopting the training task scheduling strategy and training operation mode, the task search time when training the deep learning model can be reduced, which in turn can reduce the training time of the deep learning model. In turn, the task search time can be reduced while improving the efficiency of the task search solution. Applicability makes this task search method applicable to inference scenarios and training scenarios. Finally, the first target configuration information corresponding to the target task is obtained and stored; therefore, when the server performs inference on the deep learning model, the first target configuration information can be reused, thereby reducing the task search time during model inference.

In the technical solution of this disclosure, the collection, storage, use, processing, transmission, provision and disclosure of user personal information are in compliance with relevant laws and regulations and do not violate public order and good customs.

The following are device embodiments of the present disclosure, which can be used to perform method embodiments of the present disclosure. For details not disclosed in the device embodiments of the disclosure, please refer to the method embodiments of the disclosure.

Please refer to Figure 4a, which shows a schematic structural diagram of a first task search device provided by an exemplary embodiment of the present disclosure. The task search device can be implemented as all or part of the device through software, hardware, or a combination of both. The task search device 400 includes a strategy acquisition unit 401, a method acquisition unit 402 and a task acquisition unit 403, wherein:

The policy acquisition unit 401 is used to obtain the task scheduling policy corresponding to the initial task. The task scheduling policy includes the inference duration task scheduling policy and the training task scheduling policy. The inference duration task scheduling policy is used to adjust the inference corresponding to the initial task in the inference scenario. Duration, the training task scheduling strategy is used to adjust the search duration corresponding to the initial task in the training scenario;

The mode acquisition unit 402 is used to acquire the training operation mode corresponding to the initial task;

The task acquisition unit 403 is used to search for initial tasks using the inference duration task scheduling strategy, training task scheduling strategy and training operation mode to obtain the target task.

According to some embodiments, FIG. 4b is a schematic structural diagram of a second task search device used to implement the task search method according to an embodiment of the present disclosure. As shown in Figure 4b, the policy acquisition unit 401 includes a collection display sub-unit 411, an instruction acquisition sub-unit 421 and a policy acquisition sub-unit 431. When the policy acquisition unit 4001 is used to acquire the task scheduling policy corresponding to the initial task:

The set display subunit 411 is used to display the inference duration task scheduling policy set corresponding to the initial task on the display interface;

The instruction acquisition subunit 421 is used to obtain the selection instruction input for the inference duration task scheduling policy set;

The policy acquisition subunit 431 is used to acquire the inference duration task scheduling policy corresponding to the selection instruction.

According to some embodiments, FIG. 4c is a schematic structural diagram of a third task search device used to implement the task search method according to the embodiment of the present disclosure. As shown in Figure 4c, the task acquisition unit 403 includes a potential value acquisition sub-unit 413 and a search stop sub-unit 423. The task acquisition unit 403 is used to search for the initial task using the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode. , when getting the target task:

The potential value acquisition subunit 413 is used to obtain the optimization potential value corresponding to the initial task in the training scenario when searching for the initial task using the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode;

The search stop subunit 423 is used to stop searching for the initial task when the optimization potential value is less than the potential threshold.

According to some embodiments, FIG. 4d is a schematic structural diagram of a fourth task search device used to implement the task search method according to an embodiment of the present disclosure. As shown in Figure 4d, the task search device 400 also includes an information acquisition unit 404 and a duration allocation unit 405, which are used to obtain the optimization potential value corresponding to the initial task:

Information acquisition unit 404, used to acquire time resource information corresponding to the optimization potential value;

The duration allocation unit 405 is used to allocate a search duration corresponding to the time resource information to the initial task.

According to some embodiments, FIG. 4e is a schematic structural diagram of a fifth task search device used to implement the task search method according to an embodiment of the present disclosure. As shown in Figure 4e, the task search device 400 also includes an information storage unit 406, which is used to search for the initial task using the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode, and after obtaining the target task:

The information storage unit 406 is used to obtain and store the first target configuration information corresponding to the target task.

According to some embodiments, FIG. 4f is a schematic structural diagram of a sixth task search device used to implement the task search method according to the embodiment of the present disclosure. As shown in Figure 4f, the task search device 400 also includes a code acquisition unit 407 and a code execution unit 408, which are used to obtain the task scheduling policy corresponding to the initial task:

The code acquisition unit 407 is configured to acquire the hardware code based on the second target configuration information when it is determined that the second target configuration information corresponding to the initial task exists in the cache;

The code running unit 408 is used to control the hardware to run the hardware code and obtain the running information corresponding to the initial task.

According to some embodiments, the task acquisition unit 404 is used to search for initial tasks using the inference duration task scheduling strategy, training task scheduling strategy and training operation mode, and when obtaining the target task, it is specifically used to:

The inference duration task scheduling strategy, training task scheduling strategy and training operation mode are used to search for the initial task. When the target task is obtained, the first running time for controlling the iterative running of the training sample data of the initial task is obtained, and the first running time for controlling the hardware running of the training sample data is obtained. Two running times, wherein the first running time and the second running time overlap.

It should be noted that when the task search device provided in the above embodiments performs the task search method, only the division of the above functional modules is used as an example. In practical applications, the above function allocation can be completed by different functional modules as needed. , that is, dividing the internal structure of the device into different functional modules to complete all or part of the functions described above. In addition, the task search device provided by the above embodiments and the task search method embodiments belong to the same concept. For details of the implementation process, please refer to the method embodiments, which will not be described again here.

The above serial numbers of the embodiments of the present disclosure are only for description and do not represent the advantages and disadvantages of the embodiments.

In the embodiment of the present disclosure, the task scheduling policy corresponding to the initial task is obtained through the policy acquisition unit. The task scheduling policy includes the inference duration task scheduling policy and the training task scheduling policy. The inference duration task scheduling policy is used to adjust the initial task in the inference scenario. The inference duration corresponding to the task, the training task scheduling strategy is used to adjust the search duration corresponding to the initial task in the training scenario; the method acquisition unit obtains the training operation mode corresponding to the initial task; the task acquisition unit adopts the inference duration task scheduling strategy, training task scheduling strategy and The training running mode searches the initial tasks and obtains the target tasks. Therefore, by adopting the inference time task scheduling strategy, the task search time when inferring the deep learning model can be reduced, which in turn can reduce the inference time of the deep learning model. By adopting the training task scheduling strategy and training operation mode, the task search time when training the deep learning model can be reduced, which in turn can reduce the training time of the deep learning model. Furthermore, the duration of task search can be reduced while improving the applicability of the task search solution, making the task search method applicable to inference scenarios and training scenarios.

In the technical solution of this disclosure, the acquisition, storage and application of user personal information are in compliance with relevant laws and regulations and do not violate public order and good customs.

According to embodiments of the present disclosure, the present disclosure also provides a server, an electronic device, a readable storage medium, a computer program product, and a computer program.

Figure 5 illustrates a schematic block diagram of an example server 500 that may be used to implement embodiments of the present disclosure.

As shown in FIG. 5 , the server 500 includes a computing unit 501 that can execute according to a computer program stored in a read-only memory (ROM) 502 or loaded from a storage unit 508 into a random access memory (RAM) 503 Various appropriate actions and treatments. In the RAM 503, various programs and data required for the operation of the server 500 can also be stored. Computing unit 501, ROM 502 and RAM 503 are connected to each other via bus 504. An input/output (I/O) interface 505 is also connected to bus 504.

Multiple components in the server 500 are connected to the I/O interface 505, including: input unit 506, such as keyboard, mouse, etc.; output unit 507, such as various types of displays, speakers, etc.; storage unit 508, such as magnetic disk, optical disk, etc. ; and communication unit 509, such as a network card, modem, wireless communication transceiver, etc. The communication unit 509 allows the server 500 to exchange information/data with other devices through computer networks such as the Internet and/or various telecommunications networks.

Computing unit 501 may be a variety of general and/or special purpose processing components having processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processing processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 501 performs various methods and processes described above, such as the task search method. For example, in some embodiments, the task search method may be implemented as a computer software program that is tangibly embodied in a machine-readable medium, such as storage unit 508. In some embodiments, part or all of the computer program may be loaded and/or installed on the server 500 via the ROM 502 and/or the communication unit 509. When the computer program is loaded into the RAM 503 and executed by the computing unit 501, one or more steps of the task search method described above may be performed. Alternatively, in other embodiments, the computing unit 501 may be configured to perform the task search method in any other suitable manner (eg, by means of firmware).

Various implementations of the systems and techniques described above may be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on a chip implemented in a system (SOC), load programmable logic device (CPLD), computer hardware, firmware, software, and/or a combination thereof. These various embodiments may include implementation in one or more computer programs executable and/or interpreted on a programmable system including at least one programmable processor, the programmable processor The processor, which may be a special purpose or general purpose programmable processor, may receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device. An output device.

Program code for implementing the methods of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that the program codes, when executed by the processor or controller, cause the functions specified in the flowcharts and/or block diagrams/ The operation is implemented. The program code may execute entirely on the machine, partly on the machine, as a stand-alone software package, partly on the machine and partly on a remote machine or entirely on the remote machine or server.

In the context of this disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media may include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media would include one or more wire-based electrical connections, laptop disks, hard drives, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.

To provide interaction with a user, the systems and techniques described herein may be implemented on a computer having a display device (eg, a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user ); and a keyboard and pointing device (eg, a mouse or a trackball) through which a user can provide input to the computer. Other kinds of devices may also be used to provide interaction with the user; for example, the feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and may be provided in any form, including Acoustic input, voice input or tactile input) to receive input from the user.

The systems and techniques described herein may be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., A user's computer having a graphical user interface or web browser through which the user can interact with implementations of the systems and technologies described herein), or including such backend components, middleware components, or any combination of front-end components in a computing system. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communications network). Examples of communication networks include: local area network (LAN), wide area network (WAN), the Internet, and blockchain networks.

Computer systems may include clients and servers. Clients and servers are generally remote from each other and typically interact over a communications network. The relationship of client and server is created by computer programs running on corresponding computers and having a client-server relationship with each other. The server can be a cloud server, also known as cloud computing server or cloud host. It is a host product in the cloud computing service system to solve the problem of traditional physical host and VPS service ("Virtual Private Server", or "VPS" for short) Among them, there are defects such as difficult management and weak business scalability. The server can also be a distributed system server or a server combined with a blockchain.

It should be noted that the foregoing explanations of the task search method embodiments are also applicable to the devices, servers, electronic equipment, computer-readable storage media, computer program products and computer programs in the embodiments of the present disclosure, and will not be described again here.

It should be understood that various forms of the process shown above may be used, with steps reordered, added or deleted. For example, each step described in the present disclosure can be executed in parallel, sequentially, or in a different order. As long as the desired results of the technical solution disclosed in the present disclosure can be achieved, there is no limitation here.

The above-mentioned specific embodiments do not constitute a limitation on the scope of the present disclosure. It will be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions are possible depending on design requirements and other factors. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of this disclosure shall be included in the protection scope of this disclosure.

Claims

A task search method including:

Obtain the task scheduling policy corresponding to the initial task, where the task scheduling policy includes an inference duration task scheduling policy and a training task scheduling policy. The inference duration task scheduling policy is used to adjust the inference duration corresponding to the initial task in the inference scenario. , the training task scheduling policy is used to adjust the search duration corresponding to the initial task in the training scenario;

Obtain the training operation mode corresponding to the initial task;

The initial task is searched using the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode to obtain the target task.
The method according to claim 1, wherein said obtaining the task scheduling policy corresponding to the initial task includes:

Display a set of inference duration task scheduling strategies corresponding to the initial task on the display interface;

Obtain the selection instructions input for the inference duration task scheduling policy set;

Obtain the inference duration task scheduling policy corresponding to the selection instruction.
The method according to claim 1 or 2, wherein said using the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode to search the initial task to obtain the target task includes:

When searching for the initial task using the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode, obtain the optimization potential value corresponding to the initial task in the training scenario;

In the case where the optimization potential value is less than the potential threshold, the search for the initial task is stopped.
The method according to claim 3, wherein after obtaining the optimization potential value corresponding to the initial task, it further includes:

Obtain time resource information corresponding to the optimization potential value;

The initial task is assigned a search duration corresponding to the time resource information.
The method according to any one of claims 1 to 4, wherein after using the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode to search the initial task and obtain the target task, Also includes:

Obtain and store the first target configuration information corresponding to the target task.
The method according to any one of claims 1 to 5, wherein before obtaining the task scheduling policy corresponding to the initial task, it further includes:

If it is determined that the second target configuration information corresponding to the initial task exists in the cache, obtain the hardware code based on the second target configuration information;

Control the hardware to run the hardware code and obtain the operation information corresponding to the initial task.
The method according to any one of claims 1 to 6, wherein the initial task is searched using the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode to obtain the target task, include:

The initial task is searched using the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode. When the target task is obtained, the first running time for controlling the iterative running of the training sample data of the initial task is obtained, and Control the hardware to run a second running time of the training sample data, wherein the first running time and the second running time overlap.
A task search device includes:

A policy acquisition unit is used to obtain a task scheduling policy corresponding to the initial task, wherein the task scheduling policy includes an inference duration task scheduling policy and a training task scheduling policy, and the inference duration task scheduling policy is used to adjust the inference duration task scheduling policy in the inference scenario. The inference duration corresponding to the initial task, and the training task scheduling policy is used to adjust the search duration corresponding to the initial task in the training scenario;

A mode acquisition unit, used to acquire the training operation mode corresponding to the initial task;

A task acquisition unit is used to search the initial task using the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode to obtain the target task.
The device according to claim 8, wherein the policy acquisition unit includes a collection display sub-unit, an instruction acquisition sub-unit and a policy acquisition sub-unit, and the policy acquisition unit is used to acquire the task scheduling policy corresponding to the initial task:

The set display subunit is used to display a set of inference duration task scheduling strategies corresponding to the initial task on the display interface;

The instruction acquisition subunit is used to obtain the selection instruction input for the inference duration task scheduling policy set;

The policy acquisition subunit is used to acquire the inference duration task scheduling policy corresponding to the selection instruction.
The device according to claim 8 or 9, wherein the task acquisition unit includes a potential value acquisition subunit and a search stop subunit, and the task acquisition unit is used to adopt the inference duration task scheduling strategy and the training task scheduling strategy. Search the initial task with the training operation mode and obtain the target task:

The potential value acquisition subunit is used to obtain the initial task corresponding to the training scenario when searching for the initial task using the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode. Optimization potential value;

The search stopping subunit is configured to stop searching for the initial task when the optimization potential value is less than a potential threshold.
The device according to claim 10, wherein the device further includes an information acquisition unit and a duration allocation unit, configured to: after obtaining the optimization potential value corresponding to the initial task:

The information acquisition unit is used to acquire time resource information corresponding to the optimization potential value;

The duration allocation unit is configured to allocate a search duration corresponding to the time resource information to the initial task.
The device according to any one of claims 8 to 11, wherein the device further includes an information storage unit for using the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode to perform Search for the above initial task and obtain the target task:

The information storage unit is used to obtain and store the first target configuration information corresponding to the target task.
The device according to any one of claims 8 to 12, wherein the device further includes a code acquisition unit and a code execution unit, configured to obtain the task scheduling policy corresponding to the initial task before:

The code acquisition unit is configured to acquire hardware code based on the second target configuration information when it is determined that there is second target configuration information corresponding to the initial task in the cache;

The code running unit is used to control the hardware to run the hardware code and obtain the running information corresponding to the initial task.
The device according to any one of claims 8 to 13, wherein the task acquisition unit is configured to use the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode to perform the initial task When searching and obtaining the target task, it is specifically used for:

The initial task is searched using the inference duration task scheduling strategy, the training task scheduling strategy and the training operation mode. When the target task is obtained, the first running time for controlling the iterative running of the training sample data of the initial task is obtained, and Control the hardware to run a second running time of the training sample data, wherein the first running time and the second running time overlap.
A server that includes:

at least one processor; and

A memory communicatively connected with the at least one processor; characterized in that,

The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor, so that the at least one processor can perform any one of claims 1 to 7 Methods.
A non-transitory computer-readable storage medium storing computer instructions, characterized in that the computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 7.
A computer program product comprising a computer program which, when executed by a processor, implements the method according to any one of claims 1 to 7.
An electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the implementation as described in any one of claims 1 to 7 is achieved. Methods.
A computer program, wherein the computer program includes computer program code, which when run on a computer causes the computer to perform the method according to any one of claims 1 to 7.