CN112100466A - Method, apparatus, device and storage medium for generating search space - Google Patents

Abstract

The present application discloses a method, apparatus, device and storage medium for generating a search space, in the fields of artificial intelligence such as deep learning and computer vision in computer technology. The specific implementation scheme is: the initial search space includes the search space of each layer of the target model, and the search space of each layer includes the options of all network structural units, which greatly expands the search space; all options in the search space of each layer are represented by the same The connection weights are overlapped to form the initial super-network, and the model parameters of the initial super-network and the connection weights corresponding to the options in each layer are trained and updated; the optimal search space is determined according to the connection weights corresponding to the options in each layer of the trained super-network, based on the optimal The search space can search to obtain the optimal target model, improve the performance of the obtained target model, make the target model more accurate, and the data processing speed is faster when applied to image processing, natural language processing, audio/video processing, etc.

Description

Method, apparatus, device and storage medium for generating search space

technical field

The present application relates to the field of artificial intelligence, in particular to deep learning and computer vision, and in particular to a method, apparatus, device and storage medium for generating a search space.

Background technique

In recent years, deep learning technology has achieved great success in many directions. In deep learning technology, the quality of artificial neural network structure has a very important impact on the effect of the final model. Manual design of network topology requires very rich experience and many attempts, and many parameters will produce explosive combinations. Conventional random search is almost impossible. Therefore, Neural Architecture Search (NAS) has become a research hotspot.

In NAS, the search space is very important. The search space in the existing NAS is designed manually. Given a small number of possible model structures, the number of search channels, the expansion coefficient, etc., have great limitations, and can only be limited in The optimal model structure is searched from a small number of possible model structures in the search space, and the performance of the finally found model structure is poor, and the accuracy and efficiency when used in image processing, natural language processing, audio/video processing and other data processing are relatively high. Low.

SUMMARY OF THE INVENTION

The present application provides a method, apparatus, device and storage medium for generating a search space.

According to an aspect of the present application, a method for generating a search space is provided, comprising:

Obtaining an initial search space, the initial search space includes the search space of each layer of the target model, and the search space of each layer includes the options of all network structural units;

All the options in the search space of each layer are superimposed with the same connection weight to form an initial super-network;

The initial super-network is trained, and the model parameters of the initial super-network and the connection weights corresponding to the options in each layer are updated to obtain the trained super-network;

The optimal search space is determined according to the connection weights corresponding to the options in each layer of the trained super-network, and the optimal search space is used to search to obtain the optimal target model, and the target model is used to execute the data Process tasks.

According to another aspect of the present application, an apparatus for generating a search space is provided, comprising:

an initial search space module, used to obtain an initial search space, the initial search space includes the search spaces of each layer of the target model, and the search space of each layer includes the options of all network structural units;

The search module is used for: superimposing and adding all the options in the search space of each layer with the same connection weight to form an initial super-network; training the initial super-network, and updating the model parameters of the initial super-network and each layer The connection weights corresponding to the options described in get the trained super network;

The optimal search space determination module is used to determine the optimal search space according to the connection weights corresponding to the options in each layer of the trained super network, and the optimal search space is used for searching to obtain the optimal target model , the target model is used to perform data processing tasks.

According to another aspect of the present application, an electronic device is provided, comprising:

at least one processor; and

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

The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method described above.

According to another aspect of the present application, there is provided a non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the method described above.

According to the technology of the present application, the optimal search space can be obtained by searching, and the neural network structure search based on the optimal search space can improve the accuracy and efficiency of the target model obtained by searching. According to the scene and complexity of the picture, the encoding parameters can be adjusted intelligently, the bandwidth cost requirement can be reduced under the same picture quality, and the bandwidth cost of the platform can be greatly reduced while the picture quality is improved. , super-resolution and other technologies have greatly improved the subjective visual experience.

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

Description of drawings

The accompanying drawings are used for better understanding of the present solution, and do not constitute a limitation to the present application. in:

1 is a flowchart of a method for generating a search space provided by the first embodiment of the present application;

2 is a flowchart of a method for generating a search space provided by a second embodiment of the present application;

3 is a schematic diagram of an apparatus for generating a search space provided by a third embodiment of the present application;

4 is a schematic diagram of an apparatus for generating a search space provided by a fourth embodiment of the present application;

FIG. 5 is a block diagram of an electronic device used to implement the method for generating a search space according to an embodiment of the present application.

Detailed ways

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

In recent years, deep learning technology has achieved great success in many directions. In deep learning technology, the quality of artificial neural network structure has a very important impact on the effect of the final model. Manual design of network topology requires very rich experience and many attempts, and many parameters will produce explosive combinations, and conventional random search is almost infeasible. Therefore, in recent years, Neural Architecture Search (NAS) has become a research topic. hot spot.

In NAS, the search space is very important. In the past, the search space of NAS work was designed manually, which has great limitations. Only the optimal model structure can be searched in the limited search space. Taking Google's Mnasnet as an example, the structure can only be limited to the number of search channels, expansion coefficient, etc. The structure of the model is still a mobilenet_v2-like structure. The search space determines the upper bound on the model structure that can be searched. Even if the optimal model structure can be found in an unsuitable search space, performance will be poor.

The present application provides a method, device, device and storage medium for generating a search space, which are applied to artificial intelligence fields such as deep learning and computer vision in computer technology to optimize the search space and obtain the optimal search space. Based on the optimal search space The optimal model can be obtained through NAS search, which can improve the accuracy and efficiency of data processing of the model.

The embodiments of the present application can be specifically applied to scenarios such as image processing, natural language processing, and audio/video processing. Through the methods provided by the embodiments of the present application, an optimal search space can be automatically generated, and then an optimal model can be obtained based on the optimal search space search. , to improve the accuracy and efficiency of data processing such as image processing, natural language processing, and audio/video processing by the model.

The technical solutions of the embodiments of the present application and how the technical solutions of the present application solve the above-mentioned technical problems will be described in detail below with specific examples. The following specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. Embodiments of the embodiments of the present application will be described below with reference to the accompanying drawings.

FIG. 1 is a flowchart of a method for generating a search space provided by the first embodiment of the present application. The method of this embodiment may be executed by a device for generating a search space, wherein the device for generating a search space may be specifically a client or server or server cluster (hereinafter collectively referred to as "" “Electronic device”), or, the device may also be a chip in the electronic device, etc., which is not specifically limited in this embodiment.

As shown in Figure 1, the specific steps of the method are as follows:

Step S101 , acquiring an initial search space, the initial search space includes the search spaces of each layer of the target model, and the search space of each layer includes options of all network structural units.

Among them, the target model refers to the neural network model that the user wants to search through the NAS and is suitable for the data processing of the corresponding scene. When applied to different application scenarios, the target models can be different neural network models.

The neural network model is composed of multi-layer structures. The network structure unit is the component unit that constitutes each layer of the neural network model. Each layer structure is composed of one or more network structure units according to a certain topology. A neural network model can contain multiple network structural units with the same or different structures.

The network structural unit can be the basic unit for building a neural network model, specifically a single network layer, such as a single convolutional layer or a fully connected layer; or a structural unit formed by a combination of multiple network layers, such as a convolutional layer, The block structure (block) formed by the combination of the batch normalization layer (Batch Normalization) and the nonlinear layer (such as Relu).

For example, the network structural unit may be the residual module Residual block in the residual network ResNet, or the repeating unit conv+BN+Relu (convolutional layer + normalization layer + activation layer) in the residual module; or It can be a stage in the residual network RseNet; it can also be a structural unit formed by a custom layer combination.

In this embodiment, the initial search space is a space of search spaces or a set of search spaces, including the search spaces of each layer of the neural network model. The search space of each layer contains the options of all network structural units, wherein each option can be a possible structure of a network structural unit, or can be any structure of the artificially designed network structural unit in the historical data. The search space of each layer contains any combination of all existing artificial network structural units. Each network structural unit may be a combination of existing artificially designed structures. The initial search space contains any combination of all layers.

In this way, the initial search space of each layer is many times larger than the traditional search space, and the search space exponentially increases layer by layer, which greatly expands the initial search space.

Step S102 , superimposing all options in the search space of each layer with the same connection weight to form an initial super-network.

In this embodiment, before NAS is performed, a search space is first searched based on the initial search space to generate an optimal search space, and then NAS is performed based on the optimal search space to obtain an optimal model structure.

In this step, each layer of the initial super-network contains the options of all network structural units, and all the options are superimposed with the same connection weight to form a one-layer structure of the initial super-network, and each layer of the initial super-network has the same structure, also That is, the options of all network structural units are superimposed with equal probability to form one layer, and the multi-layer superposition constitutes the initial super network.

Step S103: Train the initial super-network, update the model parameters of the initial super-network and the connection weights corresponding to the options in each layer, and obtain the trained super-network.

After obtaining the initial super network, use the training data corresponding to the specific application scenario of the target model to train the initial super network, and update the model parameters of the initial super network and the OP (Operation) connection parameters during the training process to obtain training. post-supernetwork. The OP connection parameters include the connection weights corresponding to the options in each layer.

Step S104: Determine the optimal search space according to the connection weights corresponding to the options in each layer of the trained super-network. The optimal search space is used to search to obtain the optimal target model, and the target model is used to perform data processing tasks.

After the trained super network is obtained, the options in the search space of each layer are filtered according to the connection weight corresponding to each option in each layer of the trained super network, and the option with larger connection weight is reserved to obtain the optimal search. space.

After obtaining the optimal search space, based on the optimal search space, the optimal target model can be obtained through NAS search, and then the data processing tasks of the corresponding application scenarios can be performed based on the target model.

In this embodiment, the initial search space is obtained, the initial search space includes the search space of each layer of the target model, and the search space of each layer includes the options of all network structural units, which can greatly expand the search space. All options in the search space are superimposed with the same connection weight to form the initial super-network, train the initial super-network, update the model parameters of the initial super-network and the connection weights corresponding to the options in each layer, and obtain the trained super-network; After calculating the connection weights corresponding to the options in each layer of the super network, the optimal search space can be determined. Based on the optimal search space, the optimal target model can be searched to improve the performance of the obtained target model and make the target model more accurate. And the data processing speed is faster when applied to image processing, natural language processing, audio/video processing, etc. In the image/video processing scenario, the AI model of this technical solution is used to perform depth perception learning on the picture content. According to the picture scene and complexity, the encoding parameters can be intelligently adjusted. The bandwidth cost requirement can be reduced under the same picture quality, and the increase in At the same time of image quality, the bandwidth cost of the platform is greatly reduced. At the same time, combined with image quality repair, image quality enhancement, ROI, super resolution and other technologies, the subjective visual experience is greatly improved. Further, at present, the core competitiveness of the target model obtained by training is the accuracy of the target model and the data processing speed of the target model on the hardware. In this way, based on the same hardware, the accuracy and efficiency of the target model are higher; On the premise of accuracy and efficiency, cheaper hardware can be used to reduce hardware costs.

FIG. 2 is a flowchart of a method for generating a search space provided by a second embodiment of the present application. On the basis of the above-mentioned first embodiment, in this embodiment, the initial super-network is trained, the model parameters of the initial super-network and the connection weights corresponding to the options in each layer are updated, and the trained super-network is obtained, including: The network is trained for multiple iterations, and the model parameters of the initial super-network and the connection weights corresponding to the options in each layer are updated in each iteration process, until the iterative stop condition is met, and the trained super-network is obtained.

As shown in Figure 2, the method steps are as follows:

Step S201 , obtaining an initial search space, the initial search space includes the search spaces of each layer of the target model, and the search space of each layer includes options of all network structural units.

Among them, the target model refers to the neural network model that the user wants to search through the NAS and is suitable for the data processing of the corresponding scene. When applied to different application scenarios, the target models can be different neural network models.

The neural network model is composed of multi-layer structures. The network structure unit is the component unit that constitutes each layer of the neural network model. Each layer structure is composed of one or more network structure units according to a certain topology. A neural network model can contain multiple network structural units with the same or different structures.

The network structural unit can be the basic unit for building a neural network model, specifically a single network layer, such as a single convolutional layer or a fully connected layer; or a structural unit formed by a combination of multiple network layers, such as a convolutional layer, The block structure (block) formed by the combination of the batch normalization layer (Batch Normalization) and the nonlinear layer (such as Relu).

For example, the network structural unit may be the residual module Residual block in the residual network ResNet, or the repeating unit conv+BN+Relu (convolutional layer + normalization layer + activation layer) in the residual module; or It can be a stage in the residual network RseNet; it can also be a structural unit formed by a custom layer combination.

In this embodiment, the initial search space is a space of search spaces or a set of search spaces, including the search spaces of each layer of the neural network model. The search space of each layer contains the options of all network structural units, wherein each option can be a possible structure of a network structural unit, or can be any structure of the artificially designed network structural unit in the historical data. The search space of each layer contains any combination of all existing artificial network structural units. Each network structural unit may be a combination of existing artificially designed structures. The initial search space contains any combination of all layers.

In this way, the initial search space of each layer is many times larger than the traditional search space, and the search space exponentially increases layer by layer, which greatly expands the initial search space.

Optionally, any combination of layers in the initial search space can be represented in the form of weighted superposition.

Step S202 , superimposing and adding all options in the search space of each layer with the same connection weight to form an initial super-network.

In this embodiment, before NAS is performed, a search space is first searched based on the initial search space to generate an optimal search space, and then NAS is performed based on the optimal search space to obtain an optimal model structure.

In this step, each layer of the initial super-network contains the options of all network structural units, and all the options are superimposed with the same connection weight to form a one-layer structure of the initial super-network, and each layer of the initial super-network has the same structure, also That is, the options of all network structural units are superimposed with equal probability to form one layer, and the multi-layer superposition constitutes the initial super network.

Step S203: Perform iterative training on the initial super-network, and update the model parameters of the initial super-network and the connection weights corresponding to the options in each layer.

After obtaining the initial super-network, using the training data corresponding to the specific application scenario of the target model, the initial super-network loop is trained for multiple iterations, and the model parameters and OP connection parameters of the initial super-network are updated in each iteration training process. The iteration is stopped until the iteration stopping condition is met, and the trained super network is obtained. The OP connection parameters include the connection weights corresponding to the options in each layer.

Further, in each iteration process, the model parameters of the initial super-network and the connection weights corresponding to the options in each layer are alternately updated. Specifically, in the training process, the model parameters of the super network are fixed, and the connection weights of the super network are trained using the first sample set; the connection weights of the super network are fixed, and the model parameters of the super network are trained using the second template training set. In this way, the connection weight corresponding to each network structural unit option in each layer of the super network can be realized.

Optionally, the ratio of the model parameters and connection weights of the super network to be updated alternately may be 1:1, or non-1:1, and the ratio may be configured according to the actual application scenario, or may be determined by searching in a preset search space, This embodiment is not specifically limited here.

Step S204, judging whether the iteration stop condition is satisfied.

After each iteration training is completed, it is judged whether the current iteration stop condition is satisfied.

If the iterative stop condition is satisfied, the loop iterative training is stopped to obtain the trained super network, and step S205 is executed.

If the iterative stop condition is not met, step S203 is executed to perform the next iterative training.

Optionally, the iterative stop condition may be: the current loss function of the super-network satisfies the convergence condition. The convergence condition satisfied by the loss function may be configured according to an actual application scenario, which is not specifically limited in this embodiment.

Optionally, the execution times of the iterative training are recorded, and the iterative stop condition may be: the execution times of the iterative training is greater than or equal to a preset threshold. The preset threshold may be configured according to an actual application scenario, which is not specifically limited in this embodiment.

Optionally, the iterative stop condition may be: the total duration of iterative training reaches a preset search time threshold. The preset search time threshold may be configured according to an actual application scenario, which is not specifically limited in this embodiment.

In addition, the iteration stop condition may also be configured and adjusted according to a specific application scenario, which is not specifically limited in this embodiment.

In another optional implementation manner of this embodiment, the search for the optimal search space may be searched layer by layer.

Step S205: Determine the optimal search space according to the connection weights corresponding to the options in each layer of the trained super-network.

After the trained super network is obtained, the options in the search space of each layer are filtered according to the connection weight corresponding to each option in each layer of the trained super network, and the option with larger connection weight is reserved to obtain the optimal search. space.

Exemplarily, the optimal search space includes the optimized search space of each layer.

Optionally, for each layer of the super network after training, sort the options according to the connection weights corresponding to the options in the layer in descending order, and determine the optimization of the layer according to the top k options in the sorting. Then the optimized search space of each layer is obtained, and the optimal search space is obtained. Wherein, k is a positive integer, and the value of k can be configured according to an actual application scenario, which is not specifically limited in this embodiment.

Specifically, according to the order of connection weights in each layer, the options of the top k network structural units with the largest connection weights constitute the optimized search space of this layer.

Optionally, for each layer of the super-network after training, according to the preset weight threshold of the layer, an option with a corresponding connection weight greater than the preset weight threshold can be selected to form an optimized search space for this layer, so as to obtain each layer. The search space after layer optimization is obtained to obtain the optimal search space. The preset weight thresholds of different layers may be different, and the preset weight thresholds of each layer may be configured according to actual application scenarios, which are not specifically limited in this embodiment.

For example, suppose there are 10,000 possible options in each layer of the initial search space, and each layer in the initial super-network is composed of these 10,000 options superimposed with an equal probability. For example, the connection weight of each option can be 0.0001. Iterative training is performed to obtain the trained super network. In each layer of the super network after training, determine the 50 options with the largest connection weight as the optimized search space of this layer, assuming that the target model contains 50 layers, then the optimal search space contains 50 ⁵⁰ possible networks Structural unit options.

Step S206 , perform a neural network model search in the optimal search space to obtain a target model.

After the optimal search space is determined, NAS is performed based on the optimal search space, and the optimal model is obtained by searching, which is used as the target model.

Step S207: Acquire the data to be processed, and use the target model to perform corresponding data processing on the data to be processed.

In this embodiment, the target model obtained by the search can be applied to scenarios such as image processing, natural language processing, and audio/video processing, to perform corresponding data processing tasks.

Exemplarily, when the target model is applied to an image processing scene, an image to be processed is acquired, the to-be-processed image is input into the target model, and the target model is used to perform image processing on the to-be-processed image to obtain an image processing result.

For example, the target model can be specifically applied to face recognition. In the process of generating the search space, use the training set used to train the face recognition model to train the initial model structure and record the performance to search for the most suitable face recognition scene. The optimal search space is further obtained to obtain the optimal target model for the face recognition scene. In application, the face image to be recognized is acquired, the face feature of the face image is extracted, the face feature is input into the target model, the face recognition is performed through the target model, and the face recognition result is output.

In addition, the target model can also be applied to the field of natural language processing. For example, it can be applied to speech recognition scenarios. In the process of generating the search space, use the training set used to train the speech recognition model to train the initial model structure and record the performance to search for the optimal search space for speech recognition scenarios. , and further obtain the optimal target model for the application of speech recognition scenarios. In application, the voice data to be processed is acquired, the features of the voice data are extracted, the features of the voice data are input into the target model, the voice recognition processing is performed through the target model, and the voice recognition result is output.

In this embodiment, the initial search space is obtained, the initial search space includes the search space of each layer of the target model, and the search space of each layer includes the options of all network structural units, which can greatly expand the search space. Perform multiple iterative training, update the model parameters of the initial super-network and the connection weights corresponding to the options in each layer in each iteration process, until the iterative stop condition is met, the trained super-network is obtained; All options of the super-network are superimposed with the same connection weight to form the initial super-network, and the optimal search space includes the optimized search space of each layer. Sort the options in the smallest order. According to the top k options in the sorting, the optimized search space of this layer can be determined, so as to obtain the optimal search space. Based on the optimal search space, the optimal target model can be searched and improved. The performance of the obtained target model makes the accuracy of the target model higher, and the data processing speed is faster when applied to image processing, natural language processing, audio/video processing, etc. In the image/video processing scenario, the AI model of this technical solution is used to perform depth perception learning on the picture content. According to the picture scene and complexity, the encoding parameters can be intelligently adjusted. The bandwidth cost requirement can be reduced under the same picture quality, and the increase in At the same time of image quality, the bandwidth cost of the platform is greatly reduced. At the same time, combined with image quality repair, image quality enhancement, ROI, super resolution and other technologies, the subjective visual experience is greatly improved. Further, at present, the core competitiveness of the target model obtained by training is the accuracy of the target model and the data processing speed of the target model on the hardware. In this way, based on the same hardware, the accuracy and efficiency of the target model are higher; On the premise of accuracy and efficiency, cheaper hardware can be used to reduce hardware costs.

FIG. 3 is a schematic diagram of an apparatus for generating a search space provided by a third embodiment of the present application. The apparatus for generating a search space provided by the embodiment of the present application may execute the processing flow provided by the embodiment of the method for generating a search space. As shown in FIG. 3 , the apparatus 30 for generating a search space includes: an initial search space module 301 , a search module 302 and an optimal search space determination module 303 .

Specifically, the initial search space module 301 is used to obtain the initial search space, the initial search space includes the search space of each layer of the target model, and the search space of each layer includes the options of all network structural units.

The search module 302 is used to: superimpose all options in the search space of each layer with the same connection weight to form an initial super-network; train the initial super-network, update the model parameters of the initial super-network and the connection weights corresponding to the options in each layer , to get the trained supernetwork.

The optimal search space determination module 303 is used to determine the optimal search space according to the connection weights corresponding to the options in each layer of the super network after training, the optimal search space is used to search to obtain the optimal target model, and the target model is used to execute data processing tasks.

The apparatus provided in this embodiment of the present application may be specifically configured to execute the method embodiment provided in the foregoing first embodiment, and the specific functions will not be repeated here.

In this embodiment, the initial search space is obtained, the initial search space includes the search space of each layer of the target model, and the search space of each layer includes the options of all network structural units, which can greatly expand the search space. All options in the search space are superimposed with the same connection weight to form the initial super-network, train the initial super-network, update the model parameters of the initial super-network and the connection weights corresponding to the options in each layer, and obtain the trained super-network; After calculating the connection weights corresponding to the options in each layer of the super network, the optimal search space can be determined. Based on the optimal search space, the optimal target model can be searched to improve the performance of the obtained target model and make the target model more accurate. And the data processing speed is faster when applied to image processing, natural language processing, audio/video processing, etc. In the image/video processing scenario, the AI model of this technical solution is used to perform depth perception learning on the picture content. According to the picture scene and complexity, the encoding parameters can be intelligently adjusted. The bandwidth cost requirement can be reduced under the same picture quality, and the increase in At the same time of image quality, the bandwidth cost of the platform is greatly reduced. At the same time, combined with image quality repair, image quality enhancement, ROI, super resolution and other technologies, the subjective visual experience is greatly improved. Further, at present, the core competitiveness of the target model obtained by training is the accuracy of the target model and the data processing speed of the target model on the hardware. In this way, based on the same hardware, the accuracy and efficiency of the target model are higher; On the premise of accuracy and efficiency, cheaper hardware can be used to reduce hardware costs.

FIG. 4 is a schematic diagram of an apparatus for generating a search space provided by a fourth embodiment of the present application. On the basis of the above-mentioned third embodiment, in this embodiment, the optimal search space determination module is further used for:

The optimal search space includes the optimized search space of each layer. For each layer of the super-network after training, the options are sorted according to the order of the connection weights corresponding to the options in the layer from large to small. The first k options determine the optimized search space for this layer.

In an optional implementation manner, the search module is also used for:

The initial super-network is iteratively trained for many times, and the model parameters of the initial super-network and the connection weights corresponding to the options in each layer are updated in each iteration process, until the iterative stop condition is met, and the trained super-network is obtained.

In an optional embodiment, the iterative stop condition includes:

The number of times of iterative training is greater than or equal to the preset threshold; or, the performance of the current super-network satisfies the convergence condition.

In an optional implementation manner, the optimal search space determination module is further used for:

During each iteration, the model parameters of the initial super-network and the connection weights corresponding to the options in each layer are updated alternately.

In an optional implementation manner, as shown in FIG. 4 , the apparatus 30 for generating a search space further includes:

The neural network model searching module 304 is used for: searching the neural network model in the optimal search space to obtain the target model.

In an optional implementation manner, as shown in FIG. 4 , the apparatus 30 for generating a search space further includes:

The data processing module 305 is used for: acquiring the data to be processed, and performing corresponding data processing on the data to be processed by using the target model.

In an optional implementation manner, the data processing module 305 is further configured to:

The to-be-processed image is acquired, the to-be-processed image is input into the target model, and the target model is used to perform image processing on the to-be-processed image to obtain an image processing result.

The apparatus provided in this embodiment of the present application may be specifically configured to execute the method embodiment provided in the foregoing second embodiment, and the specific functions will not be repeated here.

In this embodiment, the initial search space is obtained, the initial search space includes the search space of each layer of the target model, and the search space of each layer includes the options of all network structural units, which can greatly expand the search space. Perform multiple iterative training, update the model parameters of the initial super-network and the connection weights corresponding to the options in each layer in each iteration process, until the iterative stop condition is met, the trained super-network is obtained; All options of the super-network are superimposed with the same connection weight to form the initial super-network, and the optimal search space includes the optimized search space of each layer. Sort the options in the smallest order. According to the top k options in the sorting, the optimized search space of this layer can be determined, so as to obtain the optimal search space. Based on the optimal search space, the optimal target model can be searched and improved. The performance of the obtained target model makes the accuracy of the target model higher, and the data processing speed is faster when applied to image processing, natural language processing, audio/video processing, etc. In the image/video processing scenario, the AI model of this technical solution is used to perform depth perception learning on the picture content. According to the picture scene and complexity, the encoding parameters can be intelligently adjusted. The bandwidth cost requirement can be reduced under the same picture quality, and the increase in At the same time of image quality, the bandwidth cost of the platform is greatly reduced. At the same time, combined with image quality repair, image quality enhancement, ROI, super resolution and other technologies, the subjective visual experience is greatly improved. Further, at present, the core competitiveness of the target model obtained by training is the accuracy of the target model and the data processing speed of the target model on the hardware. In this way, based on the same hardware, the accuracy and efficiency of the target model are higher; On the premise of accuracy and efficiency, cheaper hardware can be used to reduce hardware costs.

According to the embodiments of the present application, the present application further provides an electronic device and a readable storage medium.

As shown in FIG. 5 , it is a block diagram of an electronic device of a method for generating a search space according to an embodiment of the present application. Electronic devices are intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are by way of example only, and are not intended to limit implementations of the application described and/or claimed herein.

As shown in FIG. 5, the electronic device includes: one or more processors Y01, a memory Y02, and interfaces for connecting various components, including a high-speed interface and a low-speed interface. The various components are interconnected using different buses and may be mounted on a common motherboard or otherwise as desired. The processor may process instructions executed within the electronic device, including instructions stored in or on memory to display graphical information of the GUI on an external input/output device, such as a display device coupled to the interface. In other embodiments, multiple processors and/or multiple buses may be used with multiple memories, if desired. Likewise, multiple electronic devices may be connected, each providing some of the necessary operations (eg, as a server array, a group of blade servers, or a multiprocessor system). In FIG. 5, a processor Y01 is taken as an example.

The memory Y02 is the non-transitory computer-readable storage medium provided in this application. The memory stores instructions executable by at least one processor, so that the at least one processor executes the method for generating a search space provided by the present application. The non-transitory computer-readable storage medium of the present application stores computer instructions for causing a computer to execute the method of generating a search space provided by the present application.

As a non-transitory computer-readable storage medium, the memory Y02 can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as program instructions/modules (such as program instructions/modules corresponding to the method for generating a search space in the embodiments of the present application). , the initial search space module 301, the search module 302 and the optimal search space determination module 303 shown in FIG. 3). The processor Y01 executes various functional applications and data processing of the server by running the non-transitory software programs, instructions and modules stored in the memory Y02, that is, implementing the method for generating a search space in the above method embodiments.

The memory Y02 can include a stored program area and a stored data area, wherein the stored program area can store an operating system, an application program required by at least one function; the stored data area can store data created according to the use of the electronic device that generates the search space, etc. . In addition, the memory Y02 may include high-speed random access memory, and may also include non-transitory memory, such as at least one magnetic disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory Y02 may optionally include memory located remotely from the processor Y01, and these remote memories may be connected via a network to the electronic device generating the search space. Examples of such networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

The electronic device of the method for generating a search space may further include: an input device Y03 and an output device Y04. The processor Y01, the memory Y02, the input device Y03, and the output device Y04 may be connected by a bus or in other ways, and the connection by a bus is taken as an example in FIG. 5 .

The input device Y03 can receive input numerical or character information, and generate key signal input related to user settings and function control of the electronic device generating the search space, such as touch screen, keypad, mouse, trackpad, touchpad, pointing stick, One or more input devices such as mouse buttons, trackballs, joysticks, etc. The output device Y04 may include a display device, an auxiliary lighting device (eg, LED), and a haptic feedback device (eg, a vibration motor), and the like. The display device may include, but is not limited to, a liquid crystal display (LCD), a light emitting diode (LED) display, and a plasma display. In some implementations, the display device may be a touch screen.

Various implementations of the systems and techniques described herein can be implemented in digital electronic circuitry, integrated circuit systems, application specific ASICs (application specific integrated circuits), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include being implemented in one or more computer programs executable and/or interpretable on a programmable system including at least one programmable processor that 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.

These computer programs (also referred to as programs, software, software applications, or codes) include machine instructions for programmable processors and may be implemented using high-level procedural and/or object-oriented programming languages, and/or assembly/machine languages Computer program. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and/or apparatus for providing machine instructions and/or data to a programmable processor ( For example, magnetic disks, optical disks, memories, programmable logic devices (PLDs), including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal used to provide machine instructions and/or data to a programmable processor.

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 trackball) through which a user can provide input to the computer. Other kinds of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (eg, visual feedback, auditory feedback, or tactile feedback); and can be 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 on a computing system that includes back-end components (eg, as a data server), or a computing system that includes middleware components (eg, an application server), or a computing system that includes front-end components (eg, a user computer having a graphical user interface or web browser through which a user may interact with implementations of the systems and techniques 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 (eg, a communication network). Examples of communication networks include: Local Area Networks (LANs), Wide Area Networks (WANs), and the Internet.

A computer system can include clients and servers. Clients and servers are generally remote from each other and usually interact through a communication network. The relationship of client and server arises by computer programs running on the respective computers and having a client-server relationship to each other.

According to the technical solutions of the embodiments of the present application, by obtaining the initial search space, the initial search space includes the search space of each layer of the target model, and the search space of each layer includes the options of all network structural units, which can greatly expand the search space, and further , by superimposing all options in the search space of each layer with the same connection weight to form an initial super-network, train the initial super-network, update the model parameters of the initial super-network and the connection weights corresponding to the options in each layer, and get the post-training According to the connection weights corresponding to the options in each layer of the trained super network, the optimal search space is determined. Based on the optimal search space, the optimal target model can be searched to improve the performance of the obtained target model, so that the target The accuracy of the model is higher, and the data processing speed is faster when applied to image processing, natural language processing, audio/video processing, etc. In the image/video processing scenario, the AI model of this technical solution is used to perform depth perception learning on the picture content. According to the picture scene and complexity, the encoding parameters can be intelligently adjusted. The bandwidth cost requirement can be reduced under the same picture quality, and the increase in At the same time of image quality, the bandwidth cost of the platform is greatly reduced. At the same time, combined with image quality repair, image quality enhancement, ROI, super resolution and other technologies, the subjective visual experience is greatly improved. Further, at present, the core competitiveness of the target model obtained by training is the accuracy of the target model and the data processing speed of the target model on the hardware. In this way, based on the same hardware, the accuracy and efficiency of the target model are higher; On the premise of accuracy and efficiency, cheaper hardware can be used to reduce hardware costs.

It should be understood that steps may be reordered, added or deleted using the various forms of flow shown above. For example, the steps described in the present application can be performed in parallel, sequentially or in different orders, and as long as the desired results of the technical solutions disclosed in the present application can be achieved, no limitation is imposed herein.

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

Claims (18)

1. A method of generating a search space, comprising:

acquiring an initial search space, wherein the initial search space comprises search spaces of all layers of a target model, and the search space of each layer comprises options of all network structure units;

superposing all the options in the search space of each layer by the same connection weight to form an initial hyper-network;

training the initial hyper-network, and updating model parameters of the initial hyper-network and connection weights corresponding to the options in each layer to obtain a trained hyper-network;

and determining an optimal search space according to the connection weight corresponding to the option in each layer of the trained hyper-network, wherein the optimal search space is used for searching to obtain an optimal target model, and the target model is used for executing a data processing task.

2. The method of claim 1, wherein determining an optimal search space according to connection weights corresponding to the options in each layer of the trained hyper-network comprises:

the optimal search space comprises each optimized search space, all the options in each layer of the trained hyper-network are sorted according to the descending order of the connection weights corresponding to all the options in the layer, and the optimized search space in the layer is determined according to the first k options in the sorting.

3. The method of claim 1, wherein the training the initial hyper-network, updating model parameters of the initial hyper-network and connection weights corresponding to the options in each layer, and obtaining a trained hyper-network comprises:

and carrying out repeated iterative training on the initial hyper-network, updating the model parameters of the initial hyper-network and the connection weights corresponding to the options in each layer in each iterative process, and obtaining the trained hyper-network until an iteration stop condition is met.

4. The method of claim 3, wherein the iteration stop condition comprises:

the number of times of iterative training is greater than or equal to a preset threshold value;

or,

the performance of the current super-network satisfies the convergence condition.

5. The method of claim 3, wherein said updating model parameters of said initial hyper-network and connection weights corresponding to said options in each layer during each iteration comprises:

and in each iteration process, alternately updating the model parameters of the initial hyper-network and the connection weights corresponding to the options in each layer.

6. The method according to any one of claims 1-5, wherein after determining the optimal search space according to the connection weights corresponding to the options in each layer of the trained hyper-network, further comprising:

and searching a neural network model in the optimal search space to obtain the target model.

7. The method of claim 6, wherein after the performing the neural network model search in the optimal search space to obtain the target model, further comprises:

and acquiring data to be processed, and performing corresponding data processing on the data to be processed by using the target model.

8. The method of claim 7, further comprising:

and acquiring an image to be processed, inputting the image to be processed into the target model, and performing image processing on the image to be processed by using the target model to obtain an image processing result.

9. An apparatus that generates a search space, comprising:

the initial search space module is used for acquiring an initial search space, wherein the initial search space comprises search spaces of all layers of the target model, and the search space of each layer comprises options of all network structure units;

a search module to: superposing all the options in the search space of each layer by the same connection weight to form an initial hyper-network; training the initial hyper-network, and updating model parameters of the initial hyper-network and connection weights corresponding to the options in each layer to obtain a trained hyper-network;

and the optimal search space determining module is used for determining an optimal search space according to the connection weight corresponding to the option in each layer of the trained hyper-network, wherein the optimal search space is used for searching to obtain an optimal target model, and the target model is used for executing a data processing task.

10. The apparatus of claim 9, wherein the optimal search space determination module is further configured to:

the optimal search space comprises each optimized search space, all the options in each layer of the trained hyper-network are sorted according to the descending order of the connection weights corresponding to all the options in the layer, and the optimized search space in the layer is determined according to the first k options in the sorting.

11. The apparatus of claim 9, wherein the search module is further configured to:

and carrying out repeated iterative training on the initial hyper-network, updating the model parameters of the initial hyper-network and the connection weights corresponding to the options in each layer in each iterative process, and obtaining the trained hyper-network until an iteration stop condition is met.

12. The apparatus of claim 11, wherein the iteration stop condition comprises:

the number of times of iterative training is greater than or equal to a preset threshold value;

or,

the performance of the current super-network satisfies the convergence condition.

13. The apparatus of claim 11, wherein the optimal search space determination module is further configured to:

and in each iteration process, alternately updating the model parameters of the initial hyper-network and the connection weights corresponding to the options in each layer.

14. The apparatus of any of claims 9-13, further comprising:

a neural network model search module to: and searching a neural network model in the optimal search space to obtain the target model.

15. The apparatus of claim 14, further comprising:

a data processing module to: and acquiring data to be processed, and performing corresponding data processing on the data to be processed by using the target model.

16. The apparatus of claim 15, the data processing module further to:

and acquiring an image to be processed, inputting the image to be processed into the target model, and performing image processing on the image to be processed by using the target model to obtain an image processing result.

17. An electronic device, comprising:

at least one processor; and

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

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-8.

18. A non-transitory computer readable storage medium having stored thereon computer instructions for causing the computer to perform the method of any one of claims 1-8.

Images