CN108234195B - Method, apparatus, device, medium for predicting network performance - Google Patents

Method, apparatus, device, medium for predicting network performance Download PDF

Info

Publication number
CN108234195B
CN108234195B CN201711306429.3A CN201711306429A CN108234195B CN 108234195 B CN108234195 B CN 108234195B CN 201711306429 A CN201711306429 A CN 201711306429A CN 108234195 B CN108234195 B CN 108234195B
Authority
CN
China
Prior art keywords
network
structural
layer
network layer
parameters
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201711306429.3A
Other languages
Chinese (zh)
Other versions
CN108234195A (en
Inventor
邓博洋
闫俊杰
林达华
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Sensetime Technology Development Co Ltd
Original Assignee
Beijing Sensetime Technology Development Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Sensetime Technology Development Co Ltd filed Critical Beijing Sensetime Technology Development Co Ltd
Priority to CN201711306429.3A priority Critical patent/CN108234195B/en
Publication of CN108234195A publication Critical patent/CN108234195A/en
Application granted granted Critical
Publication of CN108234195B publication Critical patent/CN108234195B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/04—Network management architectures or arrangements
    • H04L41/044—Network management architectures or arrangements comprising hierarchical management structures
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/14—Network analysis or design
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/14—Network analysis or design
    • H04L41/147—Network analysis or design for predicting network behaviour
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L43/00—Arrangements for monitoring or testing data switching networks
    • H04L43/08—Monitoring or testing based on specific metrics, e.g. QoS, energy consumption or environmental parameters
    • H04L43/0805—Monitoring or testing based on specific metrics, e.g. QoS, energy consumption or environmental parameters by checking availability
    • H04L43/0817—Monitoring or testing based on specific metrics, e.g. QoS, energy consumption or environmental parameters by checking availability by checking functioning

Abstract

The embodiment of the disclosure discloses a method, a device, equipment and a medium for predicting network performance, wherein the method comprises the following steps: acquiring network parameters of a structural network to be predicted; determining structural characteristics of the structural network based on network parameters of the structural network; determining a network performance parameter for the fabric network based on the fabric feature. Based on the method provided by the embodiment of the disclosure, the performance of the structure network is predicted through the network parameters of the structure network, and the method for predicting the network performance of the embodiment does not need to train the structure network, thereby saving a large amount of time and improving the efficiency of network performance prediction.

Description

Method, apparatus, device, medium for predicting network performance
Technical Field
The present disclosure relates to deep learning techniques, and in particular, to a method, apparatus, device, and medium for predicting network performance.
Background
The neural network is a core technology in the existing image recognition system, is an end-to-end feature learning device, can learn the feature expression of a specific task through the neural network system, and obtains a final result through a classifier. Typically a neural network requires training for a period of time to achieve satisfactory performance.
In recent years, each progress of the neural network is brought about by an improvement of the structure of the neural network, however, as the neural network becomes more complex, the design ability of human beings for the complex network structure has been a bottleneck. Accordingly, various algorithms for automatically generating a neural network structure have been proposed in succession, and these algorithms have been proven to be capable of generating a neural network having a complicated structure.
Disclosure of Invention
The disclosed embodiments provide a technique for predicting network performance.
According to an aspect of the embodiments of the present disclosure, a method for predicting network performance is provided, which is implemented by using a prediction network, and includes:
acquiring network parameters of a structural network to be predicted;
determining structural characteristics of the structural network based on network parameters of the structural network;
determining a network performance parameter for the fabric network based on the fabric feature.
In another embodiment of the above method according to the present invention, the network parameter includes at least one of:
the method comprises the steps of calculating the calculation type of each network layer in at least one network layer of the structural network, the length of a calculation core of the network layer, the width of the calculation core of the network layer, the number of channels of the network layer and the ratio of the number of output channels to the number of input channels of the network layer.
In another embodiment of the above method according to the present invention, the determining the structural characteristics of the structural network based on the network parameters of the structural network includes:
determining structure representation information of each of at least one network layer of the fabric network based on network parameters of the fabric network;
determining structural features of the structural network based on the structural representation information of each of the at least one network layer.
In another embodiment of the above method according to the present invention, the structure representing information includes a structure vector having a preset dimension.
In another embodiment of the above method according to the present invention, determining the structure representation information of each network layer of the at least one network layer of the fabric network based on the network parameters of the fabric network comprises:
determining at least one identifier for each of at least one network layer of the fabric network based on layer parameters of said each network layer;
determining structure representation information of each of the at least one network layer based on at least one identifier of the each network layer.
In another embodiment of the above method according to the present invention, determining the structure representation information of each of the at least one network layer based on at least one identifier of each of the at least one network layer comprises:
mapping each identifier in at least one identifier of each network layer to obtain a mapping result of each identifier;
and obtaining the structure representation information of each network layer based on the mapping result of each identifier in the at least one identifier.
In another embodiment of the foregoing method according to the present invention, the mapping each identifier in at least one identifier of each network layer to obtain a mapping result of each identifier includes:
and obtaining the mapping result of each identifier based on each identifier in each network layer by searching a first preset table.
In another embodiment of the above method according to the present invention, determining the structural characteristics of the structural network based on the structural representation information of each of the at least one network layer comprises:
and fusing the structure representation information of each network layer in the at least one network layer to obtain the structure characteristics of the structure network.
In another embodiment of the foregoing method based on the present invention, fusing structure representation information of each network layer in the at least one network layer to obtain the structural features of the structural network, includes:
and fusing the structure representation information of each network layer in the at least one network layer by utilizing a neural network to obtain the structural characteristics of the structural network.
In another embodiment of the above method according to the present invention, the fusing, by using a neural network, the structure representation information of each of the at least one network layer includes:
the structure representation information of each network layer in at least one network layer is input into a neural network, and the neural network fuses at least one structure representation information into a structure characteristic through a recursive algorithm.
In another embodiment of the foregoing method according to the present invention, before determining the network performance parameter of the fabric network based on the fabric feature, the method further includes:
obtaining a time vector based on a preset time point corresponding to the structural network;
determining network performance parameters of the fabric network based on the fabric feature, including:
and obtaining the network performance parameters of the structural network at the preset time point based on the structural features and the time vector.
In another embodiment of the foregoing method based on the present invention, the obtaining, based on the structural feature and the time vector, a network performance parameter corresponding to the structural network at the preset time point includes:
merging the structural features and the time vectors to obtain merged features;
and obtaining the corresponding network performance parameters of the structural network under the time vector based on the merging characteristics by utilizing a multilayer perceptron.
In another embodiment of the method according to the present invention, the obtaining the time vector based on the preset time point corresponding to the structural network includes:
and obtaining the time vector by searching a second preset table based on a preset time point corresponding to the structural network.
In another embodiment of the above method according to the present invention, the determining the network performance parameter of the fabric network based on the fabric feature includes:
determining a network performance parameter of the structural network based on the structural feature prior to training the structural network.
In another embodiment of the foregoing method according to the present invention, before the obtaining the network parameters of the structural network to be predicted, the method further includes:
training the prediction network using a plurality of sample structure networks, the sample structure networks being labeled with network performance parameters for the sample structure networks at each of at least one time point; the time point corresponds to a number of training times of the sample structure network.
In another embodiment of the above method according to the present invention, before training the prediction network using a plurality of sample structure networks, the method further includes:
carrying out network layer sampling on a plurality of preset structure networks to generate network blocks; the network block comprises at least one network layer;
constructing the sample structure network based on the network blocks.
In another embodiment of the foregoing method according to the present invention, the sampling the network layer of the multiple networks with preset structures to generate the network block includes:
sampling a plurality of preset structure networks based on a Markov chain to obtain at least one preset network layer;
and sequentially connecting the at least one preset network layer to form the network block.
In another embodiment of the above method according to the present invention, the network layer comprises any one or more of the following:
convolutional layers, max pooling layers, average pooling layers, activation layers, and batch layers.
In another embodiment of the above method according to the present invention, the constructing the sample structure network based on the network blocks includes:
connecting a plurality of the network blocks in sequence to obtain the sample structure network, wherein a first network block and a second network block in the plurality of network blocks correspond to different feature dimensions.
In another embodiment of the above method according to the present invention, a maximum pooling layer is connected between the first network block and the second network block.
In another embodiment of the foregoing method according to the present invention, sampling a plurality of networks with preset structures based on a markov chain to obtain at least one preset network layer, includes:
and sampling a plurality of preset structure networks based on the sampled network parameters of the ith network layer to obtain an ith +1 network layer, wherein i is greater than or equal to 1 and less than the number of network layers included in the network block.
According to another aspect of the embodiments of the present disclosure, there is provided an apparatus for predicting network performance, which is implemented by using a prediction network, including:
a parameter obtaining unit, configured to obtain a network parameter of a structural network to be predicted;
a structural feature unit, configured to determine a structural feature of the structural network based on a network parameter of the structural network;
a performance determining unit, configured to determine a network performance parameter of the fabric network based on the fabric feature.
In another embodiment of the above apparatus according to the present invention, the network parameter includes at least one of:
the method comprises the steps of calculating the calculation type of each network layer in at least one network layer of the structural network, the length of a calculation core of the network layer, the width of the calculation core of the network layer, the number of channels of the network layer and the ratio of the number of output channels to the number of input channels of the network layer.
In another embodiment of the above apparatus according to the present invention, the structural feature unit includes:
an information representation module for determining structure representation information of each of at least one network layer of the fabric network based on network parameters of the fabric network;
a feature determination module for determining structural features of the structural network based on the structural representation information of each of the at least one network layer.
In another embodiment of the above apparatus according to the present invention, the structure representing information includes a structure vector having a preset dimension.
In another embodiment of the above apparatus according to the present invention, the information presentation module includes:
a symbol recognition module for determining at least one identifier for each of at least one network layer of the fabric network based on layer parameters of said each network layer;
an information module to determine structural representation information for each of the at least one network layer based on at least one identifier for each of the at least one network layer.
In another embodiment of the above apparatus according to the present invention, the symbol recognition module is specifically configured to map each identifier of at least one identifier of each network layer, so as to obtain a mapping result of each identifier;
an information module, configured to obtain structure representation information of each network layer based on a mapping result of each identifier in the at least one identifier.
In another embodiment of the above apparatus according to the present invention, the symbol recognition module is further configured to obtain a mapping result of each identifier in each network layer by looking up a first preset table based on each identifier.
In another embodiment of the above apparatus according to the present invention, the characteristic determining module is specifically configured to fuse the structure representation information of each of the at least one network layer to obtain the structural characteristic of the structural network.
In another embodiment of the above apparatus according to the present invention, the feature determining module is specifically configured to fuse the structure representation information of each of the at least one network layer by using a neural network, so as to obtain the structural features of the structural network.
In another embodiment of the above apparatus according to the present invention, the feature determination module is specifically configured to input the structure representation information of each of the at least one network layer into a neural network, and the neural network fuses at least one of the structure representation information into one structure feature through a recursive algorithm.
In another embodiment of the above apparatus according to the present invention, further comprising:
the time acquisition unit is used for acquiring a time vector based on a preset time point corresponding to the structural network;
the performance determining unit is specifically configured to obtain a network performance parameter of the structural network at the preset time point based on the structural feature and the time vector.
In another embodiment of the above apparatus according to the present invention, the performance determination unit includes:
the merging module is used for merging the structural features and the time vectors to obtain merged features;
and the time performance module is used for obtaining the corresponding network performance parameters of the structural network under the time vector based on the merging characteristics by utilizing the multilayer perceptron.
In another embodiment of the apparatus according to the present invention, the time obtaining unit is specifically configured to obtain the time vector by searching a second preset table based on a preset time point corresponding to the structural network.
In another embodiment of the above apparatus according to the present invention, the performance determination unit is specifically configured to determine the network performance parameter of the structural network based on the structural feature before training the structural network.
In another embodiment of the above apparatus according to the present invention, further comprising:
a network training unit for training the prediction network using a plurality of sample structure networks, the sample structure networks being labeled with network performance parameters of the sample structure networks at each of at least one time point; the time point corresponds to a number of training times of the sample structure network.
In another embodiment of the above apparatus according to the present invention, further comprising:
the network block unit is used for sampling a plurality of preset structure networks to generate a network block; the network block comprises at least one network layer;
a network construction unit for constructing the sample structure network based on the network blocks.
In another embodiment of the above apparatus according to the present invention, the network block unit includes:
the layer sampling module is used for sampling a plurality of preset structure networks based on a Markov chain to obtain at least one preset network layer;
and the block forming module is used for sequentially connecting the at least one preset network layer to form the network block.
In another embodiment of the above apparatus according to the present invention, the network layer includes any one or more of the following:
convolutional layers, max pooling layers, average pooling layers, activation layers, and batch layers.
In another embodiment of the above apparatus according to the present invention, the network constructing unit is specifically configured to sequentially connect a plurality of network blocks to obtain the sample structure network, where a first network block and a second network block in the plurality of network blocks correspond to different feature dimensions.
In another embodiment of the above apparatus according to the present invention, a maximum pooling layer is connected between the first network block and the second network block.
In another embodiment of the above apparatus based on the present invention, the layer sampling module is specifically configured to sample a plurality of preset structure networks based on the sampled network parameters of the ith network layer to obtain an i +1 th network layer, where i is greater than or equal to 1 and is less than the number of network layers included in the network block.
According to another aspect of the embodiments of the present disclosure, there is provided an electronic device including a processor, the processor including the deployment control apparatus as described above.
According to another aspect of the embodiments of the present disclosure, there is provided an electronic device including: a memory for storing executable instructions;
and a processor in communication with the memory to execute the executable instructions to perform the operations of the content parsing method for video streams as described above.
According to another aspect of the embodiments of the present disclosure, there is provided a computer storage medium for storing computer-readable instructions which, when executed, perform the operations of the content parsing method for video streams as described above.
According to another aspect of the embodiments of the present disclosure, there is provided a computer program, which includes computer readable code, when the computer readable code runs on a device, a processor in the device executes instructions for implementing the steps in the content parsing method of video stream as described above.
Based on the method, the device, the equipment and the medium for predicting the network performance provided by the embodiment of the disclosure, the network parameters of the structural network to be predicted are obtained; determining structural characteristics of the structural network based on network parameters of the structural network; determining network performance parameters of the structural network based on the structural features; the performance of the structure network is predicted through the network parameters of the structure network, the method for predicting the network performance does not need to train the structure network, a large amount of time is saved, and the efficiency for predicting the network performance is improved.
The technical solution of the present disclosure is further described in detail by the accompanying drawings and examples.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the disclosure and together with the description, serve to explain the principles of the disclosure.
The present disclosure may be more clearly understood from the following detailed description, taken with reference to the accompanying drawings, in which:
fig. 1 is a schematic flow chart of a method for predicting network performance provided by an embodiment of the present disclosure.
FIG. 2 is a schematic diagram of a process for obtaining structural features in an embodiment of the method for predicting network performance according to the present disclosure.
Fig. 3 is a schematic structural diagram of a sample structure network generated in an embodiment of the present disclosure.
Fig. 4 is a schematic structural diagram of an example of predicting network performance of a structural network by using a prediction network according to an embodiment of the present disclosure.
Fig. 5 is a schematic structural diagram of an apparatus for predicting network performance according to an embodiment of the present disclosure.
Fig. 6 is a schematic structural diagram of an electronic device suitable for implementing the terminal device or the server according to the embodiment of the present application.
Detailed Description
Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that: the relative arrangement of the components and steps, the numerical expressions, and numerical values set forth in these embodiments do not limit the scope of the present disclosure unless specifically stated otherwise.
Meanwhile, it should be understood that the sizes of the respective portions shown in the drawings are not drawn in an actual proportional relationship for the convenience of description.
The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the disclosure, its application, or uses.
Techniques, methods, and apparatus known to those of ordinary skill in the relevant art may not be discussed in detail but are intended to be part of the specification where appropriate.
It should be noted that: like reference numbers and letters refer to like items in the following figures, and thus, once an item is defined in one figure, further discussion thereof is not required in subsequent figures.
The disclosed embodiments may be applied to computer systems/servers that are operational with numerous other general purpose or special purpose computing system environments or configurations. Examples of well known computing systems, environments, and/or configurations that may be suitable for use with the computer system/server include, but are not limited to: personal computer systems, server computer systems, thin clients, thick clients, hand-held or laptop devices, microprocessor-based systems, set top boxes, programmable consumer electronics, network pcs, minicomputer systems, mainframe computer systems, distributed cloud computing environments that include any of the above systems, and the like.
The computer system/server may be described in the general context of computer system-executable instructions, such as program modules, being executed by a computer system. Generally, program modules may include routines, programs, objects, components, logic, data structures, etc. that perform particular tasks or implement particular abstract data types. The computer system/server may be practiced in distributed cloud computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed cloud computing environment, program modules may be located in both local and remote computer system storage media including memory storage devices.
In other algorithms for automatically generating a neural network structure, each time a neural network structure is generated in a learning process, training is required to be performed on the neural network structure to know the performance of the neural network structure. Such a training may take at least 1 hour. Meanwhile, in order to make the generation algorithm learn a good generation capability, at least ten thousands of network structures need to be generated and the performance of the structures needs to be used as feedback (i.e. training each network for at least 1 hour), which causes the learning process of the generation algorithm to be inefficient.
The application provides a technical scheme for predicting the network performance, which is realized by using a prediction network and can predict the performance of various types of network structures. The embodiments of the present disclosure may be applied to a neural network having a sequential structure, and may be applied to a structural network having various different depths and/or network topology structures, which is not limited in the embodiments of the present disclosure.
Fig. 1 is a schematic flow chart of a method for predicting network performance provided by an embodiment of the present disclosure. As shown in fig. 1, the method includes:
s101, obtaining network parameters of the structural network to be predicted.
In particular, a fabric network may include multiple network layers, which may be connected in a certain order.
In one or more optional embodiments, the network parameters of the structural network may include layer parameters of at least one network layer included in the structural network, where the at least one network layer may be a part or all of the network layers of the structural network, for example, the layer parameters of each network layer of the structural network to be predicted may be obtained.
In at least one optional embodiment, the layer parameter may include any one or any combination of a computation type, a computation core (kernel) parameter, and a channel parameter, or may also include other types of layer parameters, where optionally, the computation core parameter may include a height and/or a width of the computation core (or may also be referred to as a length and a width of the computation core), the channel parameter may include a number of channels, and/or a ratio of the number of output channels to the number of input channels, or the computation core parameter and the channel parameter may also include other types of parameters, and a specific implementation of the layer parameter is not limited in the embodiments of the present application.
In an alternative example, the network parameters of the fabric network include any one or any combination of the following: the method comprises the steps of calculating the calculation type of each network layer in at least one network layer of the structure network, the length of a calculation core of each network layer, the width of the calculation core of each network layer, the number of channels of each network layer and the ratio of the number of output channels to the number of input channels of each network layer. Compared with the absolute value of the number of the channels of the network layer, the ratio of the number of the output channels to the number of the input channels of the network layer can effectively limit the numerical range of the parameters, thereby reducing the complexity of subsequent processing.
S102, determining the structural characteristics of the structural network based on the network parameters of the structural network.
In at least one alternative embodiment, the structural features may have a certain format, for example, the structural features may have pre-set dimensions and/or structures, which may enable networks having different depths and topologies to be handled in a uniform manner.
S103, determining the network performance parameters of the structural network based on the structural characteristics.
Optionally, the network performance parameter of the structural network may include accuracy of the structural network, where as an example, the network performance parameter of the structural network may include accuracy of the structural network corresponding to one or more training times, where the structural network corresponding to a certain training time may represent a structural network obtained after the structural network is trained by using the training times. Optionally, the network performance parameter may also include other parameters, which are not limited in this embodiment of the present application.
Based on the method for predicting the network performance provided by the embodiment of the disclosure, the network parameters of the structural network to be predicted are obtained; determining structural characteristics of the structural network based on network parameters of the structural network; determining network performance parameters of the structural network based on the structural features; the method for predicting the network performance realizes the prediction of the performance of the structural network through the network parameters of the structural network, does not need to train the structural network, saves a large amount of time, and improves the efficiency of predicting the network performance.
Optionally, in a specific example of the foregoing embodiment of the method for predicting network performance in the present disclosure, S102 includes:
based on the network parameters of the fabric network, structure representation information for each of at least one network layer of the fabric network is determined.
In at least one optional embodiment, the structural representation information of the network layer may be determined based on layer parameters of the network layer. In this way, the structure representation information of each of the at least one network layer of the structured network may be determined based on the layer parameters of the each network layer.
The structure representation information may indicate structural characteristics of the layer. Optionally, the structure representing information may specifically have a preset format and/or dimension, and in one or more optional embodiments, the structure representing information may include a vector having a preset dimension. Thus, different layers are distinguished through unified structural representation, and unified processing aiming at various different network structures is realized.
In an optional example, the structure representation information of a certain network layer may include a vector corresponding to a computation type of the network layer, a vector corresponding to a computation core height of the network layer, a vector corresponding to a computation core width of the network layer, and a vector corresponding to a ratio of the number of output channels to the number of input channels of the network layer. For example, the structure representing information may be obtained by performing a splicing or merging process on the four vectors, but the structure representing information may also be implemented in other manners, which is not limited in this embodiment of the present application.
In one or more alternative embodiments, the structural representation information of the network layer may be determined by the following uniform layer coding:
determining at least one identifier of the network layer based on layer parameters of the network layer;
based on at least one identifier of the network layer, structural representation information of the network layer is determined.
The identifier here may be an integer identifier. Optionally, the at least one integer, i.e. the at least one identifier, may be obtained by encoding a layer parameter of the network layer, for example, each parameter value in the layer parameter may be encoded as an integer, so that a tuple (tuple) of the at least one integer included in the network layer may be obtained. In an alternative example, the layer parameters of the network layer may include a computation type of the network layer, a width of a computation core of the network layer, a height of the computation core of the network layer, and a ratio of the number of input channels to the number of output channels of the network layer, and accordingly, each parameter value in the layer parameters may be encoded to obtain a quadruple (TY, KW, KH, CH), where TY denotes an identifier corresponding to the computation type, KW and KH denote the width and height of the computation core, respectively, and CH denotes an identifier corresponding to a ratio of the number of output channels to the number of input channels, but the embodiment of the present application is not limited thereto.
In one or more optional embodiments, a correspondence between the parameter value and the identifier, for example, a correspondence between a preset calculation type and an integer identifier, may be preset, and one or more parameter values of the network layer may be encoded based on the correspondence, but the embodiments of the present application are not limited thereto.
In one or more optional embodiments, when encoding the ratio of the number of output channels to the number of input channels, the ratio may be quantized to a preset value, and then the identifier corresponding to the ratio may be determined based on the quantization result of the ratio of the number of output channels to the number of input channels. For example, it may be preset 8 intervals (bins), whose center points are 0.25,0.5,0.75,1.0,1.5,2.0,2.5, and 3.0, respectively, and then quantize the ratio of the number of input channels to the number of input channels of the network layer to one of the 8 intervals, and use the number or index of the determined quantization interval as the identifier corresponding to the ratio of the number of output channels to the number of input channels. Optionally, the identifier may also be determined in other manners, which is not limited in this application embodiment.
In the embodiment of the application, after obtaining at least one identifier of the network layer, the obtaining determines the structure representation information of the network layer based on the at least one identifier. In one or more optional embodiments, each of the at least one identifier of the network layer may be mapped, a mapping result of each identifier may be obtained, and the structure representation information of the network layer may be obtained based on the mapping result of each of the at least one identifier.
In an alternative embodiment, the mapping may be implemented by looking up a table, for example, a first preset table may be looked up based on a certain identifier corresponding to the network layer, and a mapping result of the identifier is obtained. Optionally, the first preset table here may be obtained in a process of training the prediction network, for example, an initial value in the first preset table may be a random number, and then the value in the first preset table is updated by training the prediction network, so as to obtain the first preset table finally, but the embodiment of the present application does not limit this.
Optionally, each identifier corresponding to the network layer may be mapped by using a lookup table, where as shown in fig. 2, identifiers corresponding to different parameter values may be mapped by using different tables, for example, identifiers (or integer numbers) corresponding to a calculation type, a calculation core width, a calculation core height, and an output-input channel number ratio may be mapped by using different tables, so as to obtain a type vector, a core width vector, a core height vector, and a channel vector. Thus, higher latitude vectors are obtained by looking up the table, and the higher latitude vectors can express richer features of each layer.
In this embodiment of the present disclosure, after the mapping result of each identifier in the at least one identifier is obtained, the mapping results of the at least one identifier may be spliced or combined to obtain the structure indicating information of the network layer. For example, the type vector, the kernel width vector, the kernel height vector, and the channel vector of the network layer may be subjected to a stitching process to obtain a vector representation of the network layer, and at this time, the structure representation information is a vector of a fixed dimension, but the embodiment of the present application is not limited thereto.
Optionally, S120 may further include: structural features of the structural network are determined based on the structural representation information of each of the at least one network layer.
In one or more optional examples, the structural representation information of each of the at least one network layer may be fused or spliced to obtain the structural features of the structural network. For example, the structure representation information of each network layer in all network layers of the structure network may be subjected to a splicing process to obtain the structure characteristics of the structure network. Alternatively, the structural characteristics of the structural network may also be determined in other ways, which is not limited in this application.
Alternatively, the neural network may be used to fuse or splice the structure representation information of each of the at least one network layer. The neural network has unique advantages in the aspect of information processing, can be trained to obtain a targeted neural network aiming at different tasks, and can fuse the structure representation information based on the trained neural network, so that a better fusion effect can be achieved.
In one or more optional embodiments, fusing the structure representation information of each of the at least one network layer using the neural network comprises: the structural representation information of each network layer in at least one network layer is input into the neural network processing.
The neural network may utilize one or more algorithms to fuse the input structural representation information of the one or more network layers, for example, the structural representation information of the at least one network layer may be fused along the sequence of the at least one network layer. Alternatively, the neural network may fuse at least one structure representation information into a structural feature by a recursive algorithm. For example, the above fusion or splicing process may be performed by using a Long-Short Term Memory (LSTM) network, but the embodiment of the present application is not limited to the specific implementation of the neural network.
In one or more alternative embodiments, the LSTM network may include at least one LSTM element, each LSTM element corresponding to a network layer. Alternatively, the LSTM may include maintaining a hidden state htAnd a unit memory ctAnd using input gates itOutput gate otAnd a forgetting gate (forget gate) ftTo control the flow of information. At each step, it needs to input xtDetermining the values of all gates to produce an output utAnd updates hidden state (hiddenstate) htAnd cell memory ctAs shown in formulas (1) to (6):
it＝σ(W(i)xt+U(i)ht-1+b(i)) Formula (1)
ot＝σ(W(o)xt+U(o)ht+b(o)) Formula (2)
ft＝σ(W(f)xt+U(f)ht+b(f)) Formula (3)
ut＝tanh(W(u)xt+U(u)ht-1+b(u)) Formula (4)
ct＝it⊙ut+ft⊙ct-1Formula (5)
ht＝ot⊙tanh(ct) Formula (6)
Wherein σ represents a Sigmoid function, and-indicates an element and or operation, and the LSTM layer-by-layer information gradually integrated into the network is in a hidden state from a low level to a high level. In the last step, i.e. the classification of fully connected layers before a layer, the LSTM elements of the hidden state are extracted to represent the structure of the whole network, called structural features.
In this embodiment of the present disclosure, optionally, the network performance parameter predicted in S103 may be a predicted network performance parameter of the structural network after the training is completed, or may also be a predicted network performance parameter corresponding to a certain time point of the structural network in the training process, where the time point may correspond to a period of completion of a certain training. In one or more optional embodiments, before S103, a time vector may also be obtained based on a preset time point corresponding to the structural network.
Specifically, the preset time point may be represented as an Epoch ID, which represents the number of times that the structural network needs to pass through, for example, the number of times that the structural network repeatedly traverses the entire training set in normal training. Alternatively, the time vector may be obtained by encoding or mapping the preset time point. For example, the time vector is obtained by querying a second preset table by using the preset time point, where optionally, the second preset table may be obtained by training the prediction network, and specifically, the above description of the first preset table may be referred to, and the embodiment of the present application is not limited to the specific implementation thereof.
In one or more optional embodiments, after obtaining the time vector, in S103, a network performance parameter of the structural network at a preset time point may be obtained based on the structural feature and the time vector.
Optionally, the structural features of the structural network and the time vector may be subjected to splicing or fusion processing to obtain a splicing or fusion result, and based on the splicing or fusion result, the network performance parameters of the structural network at the preset time point are obtained.
In one or more alternative examples, the structural features and the time vector may be combined to obtain combined features, and the combination may be implemented by means of concatenation. Optionally, the merging result may be processed by using a multilayer sensor, so as to obtain a network performance parameter of the structural network at a preset time point.
Specifically, the multi-layer perceptron may be obtained through training, and the multi-layer perceptron obtained through training may obtain network performance parameters based on the structural features and the time vectors.
In one or more optional embodiments, the multilayer perceptron may include one or more of a fully-connected layer, a Batch Normalization (BN) layer, and an activation layer. For example, the multi-layer sensor may include three fully-connected layers, a BN layer, and a modified Linear Unit (ReLU) activation layer, but the embodiment of the present application is not limited to the specific implementation of the multi-layer sensor.
In the embodiment of the present disclosure, optionally, before the structural network is trained, the network performance parameters of the structural network may be determined through the above procedure, and the accuracy of the structural network at different time points may be obtained without training the structural network, so that the time for predicting the network performance is greatly shortened (which may be shortened to 0.01 second), the network performance of the structural network generated by an automatic generation algorithm and the like may be predicted more quickly, and the probability of obtaining a neural network with excellent performance is improved.
Optionally, before S101, the prediction network may be trained by using sample data, where the sample data includes a plurality of sample structure networks and network performance parameters of the sample structure networks at each of at least one time point.
In this way, the prediction network can be trained using a plurality of sample structure networks labeled with network performance parameters (e.g., accuracy), resulting in a training network for use.
Optionally, in the process of training the prediction network, the training times of the sample structure network may include each training of the sample structure network in the training process, and when the network performance parameters of the structure network are predicted by using the prediction network, only the network performance parameters corresponding to the structure network when the training is completed may be obtained.
In embodiments of the present invention, the network of sample structures may be generated in a block-based manner. In particular, the sample structure network may be constructed by stacking similarly structured blocks. Each block can be constructed first, and then the constructed blocks are stacked in a certain network architecture to form a sample structure network.
Optionally, before training the prediction network by using a plurality of sample structure networks, the method further includes:
and carrying out network layer sampling on the plurality of preset structure networks to generate network blocks. Specifically, a plurality of network layers obtained by sampling are connected to form a network block, wherein the connection relationship between the network layers is determined by network parameters in the network layers, such as: and obtaining network parameters of the j layer network layer, and determining what network layer the j +1 layer network layer is based on the network parameters. Wherein, the network layer comprises any one or more of the following: convolutional layers, max pooling layers, average pooling layers, activation layers, and batch layers.
Specifically, the network block includes at least one network layer; the preset structure network can be obtained based on a network or a large database, the obtained structure network is usually a neural network, a network layer is obtained by sampling the neural network, the network layer is combined to obtain a network block, and the sample structure network obtained by combining the network blocks can overcome the defects of the neural network obtained by directly combining the network layers in the prior art, wherein the defects of the prior art comprise: network layers which cannot be combined are combined together, and network layers which cannot be connected or are meaningless after connection are connected together, at this time, the neural network obtained based on the prior art cannot be applied to training.
In one or more embodiments, a sample structure network is constructed based on the network blocks constructed in any of the above embodiments.
In the case of a structured sample structure network, in addition to the network blocks, an averaging pooling layer and a linear layer are usually included in the end. The input to the pooling layer is generally derived from the last convolutional layer, and the main function is to provide strong robustness (for example, max-pooling takes the maximum value in a small region, and if other values in this region slightly change or the image slightly shifts, the result after pooling still does not change), and reduce the number of parameters to prevent the over-fitting phenomenon. The pooling layer generally has no parameters, so that only derivation is needed for input parameters during back propagation, and weight updating is not needed.
Specifically, the process of sampling a preset structure network and generating a network block includes:
sampling a plurality of preset structure networks based on a Markov chain to obtain at least one preset network layer;
and connecting at least one preset network layer in sequence to form a network block. The network layers that make up a network block typically do not exceed 10 layers, whereas the first layer of a network block is typically a convolutional layer.
Optionally, constructing a sample structure network based on the network blocks includes:
sequentially connecting a plurality of network blocks to obtain a sample structure network, or stacking a plurality of network blocks to obtain a sample structure network, wherein a first network block and a second network block in the plurality of network blocks correspond to different feature dimensions; there is no network block between the first network block and the second network block.
The maximum pooling layer is connected between the first network block and the second network block.
Optionally, the markov chain sampling a plurality of preset structure networks to obtain at least one preset network layer, including:
and sampling a plurality of preset structure networks based on the sampled network parameters of the ith network layer to obtain an (i + 1) th network layer.
At each step of the Markov chain, the system may change from one state to another state, or may maintain the current state, according to a probability distribution. The change of state is called a transition and the probability associated with a different state change is called a transition probability. Specifically, sampling is performed based on a markov chain, and what network layer the next layer can be connected to can be determined based on the network parameters of the previous network layer, so that a network block can be quickly constructed.
The obtaining of the network block is based on the condition that the previous network layer determines the next network layer, and may be determined according to a probability between preset transition layer types, wherein the probability between the transition layer types is obtained from an empirical estimation of an actual network. For example, the next network layer of a convolutional layer has a high probability to be a batch layer and a nonlinear activation layer. The next layer of the active layer has a higher probability of connecting the convolutional or pool layers.
Fig. 3 is a schematic structural diagram of a sample structure network generated in an embodiment of the present disclosure.
The sample structure network formed in fig. 3 includes three regions, the three regions correspond to different feature dimensions, each region includes two network blocks, a maximum pooling layer is connected between every two regions, and each region corresponds to a different feature dimension, and in this embodiment, the feature dimensions corresponding to the three regions are 16, 32, and 64, respectively.
Specifically, fig. 4 is a schematic structural diagram of an example of predicting network performance of a structural network by using a prediction network according to an embodiment of the present disclosure. Firstly, each Layer in a Block Architecture (i.e., Block Architecture) of a structural network may be input to a prediction network, and the prediction network may encode and map (or is called embedding) each Layer in the Block Architecture, specifically, the calculation type, the width and the height of a calculation core, and the output-to-input channel ratio of each Layer may be encoded, respectively, to obtain a corresponding integer Code, so that a tuple (TY, KW, KH, CH) including four integer codes corresponding to each Layer may be obtained (i.e., each row in a Block Layer Code); then, Layer Embedding (i.e., Layer Embedding) may be performed on the tuple corresponding to each Layer, specifically, mapping may be performed on each integer code in the tuple to obtain a vector corresponding to each integer code, and the vectors corresponding to the four integer codes are combined into one 4-dimensional vector, where the fixed-length vector may be represented as a code vector of the Layer. Then, the 4-dimensional vectors corresponding to each layer may be respectively input to an LSTM unit of the LSTM, and the LSTM may process the input 4-dimensional vectors to obtain an overall representation of the entire Structural network, i.e., Structural features (Structural features) of the Structural network. In addition, a time period identifier (Epoch ID) corresponding to a time point to be predicted may be input into the prediction network, and the prediction network may obtain a time Vector or a time period Vector (Epoch Vector) by means of time period Embedding (i.e., Epoch Embedding), where the time period Embedding may be implemented by means of table lookup, similar to the layer Embedding process described above. Finally, the structural features and the epoch vectors may be combined and the combined result input to a Multi-Layer Perceptron (MLP) that outputs network performance parameters of the structural network, e.g., the accuracy of the structural network predicted by the prediction network at the point in time.
It should be understood that the example shown in fig. 4 is only used to help understand the technical solutions of the embodiments of the present disclosure, and should not be understood as a limitation to the embodiments of the present disclosure, and the embodiments of the present disclosure may also be implemented in other ways.
Those of ordinary skill in the art will understand that: all or part of the steps for implementing the method embodiments may be implemented by hardware related to program instructions, and the program may be stored in a computer readable storage medium, and when executed, the program performs the steps including the method embodiments; and the aforementioned storage medium includes: various media that can store program codes, such as ROM, RAM, magnetic or optical disks.
Those of ordinary skill in the art will understand that: all or part of the steps for implementing the method embodiments may be implemented by hardware related to program instructions, and the program may be stored in a computer readable storage medium, and when executed, the program performs the steps including the method embodiments; and the aforementioned storage medium includes: various media that can store program codes, such as ROM, RAM, magnetic or optical disks.
Fig. 5 is a schematic structural diagram of an apparatus for predicting network performance according to an embodiment of the present disclosure. The apparatus of this embodiment may be used to implement the method embodiments of the present disclosure described above. As shown in fig. 5, the apparatus of this embodiment is implemented using a predictive network, and includes:
a parameter obtaining unit 51, configured to obtain a network parameter of the structural network to be predicted.
Specifically, the network parameters include at least one of:
the method comprises the steps of calculating the type of each network layer in at least one network layer of the structure network, the length of a calculation core of the network layer, the width of the calculation core of the network layer, the number of channels of the network layer and the ratio of the number of output channels to the number of input channels of the network layer.
A structural feature unit 52, configured to determine a structural feature of the structural network based on the network parameter of the structural network.
And a performance determining unit 53, configured to determine a network performance parameter of the fabric network based on the fabric feature.
Based on the device for predicting network performance provided by the embodiment of the disclosure, network parameters of a structural network to be predicted are obtained; determining structural characteristics of the structural network based on network parameters of the structural network; determining network performance parameters of the structural network based on the structural features; the performance of the structure network is predicted through the network parameters of the structure network, the method for predicting the network performance does not need to train the structure network, a large amount of time is saved, and the efficiency for predicting the network performance is improved.
A structural feature cell 52 comprising:
an information representation module for determining structure representation information of each network layer of at least one network layer of the structural network based on network parameters of the structural network;
optionally, the structure representation information includes a structure vector having a preset dimension.
A feature determination module for determining structural features of the structural network based on the structural representation information of each of the at least one network layer.
Specifically, the information representation module may include a symbol recognition module and an information module;
a symbol recognition module for determining at least one identifier for each network layer of at least one network layer of the fabric network based on layer parameters of each network layer;
an information module for determining structure representation information for each of the at least one network layer based on at least one identifier for each of the at least one network layer.
Optionally, in a specific example, the symbol recognition module is specifically configured to map each identifier of the at least one identifier of each network layer, so as to obtain a mapping result of each identifier;
and the information module is specifically used for obtaining the structure representation information of each network layer based on the mapping result of each identifier in the at least one identifier.
Specifically, the symbol recognition module obtains the mapping result based on the identifier by searching a first preset table to obtain the mapping result of each identifier. The first preset table is determined in a training process through a prediction network, numerical values in the initial first preset table are randomly generated, and numerical value information specific to a specific task is obtained through the training process.
Optionally, in a specific example, the feature determining module is specifically configured to fuse the structure representation information of each network layer in the at least one network layer to obtain the structural feature of the structural network.
Specifically, the fusion structure representing information may be obtained by fusing the structure representing information of each network layer in at least one network layer by using a neural network, to obtain the structural features of the structural network.
The information is represented by fusing the structure by the neural network, specifically, the structure representation information of each network layer in at least one network layer is input into the neural network, and the neural network fuses the at least one structure representation information into one structure characteristic by a recursive algorithm. Wherein a recursive algorithm is a procedure that directly or indirectly invokes its own algorithm. Recursive algorithms are effective in solving a large class of problems in computer programming, which tends to make the description of the algorithms concise and easy to understand.
On the basis of the foregoing embodiments, another embodiment of the apparatus for predicting network performance according to the present disclosure may further include:
and the time acquisition unit is used for acquiring a time vector based on a preset time point corresponding to the structural network.
Specifically, a preset time point Epoch ID represents the number of times that a structural network needs to be trained, a time vector is obtained through the number of times in a mapping table look-up mode, the time vector is a high latitude vector, and a mapping table of the time vector is obtained through training; and expressing the traversal times of the structural network repeatedly traversing the whole training set in normal training through the defined time point.
At this time, the performance determining unit 53 is specifically configured to obtain the network performance parameter of the structural network at the preset time point based on the structural feature and the time vector.
Specifically, after the structural features and the time vectors are combined, a multi-layer perceptron is used for processing, and network performance parameters are obtained, and are generally expressed through the accuracy of the structural network.
Merging the structural features and the time vectors to obtain merged features; the method for combining the structural features and the time vectors can be realized in a splicing mode, the spliced combined features can reflect the structural features and simultaneously correspond the structural features to the time points, the combined feature expression obtained at the moment corresponds to the time points, and the accuracy obtained based on the combined features corresponds to the time points of the time vectors.
Optionally, in a specific example, the time obtaining unit is specifically configured to obtain the time vector by searching a second preset table based on a preset time point corresponding to the structural network.
Specifically, the method implemented by the device of the present disclosure is performed before the structure network is trained. By obtaining the network performance parameters (such as accuracy) of the structural network before the structural network is trained, the time for predicting the network performance can be greatly shortened, and the efficiency of performance prediction is improved.
On the basis of the foregoing embodiments, the present disclosure further includes:
the network training unit is used for training the prediction network by utilizing a plurality of sample structure networks, and the sample structure networks are marked with network performance parameters of the sample structure networks at each time point in at least one time point; the time point corresponds to the number of training times of the sample structure network.
The time point corresponds to the training times of the sample structure network, the training times are performed according to each training time (each time required for training completion) in the training process, and when the network is predicted in specific application, the network performance parameter corresponding to the time point of the last training time is only required to be obtained.
Optionally, before the network training unit, the method further includes:
and the network block unit is used for sampling the network layers of the plurality of preset structure networks to generate network blocks.
Specifically, the network block unit may include:
the layer sampling module is used for sampling a plurality of preset structure networks based on a Markov chain to obtain at least one preset network layer;
and the block forming module is used for sequentially connecting at least one preset network layer to form a network block.
Wherein, the network block comprises at least one network layer; the preset structure network can be obtained based on a network or a large database, the obtained structure network is usually a neural network, a network layer is obtained by sampling the neural network, the network layer is combined to obtain a network block, and the sample structure network obtained by combining the network blocks can overcome the defects of the neural network obtained by directly combining the network layers in the prior art, wherein the defects of the prior art comprise: network layers which cannot be combined are combined together, and network layers which cannot be connected or are meaningless after connection are connected together, at this time, the neural network obtained based on the prior art cannot be applied to training.
And the network construction unit is used for constructing the sample structure network based on the network blocks. The method is particularly applicable to sequentially connecting a plurality of network blocks to obtain a sample structure network, wherein a first network block and a second network block in the plurality of network blocks correspond to different feature dimensions.
Specifically, the network layer includes any one or more of: convolutional layers, max pooling layers, average pooling layers, activation layers, and batch layers.
Optionally, a max-pooling layer is connected between the first network block and the second network block. The network layers that make up a network block typically do not exceed 10 layers, whereas the first layer of a network block is typically a convolutional layer.
Optionally, the layer sampling module is specifically configured to sample a plurality of preset structure networks based on the sampled network parameters of the ith network layer to obtain an (i + 1) th network layer, where i is greater than or equal to 1 and is less than the number of network layers included in the network block.
According to another aspect of the embodiments of the present disclosure, there is provided an electronic device including a processor, where the processor includes the apparatus for predicting network performance according to any one of the embodiments of the present disclosure.
According to another aspect of the embodiments of the present disclosure, there is provided an electronic device including: a memory for storing executable instructions;
and a processor in communication with the memory for executing the executable instructions to perform the operations of any of the above embodiments of the method of predicting network performance of the present disclosure.
According to another aspect of the embodiments of the present disclosure, a computer storage medium is provided for storing computer-readable instructions that, when executed, perform the operations of any one of the above embodiments of the method for predicting network performance of the present disclosure.
According to another aspect of an embodiment of the present disclosure, there is provided a computer program comprising computer readable code which, when run on a device, executes instructions of the steps of the method of predicting network performance of the present disclosure.
According to yet another aspect of the embodiments of the present disclosure, there is provided a computer program product for storing computer readable instructions, which when executed, cause a computer to execute the human key point detection method described in any one of the above possible implementation manners.
In an alternative embodiment the computer program product is embodied as a computer storage medium, and in another alternative embodiment the computer program product is embodied as a software product, such as an SDK.
In addition, the embodiment of the present disclosure also provides a computer storage medium for storing computer readable instructions, which when executed, implement the method for predicting network performance of any of the above embodiments of the present disclosure.
In addition, the embodiment of the present disclosure also provides a computer program, which includes computer readable instructions, and when the computer readable instructions are run in a device, a processor in the device executes a method for implementing the method for predicting network performance according to any one of the above embodiments of the present disclosure.
In an alternative embodiment, the computer program is embodied in a Software product, such as a Software Development Kit (SDK), or the like.
In one or more optional implementation manners, the present disclosure also provides a computer program product for storing computer readable instructions, which when executed, cause a computer to execute the method for predicting network performance of any of the above embodiments.
The computer program product may be embodied in hardware, software or a combination thereof. In an alternative example, the computer program product is embodied as a computer storage medium, and in another alternative example, the computer program product is embodied as a software product, such as an SDK or the like.
In one or more optional implementation manners, the present disclosure also provides a method for predicting network performance, and a corresponding apparatus and electronic device, a computer storage medium, a computer program, and a computer program product, where the method includes: the first device sending an indication of predicted network performance to the second device, the indication causing the second device to perform the method of predicting network performance in any of the possible embodiments described above; the first device receives the network performance parameters sent by the second device.
In some embodiments, the human body key point detection indication may be embodied as a call instruction, and the first device may instruct the second device to perform the prediction of the network performance by calling, and accordingly, in response to receiving the call instruction, the second device may perform the steps and/or processes in any embodiment of the method for predicting the network performance.
The embodiment of the disclosure also provides an electronic device, which may be a mobile terminal, a Personal Computer (PC), a tablet computer, a server, and the like. Referring now to fig. 6, there is shown a schematic diagram of an electronic device 600 suitable for use in implementing a terminal device or server of an embodiment of the present application: as shown in fig. 6, computer system 600 includes one or more processors, communications, etc., such as: one or more Central Processing Units (CPUs) 601, and/or one or more image processors (GPUs) 613, etc., which may perform various appropriate actions and processes according to executable instructions stored in a Read Only Memory (ROM)602 or loaded from a storage section 608 into a Random Access Memory (RAM) 603. Communications portion 612 may include, but is not limited to, a network card, which may include, but is not limited to, an IB (Infiniband) network card.
The processor may communicate with the read-only memory 602 and/or the random access memory 630 to execute the executable instructions, connect with the communication part 612 through the bus 604, and communicate with other target devices through the communication part 612, so as to complete the operations corresponding to any method provided by the embodiments of the present application, for example, obtaining the network parameters of the structural network to be predicted; determining structural characteristics of the structural network based on network parameters of the structural network; and determining and obtaining the network performance parameters of the structural network based on the structural characteristics.
In addition, in the RAM603, various programs and data necessary for the operation of the device can also be stored. The CPU601, ROM602, and RAM603 are connected to each other via a bus 604. The ROM602 is an optional module in case of the RAM 603. The RAM603 stores or writes executable instructions into the ROM602 at runtime, and the executable instructions cause the processor 601 to perform operations corresponding to the above-described communication method. An input/output (I/O) interface 605 is also connected to bus 604. The communication unit 612 may be integrated, or may be provided with a plurality of sub-modules (e.g., a plurality of IB network cards) and connected to the bus link.
The following components are connected to the I/O interface 605: an input portion 606 including a keyboard, a mouse, and the like; an output portion 607 including a display such as a Cathode Ray Tube (CRT), a Liquid Crystal Display (LCD), and the like, and a speaker; a storage section 608 including a hard disk and the like; and a communication section 609 including a network interface card such as a LAN card, a modem, or the like. The communication section 609 performs communication processing via a network such as the internet. The driver 610 is also connected to the I/O interface 605 as needed. A removable medium 611 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, or the like is mounted on the drive 610 as necessary, so that a computer program read out therefrom is mounted in the storage section 608 as necessary.
It should be noted that the architecture shown in fig. 6 is only an optional implementation manner, and in a specific practical process, the number and types of the components in fig. 6 may be selected, deleted, added or replaced according to actual needs; in different functional component settings, separate settings or integrated settings may also be used, for example, the GPU and the CPU may be separately set or the GPU may be integrated on the CPU, the communication part may be separately set or integrated on the CPU or the GPU, and so on. These alternative embodiments are all within the scope of the present disclosure.
In particular, according to an embodiment of the present disclosure, the processes described above with reference to the flowcharts may be implemented as computer software programs. For example, embodiments of the present disclosure include a computer program product comprising a computer program tangibly embodied on a machine-readable medium, the computer program comprising program code for performing the method illustrated in the flowchart, the program code may include instructions corresponding to performing the method steps provided by embodiments of the present disclosure, e.g., obtaining network parameters of a structural network to be predicted; determining structural characteristics of the structural network based on network parameters of the structural network; and determining and obtaining the network performance parameters of the structural network based on the structural characteristics. In such an embodiment, the computer program may be downloaded and installed from a network through the communication section 609, and/or installed from the removable medium 611. The computer program performs the above-described functions defined in the method of the present application when executed by a Central Processing Unit (CPU) 601.
The methods and apparatus, devices of the present disclosure may be implemented in a number of ways. For example, the methods and apparatuses, devices of the present disclosure may be implemented by software, hardware, firmware, or any combination of software, hardware, firmware. The above-described order for the steps of the method is for illustration only, and the steps of the method of the present disclosure are not limited to the order specifically described above unless specifically stated otherwise. Further, in some embodiments, the present disclosure may also be embodied as programs recorded in a recording medium, the programs including machine-readable instructions for implementing the methods according to the present disclosure. Thus, the present disclosure also covers a recording medium storing a program for executing the method according to the present disclosure.
The description of the present disclosure has been presented for purposes of illustration and description, and is not intended to be exhaustive or limited to the disclosure in the form disclosed. Many modifications and variations will be apparent to practitioners skilled in this art. The embodiment was chosen and described in order to best explain the principles of the disclosure and the practical application, and to enable others of ordinary skill in the art to understand the disclosure for various embodiments with various modifications as are suited to the particular use contemplated.

Claims (41)

1. A method of predicting network performance, implemented using a predictive network, comprising:
acquiring network parameters of a structural network to be predicted; the network parameters of the fabric network comprise layer parameters of at least one network layer comprised by the fabric network;
determining structural characteristics of the structural network based on network parameters of the structural network;
determining network performance parameters of the structural network based on the structural features prior to training the structural network;
the determining the structural characteristics of the structural network based on the network parameters of the structural network comprises:
determining structure representation information of each of at least one network layer of the fabric network based on network parameters of the fabric network;
determining structural features of the structural network based on the structural representation information of each of the at least one network layer.
2. The method of claim 1, wherein the layer parameters comprise at least one of:
the method comprises the steps of calculating the calculation type of each network layer in at least one network layer of the structural network, the length of a calculation core of the network layer, the width of the calculation core of the network layer, the number of channels of the network layer and the ratio of the number of output channels to the number of input channels of the network layer.
3. The method of claim 1, wherein the structure representation information comprises a structure vector having preset dimensions.
4. The method of claim 1, wherein determining the structure representation information for each of the at least one network layer of the fabric network based on the network parameters of the fabric network comprises:
determining at least one identifier for each of at least one network layer of the fabric network based on layer parameters of said each network layer;
determining structure representation information of each of the at least one network layer based on at least one identifier of the each network layer.
5. The method of claim 4, wherein determining the structural representation information of each of the at least one network layer based on at least one identifier of the each network layer comprises:
mapping each identifier in at least one identifier of each network layer to obtain a mapping result of each identifier;
and obtaining the structure representation information of each network layer based on the mapping result of each identifier in the at least one identifier.
6. The method of claim 5, wherein the mapping each identifier of the at least one identifier of each network layer to obtain a mapping result of each identifier comprises:
and obtaining the mapping result of each identifier based on each identifier in each network layer by searching a first preset table.
7. The method of claim 1, wherein determining the structural characteristics of the structural network based on the structural representation information of each of the at least one network layer comprises:
and fusing the structure representation information of each network layer in the at least one network layer to obtain the structure characteristics of the structure network.
8. The method according to claim 7, wherein fusing the structure representation information of each of the at least one network layer to obtain the structure features of the structure network comprises:
and fusing the structure representation information of each network layer in the at least one network layer by utilizing a neural network to obtain the structural characteristics of the structural network.
9. The method of claim 8, wherein said fusing the structural representation information of each of the at least one network layer using a neural network comprises:
the structure representation information of each network layer in at least one network layer is input into a neural network, and the neural network fuses at least one structure representation information into a structure characteristic through a recursive algorithm.
10. The method according to any of claims 1-9, wherein prior to determining the network performance parameter of the fabric network based on the fabric characteristic, further comprising:
obtaining a time vector based on a preset time point corresponding to the structural network;
determining network performance parameters of the fabric network based on the fabric feature, including:
and obtaining the network performance parameters of the structural network at the preset time point based on the structural features and the time vector.
11. The method of claim 10, wherein the deriving the network performance parameter of the structural network at the preset time point based on the structural feature and the time vector comprises:
merging the structural features and the time vectors to obtain merged features;
and obtaining the corresponding network performance parameters of the structural network under the time vector based on the merging characteristics by utilizing a multilayer perceptron.
12. The method according to claim 10, wherein obtaining the time vector based on the preset time point corresponding to the structural network comprises:
and obtaining the time vector by searching a second preset table based on a preset time point corresponding to the structural network.
13. The method according to any of claims 1-9, further comprising, before said obtaining network parameters of the structural network to be predicted:
training the prediction network using a plurality of sample structure networks, the sample structure networks being labeled with network performance parameters for the sample structure networks at each of at least one time point; the time point corresponds to a number of training times of the sample structure network.
14. The method of claim 13, wherein prior to training the predictive network using the plurality of sample structure networks, further comprising:
carrying out network layer sampling on a plurality of preset structure networks to generate network blocks; the network block comprises at least one network layer;
constructing the sample structure network based on the network blocks.
15. The method of claim 14, wherein the network layer sampling the plurality of pre-configured networks to generate the network block comprises:
sampling a plurality of preset structure networks based on a Markov chain to obtain at least one preset network layer;
and sequentially connecting the at least one preset network layer to form the network block.
16. The method of claim 14, wherein the network layer comprises any one or more of:
convolutional layers, max pooling layers, average pooling layers, activation layers, and batch layers.
17. The method of claim 15, wherein the constructing the sample structure network based on the network blocks comprises:
connecting a plurality of the network blocks in sequence to obtain the sample structure network, wherein a first network block and a second network block in the plurality of network blocks correspond to different feature dimensions.
18. The method of claim 17, wherein a max-pooling layer is connected between the first network block and the second network block.
19. The method of claim 15, wherein sampling a plurality of pre-configured networks based on a markov chain to obtain at least one pre-configured network layer comprises:
and sampling a plurality of preset structure networks based on the sampled network parameters of the ith network layer to obtain an ith +1 network layer, wherein i is greater than or equal to 1 and less than the number of network layers included in the network block.
20. An apparatus for predicting network performance, implemented using a predictive network, comprising:
a parameter obtaining unit, configured to obtain a network parameter of a structural network to be predicted; the network parameters of the fabric network comprise layer parameters of at least one network layer comprised by the fabric network;
a structural feature unit, configured to determine a structural feature of the structural network based on a network parameter of the structural network;
a performance determination unit, configured to determine a network performance parameter of the structural network based on the structural feature before training the structural network;
the structural feature cell includes:
an information representation module for determining structure representation information of each of at least one network layer of the fabric network based on network parameters of the fabric network;
a feature determination module for determining structural features of the structural network based on the structural representation information of each of the at least one network layer.
21. The apparatus of claim 20, wherein the layer parameters comprise at least one of:
the method comprises the steps of calculating the calculation type of each network layer in at least one network layer of the structural network, the length of a calculation core of the network layer, the width of the calculation core of the network layer, the number of channels of the network layer and the ratio of the number of output channels to the number of input channels of the network layer.
22. The apparatus of claim 20, wherein the structure representation information comprises a structure vector having preset dimensions.
23. The apparatus of claim 20, wherein the information presentation module comprises:
a symbol recognition module for determining at least one identifier for each of at least one network layer of the fabric network based on layer parameters of said each network layer;
an information module to determine structural representation information for each of the at least one network layer based on at least one identifier for each of the at least one network layer.
24. The apparatus according to claim 23, wherein the symbol recognition module is specifically configured to map each identifier of at least one identifier of each network layer to obtain a mapping result of each identifier;
an information module, configured to obtain structure representation information of each network layer based on a mapping result of each identifier in the at least one identifier.
25. The apparatus of claim 24, wherein the symbol recognition module is further configured to obtain a mapping result of each identifier in each network layer by looking up a first preset table based on each identifier in each network layer.
26. The apparatus according to claim 20, wherein the feature determining module is specifically configured to fuse the structure representation information of each of the at least one network layer to obtain the structural feature of the structural network.
27. The apparatus according to claim 26, wherein the feature determining module is specifically configured to fuse the structure representation information of each of the at least one network layer by using a neural network to obtain the structural features of the structural network.
28. The apparatus according to claim 27, wherein the feature determination module is specifically configured to input the structure representation information of each of the at least one network layer into a neural network, and the neural network fuses at least one of the structure representation information into one structure feature through a recursive algorithm.
29. The apparatus of any one of claims 20-28, further comprising:
the time acquisition unit is used for acquiring a time vector based on a preset time point corresponding to the structural network;
the performance determining unit is specifically configured to obtain a network performance parameter of the structural network at the preset time point based on the structural feature and the time vector.
30. The apparatus of claim 29, wherein the performance determination unit comprises:
the merging module is used for merging the structural features and the time vectors to obtain merged features;
and the time performance module is used for obtaining the corresponding network performance parameters of the structural network under the time vector based on the merging characteristics by utilizing the multilayer perceptron.
31. The apparatus according to claim 29, wherein the time obtaining unit is specifically configured to obtain the time vector by searching a second preset table based on a preset time point corresponding to the structural network.
32. The apparatus of any one of claims 20-28, further comprising:
a network training unit for training the prediction network using a plurality of sample structure networks, the sample structure networks being labeled with network performance parameters of the sample structure networks at each of at least one time point; the time point corresponds to a number of training times of the sample structure network.
33. The apparatus of claim 32, further comprising:
the network block unit is used for sampling a plurality of preset structure networks to generate a network block; the network block comprises at least one network layer;
a network construction unit for constructing the sample structure network based on the network blocks.
34. The apparatus of claim 33, wherein the network block unit comprises:
the layer sampling module is used for sampling a plurality of preset structure networks based on a Markov chain to obtain at least one preset network layer;
and the block forming module is used for sequentially connecting the at least one preset network layer to form the network block.
35. The apparatus of claim 33, wherein the network layer comprises any one or more of:
convolutional layers, max pooling layers, average pooling layers, activation layers, and batch layers.
36. The apparatus according to claim 34, wherein the network construction unit is specifically configured to sequentially connect a plurality of network blocks to obtain the sample structure network, wherein a first network block and a second network block of the plurality of network blocks correspond to different feature dimensions.
37. The apparatus of claim 36, wherein a max-pooling layer is connected between the first network block and the second network block.
38. The apparatus according to claim 34, wherein the layer sampling module is specifically configured to sample a plurality of preset-structure networks based on the sampled network parameters of an ith network layer to obtain an (i + 1) th network layer, where i is greater than or equal to 1 and less than the number of network layers included in the network block.
39. An electronic device comprising a processor including the apparatus for predicting network performance of any one of claims 20-38.
40. An electronic device, comprising: a memory for storing executable instructions;
and a processor in communication with the memory for executing the executable instructions to perform the operations of the method of predicting network performance of any one of claims 1 to 19.
41. A computer storage medium storing computer readable instructions that, when executed, perform the operations of the method of predicting network performance of any one of claims 1 to 19.
CN201711306429.3A 2017-12-08 2017-12-08 Method, apparatus, device, medium for predicting network performance Active CN108234195B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201711306429.3A CN108234195B (en) 2017-12-08 2017-12-08 Method, apparatus, device, medium for predicting network performance

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201711306429.3A CN108234195B (en) 2017-12-08 2017-12-08 Method, apparatus, device, medium for predicting network performance

Publications (2)

Publication Number Publication Date
CN108234195A CN108234195A (en) 2018-06-29
CN108234195B true CN108234195B (en) 2021-08-31

Family

ID=62653467

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201711306429.3A Active CN108234195B (en) 2017-12-08 2017-12-08 Method, apparatus, device, medium for predicting network performance

Country Status (1)

Country Link
CN (1) CN108234195B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112561174B (en) * 2020-12-18 2023-05-02 西南交通大学 LSTM and MLP-based superimposed neural network prediction geothermal energy production method
CN115396929A (en) * 2022-08-15 2022-11-25 中国联合网络通信集团有限公司 Performance data prediction method, device and storage medium
CN115860055B (en) * 2022-11-23 2024-01-02 北京百度网讯科技有限公司 Performance determination method, performance optimization method, device, electronic equipment and medium

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101675648A (en) * 2007-03-08 2010-03-17 Lm爱立信电话有限公司 Structure relevant and method with performance monitoring
CN102567786A (en) * 2011-12-13 2012-07-11 北京交通大学 Method for predicting derailment coefficients
CN104091045A (en) * 2014-06-16 2014-10-08 华南理工大学 Predicting method for long-term performance of air conditioner based on BP neural network

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9323639B2 (en) * 2009-09-10 2016-04-26 International Business Machines Corporation Dynamic recommendation framework for information technology management
US9338622B2 (en) * 2012-10-04 2016-05-10 Bernt Erik Bjontegard Contextually intelligent communication systems and processes

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101675648A (en) * 2007-03-08 2010-03-17 Lm爱立信电话有限公司 Structure relevant and method with performance monitoring
CN102567786A (en) * 2011-12-13 2012-07-11 北京交通大学 Method for predicting derailment coefficients
CN104091045A (en) * 2014-06-16 2014-10-08 华南理工大学 Predicting method for long-term performance of air conditioner based on BP neural network

Also Published As

Publication number Publication date
CN108234195A (en) 2018-06-29

Similar Documents

Publication Publication Date Title
US11295208B2 (en) Robust gradient weight compression schemes for deep learning applications
KR20200109230A (en) Method and apparatus for generating neural network
CN113553864B (en) Translation model training method and device, electronic equipment and storage medium
CN112035743B (en) Data recommendation method and device, computer equipment and storage medium
KR102420715B1 (en) System reinforcement learning method and apparatus, electronic device, computer storage medium
CN108234195B (en) Method, apparatus, device, medium for predicting network performance
CN114820871B (en) Font generation method, model training method, device, equipment and medium
CN111340221A (en) Method and device for sampling neural network structure
CN110659657A (en) Method and device for training model
CN113327599B (en) Voice recognition method, device, medium and electronic equipment
US9436912B1 (en) Symmetric schema instantiation method for use in a case-based reasoning system
CN113627536A (en) Model training method, video classification method, device, equipment and storage medium
WO2022188534A1 (en) Information pushing method and apparatus
CN114120172B (en) Video-based target detection method and device, electronic equipment and storage medium
WO2024051707A1 (en) Recommendation model training method and apparatus, and resource recommendation method and apparatus
CN111090740B (en) Knowledge graph generation method for dialogue system
CN113468857B (en) Training method and device for style conversion model, electronic equipment and storage medium
CN112784967B (en) Information processing method and device and electronic equipment
CN114638344A (en) Model quantization method and related device
CN111814368B (en) Tensor-based land utilization simulation method, system, equipment and storage medium
CN114792097A (en) Method and device for determining prompt vector of pre-training model and electronic equipment
CN113239215A (en) Multimedia resource classification method and device, electronic equipment and storage medium
CN115049730B (en) Component mounting method, component mounting device, electronic apparatus, and storage medium
CN117808083B (en) Distributed training communication method, device, system, equipment and storage medium
CN113836438B (en) Method, electronic device, and storage medium for post recommendation

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant