CN110533179A - Network structure searching method and device, readable storage medium storing program for executing, electronic equipment - Google Patents
Network structure searching method and device, readable storage medium storing program for executing, electronic equipment Download PDFInfo
- Publication number
- CN110533179A CN110533179A CN201910636695.5A CN201910636695A CN110533179A CN 110533179 A CN110533179 A CN 110533179A CN 201910636695 A CN201910636695 A CN 201910636695A CN 110533179 A CN110533179 A CN 110533179A
- Authority
- CN
- China
- Prior art keywords
- network
- parameter
- layer
- default
- obtains
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Data Exchanges In Wide-Area Networks (AREA)
Abstract
The embodiment of the present disclosure discloses a kind of network structure searching method and device, readable storage medium storing program for executing, electronic equipment, wherein method includes: to obtain the second network based on the sorter network of known network structure and network parameter;Second network is trained based on the corresponding training set of goal task, determines the connection weight in second network between every two network layer;The corresponding target network of the goal task is obtained based on the connection weight.The present embodiment belongs to the second network of supernet by establishing, realize the extension of network structure, and the second network obtained after structure extension is trained, third network is obtained from trained second network, since the second network is more suitable for default training mission by the network parameter that training obtains, therefore the third network that corresponding default training mission is obtained based on the second web search, is realized quickening search speed, improves the accuracy of search.
Description
Technical field
This disclosure relates to nerual network technique, especially a kind of network structure searching method and device, readable storage medium storing program for executing,
Electronic equipment.
Background technique
Existing network structure searching method mainly includes four kinds of methods: intensified learning, evolution algorithm.Intensified learning and into
Changing algorithm is first to generate a kind of structure, then the structure that training obtains obtains its performance indicator, goes to generate more according to current experience
Good structure.
Existing network structure searching method is all to start from scratch to construct search space, and search speed is slow.
Summary of the invention
In order to solve the slow technical problem of above-mentioned search speed, the disclosure is proposed.Embodiment of the disclosure provides one
Kind network structure searching method and device, readable storage medium storing program for executing, electronic equipment.
According to the one aspect of the embodiment of the present disclosure, a kind of network structure searching method is provided, comprising:
The second network is determined based on default network structure and the first network with default network parameter;
Second network is trained based on the corresponding training set sample of default training mission, determines second net
Connection weight in network between every two network layer;
Based on the connection weight, the corresponding third network of the default training mission is determined.
According to the another aspect of the embodiment of the present disclosure, a kind of network structure searcher is provided, comprising:
Second network determining module, for being determined based on default network structure with the first network with default network parameter
Second network;
Weight determination module, for determining mould to second network based on the corresponding training set sample of default training mission
The second network that block determines is trained, and determines the connection weight in second network between every two network layer;
Third network determining module, the connection weight for being determined based on the weight determination module are determined described default
The corresponding third network of training mission.
According to the another aspect of the embodiment of the present disclosure, a kind of computer readable storage medium, the storage medium are provided
It is stored with computer program, the computer program is for executing network structure searching method described in above-described embodiment.
According to the also one side of the embodiment of the present disclosure, a kind of electronic equipment is provided, the electronic equipment includes:
Processor;
For storing the memory of the processor-executable instruction;
The processor for reading the executable instruction from the memory, and executes described instruction to realize
Network structure searching method described in above-described embodiment.
Based on a kind of disclosure network structure searching method provided by the above embodiment and device, readable storage medium storing program for executing, electricity
Sub- equipment obtains the second network based on the sorter network of known network structure and network parameter;Based on the corresponding instruction of goal task
Practice collection to be trained second network, determines the connection weight in second network between every two network layer;It is based on
The connection weight obtains the corresponding target network of the goal task.The present embodiment belongs to the second of supernet by establishing
Network realizes the extension of network structure, and is trained to the second network obtained after structure extension, from trained
Third network is obtained in two networks, since the second network is more suitable for default training mission by the network parameter that training obtains, because
This obtains the third network of corresponding default training mission based on the second web search, realizes quickening search speed, improves and search
The accuracy of rope.
Below by drawings and examples, the technical solution of the disclosure is described in further detail.
Detailed description of the invention
The embodiment of the present disclosure is described in more detail in conjunction with the accompanying drawings, the above-mentioned and other purposes of the disclosure,
Feature and advantage will be apparent.Attached drawing is used to provide to further understand the embodiment of the present disclosure, and constitutes explanation
A part of book is used to explain the disclosure together with the embodiment of the present disclosure, does not constitute the limitation to the disclosure.In the accompanying drawings,
Identical reference label typically represents same parts or step.
Fig. 1 is a kind of flow diagram for the network structure searching method that the disclosure provides.
Fig. 2 a is a structural schematic diagram of first network in step 101 in the embodiment of Fig. 1 offer.
Fig. 2 b is a structural schematic diagram of supernet in step 102 in the embodiment of Fig. 1 offer.
Fig. 2 c is a structural schematic diagram of the sub-network obtained in step 103 in the embodiment of Fig. 1 offer.
Fig. 3 a shows a kind of structural schematic diagram of MBConv structure.
Fig. 3 b shows another structural schematic diagram of MBConv structure.
Fig. 4 demonstrates the structural schematic diagram that optimal minor structure is obtained based on gradient.
Fig. 5 is the flow diagram for the network structure searching method that one exemplary embodiment of the disclosure provides.
Fig. 6 is a flow diagram of step 501 in disclosure embodiment shown in fig. 5.
Fig. 7 is a flow diagram of step 5011 in disclosure embodiment shown in fig. 6.
Fig. 8 is a flow diagram of step 5012 in disclosure embodiment shown in fig. 6.
Fig. 9 is a flow diagram of step 503 in disclosure embodiment shown in fig. 5.
Figure 10 is the structural schematic diagram for the network structure searcher that one exemplary embodiment of the disclosure provides.
Figure 11 is the structural schematic diagram for the network structure searcher that disclosure another exemplary embodiment provides.
Figure 12 is the structure chart for the electronic equipment that one exemplary embodiment of the disclosure provides.
Specific embodiment
In the following, will be described in detail by referring to the drawings according to an example embodiment of the present disclosure.Obviously, described embodiment is only
It is only a part of this disclosure embodiment, rather than the whole embodiments of the disclosure, it should be appreciated that the disclosure is not by described herein
The limitation of example embodiment.
It should also be noted that unless specifically stated otherwise, the opposite cloth of the component and step that otherwise illustrate in these embodiments
It sets, numerical expression and the unlimited the scope of the present disclosure processed of numerical value.
It will be understood by those skilled in the art that the terms such as " first ", " second " in the embodiment of the present disclosure are only used for distinguishing
Different step, equipment or module etc., neither represent any particular technology meaning, also do not indicate that the inevitable logic between them is suitable
Sequence.
It should also be understood that in the embodiments of the present disclosure, " multiple " can refer to two or more, and "at least one" can refer to
One, two or more.
It should also be understood that for the either component, data or the structure that are referred in the embodiment of the present disclosure, clearly limit no or
Person may be generally understood to one or more in the case where context provides opposite enlightenment.
In addition, term "and/or" in the disclosure, only a kind of incidence relation for describing affiliated partner, expression can be deposited
In three kinds of relationships, for example, A and/or B, can indicate: individualism A exists simultaneously A and B, these three situations of individualism B.
In addition, character "/" in the disclosure, typicallys represent the relationship that forward-backward correlation object is a kind of "or".
It should also be understood that the disclosure highlights the difference between each embodiment to the description of each embodiment,
Same or similar place can be referred to mutually, for sake of simplicity, no longer repeating one by one.
Simultaneously, it should be appreciated that for ease of description, the size of various pieces shown in attached drawing is not according to reality
Proportionate relationship draw.
Be to the description only actually of at least one exemplary embodiment below it is illustrative, never as to the disclosure
And its application or any restrictions used.
Technology, method and apparatus known to person of ordinary skill in the relevant may be not discussed in detail, but suitable
In the case of, the technology, method and apparatus should be considered as part of specification.
It should also be noted that similar label and letter indicate similar terms in following attached drawing, therefore, once a certain Xiang Yi
It is defined in a attached drawing, then in subsequent attached drawing does not need that it is further discussed.
The embodiment of the present disclosure can be applied to the electronic equipments such as terminal device, computer system, server, can with it is numerous
Other general or specialized computing system environments or configuration operate together.Suitable for electric with terminal device, computer system, server etc.
The example of well-known terminal device, computing system, environment and/or configuration that sub- equipment is used together includes but is not limited to:
Personal computer system, thin client, thick client computer, hand-held or laptop devices, is based on microprocessor at server computer system
System, set-top box, programmable consumer electronics, NetPC Network PC, minicomputer system, large computer system and
Distributed cloud computing technology environment, etc. including above-mentioned any system.
The electronic equipments such as terminal device, computer system, server can be in the department of computer science executed by computer system
It is described under the general context of system executable instruction (such as program module).In general, program module may include routine, program, mesh
Beacon course sequence, component, logic, data structure etc., they execute specific task or realize specific abstract data type.Meter
Calculation machine systems/servers can be implemented in distributed cloud computing environment, and in distributed cloud computing environment, task is by by logical
What the remote processing devices of communication network link executed.In distributed cloud computing environment, it includes storage that program module, which can be located at,
On the Local or Remote computing system storage medium of equipment.
Application is summarized
In implementing the present disclosure, inventors have found that existing network structure searching method is all structure of starting from scratch
Build search space, but this method needs to expend many resources the prior art has at least the following problems: search speed is very slow.
Exemplary system
Fig. 1 is a kind of flow diagram for the network structure searching method that the disclosure provides.As shown in Figure 1, this method packet
It includes: step 101, obtaining a first network.The first network is the existing network of network structure known in the state of the art, usually
It include the pre-training parameter in classification or dense prediction task in these existing networks, wherein dense prediction task is index
Outpour the object type of each pixel in image, it is desirable that not only provide the position of objectives, also to describe the boundary of object
Task, usually predicted on the image of big resolution ratio, for example, the tasks such as image segmentation, object detection, edge detection.It is above-mentioned
Existing network in addition to include in the pre-training parameter of classification or dense prediction task other than, can also include such as recognition of face,
The pre-training parameter of the task in the non-field CV such as Attitude estimation.In an optional example, the first network which obtains can
With as shown in Figure 2 a, Fig. 2 a is a structural schematic diagram of first network in step 101 in embodiment that Fig. 1 is provided, in Fig. 2 a
Each of circle indicate a network layer, arrow indicates the connection between two network layers, and the multiple circles for including in Fig. 2 a indicate phases
Same or different types of network layer (e.g., convolutional layer, full articulamentum etc.).
Step 102, a supernet is generated according to first network.The first network obtained due to above-mentioned steps 101
Structure is all only to adapt to single task role (for example, classification task) mostly;Therefore, the present embodiment to the network structure of first network into
Row extension, generates the network structure of a supernet (supernet), and the pre-training parameter of first network is moved to super
In grade network, the sub-network suitable for other tasks, optionally, different tasks can be obtained based on the supernet to realize
There is different super father figures;Wherein extended operation can include but is not limited to: 3*3 (is such as changed to by the size for changing convolution kernel
5*5,7*7 etc.);The variation (such as becoming 3,6) of the coefficient of expansion;Be added different operations such as jump connection (jump connection refers to
It skips next network layer and is directly connected to next one network layer, jump connection is only in the immovable situation of I/O channel number
It is lower just to add);Increase new network layer (one in first network after at least one node in multiple stages of first network
A stage includes at least one indeclinable network layer of port number);Increase MBConv structure.
Wherein, it includes two kinds that MBConv structure, which has altogether, as shown in Figure 3a and Figure 3b shows.When the change for having port number (channels)
Using structure shown in Fig. 3 a when change, otherwise using structure shown in Fig. 3 b.
The input channel number of the structure in parallel layer shown in Fig. 3 a is C, and output channel number is C ', and C and C ' can be with
It is not identical.Structure in Fig. 3 a includes 3 parts.Leftmost ladder-shaped frame indicate the convolution operation that convolution kernel size is 1 × 1 and
ReLU6 activation primitive, input channel number are C, and output channel number is tC, and t therein is the coefficient of expansion.Intermediate rectangle frame
Indicate that step-length is 2, depth that convolution kernel size is k × k separates convolution and ReLU6 activation primitive, input channel number and
Output channel number is tC.The ladder-shaped frame of the rightmost side indicates that convolution kernel size is 1 × 1 convolution operation number, input channel number
For tC, output channel number is C.T and k × k therein depend on candidate operations.
It include three parts in Fig. 3 b.Leftmost ladder-shaped frame indicate the convolution operation that convolution kernel size is 1 × 1 and
ReLU6 activation primitive.Intermediate rectangle frame indicates that the depth that convolution kernel size is k × k separates convolution and ReLU6 activation
Function.The ladder-shaped frame of the rightmost side indicates that convolution kernel size is 1 × 1 convolution operation number.Have jump connection, will input with
Output is added.Input channel number and output channel number in Fig. 3 b are C.
It further include that assignment is carried out to the network parameter in supernet, the present embodiment is used the after structure extension
Parameter in one network moves in supernet, and specifically, the network structure part on network structure can correspond to directly will
Network parameter is loaded into corresponding network structure, the unmatched part of network structure can intercept parameter to matching (for example,
Convolution kernel in first network is 5*5, and the convolution kernel size of corresponding position is 3*3 in supernet, intercepts 5*5 centre bit at this time
The convolution kernel value for the 3*3 size set carries out assignment to the convolution kernel of supernet), if the network structure in supernet
Corresponding network structure is not present in first network, acquisition network parameter can not be also intercepted, at this point it is possible to select to the network
Structure carries out the parameter before random initializtion or multiplexing.
Structure extension is carried out to the existing network that Fig. 2 a is provided by step 102, the supernet of acquisition can be such as Fig. 2 b
Shown, Fig. 2 b is a structural schematic diagram of supernet in step 102 in the embodiment of Fig. 1 offer, Fig. 2 b similar with Fig. 2 a
Each of circle indicate a network layer, arrow indicates the connection between two network layers, and the multiple circles for including in Fig. 2 b indicate
The network layer (e.g., convolutional layer, full articulamentum etc.) of identical or different type.The structure level number and the first net that the supernet includes
The structure level number that network includes may be identical or different, and multiple networks are respectively included in every layer in the structure sheaf of supernet
Layer;Wherein, the multiple network layers in every layer of structure can be identical function but the different multiple network layers of parameter, be also possible to not
The multiple network layers of congenerous, for example, including 3 convolution kernels convolutional layer of different sizes in one layer of structure;The present embodiment is unlimited
The specific structure of supernet processed.
Step 103, supernet is trained, it is more suitable using the method search fine tuning based on gradient in supernet
The structure for closing dense prediction task, finally selects the sub-network for being more suitable for dense prediction task.Search process, that is, network instruction
Practice process, the sub-network for searching for acquisition can be as shown in Figure 2 c, and Fig. 2 c is the son obtained in step 103 in the embodiment of Fig. 1 offer
One structural schematic diagram of network.It include multiple sub-networks in supernet, minor structure shown in Fig. 2 c is based on connection weight
The sub-network that (structural parameters) are selected from supernet, each circle same table in Fig. 2 c similar with Fig. 2 a and Fig. 2 b
Show a network layer, arrow indicates the connection between two network layers, and the multiple circles for including in Fig. 2 c indicate identical or different class
The network layer (e.g., convolutional layer, full articulamentum etc.) of type.Because being loaded with parameter in supernet, search when
Wait speed faster, accuracy is higher.
Fig. 4 demonstrates the structural schematic diagram that optimal minor structure is obtained based on gradient.As shown in figure 4, father's network
Include all sub-networks inside (supernet e.g., provided in Fig. 1 embodiment), includes 0,1,2,3 four in father's network
Part, intermediate portion is similar with middle layer shown in Fig. 2 b, is respectively provided with multiple network layers, therefore, between every two part
Including a variety of connections, the corresponding connection weight of each connection, each sub-network has corresponding connection weight (in supernet
In initial connection weight can be that not have all connection weights between double-layer structure be 1, the connection weight of each connection is equal
Value, for example, including 4 connection weights between double-layer structure, then the corresponding initial connection weight of each connection is in 4 connections
0.25), connection weight can change according to gradient during training, have the maximum subnet of connection weight after training
Network, last test leave behind the maximum sub-network of connection weight, are exactly optimal sub-network.
Illustrative methods
Fig. 5 is the flow diagram for the network structure searching method that one exemplary embodiment of the disclosure provides.The present embodiment
It can be applicable on electronic equipment, as shown in figure 5, including the following steps:
Step 501, the first network based on default network structure and with default network parameter determines the second network.
Wherein, the structure of first network can be with reference to shown in Fig. 2 a, for showing for a known network structure and network parameter
There is network, the corresponding setting task of the existing network, the existing network can be for example sorter network etc., and setting task is for example
Correspond to classification task.The identical network layer in part may be present in second network and first network, optionally, first network can be with
It is a sub-network in the second network.
Step 502, the second network is trained based on the corresponding training set sample of default training mission, determines the second net
Connection weight in network between every two network layer.
In one embodiment, default training mission is the goal task of the disclosure, can be dense prediction task dispatching,
In, the second network is supernet, and the second network is to be obtained based on first network by structure extension and/or transformation, knot
Structure can refer to shown in Fig. 2 b.
Step 503, it is based on connection weight, determines the default corresponding third network of training mission.
Optionally, by the training to the second network, multiple companies in second network between every double-layer structure be can get
The structure of the connection weight connect, the third network based on connection weight acquisition can refer to shown in Fig. 2 c, and Fig. 2 c indicates the second network
The sub-network of middle weight maximum (sums of all connection weights in automatic network).
Disclosure network structure searching method provided by the above embodiment, point based on known network structure and network parameter
Class network obtains the second network;Second network is trained based on the corresponding training set of goal task, determines described
Connection weight in two networks between every two network layer;The corresponding target of the goal task is obtained based on the connection weight
Network.The present embodiment belongs to the second network of supernet by establishing, and realizes the extension of network structure, and to structure extension
The second network obtained afterwards is trained, and obtains third network from trained second network, since the second network passes through
The network parameter that training obtains is more suitable for default training mission, therefore obtains corresponding default training mission based on the second web search
Third network, realize quickening search speed, improve the accuracy of search.
As shown in fig. 6, step 501 may include following steps on the basis of above-mentioned embodiment illustrated in fig. 5:
Step 5011, acquisition expansion structure is extended by the default network structure to first network.
Optionally, the structure extension of first network can be realized by the step 102 in above-mentioned embodiment shown in FIG. 1,
Extending obtained expansion structure can refer to shown in Fig. 2 b, and Fig. 2 b increases the network layer of multiple and different types, make figure with respect to Fig. 2 a
It include multiple sub-networks in 2b, at this point, Fig. 2 a can be a sub-network in Fig. 2 b.
Step 5012, assignment is carried out based on parameter of the network parameter of first network to the network layer in expansion structure, obtained
Obtain the second network.
Optionally, the process for carrying out parameter assignment to expansion structure can refer to the step in above-mentioned embodiment shown in FIG. 1
The parameter transition process provided in 102.
The present embodiment is by carrying out structure extension and parameter assignment, being loaded with based on first network
Second network of initial network parameter includes multiple sub-networks in the structure of second network, at this point, based on second network into
Row search is more suitable for the second network, obtains optimal son since wherein initial network parameter is first network by training acquisition
The speed of network (i.e. third network), it is opposite to be obtained based on the supernet of arbitrary network structure and random initializtion network parameter
Faster, accuracy is higher for the speed of optimal sub-network.
As shown in fig. 7, step 5011 may include following steps on the basis of above-mentioned embodiment illustrated in fig. 6:
Step 701, at least two branches are increased by least two operations to each network layer in first network respectively
Network layer obtains the first extended network of first network.
Optionally, a variety of expansions that the operation in the present embodiment can be provided using step 102 in above-mentioned embodiment illustrated in fig. 1
Open up two or more in operation;By these extended operations, will can be all extended between every two-tier network structure including
Multiple connections, i.e., with the structure of supernet.
In one alternate embodiment, at least two operations include following at least two: changing the network ginseng in network layer
Number obtains at least one branched network network layers;The action type of change network layer obtains at least one branched network network layers;Skip network
Layer.
Step 702, it is newly-increased to increase at least one after at least one stage in multiple stages that first network includes
Network layer obtains the second extended network of first network.
Optionally, each stage refers to the constant multiple network layers of port number, increases at least one after at least one stage
Network layer, the network layer can be some or the multiple layers of duplication on last stage.
Step 703, it is based on the first extended network and the second extended network, obtains the expansion structure of first network.
The present embodiment obtains the first extension net by being expanded laterally at least Liang Ge branch by least one means at every layer
Network is realized to the Longitudinal Extension (number of plies for increasing network) of network structure by increasing network layer after each stage, makes to obtain
The structure of the second network have more diversity, including sub-network it is more, the third net obtained from these sub-networks
Network performance is more optimized.
As shown in figure 8, step 5012 may include following steps on the basis of above-mentioned embodiment illustrated in fig. 6:
Step 801, based on the network parameter in first network, in expansion structure, there are corresponding relationships with first network
The parameter of network layer carries out assignment.
Step 802, random assignment is carried out to the network parameter of branched network network layers and newly-increased network layer in expansion structure, obtained
Obtain the second network.
Optionally, the process for carrying out assignment to the parameter in the second network can refer to step 102 in embodiment shown in FIG. 1
In carry out the process of parameter migration to the network structure after extension, the present embodiment passes through parameter Direct Transfer, parameter interception, random
The means such as the parameter before initialization, multiplexing are realized to the parameter assignment of the second network, and the ginseng of the second network before training is made
Number meets the corresponding task of first network, since first network is trained, network parameter based on its corresponding task
It is more suitably applied to the task, the present embodiment is suitably applied the initial network parameter in the second network relatively by parameter assignment
The corresponding task of first network, the second network of opposite arbitrary initial network parameter, the second internetworking that the present embodiment obtains
More it can be more suitable for searching for third network.
In some alternative embodiments, step 502 is being based on the default corresponding training set sample of training mission to second
Network is trained, and while determining the connection weight in the second network between every two network layer, is also determined in the second network
Network parameter in each network layer.
In the present embodiment, during being trained to the second network, in addition to being adjusted based on gradient to connection weight,
It can also realize and network parameter is adjusted based on gradient, realize and be more suitable for using the method removal search fine tuning based on gradient
The structure and network parameter of dense prediction task.
As shown in figure 9, step 503 may include following steps on the basis of above-mentioned embodiment illustrated in fig. 5:
Step 5031, true based on the maximum connection weight in multiple connection weights between every double-layer structure in the second network
Determine a connection between double-layer structure, obtains the intermediate structure for only having a connection between every double-layer structure.
Optionally, the process for obtaining intermediate structure can refer to embodiment illustrated in fig. 4, from multiple companies between every double-layer structure
The maximum connection weight connect in weight is determined as connecting, and will obtain the maximum subnet of connection weight based on all maximum connection weights
Network, i.e., best sub-network.
Step 5032, network parameter assignment is carried out to intermediate structure by the network parameter of the second network or first network,
Determine third network.
In the present embodiment, by carrying out parameter assignment to the intermediate structure of acquisition, make the network in the third network obtained
Parameter is more suitable for default training mission.
In some alternative embodiments, step 5032 may include:
Based on there are the network parameters of the network layer of corresponding relationship to intermediate structure progress with intermediate structure in the second network
Assignment obtains third network;
Alternatively, in intermediate structure, there are the networks of corresponding relationship with first network based on the network parameter in first network
Layer carries out assignment, and in intermediate structure, there is no the network layers of corresponding relationship to carry out random assignment with first network, determines the
Three networks.
It, can be based on the second network (supernet) or first network (for example, classification after obtaining structure based on weight
Network) in network parameter carry out assignment, for wherein cannot matched network structure can carry out random assignment, wherein due to the
Include multiple sub-networks in two networks, and third network is one of sub-network, therefore, with the second network to third
When network carries out parameter assignment, directly network parameter corresponding in third network can be moved in third network;And when with the
When the network parameter of one network carries out assignment to third network, it is understood that there may be the unmatched situation of structure, at this point for mismatch
The parameter of network structure can random initializtion.
Before application third network processes preset training mission, can also using default training mission for training set
Sample is trained third network, so that the third network after training is more suitable for handling default training mission.
Any network structure searching method that the embodiment of the present disclosure provides can have data processing by any suitable
The equipment of ability executes, including but not limited to: terminal device and server etc..Alternatively, embodiment of the present disclosure offer is any
Network structure searching method can be executed by processor, as processor executes sheet by the command adapted thereto for calling memory to store
Any network structure searching method that open embodiment refers to.Hereafter repeat no more.
Exemplary means
Figure 10 is the structural schematic diagram for the network structure searcher that one exemplary embodiment of the disclosure provides.Such as Figure 10 institute
Show, the present embodiment includes:
Second network determining module 11, it is true for the first network based on default network structure and with default network parameter
Fixed second network.
Weight determination module 12, for being based on the default corresponding training set sample of training mission to the second network determining module
11 the second networks determined are trained, and determine the connection weight in the second network between every two network layer.
Third network determining module 13, the connection weight for being determined based on weight determination module 12 determine default training
The corresponding third network of task.
Disclosure network structure searcher provided by the above embodiment, point based on known network structure and network parameter
Class network obtains the second network;Second network is trained based on the corresponding training set of goal task, determines described
Connection weight in two networks between every two network layer;The corresponding target of the goal task is obtained based on the connection weight
Network.The present embodiment belongs to the second network of supernet by establishing, and realizes the extension of network structure, and to structure extension
The second network obtained afterwards is trained, and obtains third network from trained second network, since the second network passes through
The network parameter that training obtains is more suitable for default training mission, therefore obtains corresponding default training mission based on the second web search
Third network, realize quickening search speed, improve the accuracy of search.
Figure 11 is the structural schematic diagram for the network structure searcher that disclosure another exemplary embodiment provides.Such as Figure 11
Shown, in the present embodiment, the second network determining module 11 includes:
Structure extending unit 111, for being extended acquisition expansion structure by the default network structure to first network.
Optionally, structure extending unit 111, specifically for passing through at least two to each network layer in first network respectively
Kind operation increases by least two branched network network layers, obtains the first extended network of first network;First network include it is multiple
Increase at least one newly-increased network layer after at least one stage in stage, obtains the second extended network of first network;Base
In the first extended network and the second extended network, the expansion structure of first network is obtained.
Wherein, at least two operations include following at least two: the network parameter changed in network layer obtains at least one
Branched network network layers;The action type of change network layer obtains at least one branched network network layers;Skip network layer.
Network assignment unit 112, for the network parameter based on first network to the parameter of the network layer in expansion structure
Assignment is carried out, the second network is obtained.
Optionally, network assignment unit 112, specifically for based on the network parameter in first network in expansion structure with
There are the parameters of the network layer of corresponding relationship to carry out assignment for first network;To in expansion structure branched network network layers and newly-increased network
The network parameter of layer carries out random assignment, obtains the second network.
Optionally, weight determination module 12 are also used to determine the network parameter in the second network in each network layer.
In the present embodiment, third network determining module 13 includes:
Connect screening unit 131, for based in multiple connection weights between every double-layer structure in the second network most
Big connection weight determines a connection between double-layer structure, obtains the intermediate knot for only having a connection between every double-layer structure
Structure;
Parameter migration units 132 carry out net to intermediate structure for the network parameter by the second network or first network
Network parameter assignment determines third network.
Optionally, parameter migration units 132, specifically for based on there are corresponding relationships with intermediate structure in the second network
The network parameter of network layer carries out assignment to intermediate structure, obtains third network;
Alternatively, in intermediate structure, there are the networks of corresponding relationship with first network based on the network parameter in first network
Layer carries out assignment, and in intermediate structure, there is no the network layers of corresponding relationship to carry out random assignment with first network, determines the
Three networks.
Example electronic device
In the following, being described with reference to Figure 12 the electronic equipment according to the embodiment of the present disclosure.The electronic equipment can be first and set
Standby 100 and second any of equipment 200 or both or with their independent stand-alone devices, which can be with the
One equipment and the second equipment are communicated, to receive the collected input signal of institute from them.
Figure 12 illustrates the block diagram of the electronic equipment according to the embodiment of the present disclosure.
As shown in figure 12, electronic equipment 120 includes one or more processors 121 and memory 122.
Processor 121 can be central processing unit (CPU) or have data-handling capacity and/or instruction execution capability
Other forms processing unit, and can control the other assemblies in electronic equipment 120 to execute desired function.
Memory 122 may include one or more computer program products, and the computer program product may include
Various forms of computer readable storage mediums, such as volatile memory and/or nonvolatile memory.The volatibility is deposited
Reservoir for example may include random access memory (RAM) and/or cache memory (cache) etc..It is described non-volatile
Memory for example may include read-only memory (ROM), hard disk, flash memory etc..It can be on the computer readable storage medium
One or more computer program instructions are stored, processor 121 can run described program instruction, to realize sheet described above
The network structure searching method of disclosed each embodiment and/or other desired functions.It computer-readable is deposited described
The various contents such as input signal, signal component, noise component(s) can also be stored in storage media.
In one example, electronic equipment 120 can also include: input unit 123 and output device 124, these components
It is interconnected by bindiny mechanism's (not shown) of bus system and/or other forms.
For example, the input unit 123 can be above-mentioned when the electronic equipment is the first equipment 100 or the second equipment 200
Microphone or microphone array, for capturing the input signal of sound source.When the electronic equipment is stand-alone device, input dress
Setting 123 can be communication network connector, for receiving input signal collected from the first equipment 100 and the second equipment 200.
In addition, the input equipment 123 can also include such as keyboard, mouse etc..
The output device 124 can be output to the outside various information, including range information, the directional information etc. determined.
The output equipment 124 may include such as display, loudspeaker, printer and communication network and its be connected long-range defeated
Equipment etc. out.
Certainly, to put it more simply, illustrating only in the electronic equipment 120 one in component related with the disclosure in Figure 12
A bit, the component of such as bus, input/output interface etc. is omitted.In addition to this, according to concrete application situation, electronic equipment
120 can also include any other component appropriate.
Illustrative computer program product and computer readable storage medium
Other than the above method and equipment, embodiment of the disclosure can also be computer program product comprising meter
Calculation machine program instruction, it is above-mentioned that the computer program instructions make the processor execute this specification when being run by processor
According to the step in the network structure searching method of the various embodiments of the disclosure described in " illustrative methods " part.
The computer program product can be write with any combination of one or more programming languages for holding
The program code of row embodiment of the present disclosure operation, described program design language includes object oriented program language, such as
Java, C++ etc. further include conventional procedural programming language, such as " C " language or similar programming language.Journey
Sequence code can be executed fully on the user computing device, partly execute on a user device, be independent soft as one
Part packet executes, part executes on a remote computing or completely in remote computing device on the user computing device for part
Or it is executed on server.
In addition, embodiment of the disclosure can also be computer readable storage medium, it is stored thereon with computer program and refers to
It enables, the computer program instructions make the processor execute above-mentioned " the exemplary side of this specification when being run by processor
According to the step in the network structure searching method of the various embodiments of the disclosure described in method " part.
The computer readable storage medium can be using any combination of one or more readable mediums.Readable medium can
To be readable signal medium or readable storage medium storing program for executing.Readable storage medium storing program for executing for example can include but is not limited to electricity, magnetic, light, electricity
Magnetic, the system of infrared ray or semiconductor, device or device, or any above combination.Readable storage medium storing program for executing it is more specific
Example (non exhaustive list) includes: the electrical connection with one or more conducting wires, portable disc, hard disk, random access memory
Device (RAM), read-only memory (ROM), erasable programmable read only memory (EPROM or flash memory), optical fiber, portable compact disc
Read-only memory (CD-ROM), light storage device, magnetic memory device or above-mentioned any appropriate combination.
The basic principle of the disclosure is described in conjunction with specific embodiments above, however, it is desirable to, it is noted that in the disclosure
The advantages of referring to, advantage, effect etc. are only exemplary rather than limitation, must not believe that these advantages, advantage, effect etc. are the disclosure
Each embodiment is prerequisite.In addition, detail disclosed above is merely to exemplary effect and the work being easy to understand
With, rather than limit, it is that must be realized using above-mentioned concrete details that above-mentioned details, which is not intended to limit the disclosure,.
Each embodiment in this specification is described in a progressive manner, the highlights of each of the examples are with its
The difference of its embodiment, the same or similar part cross-reference between each embodiment.For system embodiment
For, since it is substantially corresponding with embodiment of the method, so being described relatively simple, referring to the portion of embodiment of the method in place of correlation
It defends oneself bright.
Device involved in the disclosure, device, equipment, system block diagram only as illustrative example and be not intended to
It is required that or hint must be attached in such a way that box illustrates, arrange, configure.As those skilled in the art will appreciate that
, it can be connected by any way, arrange, configure these devices, device, equipment, system.Such as "include", "comprise", " tool
" etc. word be open vocabulary, refer to " including but not limited to ", and can be used interchangeably with it.Vocabulary used herein above
"or" and "and" refer to vocabulary "and/or", and can be used interchangeably with it, unless it is not such that context, which is explicitly indicated,.Here made
Vocabulary " such as " refers to phrase " such as, but not limited to ", and can be used interchangeably with it.
Disclosed method and device may be achieved in many ways.For example, can by software, hardware, firmware or
Software, hardware, firmware any combination realize disclosed method and device.The said sequence of the step of for the method
Merely to be illustrated, the step of disclosed method, is not limited to sequence described in detail above, special unless otherwise
It does not mentionlet alone bright.In addition, in some embodiments, also the disclosure can be embodied as to record program in the recording medium, these programs
Including for realizing according to the machine readable instructions of disclosed method.Thus, the disclosure also covers storage for executing basis
The recording medium of the program of disclosed method.
It may also be noted that each component or each step are can to decompose in the device of the disclosure, device and method
And/or reconfigure.These decompose and/or reconfigure the equivalent scheme that should be regarded as the disclosure.
The above description of disclosed aspect is provided so that any person skilled in the art can make or use this
It is open.Various modifications in terms of these are readily apparent to those skilled in the art, and are defined herein
General Principle can be applied to other aspect without departing from the scope of the present disclosure.Therefore, the disclosure is not intended to be limited to
Aspect shown in this, but according to principle disclosed herein and the consistent widest range of novel feature.
In order to which purpose of illustration and description has been presented for above description.In addition, this description is not intended to the reality of the disclosure
It applies example and is restricted to form disclosed herein.Although already discussed above multiple exemplary aspects and embodiment, this field skill
Its certain modifications, modification, change, addition and sub-portfolio will be recognized in art personnel.
Claims (11)
1. a kind of network structure searching method, comprising:
The second network is determined based on default network structure and the first network with default network parameter;
Second network is trained based on the corresponding training set sample of default training mission, is determined in second network
Connection weight between every two network layer;
Based on the connection weight, the corresponding third network of the default training mission is determined.
2. according to the method described in claim 1, wherein, based on default network structure and with default network parameter
One network determines the second network, comprising:
Acquisition expansion structure is extended by the default network structure to the first network;
Network parameter based on the first network carries out assignment to the parameter of the network layer in the expansion structure, described in acquisition
Second network.
3. according to the method described in claim 2, wherein, the default network structure by the first network expands
Exhibition obtains expansion structure, comprising:
At least two branched network network layers are increased by least two operations to each network layer in the first network respectively, are obtained
To the first extended network of the first network;
Increase at least one newly-increased network layer after at least one stage in multiple stages that the first network includes, obtains
To the second extended network of the first network;
Based on first extended network and second extended network, the expansion structure of first network is obtained.
4. according to the method described in claim 3, wherein, at least two operation includes following at least two: described in change
Network parameter in network layer obtains at least one branched network network layers;The action type for changing the network layer obtains at least one
Branched network network layers;Skip the network layer.
5. according to the method described in claim 4, wherein, the network parameter based on the first network ties the extension
The parameter of network layer in structure carries out assignment, obtains second network, comprising:
Based on the network parameter in the first network, in the expansion structure, there are corresponding relationships with the first network
The parameter of network layer carries out assignment;
Random assignment is carried out to the network parameters of branched network network layers and newly-increased network layer in the expansion structure, obtains described the
Two networks.
6. -5 any method according to claim 1 is being based on the connection weight, is determining the default training mission pair
Before the third network answered, further includes:
Second network is trained based on the corresponding training set sample of default training mission, is determined in second network
Network parameter in each network layer.
7. it is described to be based on the connection weight according to the method described in claim 6, wherein, determine the default training mission
Corresponding third network, comprising:
Described two are determined based on the maximum connection weight in multiple connection weights between every double-layer structure in second network
A connection between layer structure, obtains the intermediate structure for only having a connection between every double-layer structure;
Network parameter assignment is carried out to the intermediate structure by the network parameter of second network or the first network, really
The fixed third network.
8. the network of second network or the first network that passes through is joined according to the method described in claim 7, wherein
It is several that network parameter assignment is carried out to the intermediate structure, determine the third network, comprising:
Based on there are the network parameters of the network layer of corresponding relationship to the centre with the intermediate structure in second network
Structure carries out assignment, obtains third network;
Alternatively, being closed with the first network there are corresponding based on the network parameter in the first network in the intermediate structure
The network layer of system carries out assignment, and there is no the network layers of corresponding relationship to carry out with the first network in the intermediate structure
Random assignment determines the third network.
9. a kind of network structure searcher, comprising:
Second network determining module, for based on default network structure and the first network with default network parameter determines second
Network;
Weight determination module, for true to the second network determining module based on the default corresponding training set sample of training mission
The second fixed network is trained, and determines the connection weight in second network between every two network layer;
Third network determining module, the connection weight for being determined based on the weight determination module, determines the default training
The corresponding third network of task.
10. a kind of computer readable storage medium, the storage medium is stored with computer program, and the computer program is used for
Execute any network structure searching method of the claims 1-8.
11. a kind of electronic equipment, the electronic equipment include:
Processor;
For storing the memory of the processor-executable instruction;
The processor, for reading the executable instruction from the memory, and it is above-mentioned to realize to execute described instruction
Any network structure searching method of claim 1-8.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910636695.5A CN110533179A (en) | 2019-07-15 | 2019-07-15 | Network structure searching method and device, readable storage medium storing program for executing, electronic equipment |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910636695.5A CN110533179A (en) | 2019-07-15 | 2019-07-15 | Network structure searching method and device, readable storage medium storing program for executing, electronic equipment |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110533179A true CN110533179A (en) | 2019-12-03 |
Family
ID=68660287
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910636695.5A Pending CN110533179A (en) | 2019-07-15 | 2019-07-15 | Network structure searching method and device, readable storage medium storing program for executing, electronic equipment |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110533179A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112101553A (en) * | 2020-11-10 | 2020-12-18 | 鹏城实验室 | Network structure searching method and device, equipment and storage medium |
CN112259122A (en) * | 2020-10-20 | 2021-01-22 | 北京小米松果电子有限公司 | Audio type identification method and device and storage medium |
CN113408692A (en) * | 2020-03-16 | 2021-09-17 | 顺丰科技有限公司 | Network structure searching method, device, equipment and storage medium |
EP4080416A4 (en) * | 2020-01-15 | 2023-02-01 | Huawei Technologies Co., Ltd. | Adaptive search method and apparatus for neural network |
WO2023024577A1 (en) * | 2021-08-27 | 2023-03-02 | 之江实验室 | Edge computing-oriented reparameterization neural network architecture search method |
-
2019
- 2019-07-15 CN CN201910636695.5A patent/CN110533179A/en active Pending
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP4080416A4 (en) * | 2020-01-15 | 2023-02-01 | Huawei Technologies Co., Ltd. | Adaptive search method and apparatus for neural network |
CN113408692A (en) * | 2020-03-16 | 2021-09-17 | 顺丰科技有限公司 | Network structure searching method, device, equipment and storage medium |
CN112259122A (en) * | 2020-10-20 | 2021-01-22 | 北京小米松果电子有限公司 | Audio type identification method and device and storage medium |
CN112101553A (en) * | 2020-11-10 | 2020-12-18 | 鹏城实验室 | Network structure searching method and device, equipment and storage medium |
CN112101553B (en) * | 2020-11-10 | 2021-02-23 | 鹏城实验室 | Network structure searching method and device, equipment and storage medium |
WO2023024577A1 (en) * | 2021-08-27 | 2023-03-02 | 之江实验室 | Edge computing-oriented reparameterization neural network architecture search method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110533179A (en) | Network structure searching method and device, readable storage medium storing program for executing, electronic equipment | |
EP3369045B1 (en) | Determining orders of execution of a neural network | |
Wang et al. | Efficient video transformers with spatial-temporal token selection | |
CN111582453B (en) | Method and device for generating neural network model | |
US10796244B2 (en) | Method and apparatus for labeling training samples | |
US20200005122A1 (en) | Multiscale feature representations for object recognition and detection | |
Marble et al. | Asymptotically near-optimal is good enough for motion planning | |
CN111582454B (en) | Method and device for generating neural network model | |
CN109685204B (en) | Image processing method and device, storage medium and electronic equipment | |
CN112990203B (en) | Target detection method and device, electronic equipment and storage medium | |
CN108776833B (en) | Data processing method, system and computer readable storage medium | |
JP2023527700A (en) | Dynamic automation of pipeline artifact selection | |
CN111368973B (en) | Method and apparatus for training a super network | |
US20210279279A1 (en) | Automated graph embedding recommendations based on extracted graph features | |
JP2021192286A (en) | Model training, image processing method and device, storage medium, and program product | |
EP4040378A1 (en) | Burst image-based image restoration method and apparatus | |
CN112001485B (en) | Group convolution number searching method and device | |
CN110533180A (en) | Network structure searching method and device, readable storage medium storing program for executing, electronic equipment | |
CN110309836A (en) | Image characteristic extracting method, device, storage medium and equipment | |
CN114282581A (en) | Training sample obtaining method and device based on data enhancement and electronic equipment | |
CN111274981A (en) | Target detection network construction method and device and target detection method | |
CN114419363A (en) | Target classification model training method and device based on label-free sample data | |
CN111582477A (en) | Training method and device of neural network model | |
CN109871891A (en) | A kind of object identification method, device and storage medium | |
CN111652354B (en) | Method, apparatus, device and storage medium for training super network |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191203 |
|
RJ01 | Rejection of invention patent application after publication |