CN112819138A - Optimization method and device of image neural network structure - Google Patents
Optimization method and device of image neural network structure Download PDFInfo
- Publication number
- CN112819138A CN112819138A CN202110102292.XA CN202110102292A CN112819138A CN 112819138 A CN112819138 A CN 112819138A CN 202110102292 A CN202110102292 A CN 202110102292A CN 112819138 A CN112819138 A CN 112819138A
- Authority
- CN
- China
- Prior art keywords
- neural network
- network structure
- image
- template
- target
- 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
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 751
- 238000000034 method Methods 0.000 title claims abstract description 46
- 238000005457 optimization Methods 0.000 title claims abstract description 44
- 230000015654 memory Effects 0.000 claims description 26
- 238000003860 storage Methods 0.000 claims description 16
- 238000012545 processing Methods 0.000 claims description 14
- 238000012216 screening Methods 0.000 claims description 13
- 238000013473 artificial intelligence Methods 0.000 abstract description 3
- 238000010586 diagram Methods 0.000 description 37
- 238000013135 deep learning Methods 0.000 description 8
- 238000013461 design Methods 0.000 description 8
- 230000006870 function Effects 0.000 description 8
- 238000004590 computer program Methods 0.000 description 7
- 230000008569 process Effects 0.000 description 7
- 238000012986 modification Methods 0.000 description 5
- 230000004048 modification Effects 0.000 description 5
- 238000001514 detection method Methods 0.000 description 4
- 230000003287 optical effect Effects 0.000 description 4
- 230000004913 activation Effects 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 239000013307 optical fiber Substances 0.000 description 2
- 230000000644 propagated effect Effects 0.000 description 2
- 230000002787 reinforcement Effects 0.000 description 2
- 230000004075 alteration Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 238000010845 search algorithm Methods 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
Images
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
- G06N3/04—Architecture, e.g. interconnection topology
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)
- Image Analysis (AREA)
Abstract
The application discloses an optimization method and device of an image neural network structure, which belong to the technical field of artificial intelligence, and the method comprises the following steps: generating a first image neural network structure according to the structure description information contained in the received image neural network structure optimization request, comparing the first image neural network structure with each neural network structure template, to determine a target neural network structure template for use by the first image neural network structure and a target sub-neural network structure in the first image neural network structure that corresponds to the target neural network structure template, then, using a plurality of preset image neural network structures corresponding to each target neural network structure template to respectively replace target sub-neural network structures corresponding to the target neural network structure template in the first image neural network structure to obtain a candidate neural network structure set, and searching the neural network structure in the candidate neural network structure set to obtain the optimized second image neural network structure.
Description
Technical Field
The application relates to the technical field of artificial intelligence, in particular to an optimization method and device of an image neural network structure.
Background
In the technical field of artificial intelligence, a deep learning technology is an important branch, and in the development process of the deep learning technology, the structural design of an image neural network plays a crucial role all the time.
In recent years, the image neural network structure designed by field experts has achieved great success on various tasks. However, with the increasing diversification of actual business requirements, the existing image Neural network structure is more and more difficult to meet the actual business requirements, so that a Neural network structure search (NAS) is presented, and the purpose of the Neural network structure search is to automatically design the image Neural network structure, so as to save the labor cost and the time cost. However, the existing neural network structure search schemes all require a user to have some knowledge about the neural network structure search, and require the user to manually design a candidate neural network structure set (i.e., a search space), so that the degree of automation is low, and the use threshold is high.
Disclosure of Invention
The embodiment of the application provides an optimization method and device of an image neural network structure, which are used for solving the problems that in the prior art, the automation degree of a neural network structure search scheme is low, the use threshold is high, and the rapid acquisition of the image neural network structure is not facilitated.
In a first aspect, an embodiment of the present application provides an optimization method for an image neural network structure, including:
receiving an image neural network structure optimization request, wherein the image neural network structure optimization request comprises structure description information of a first image neural network structure;
generating a first image neural network structure according to the structure description information;
comparing the first image neural network structure with each neural network structure template to determine template use information of the first image neural network structure, wherein the template use information comprises a target neural network structure template used by the first image neural network structure and a target sub-neural network structure corresponding to the target neural network structure template in the first image neural network structure;
acquiring a plurality of preset image neural network structures corresponding to each target neural network structure template;
replacing target sub-neural network structures corresponding to the target neural network structure template in the first image neural network structure respectively by using multiple preset image neural network structures corresponding to each target neural network structure template to obtain a candidate neural network structure set;
and searching a neural network structure in the candidate neural network structure set, and determining the searched neural network structure as a second image neural network structure after the first image neural network structure is optimized.
In one possible embodiment, comparing the first image neural network structure with each neural network structure template to determine template usage information of the first image neural network structure includes:
screening a sub-neural network structure which is completely or partially identical with the neural network structure of each neural network structure template from the first image neural network structure;
and if the sub-neural network structure is screened out, determining the neural network structure template as a target neural network structure template used by the first image neural network structure, and determining the screened sub-neural network structure as a target sub-neural network structure corresponding to the target neural network structure template in the first image neural network structure.
In one possible embodiment, the screening of the neural network structure from the first image neural network structure for the same sub-neural network structure as the neural network structure part of each neural network structure template comprises:
and screening a sub neural network structure which is the same as the designated neural network structure of each neural network structure template from the first image neural network structure, wherein the designated neural network structure is used for representing the structural characteristics of the neural network structure template.
In a possible implementation manner, using multiple preset image neural network structures corresponding to each target neural network structure template to respectively replace target sub-neural network structures corresponding to the target neural network structure template in the first image neural network structure, so as to obtain a candidate neural network structure set, including:
taking the first image neural network structure as a reference neural network structure, and replacing the ith target sub-neural network structure in the reference neural network structure by using each preset image neural network structure corresponding to the corresponding target neural network structure template to obtain an intermediate neural network structure set, wherein i is an integer starting from 1;
if i is determined to be smaller than the total number of the target sub-neural network structures, updating i to i +1, taking each neural network structure in the intermediate neural network structure set as a new reference neural network structure, executing the step of replacing the ith target sub-neural network structure in the reference neural network structure by using each preset image neural network structure corresponding to the corresponding target neural network structure template to obtain an intermediate neural network structure set;
if it is determined that i is equal to the total number of target sub-neural network structures, determining the set of intermediate neural network structures as a set of candidate neural network structures.
In a possible implementation, before generating the first image neural network structure according to the structure description information, the method further includes:
and if the structural description information is determined not to be represented by the specified data format, converting the structural description information into the structural description information represented by the specified data format.
In a second aspect, an embodiment of the present application provides an apparatus for optimizing an image neural network structure, including:
the image neural network structure optimization system comprises a receiving module, a processing module and a processing module, wherein the receiving module is used for receiving an image neural network structure optimization request which contains structure description information of a first image neural network structure;
the generating module is used for generating a first image neural network structure according to the structure description information;
a determining module, configured to compare the first image neural network structure with each neural network structure template to determine template usage information of the first image neural network structure, where the template usage information includes a target neural network structure template used by the first image neural network structure and a target sub-neural network structure in the first image neural network structure corresponding to the target neural network structure template;
the acquisition module is used for acquiring a plurality of preset image neural network structures corresponding to each target neural network structure template;
the replacing module is used for replacing target sub-neural network structures corresponding to the target neural network structure template in the first image neural network structure by using a plurality of preset image neural network structures corresponding to each target neural network structure template to obtain a candidate neural network structure set;
and the searching module is used for searching the neural network structure in the candidate neural network structure set and determining the searched neural network structure as the second image neural network structure after the first image neural network structure is optimized.
In a possible implementation, the determining module is specifically configured to:
screening a sub-neural network structure which is completely or partially identical with the neural network structure of each neural network structure template from the first image neural network structure;
and if the sub-neural network structure is screened out, determining the neural network structure template as a target neural network structure template used by the first image neural network structure, and determining the screened sub-neural network structure as a target sub-neural network structure corresponding to the target neural network structure template in the first image neural network structure.
In a possible implementation, the determining module is specifically configured to:
and screening a sub neural network structure which is the same as the designated neural network structure of each neural network structure template from the first image neural network structure, wherein the designated neural network structure is used for representing the structural characteristics of the neural network structure template.
In a possible implementation, the replacement module is specifically configured to:
taking the first image neural network structure as a reference neural network structure, and replacing the ith first image neural network structure in the reference neural network structure by using each preset image neural network structure corresponding to the ith first image neural network structure to obtain an intermediate neural network structure set, wherein i is an integer starting from 1;
if i is determined to be smaller than the total number of the target sub-neural network structures, updating i to i +1, taking each neural network structure in the intermediate neural network structure set as a reference neural network structure, and executing a step of replacing the ith first image neural network structure in the reference neural network structure by using each preset image neural network structure corresponding to the ith first image neural network structure to obtain an intermediate neural network structure set;
if it is determined that i is equal to the total number of target sub-neural network structures, determining the set of intermediate neural network structures as a set of candidate neural network structures.
In a possible implementation, the system further includes a conversion module configured to:
before generating a first image neural network structure according to the structure description information, if the structure description information is determined not to be represented by a specified data format, converting the structure description information into the structure description information represented by the specified data format.
In a third aspect, an embodiment of the present application provides an electronic device, including: at least one processor, and a memory communicatively coupled to the at least one processor, wherein:
the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the above-described method of optimization of an image neural network structure.
In a fourth aspect, the present application provides a storage medium, and when instructions in the storage medium are executed by a processor of an electronic device, the electronic device is capable of executing the optimization method of the image neural network structure.
In the embodiment of the application, an image neural network structure optimization request is received, a first image neural network structure is generated according to structure description information of the first image neural network structure contained in the image neural network structure optimization request, the first image neural network structure is compared with each neural network structure template to determine a target neural network structure template used by the first image neural network structure and a target sub-neural network structure corresponding to the target neural network structure template in the first image neural network structure, a plurality of preset image neural network structures corresponding to each target neural network structure template are obtained, the plurality of preset image neural network structures corresponding to each target neural network structure template are used to respectively replace the target sub-neural network structure corresponding to the target neural network structure template in the first image neural network structure, and obtaining a candidate neural network structure set, further searching a neural network structure in the candidate neural network structure set, and determining the searched neural network structure as a second image neural network structure after the first image neural network structure is optimized. Therefore, the candidate neural network structure set is automatically generated, a user does not need to manually design the candidate neural network structure set and know more neural network structure searching knowledge, the first image neural network structure can be quickly optimized to obtain the second image neural network structure, the automation degree of neural network structure searching is improved, and the use convenience is improved.
Drawings
The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the application and together with the description serve to explain the application and not to limit the application. In the drawings:
fig. 1 is a flowchart of an optimization method for an image neural network structure according to an embodiment of the present disclosure;
FIG. 2 is a flowchart of a method for generating a set of candidate neural network structures according to an embodiment of the present application;
fig. 3 is a flowchart of a further optimization method for an image neural network structure according to an embodiment of the present disclosure;
fig. 4 is a schematic diagram illustrating a manner of representing structure description information of a first image neural network structure according to an embodiment of the present application;
fig. 5 is a schematic diagram of a first image neural network structure according to an embodiment of the present disclosure;
fig. 6 is a schematic diagram of a neural network structure template BottleNeck provided in an embodiment of the present application;
fig. 7 is a schematic diagram of a neural network structure template, BasicBlock, provided by an embodiment of the present application;
fig. 8 is a schematic diagram of a neural network structure template DoubleConv provided in an embodiment of the present application;
fig. 9 is a schematic diagram of a neural network structure template SingleConv provided in an embodiment of the present application;
fig. 10 is a schematic diagram of a preset image neural network structure corresponding to BottleNeck according to an embodiment of the present application;
fig. 11 is a schematic diagram of a preset image neural network structure corresponding to BasicBlock according to an embodiment of the present application;
fig. 12 is a schematic diagram of a neural network structure of a preset image corresponding to a DoubleConv according to an embodiment of the present application;
fig. 13 is a schematic diagram of a preset image neural network structure corresponding to SingleConv according to an embodiment of the present disclosure;
FIG. 14 is a diagram illustrating template usage information of a first image neural network structure according to an embodiment of the present disclosure;
FIG. 15 shows a schematic diagram of a first image neural network structure after replacing a target sub-neural network structure corresponding to SingleConv;
fig. 16 shows a schematic diagram of a further alternative of the target sub-neural network structure corresponding to SingleConv in the first image neural network structure;
FIG. 17 is a schematic diagram showing a first image neural network structure after replacement of all target sub-neural network structures;
FIG. 18 is a schematic diagram showing yet another alternative for replacing all target sub-neural network structures in the first image neural network structure;
FIG. 19 is a schematic diagram showing yet another alternative for replacing all target sub-neural network structures in the first image neural network structure;
FIG. 20 is a schematic diagram showing yet another alternative for replacing all target sub-neural network structures in the first image neural network structure;
fig. 21 is a schematic structural diagram of an optimization apparatus for an image neural network structure according to an embodiment of the present disclosure;
fig. 22 is a hardware structural diagram of an electronic device for implementing an optimization method of an image neural network structure according to an embodiment of the present application.
Detailed Description
In order to solve the problems that in the prior art, a neural network structure search scheme has a low degree of automation and a high use threshold, and is not beneficial to quickly obtaining an image neural network structure, the embodiment of the application provides an optimization method and device of the image neural network structure.
The preferred embodiments of the present application will be described below with reference to the accompanying drawings of the specification, it should be understood that the preferred embodiments described herein are merely for illustrating and explaining the present application, and are not intended to limit the present application, and that the embodiments and features of the embodiments in the present application may be combined with each other without conflict.
The Neural network structure search (NAS) means that a recurrent Neural network is used as a controller to generate a sub-Neural network under the condition that the length and the structure of the Neural network are uncertain, the sub-Neural network is trained and evaluated to obtain the network performance (such as accuracy), and the network performance of the sub-Neural network is used as a feedback signal of the controller to update the controller, so that the controller generates a Neural network with higher accuracy next time. In this way, the controller will improve the search effect of the neural network by continuously learning.
Fig. 1 is a flowchart of an optimization method of an image neural network structure according to an embodiment of the present application, including the following steps:
s101: and receiving an image neural network structure optimization request, wherein the image neural network structure optimization request comprises structure description information of a first image neural network structure.
The structure description information of the first image neural network structure may be represented in a (JavaScript Object notification, JSON) data format.
S102: and generating a first image neural network structure according to the structure description information.
S103: and comparing the first image neural network structure with each neural network structure template to determine template use information of the first image neural network structure, wherein the template use information comprises a target neural network structure template used by the first image neural network structure and a target sub-neural network structure corresponding to the target neural network structure template in the first image neural network structure.
Wherein each neural network structure template can be predetermined by a skilled person according to experience or experimental data.
In specific implementation, a sub-neural network structure which is completely or partially identical to the neural network structure of each neural network structure template can be screened from the first image neural network structure, if the sub-neural network structure is screened, the first image neural network structure uses the neural network structure template, the neural network structure template is determined as a target neural network structure template, and the screened sub-neural network structure is determined as a target sub-neural network structure corresponding to the target neural network structure template in the first image neural network structure; and if the sub-neural network structure is not screened out, the first image neural network structure does not use the neural network structure template.
In the above process, when a sub-neural network structure identical to the neural network structure part of each neural network structure template is selected from the first image neural network structure, a sub-neural network structure identical to a designated neural network structure of the neural network structure template may be selected from the first image neural network structure, wherein the designated neural network structure is a part of the neural network structure template and is used for characterizing the structural features of the neural network structure template. That is, the specified neural network structure is an important neural network structure in the neural network structure template, and the first image neural network structure is considered to use the neural network structure template as long as there is a sub-neural network structure in the first image neural network structure that is the same as the important neural network structure in the neural network structure template.
In addition, the first image neural network structure may use one, two, three, or more neural network structure templates, and the first image neural network structure may be used one, two, three, or more times for the same neural network structure template. That is, there may be more than one target neural network structure template, and there may also be more than one target sub-neural network structure in the first image neural network structure corresponding to one target neural network structure template.
S104: and acquiring multiple preset image neural network structures corresponding to each target neural network structure template.
In specific implementation, each target neural network structure template can correspond to multiple preset image neural network structures, the multiple preset image neural network structures can be obtained by deforming the target neural network structure templates, and the multiple preset image neural network structures have the same structural characteristics.
In addition, the plurality of preset image neural network structures can also comprise a neural network structure which is completely the same as the neural network structure of the target neural network structure template, so that a candidate neural network structure set obtained subsequently can be enriched as much as possible.
S105: and replacing the target sub-neural network structures corresponding to the target neural network structure template in the first image neural network structure respectively by using multiple preset image neural network structures corresponding to each target neural network structure template to obtain a candidate neural network structure set.
In particular implementation, the candidate neural network structure set may be generated according to the process shown in fig. 2, where the process includes the following steps:
s201 a: the first image neural network structure is taken as a reference neural network structure.
It should be noted that, for the first image neural network structure, no matter how the target sub-neural network structure in the first image neural network structure is subjected to the replacement processing, the total number of the target sub-neural network structures in the first image neural network structure is constant.
S202 a: and respectively replacing the ith target sub-neural network structure in the reference neural network structure by using each preset image neural network structure corresponding to the corresponding target neural network structure template to obtain an intermediate neural network structure set, wherein i is an integer starting from 1.
The corresponding target neural network structure template refers to a target neural network structure template corresponding to the ith target sub-neural network structure in the reference neural network structure.
S203 a: judging whether i is smaller than the total number of the target sub-neural network structures, if so, entering S204 a; if not, the process proceeds to S205 a.
Since the first image neural network structure may be used more than once for the same target neural network structure template, the total number of target neural network structure templates used by the first image neural network structure is not necessarily equal to the total number of target sub-neural network structures in the first image neural network structure.
S204 a: updating i to i +1, and taking each neural network structure in the intermediate neural network structure set as a reference neural network structure, and returning to S202 a.
S205 a: determining the set of intermediate neural network structures as a set of candidate neural network structures.
S106: and searching the neural network structure in the candidate neural network structure set, and determining the searched neural network structure as a second image neural network structure after the first image neural network structure is optimized.
In specific implementation, algorithms such as random search, bayesian optimization, reinforcement learning and the like can be adopted to search the neural network structures in the candidate neural network structure set, and then the searched neural network structure is determined as the second image neural network structure after the first image neural network structure is optimized. Subsequently, a second image neural network structure can be used to perform image processing tasks (e.g., face recognition, object detection, etc.).
In practical applications, the deep learning frameworks used by different users when constructing the first image neural network structure may be different, for example, some users construct the first image neural network structure using TensorFlow, some users construct the first image neural network structure using Pytorch, and some users construct the first image neural network structure using Mxnet. And the structural description information obtained after the first image neural network structure is built by using different depth learning frameworks has different representation modes.
In specific implementation, a deep learning framework used by a user in constructing the first image neural network structure may not be limited, that is, the structure description information of the first image neural network structure carried in the neural network structure optimization request may be generated by the user using any deep learning framework, and then the structure description information carried in the neural network structure optimization request is converted into a uniform representation mode, and a subsequent operation is performed based on the structure description information of the uniform representation mode, so as to reduce a technical threshold of neural network structure search as much as possible.
Fig. 3 is a flowchart of a further method for optimizing an image neural network structure according to an embodiment of the present application, including the following steps:
s301: and receiving an image neural network structure optimization request, wherein the image neural network structure optimization request comprises structure description information of a first image neural network structure.
In a specific implementation, the structural description information of the first image neural network structure may be defined by a deep learning framework such as tensorblow, pytorch, mxnet, or may be defined by a (JavaScript Object notification, JSON) data format.
S302: and if the structural description information is determined not to be represented by the specified data format, converting the structural description information into the structural description information represented by the specified data format.
Wherein the specified data format may be a JSON data format.
S303: and generating a first image neural network structure according to the structure description information expressed by adopting a specified data format.
S304: and comparing the first image neural network structure with each neural network structure template to determine template use information of the first image neural network structure, wherein the template use information comprises a target neural network structure template used by the first image neural network structure and a target sub-neural network structure corresponding to the target neural network structure template in the first image neural network structure.
S305: and acquiring multiple preset image neural network structures corresponding to each target neural network structure template.
S306: and replacing the target sub-neural network structures corresponding to the target neural network structure template in the first image neural network structure respectively by using multiple preset image neural network structures corresponding to each target neural network structure template to obtain a candidate neural network structure set.
The implementation of this step can be referred to the implementation of S105, and is not described herein again.
S307: and searching the neural network structure in the candidate neural network structure set, and determining the searched neural network structure as a second image neural network structure after the first image neural network structure is optimized.
Taking object detection as an example, in an object detection task, the design of the image neural network structure has a great influence on the final object detection performance. The most direct design method in the prior art of the image neural network structure is to replace the most advanced image neural network structure (manually designed by field experts), but the existing image neural network structure cannot realize the optimal balance of precision and speed. Therefore, another method is to use the neural network structure search to obtain an image neural network structure with optimal precision and speed, but the existing neural network structure search scheme requires users to know more neural network structure search knowledge, manually designs a search space, and is time-consuming and labor-consuming.
In order to solve the above problem, embodiments of the present application provide a scheme that can automatically generate a search space. The method comprises the steps of firstly obtaining structure description information of a first image neural network structure (manually designed by a user) to be optimized, generating the first image neural network structure according to the structure description information, then matching and marking a target neural network structure template used by the first image neural network structure in each preset neural network structure template, determining a target sub-neural network structure corresponding to the target neural network structure template in the first image neural network structure, then replacing the target sub-neural network structure in the first image neural network structure with a plurality of preset image neural network structures corresponding to the target neural network structure template respectively to automatically generate a search space, and finally searching the neural network structure in the search space to obtain a second image neural network structure after the first image neural network structure is optimized.
Therefore, the technical threshold of searching the neural network structure is reduced, all links of searching the neural network structure are automated, a user does not need to know more neural network structure searching knowledge, and time cost and labor cost can be greatly saved.
The technical solution of the present application is described below with reference to specific embodiments.
In specific implementation, after receiving the image neural network structure optimization request, the structure description information of the first image neural network structure may be obtained from the neural network structure optimization request, and if it is determined that the structure description information of the first image neural network structure is not represented in the JSON data format, the structure description information of the first image neural network structure may be converted into the structure description information represented in the JSON data format.
Fig. 4 is a schematic diagram of an expression manner of structure description information of a first image neural network structure provided in an embodiment of the present application, where the structure description information on the upper left side and the structure description information on the lower left side are expressed by a pytorch and an mxnet, respectively, it can be seen that the structure description information of the first image neural network structure expressed by the two deep learning frameworks is different, and after the structure description information of the two different expression forms is converted into a JSON data format, the structure description information of the first image neural network structure is both in an expression form on the right side.
Therefore, the structural description information of the first image neural network structure in different expression forms is converted into a uniform expression form, so that a user can freely select to construct a deep learning framework of the first image neural network structure, the technical threshold of searching the neural network structure can be further reduced, and the user experience is improved.
The first image neural network structure may then be generated from the structure description information for the first image neural network structure represented in the JSON data format.
Assume that the first image neural network structure is as shown in fig. 5, where Conv3 × 3 denotes a convolution layer with a convolution kernel size of 3 × 3, BN denotes a Batch Normalization (BN) layer, ReLU denotes an activation layer using a ReLU activation function, Pool denotes a pooling layer, Conv1 × 1 denotes a convolution layer with a convolution kernel size of 1 × 1, and Add denotes an Add operation. And assume that there are 4 neural network structure templates, wherein the neural network structure template 1 may be abbreviated as BottleNeck, whose neural network structure is shown in fig. 6, the neural network structure template 2 may be abbreviated as BasicBlock, whose neural network structure is shown in fig. 7, the neural network structure template 3 may be abbreviated as DoubleConv, whose neural network structure is shown in fig. 8, the neural network structure template 4 may be abbreviated as SingleConv, whose neural network structure is shown in fig. 9.
Further, assuming that a preset image neural network structure corresponding to the BottleNeck is as shown in fig. 10, where r1 is a preset constant, changing the value of r1 can obtain different preset image neural network structures corresponding to the BottleNeck, and a represents the number of channels of the input channel. Assuming that the preset image neural network structure corresponding to BasicBlock is as shown in fig. 11, where r2 is a preset constant, the value of r2 is changed to obtain different preset image neural network structures corresponding to BasicBlock, and b represents the number of channels of the input channel. Assuming that the preset image neural network structure corresponding to the DoubleConv is as shown in fig. 12, where r3 is a preset constant, the value of r3 is changed to obtain different preset image neural network structures corresponding to the DoubleConv, and c represents the number of channels of the input channel. Assuming that the preset image neural network structure corresponding to SingleConv is as shown in fig. 13, where r4 is a preset constant, the value of r4 is changed to obtain different preset image neural network structures corresponding to SingleConv, and d represents the number of channels of the input channel.
And then, screening a sub-neural network structure which is the same as the neural network structure of each neural network structure template from the first image neural network structure, if the sub-neural network structure is screened out, indicating that the first image neural network structure uses the neural network structure template, further determining the neural network structure template as a target neural network structure template, and determining the screened sub-neural network structure as the target sub-neural network structure corresponding to the target neural network structure template in the first image neural network structure.
Fig. 14 is a schematic diagram of template usage information of a first image neural network structure provided in an embodiment of the present application, where target neural network structure templates used by the first image neural network structure are SingleConv and BasicBlock, a target sub-neural network structure in the first image neural network structure corresponding to SingleConv is a sub-neural network structure in an upper dotted frame, and a target sub-neural network structure in the first image neural network structure corresponding to BasicBlock is a sub-neural network structure in a lower dotted frame.
Further, a preset image neural network structure corresponding to singleConv is obtained, and a preset image neural network structure corresponding to BasicBlock is obtained.
Assume that there are 2 preset image neural network structures corresponding to SingleConv, for example, r4 in fig. 13 takes values of 2 and 3, respectively, and assume that there are 2 preset image neural network structures corresponding to BasicBlock, for example, r2 in fig. 11 takes values of 2 and 3, respectively.
Then, 2 kinds of preset image neural network structures corresponding to the SingleConv in the first image neural network structure can be used to replace the target sub-neural network structure corresponding to the SingleConv in the first image neural network structure, fig. 15 shows a schematic diagram after replacing the target sub-neural network structure corresponding to the SingleConv in the first image neural network structure, where a represents the number of channels of an input channel, and fig. 16 shows a schematic diagram after replacing the target sub-neural network structure corresponding to the SingleConv in the first image neural network structure.
Further, two preset image neural network structures corresponding to BasicBlock are used to respectively replace the target sub-neural network structure corresponding to BottleNeck in fig. 15, fig. 17 shows a schematic diagram after replacing all target sub-neural network structures in the first image neural network structure, where B represents the number of channels of the input channel, and fig. 18 shows a further schematic diagram after replacing all target sub-neural network structures in the first image neural network structure. And two preset image neural network structures corresponding to BasicBlock are used to respectively replace the target sub-neural network structure corresponding to BottleNeck in fig. 16, fig. 19 shows a schematic diagram of a first image neural network structure after replacing all target sub-neural network structures, and fig. 20 shows a schematic diagram of a first image neural network structure after replacing all target sub-neural network structures.
So far, the neural network structures shown in fig. 17 to 20 are candidate neural network structures in the candidate neural network structure set.
It should be noted that, in practical application, each neural network structure template corresponds to a relatively large number of preset image neural network structures, and for simplification, only 2 preset image neural network structures corresponding to each neural network structure template are taken as an example for description here.
Further, algorithms such as reinforcement learning and the like can be adopted to search the neural network structures in the candidate neural network structure set, and the searched neural network structure is determined to be the second image neural network structure after the first image neural network structure is optimized.
In addition, in the above process, a sub neural network structure that is the same as the neural network structure part of each neural network structure template may be screened from the first image neural network structure, and if the sub neural network structure is screened, it is indicated that the neural network structure template is used by the first image neural network structure, and then the neural network structure template may be determined as a target neural network structure template, and the screened sub neural network structure is determined as a target sub neural network structure corresponding to the target neural network structure template in the first image neural network structure.
In particular, a sub-neural network structure identical to the designated neural network structure of each neural network structure template can be screened from the first image neural network structure, wherein the designated neural network structure is a part of the neural network structures of the neural network structure templates and is used for representing the structural features of the neural network structure templates. That is, the specified neural network structure is an important neural network structure in the neural network structure template, and the first image neural network structure is considered to use the neural network structure template as long as there is a sub-neural network structure in the first image neural network structure that is the same as the important neural network structure in the neural network structure template.
Taking the BottleNeck in fig. 6 as an example, the designated neural network structure may be a top-to-bottom structure of Conv1 × 1, Conv3 × 3, Conv1 × 1, Add; taking BasicBlock in fig. 7 as an example, the designated neural network structure may be a top-down Conv3 × 3, Conv3 × 3, Add structure; taking the DoubleConv in fig. 8 as an example, the designated neural network structure may be a top-to-bottom Conv3 × 3, Conv3 × 3 structure; taking SingleConv in fig. 9 as an example, the designated neural network structure may be a structure of Conv3 × 3.
In specific implementation, after determining the target neural network structure template used by the first image neural network structure and the target sub-neural network structure in the first image neural network structure corresponding to the target neural network structure template, the subsequent processing procedure is similar to the above-mentioned processing procedure, and details are not repeated here.
The optimization method of the image neural network structure provided by the embodiment of the application can automatically generate a search space (namely a candidate neural network structure set), so that each link of the neural network structure design is automated, the artificial participation can be greatly reduced in actual services, and the technical threshold of the neural network structure search can be reduced. In addition, the optimization method of the image neural network structure provided by the embodiment of the application can be combined with various existing neural network structure search algorithms, and the flexibility is good.
When the method provided in the embodiments of the present application is implemented in software or hardware or a combination of software and hardware, a plurality of functional modules may be included in the electronic device, and each functional module may include software, hardware or a combination of software and hardware.
Fig. 21 is a schematic structural diagram of an optimization apparatus for an image neural network structure according to an embodiment of the present application, and includes a receiving module 2101, a generating module 2102, a determining module 2103, an acquiring module 2104, a replacing module 2105, and a searching module 2106.
A receiving module 2101, configured to receive an image neural network structure optimization request, where the image neural network structure optimization request includes structure description information of a first image neural network structure;
a generating module 2102 configured to generate a first image neural network structure according to the structure description information;
a determining module 2103, configured to compare the first image neural network structure with each neural network structure template, so as to determine template usage information of the first image neural network structure, where the template usage information includes a target neural network structure template used by the first image neural network structure and a target sub-neural network structure in the first image neural network structure corresponding to the target neural network structure template;
an obtaining module 2104 for obtaining a plurality of preset image neural network structures corresponding to each target neural network structure template;
a replacing module 2105, configured to use multiple preset image neural network structures corresponding to each target neural network structure template to respectively replace target sub-neural network structures, corresponding to the target neural network structure template, in the first image neural network structure, so as to obtain a candidate neural network structure set;
a searching module 2106, configured to perform neural network structure search in the candidate neural network structure set, and determine the searched neural network structure as the second image neural network structure after the first image neural network structure is optimized.
In a possible implementation, the determining module 2103 is specifically configured to:
screening a sub-neural network structure which is completely or partially identical with the neural network structure of each neural network structure template from the first image neural network structure;
and if the sub-neural network structure is screened out, determining the neural network structure template as a target neural network structure template used by the first image neural network structure, and determining the screened sub-neural network structure as a target sub-neural network structure corresponding to the target neural network structure template in the first image neural network structure.
In a possible implementation, the determining module 2103 is specifically configured to:
and screening a sub neural network structure which is the same as the designated neural network structure of each neural network structure template from the first image neural network structure, wherein the designated neural network structure is used for representing the structural characteristics of the neural network structure template.
In a possible implementation, the replacement module 2105 is specifically configured to:
taking the first image neural network structure as a reference neural network structure, and replacing the ith target sub-neural network structure in the reference neural network structure by using each preset image neural network structure corresponding to the corresponding target neural network structure template to obtain an intermediate neural network structure set, wherein i is an integer starting from 1;
if i is determined to be smaller than the total number of the target sub-neural network structures, updating i to i +1, taking each neural network structure in the intermediate neural network structure set as a new reference neural network structure, executing the step of replacing the ith target sub-neural network structure in the reference neural network structure by using each preset image neural network structure corresponding to the corresponding target neural network structure template to obtain an intermediate neural network structure set;
if it is determined that i is equal to the total number of target sub-neural network structures, determining the set of intermediate neural network structures as a set of candidate neural network structures.
In a possible implementation, the system further includes a conversion module 2107 configured to:
before generating a first image neural network structure according to the structure description information, if the structure description information is determined not to be represented by a specified data format, converting the structure description information into the structure description information represented by the specified data format.
The division of the modules in the embodiments of the present application is schematic, and only one logical function division is provided, and in actual implementation, there may be another division manner, and in addition, each functional module in each embodiment of the present application may be integrated in one processor, may also exist alone physically, or may also be integrated in one module by two or more modules. The coupling of the various modules to each other may be through interfaces that are typically electrical communication interfaces, but mechanical or other forms of interfaces are not excluded. Thus, modules described as separate components may or may not be physically separate, may be located in one place, or may be distributed in different locations on the same or different devices. The integrated module can be realized in a hardware mode, and can also be realized in a software functional module mode.
Fig. 22 is a schematic structural diagram of an electronic device according to an embodiment of the present disclosure, where the electronic device includes physical devices such as a transceiver 2201 and a processor 2202, where the processor 2202 may be a Central Processing Unit (CPU), a microprocessor, an application specific integrated circuit, a programmable logic circuit, a large scale integrated circuit, or a digital Processing Unit. The transceiver 2201 is used for data transmission and reception between the electronic device and other devices.
The electronic device may also include a memory 2203 for storing software instructions executed by the processor 2202, but may also store some other data required by the electronic device, such as identification information of the electronic device, encryption information of the electronic device, user data, etc. The Memory 2203 may be a Volatile Memory (Volatile Memory), such as a Random-Access Memory (RAM); the Memory 2203 may also be a Non-Volatile Memory (Non-Volatile Memory), such as a Read-Only Memory (ROM), a Flash Memory (Flash Memory), a Hard Disk Drive (HDD) or a Solid-State Drive (SSD), or the Memory 2203 may be any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer, but is not limited thereto. The memory 2203 may be a combination of the above memories.
The specific connection medium among the processor 2202, the memory 2203 and the transceiver 2201 is not limited in the embodiments of the present application. In the embodiment of the present application, only the memory 2203, the processor 2202, and the transceiver 2201 are connected by the bus 2204 in fig. 22 for explanation, the bus is shown by a thick line in fig. 22, and the connection manner between other components is only for illustrative purpose and is not limited thereto. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, only one thick line is shown in FIG. 22, but this does not indicate only one bus or one type of bus.
The processor 2202 may be dedicated hardware or a processor running software, and when the processor 2202 may run software, the processor 2202 reads software instructions stored in the memory 2203 and, under the drive of the software instructions, executes the optimization method of the image neural network structure involved in the foregoing embodiments.
The embodiment of the present application also provides a storage medium, and when instructions in the storage medium are executed by a processor of an electronic device, the electronic device can execute the optimization method of the image neural network structure involved in the foregoing embodiments.
In some possible embodiments, the aspects of the optimization method of the image neural network structure provided in the present application may also be implemented in the form of a program product, where the program product includes program code, and when the program product runs on an electronic device, the program code is used to make the electronic device execute the optimization method of the image neural network structure referred to in the foregoing embodiments.
The program product may employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the foregoing. More specific examples (a non-exhaustive list) of the readable storage medium include: an electrical connection having one or more wires, a portable Disk, a hard Disk, a RAM, a ROM, an Erasable Programmable Read-Only Memory (EPROM), a flash Memory, an optical fiber, a Compact Disk Read-Only Memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.
The program product for optimization of an image neural network structure in the embodiments of the present application may employ a CD-ROM and include program code, and may be run on a computing device. However, the program product of the present application is not limited thereto, and in this document, a readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.
A readable signal medium may include a propagated data signal with readable program code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated data signal may take many forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A readable signal medium may also be any readable medium that is not a readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.
Program code embodied on a readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, Radio Frequency (RF), etc., or any suitable combination of the foregoing.
Program code for carrying out operations of the present application may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C + + or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code may execute entirely on the user's computing device, partly on the user's device, as a stand-alone software package, partly on the user's computing device and partly on a remote computing device, or entirely on the remote computing device or server. In situations involving remote computing devices, the remote computing devices may be connected to the user computing device through any kind of Network, including a Local Area Network (LAN) or Wide Area Network (WAN), or may be connected to an external computing device (e.g., through the internet using an internet service provider).
It should be noted that although several units or sub-units of the apparatus are mentioned in the above detailed description, such division is merely exemplary and not mandatory. Indeed, the features and functions of two or more units described above may be embodied in one unit, according to embodiments of the application. Conversely, the features and functions of one unit described above may be further divided into embodiments by a plurality of units.
Further, while the operations of the methods of the present application are depicted in the drawings in a particular order, this does not require or imply that these operations must be performed in this particular order, or that all of the illustrated operations must be performed, to achieve desirable results. Additionally or alternatively, certain steps may be omitted, multiple steps combined into one step execution, and/or one step broken down into multiple step executions.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
While the preferred embodiments of the present application have been described, additional variations and modifications in those embodiments may occur to those skilled in the art once they learn of the basic inventive concepts. Therefore, it is intended that the appended claims be interpreted as including preferred embodiments and all alterations and modifications as fall within the scope of the application.
It will be apparent to those skilled in the art that various changes and modifications may be made in the present application without departing from the spirit and scope of the application. Thus, if such modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is intended to include such modifications and variations as well.
Claims (10)
1. A method for optimizing an image neural network structure is characterized by comprising the following steps:
receiving an image neural network structure optimization request, wherein the image neural network structure optimization request comprises structure description information of a first image neural network structure;
generating a first image neural network structure according to the structure description information;
comparing the first image neural network structure with each neural network structure template to determine template use information of the first image neural network structure, wherein the template use information comprises a target neural network structure template used by the first image neural network structure and a target sub-neural network structure corresponding to the target neural network structure template in the first image neural network structure;
acquiring a plurality of preset image neural network structures corresponding to each target neural network structure template;
replacing target sub-neural network structures corresponding to the target neural network structure template in the first image neural network structure respectively by using multiple preset image neural network structures corresponding to each target neural network structure template to obtain a candidate neural network structure set;
and searching a neural network structure in the candidate neural network structure set, and determining the searched neural network structure as a second image neural network structure after the first image neural network structure is optimized.
2. The method of claim 1, wherein comparing the first image neural network structure to neural network structure templates to determine template usage information for the first image neural network structure comprises:
screening a sub-neural network structure which is completely or partially identical with the neural network structure of each neural network structure template from the first image neural network structure;
and if the sub-neural network structure is screened out, determining the neural network structure template as a target neural network structure template used by the first image neural network structure, and determining the screened sub-neural network structure as a target sub-neural network structure corresponding to the target neural network structure template in the first image neural network structure.
3. The method of claim 2, wherein screening the first image neural network structure for sub-neural network structures that are identical to the neural network structure portion of each neural network structure template comprises:
and screening a sub neural network structure which is the same as the designated neural network structure of each neural network structure template from the first image neural network structure, wherein the designated neural network structure is used for representing the structural characteristics of the neural network structure template.
4. The method according to any one of claims 1 to 3, wherein the step of replacing the target sub-neural network structure corresponding to the target neural network structure template in the first image neural network structure with a plurality of preset image neural network structures corresponding to each target neural network structure template to obtain a candidate neural network structure set comprises:
taking the first image neural network structure as a reference neural network structure, and replacing the ith target sub-neural network structure in the reference neural network structure by using each preset image neural network structure corresponding to the corresponding target neural network structure template to obtain an intermediate neural network structure set, wherein i is an integer starting from 1;
if i is determined to be smaller than the total number of the target sub-neural network structures, updating i to i +1, taking each neural network structure in the intermediate neural network structure set as a new reference neural network structure, executing the step of replacing the ith target sub-neural network structure in the reference neural network structure by using each preset image neural network structure corresponding to the corresponding target neural network structure template to obtain an intermediate neural network structure set;
if it is determined that i is equal to the total number of target sub-neural network structures, determining the set of intermediate neural network structures as a set of candidate neural network structures.
5. The method of any of claims 1-3, further comprising, prior to generating the first image neural network structure from the structure description information:
and if the structural description information is determined not to be represented by the specified data format, converting the structural description information into the structural description information represented by the specified data format.
6. An apparatus for optimizing an image neural network structure, comprising:
the image neural network structure optimization system comprises a receiving module, a processing module and a processing module, wherein the receiving module is used for receiving an image neural network structure optimization request which contains structure description information of a first image neural network structure;
the generating module is used for generating a first image neural network structure according to the structure description information;
a determining module, configured to compare the first image neural network structure with each neural network structure template to determine template usage information of the first image neural network structure, where the template usage information includes a target neural network structure template used by the first image neural network structure and a target sub-neural network structure in the first image neural network structure corresponding to the target neural network structure template;
the acquisition module is used for acquiring a plurality of preset image neural network structures corresponding to each target neural network structure template;
the replacing module is used for replacing target sub-neural network structures corresponding to the target neural network structure template in the first image neural network structure by using a plurality of preset image neural network structures corresponding to each target neural network structure template to obtain a candidate neural network structure set;
and the searching module is used for searching the neural network structure in the candidate neural network structure set and determining the searched neural network structure as the second image neural network structure after the first image neural network structure is optimized.
7. The apparatus of claim 6, wherein the determination module is specifically configured to:
screening a sub-neural network structure which is completely or partially identical with the neural network structure of each neural network structure template from the first image neural network structure;
and if the sub-neural network structure is screened out, determining the neural network structure template as a target neural network structure template used by the first image neural network structure, and determining the screened sub-neural network structure as a target sub-neural network structure corresponding to the target neural network structure template in the first image neural network structure.
8. The apparatus of claim 7, wherein the determination module is specifically configured to:
and screening a sub neural network structure which is the same as the designated neural network structure of each neural network structure template from the first image neural network structure, wherein the designated neural network structure is used for representing the structural characteristics of the neural network structure template.
9. An electronic device, comprising: at least one processor, and a memory communicatively coupled to the at least one processor, wherein:
the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-5.
10. A storage medium, wherein instructions in the storage medium, when executed by a processor of an electronic device, enable the electronic device to perform the method of any of claims 1-5.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110102292.XA CN112819138A (en) | 2021-01-26 | 2021-01-26 | Optimization method and device of image neural network structure |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110102292.XA CN112819138A (en) | 2021-01-26 | 2021-01-26 | Optimization method and device of image neural network structure |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112819138A true CN112819138A (en) | 2021-05-18 |
Family
ID=75859262
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110102292.XA Pending CN112819138A (en) | 2021-01-26 | 2021-01-26 | Optimization method and device of image neural network structure |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112819138A (en) |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH06333054A (en) * | 1993-04-27 | 1994-12-02 | Sony Electron Inc | System for detecting target pattern within image |
| CN109919304A (en) * | 2019-03-04 | 2019-06-21 | 腾讯科技（深圳）有限公司 | Neural network searching method, device, readable storage medium storing program for executing and computer equipment |
| WO2020092810A1 (en) * | 2018-10-31 | 2020-05-07 | Movidius Ltd. | Automated generation of neural networks |
| CN111178512A (en) * | 2019-12-31 | 2020-05-19 | 中国科学院自动化研究所南京人工智能芯片创新研究院 | Device operation neural network test method and device |
| CN111406267A (en) * | 2017-11-30 | 2020-07-10 | 谷歌有限责任公司 | Neural architecture search using performance-predictive neural networks |
| CN111406263A (en) * | 2018-11-28 | 2020-07-10 | 深圳市大疆创新科技有限公司 | Method and device for searching neural network architecture |
| CN111967569A (en) * | 2020-06-29 | 2020-11-20 | 北京百度网讯科技有限公司 | Neural network structure generation method and device, storage medium and electronic equipment |
| CN111985644A (en) * | 2020-08-28 | 2020-11-24 | 北京市商汤科技开发有限公司 | Neural network generation method and device, electronic device and storage medium |
| CN112000450A (en) * | 2020-08-18 | 2020-11-27 | ***股份有限公司 | Neural network architecture searching method and device |
-
2021
- 2021-01-26 CN CN202110102292.XA patent/CN112819138A/en active Pending
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH06333054A (en) * | 1993-04-27 | 1994-12-02 | Sony Electron Inc | System for detecting target pattern within image |
| CN111406267A (en) * | 2017-11-30 | 2020-07-10 | 谷歌有限责任公司 | Neural architecture search using performance-predictive neural networks |
| WO2020092810A1 (en) * | 2018-10-31 | 2020-05-07 | Movidius Ltd. | Automated generation of neural networks |
| CN111406263A (en) * | 2018-11-28 | 2020-07-10 | 深圳市大疆创新科技有限公司 | Method and device for searching neural network architecture |
| CN109919304A (en) * | 2019-03-04 | 2019-06-21 | 腾讯科技（深圳）有限公司 | Neural network searching method, device, readable storage medium storing program for executing and computer equipment |
| CN111178512A (en) * | 2019-12-31 | 2020-05-19 | 中国科学院自动化研究所南京人工智能芯片创新研究院 | Device operation neural network test method and device |
| CN111967569A (en) * | 2020-06-29 | 2020-11-20 | 北京百度网讯科技有限公司 | Neural network structure generation method and device, storage medium and electronic equipment |
| CN112000450A (en) * | 2020-08-18 | 2020-11-27 | ***股份有限公司 | Neural network architecture searching method and device |
| CN111985644A (en) * | 2020-08-28 | 2020-11-24 | 北京市商汤科技开发有限公司 | Neural network generation method and device, electronic device and storage medium |
Non-Patent Citations (3)
| Title |
|---|
| THOMAS ELSKEN ET AL.: "Neural Architecture Search: A Survey", 《ARXIV》 * |
| 孟子尧 等: "深度神经架构搜索综述", 《计算机研究与发展》 * |
| 李涛;: "神经网络搜索综述", 电脑知识与技术, no. 17 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110807515B (en) | Model generation method and device | |
| CN110852438B (en) | Model generation method and device | |
| CN110852421B (en) | Model generation method and device | |
| US20190325307A1 (en) | Estimation of resources utilized by deep learning applications | |
| US20190279088A1 (en) | Training method, apparatus, chip, and system for neural network model | |
| CN111340220B (en) | Method and apparatus for training predictive models | |
| CN111406264A (en) | Neural architecture search | |
| CN111368973B (en) | Method and apparatus for training a super network | |
| CN108304376B (en) | Text vector determination method and device, storage medium and electronic device | |
| CN111133458B (en) | Enhanced neural network | |
| CN116502807B (en) | Industrial chain analysis application method and device based on scientific and technological knowledge graph | |
| CN111353601A (en) | Method and apparatus for predicting delay of model structure | |
| CN116129452A (en) | Method, application method, device, equipment and medium for generating document understanding model | |
| CN110782016A (en) | Method and apparatus for optimizing neural network architecture search | |
| CN112712795B (en) | Labeling data determining method, labeling data determining device, labeling data determining medium and electronic equipment | |
| CN115186738B (en) | Model training method, device and storage medium | |
| CN112819138A (en) | Optimization method and device of image neural network structure | |
| CN116092092A (en) | Matching method, device, medium and electronic equipment | |
| CN114187259A (en) | Creation method of video quality analysis engine, video quality analysis method and equipment | |
| CN111582456B (en) | Method, apparatus, device and medium for generating network model information | |
| CN112948251A (en) | Automatic software testing method and device | |
| CN113761289A (en) | Method, frame, computer system and readable storage medium for drawing learning | |
| CN116107666B (en) | Program service flow information generation method, device, electronic equipment and computer medium | |
| CN111461310A (en) | Neural network device, neural network system and method for processing neural network model | |
| CN111406267B (en) | Neural architecture search using performance prediction neural networks |
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 |