CN110569972A - search space construction method and device of hyper network and electronic equipment - Google Patents
search space construction method and device of hyper network and electronic equipment Download PDFInfo
- Publication number
- CN110569972A CN110569972A CN201910860577.2A CN201910860577A CN110569972A CN 110569972 A CN110569972 A CN 110569972A CN 201910860577 A CN201910860577 A CN 201910860577A CN 110569972 A CN110569972 A CN 110569972A
- Authority
- CN
- China
- Prior art keywords
- layer
- channels
- input
- output
- channel number
- 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
- 238000010276 construction Methods 0.000 title claims abstract description 24
- 238000010586 diagram Methods 0.000 claims abstract description 83
- 238000006243 chemical reaction Methods 0.000 claims abstract description 38
- 230000003044 adaptive effect Effects 0.000 claims abstract description 15
- 230000009466 transformation Effects 0.000 claims abstract description 6
- 238000000034 method Methods 0.000 claims description 23
- 230000015654 memory Effects 0.000 claims description 17
- 230000006978 adaptation Effects 0.000 claims description 3
- 238000013528 artificial neural network Methods 0.000 abstract description 20
- 238000013461 design Methods 0.000 abstract description 9
- 230000008859 change Effects 0.000 abstract description 6
- 230000008569 process Effects 0.000 description 6
- 238000004422 calculation algorithm Methods 0.000 description 4
- 238000004891 communication Methods 0.000 description 4
- 238000004590 computer program Methods 0.000 description 4
- 230000006870 function Effects 0.000 description 4
- 230000004048 modification Effects 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 239000004973 liquid crystal related substance Substances 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 210000004027 cell Anatomy 0.000 description 1
- 230000001413 cellular effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000010295 mobile communication Methods 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000001151 other effect Effects 0.000 description 1
- 230000002787 reinforcement Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000001953 sensory effect Effects 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
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
- G06N3/045—Combinations of networks
-
- 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
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
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 a search space construction method and device of a super network and electronic equipment, and relates to the field of neural network search. The specific implementation scheme is as follows: acquiring a plurality of output channels of the L-th layer characteristic diagram, wherein L is greater than or equal to 1; inputting a plurality of output channel numbers of the L-th layer feature diagram to an adaptive layer for channel number conversion to obtain a plurality of input channel numbers of the L + 1-th layer feature diagram; and constructing a search space of the super network according to the number of a plurality of output channels of the L-th layer characteristic diagram and the number of a plurality of input channels of the L + 1-th layer characteristic diagram. By the self-adaptive hierarchical connection design of channel number transformation, the width of a search space structure is expanded, the technical problem that the number of network convolution channels is invariable is solved, parameter sharing under the scene of supporting the size change of a deep neural network structure is realized, and efficient automatic search of the model structure can be realized under the constraint conditions of any model size and model speed.
Description
Technical Field
The present application relates to the field of computer vision, and more particularly to the field of neural network searching.
background
Deep learning techniques have enjoyed tremendous success in many directions, and NAS technology (Neural Architecture Search) has become a research hotspot in recent years. The NAS is a neural network architecture which is automatically searched out in a massive search space by replacing fussy manual operation with an algorithm. The step of conducting an architectural search of the neural network includes: first, a search space is defined and determined. Then, a search strategy is determined according to the adopted optimization algorithm, such as an algorithm of reinforcement learning, an evolutionary algorithm, Bayesian optimization and the like. And finally, searching to obtain the speed of the model structure and the performance of the model.
the search space based on the super network is typically based on a stack of cell blocks. The unit structure comprises two basic unit structures, namely a common unit structure and a downsampling unit structure. Based on the super network search space design of unit structure stacking, the size of the deep neural network comprises the size of a convolution kernel, the number of channels and the depth. The size of the deep neural network is critical to the speed of the super network, the size of the model, and the performance of the model. The number of channels of the originally input picture sample depends on the picture type, such as RGB; the number of output channels after the convolution operation is completed, i.e., the number of output channels, depends on the number of convolution kernels. The number of output channels at this time is also used as the number of input channels of the convolution kernel at the next convolution. However, the number of output channels of the previous convolution and the number of input channels of the next convolution must be fixed, otherwise, parameter sharing cannot be realized. The fixed number of network channels places a significant constraint on the search space, resulting in fewer deep neural networks being selectable within the search space.
Disclosure of Invention
The embodiment of the application provides a search space construction method and device of a hyper network and electronic equipment, and aims to solve at least one technical problem.
in a first aspect, an embodiment of the present application provides a method for constructing a search space of a super network, including:
Acquiring a plurality of output channels of the L-th layer characteristic diagram, wherein L is greater than or equal to 1;
Inputting a plurality of output channel numbers of the L-th layer feature diagram to an adaptive layer for channel number conversion to obtain a plurality of input channel numbers of the L + 1-th layer feature diagram;
And constructing a search space of the super network according to the number of a plurality of output channels of the L-th layer characteristic diagram and the number of a plurality of input channels of the L + 1-th layer characteristic diagram.
in the embodiment, the width of a search space structure is expanded through the adaptive hierarchical connection design of channel number conversion, parameter sharing under the scene of supporting the size change of a deep neural network structure is realized, and efficient automatic search of the model structure can be realized under the constraint conditions of any model size and model speed.
In one embodiment, inputting a plurality of output channels of the L-th layer feature map to an adaptive layer for channel number conversion to obtain a plurality of input channels of the L + 1-th layer feature map includes:
Inputting a plurality of output channel numbers of the L-th layer feature diagram into a plurality of operation layers to obtain a plurality of input channel numbers of the L + 1-th layer feature diagram;
the number of output channels of the L-th layer feature diagram is the same as the number of input channels of the L + 1-th layer feature diagram.
in the embodiment, the conversion of a plurality of output channels into a plurality of input channels is realized, so that the number of network channels is variable, the width of a search space structure is expanded, and the number of deep neural networks in a space is expanded.
In one embodiment, inputting a plurality of output channels of the L-th layer feature map into a plurality of operation layers to obtain a plurality of input channels of the L + 1-th layer feature map includes:
Inputting the Mth branch input channel number in the plurality of output channel numbers of the L-th layer characteristic diagram into different convolution operations to obtain a plurality of branch input channel numbers, wherein M is greater than or equal to 1;
and determining the number of all branch input channels obtained according to the number of all branch output channels as a plurality of input channels of the L + 1-th layer characteristic diagram.
In the embodiment, since the design of the 2-level structure is adopted, the number of channels can be selected from the L-th layer characteristic diagram to the L + 1-th layer characteristic diagram, so that the problem that the number of channels is not variable is solved.
In one embodiment, inputting a plurality of output channels of the L-th layer feature map to an adaptive layer for channel number conversion to obtain a plurality of input channels of the L + 1-th layer feature map includes:
Inputting the Mth branch output channel number in the plurality of output channel numbers of the L-th layer characteristic diagram into the 1 x 1 convolution layer to obtain the Mth branch input channel number, wherein M is greater than or equal to 1;
The number of the Mth branch output channels is different from the number of the Mth branch input channels.
in the present embodiment, since 1 × 1 convolution is used, an arbitrary number of output channels and an arbitrary number of input channels can be directly selected, which not only reduces the amount of calculation but also improves the search speed and efficiency.
in a second aspect, an embodiment of the present application provides a search space construction apparatus for a super network, including:
the output channel number acquisition module is used for acquiring a plurality of output channel numbers of the L-th layer feature diagram, wherein L is greater than or equal to 1;
The channel number conversion module is used for inputting a plurality of output channel numbers of the L-th layer characteristic diagram to an adaptive layer for channel number conversion to obtain a plurality of input channel numbers of the L + 1-th layer characteristic diagram;
And the search space construction module is used for constructing a search space of the super network according to the number of a plurality of output channels of the L-th layer characteristic diagram and the number of a plurality of input channels of the L + 1-th layer characteristic diagram.
In one embodiment, a channel number conversion module includes:
The first conversion submodule is used for inputting a plurality of output channel numbers of the L-th layer feature diagram into a plurality of operation layers to obtain a plurality of input channel numbers of the L + 1-th layer feature diagram, wherein the plurality of output channel numbers of the L-th layer feature diagram are the same as the plurality of input channel numbers of the L + 1-th layer feature diagram.
in one embodiment, the first conversion submodule includes:
The branch channel number conversion unit is used for inputting the Mth branch input channel number in the output channel numbers of the L-th layer characteristic diagram to different convolution operations to obtain a plurality of branch input channel numbers, wherein M is greater than or equal to 1;
And the input channel number determining unit is used for determining the number of all branch input channels obtained according to the number of the branch output channels as a plurality of input channels of the L + 1-th layer characteristic diagram.
In one embodiment, a channel number conversion module includes:
and the second conversion submodule is used for inputting the Mth branch output channel number in the output channel numbers of the L-th layer characteristic diagram into the 1 x 1 convolution layer to obtain the Mth branch input channel number, wherein M is greater than or equal to 1, and the Mth branch output channel number is different from the Mth branch input channel number.
One embodiment in the above application has the following advantages or benefits: because the technical means of self-adaptive hierarchical connection of channel number transformation is adopted, the technical problem that the number of network convolution channels in the structural dimension of the deep neural network is invariable is solved, the quantity of the deep neural networks in a search space is further improved, parameter sharing is realized under the scene of supporting the structural dimension change of the deep neural network, and the technical effect of efficient automatic searching of the model structure is realized under the constraint conditions of any model size and model speed.
other effects of the above-described alternative will be described below with reference to specific embodiments.
drawings
The drawings are included to provide a better understanding of the present solution and are not intended to limit the present application. Wherein:
FIG. 1 is a schematic flow chart of a search space construction method of a hyper-network according to an embodiment of the present application;
FIG. 2 is a schematic flow chart of a search space construction method of another hyper-network according to an embodiment of the present application;
FIG. 3 is a scene diagram of a search space construction method of a hyper-network, in which embodiments of the present application may be implemented;
FIG. 4 is a diagram illustrating a search space construction method for another super network that can implement embodiments of the present application;
FIG. 5 is a block diagram of a search space construction apparatus for a super network, which can implement an embodiment of the present application;
FIG. 6 is a block diagram of another search space construction apparatus for a super network, in which embodiments of the present application can be implemented;
FIG. 7 is a block diagram of an electronic device for implementing a method for search space construction for a hyper-network according to an embodiment of the present application.
Detailed Description
The following description of the exemplary embodiments of the present application, taken in conjunction with the accompanying drawings, includes various details of the embodiments of the application for the understanding of the same, which are to be considered exemplary only. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the present application. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.
EXAMPLE I …
in one embodiment, as shown in fig. 1, a method for constructing a search space of a super network is provided, which includes:
step S10: acquiring a plurality of output channels of the L-th layer characteristic diagram, wherein L is greater than or equal to 1;
Step S20: inputting a plurality of output channel numbers of the L-th layer feature diagram to an adaptive layer for channel number conversion to obtain a plurality of input channel numbers of the L + 1-th layer feature diagram;
step S30: and constructing a search space of the super network according to the number of a plurality of output channels of the L-th layer characteristic diagram and the number of a plurality of input channels of the L + 1-th layer characteristic diagram.
In one example, any integer can be used as the number of output channels of a feature map of a certain layer. Usually, for better performance, the number of channels is designed to be a multiple of 4, for example, a plurality of channels in the channel number set [3,4,8,12,16,24,32,48,64,80,96,128,144,160,192,224,256,320,384,512,1024,2048] may be selected to be determined as a plurality of output channels of the lth feature map. The determination of the number of the output channels expands the width of a search space structure, so that the number of the output channels of the feature map of the previous layer has multiple choices when convolution operation is carried out between every two layers of feature maps.
However, in the parallel branch with only 1-level hierarchy, the number of output channels of the feature map of the previous layer and the number of input channels of the next layer must be the same and fixed. For example, the deep neural network can only be coded [ 2335 ], i.e., the number of output channels of the second layer feature map is 12, and the number of input channels of the third layer feature map is only 12. In this embodiment, if the number of output channels of the second layer feature diagram is 16 and 32, and the number of input channels of the third layer feature diagram is also 16 and 32, two different operation layers may be used to act on the number of output channels 16 of the second layer feature diagram to obtain 16 and 32, and another two different operation layers may be used to act on the number of output channels 32 of the second layer feature diagram to obtain 32 and 16. The obtained 16 and 32 are used as the input channel number of the third layer characteristic diagram. Through the process, the conversion of the number of channels is realized, the deep neural network can be searched according to the number of the channels selected in various ways, and the number of the deep neural networks in a search space is increased. It should be noted that the four operation layers preset in the above process are used to implement channel number conversion, and the adaptive layer includes the four operation layers preset in the above process. Of course, the selection of the number and size of the output network channels may be adaptively adjusted according to actual needs, and all are within the protection scope of the present embodiment.
The search space of the super network provided by the embodiment breaks through the bottleneck that the traditional super network search space based on unit structure stacking cannot search the size of the deep neural network. Through the adaptive hierarchical connection design of channel number conversion, parameter sharing under the scene of supporting the size change of the deep neural network structure is realized, and efficient automatic searching of the model structure can be realized under the constraint conditions of any model size and model speed.
in one embodiment, as shown in fig. 2, step S20 includes:
Step S201: and inputting a plurality of output channels of the L-th layer feature diagram into a plurality of operation layers to obtain a plurality of input channels of the L + 1-th layer feature diagram, wherein the plurality of output channels of the L-th layer feature diagram are the same as the plurality of input channels of the L + 1-th layer feature diagram.
in this embodiment, the adaptive layer for channel number conversion includes a plurality of operation layers, and realizes conversion from a plurality of output channel numbers to a plurality of input channel numbers, so that the number of network channels is variable, the width of a search space structure is expanded, and the number of deep neural networks in a space is expanded.
In one embodiment, as shown in fig. 2, step S201: the method comprises the following steps:
Step S2011: inputting the Mth branch input channel number in the plurality of output channel numbers of the L-th layer characteristic diagram into different convolution operations to obtain a plurality of branch input channel numbers, wherein M is greater than or equal to 1;
step S2012: and determining the number of all branch input channels obtained according to the number of all branch output channels as a plurality of input channels of the L + 1-th layer characteristic diagram.
In one example, as shown in fig. 3, the number of output channels of the L-th layer feature map includes C1-8, C2-12, C3-16, and C4-24. C1 is input to the 1 st convolutional layer to the 4 th convolutional layer in sequence as the first branch input channel number, and after the convolution operations of the four convolutional layers, the branch input channel numbers 8,12,16 and 24 are output respectively. C2 is input to the 5 th convolutional layer to the 8 th convolutional layer in sequence as the second branch input channel number of 12, and after the convolution operations of the four convolutional layers, the branch input channel numbers of 8,12,16 and 24 are output respectively. C3 is input to the 9 th convolutional layer to the 12 th convolutional layer in sequence as the third branch input channel number, and after the convolution operations of the four convolutional layers, the branch input channel numbers 8,12,16 and 24 are output respectively. C4 is input to the 13 th convolution layer to the 16 th convolution layer in sequence as the fourth branch input channel number 24, and the branch input channel numbers 8,12,16 and 24 are output, respectively. The operation process in each branch is described by taking C2-12 as an input into the 8 th convolutional layer and the output 24 as an example. The number of branch input channels, C2-12, is input into the 8 th convolutional layer, which contains N different convolution operations of C2-OP1 … … C2-OPN, and the number of branch output channels obtained by whichever convolution operation is used is fixed, and is 24.
By analogy, C2 is input to the 17 th to 20 th convolutional layers in turn, and 8,12,16, and 24 are output, respectively. C2 is sequentially input to the 21 st to 24 th convolutional layers, and output 8,12,16, and 24, respectively. C2 is sequentially input to 25 th to 28 th convolutional layers, and outputs 8,12,16, and 24, respectively. Similarly, each convolution operation in the same convolution layer is different, and the obtained branch output channels are consistent in number. A similar process continues until C4 ═ 24 is input to the different convolutional layers, respectively, outputting 8,12,16, and 24. By analogy, the number of CM branch input channels is output after N different convolution operations of CM-OP1 … … CM-OPN.
Due to the adoption of a 2-level structure design, the number of channels corresponding to a plurality of branch outputs of the characteristic diagram is different at the 1 st level. And in the 2 nd level, performing multiple convolution operations on each branch output channel number to obtain multiple branch input channel numbers, wherein each convolution operation obtains the branch input channel number. Through the design, due to the fact that a plurality of different branches are arranged, a plurality of different channel numbers can be selected from the L-th layer characteristic diagram to the L + 1-th layer characteristic diagram, and therefore the problem that the channel numbers are not variable is solved.
in one embodiment, as shown in fig. 2, step S20: the method comprises the following steps:
step S202: inputting the Mth branch output channel number in the plurality of output channel numbers of the L-th layer characteristic diagram into the 1 x 1 convolution layer to obtain the Mth branch input channel number, wherein M is greater than or equal to 1, and the Mth branch output channel number is different from the Mth branch input channel number.
in one example, as shown in fig. 4, the adaptive layer for channel number transformation includes 1 × 1 convolution layer, and the convolution with 1 × 1 can deepen the widened network structure. And converting any output channel number into any input channel number. The signatures of (h1, w1, c1) become signatures of (h1, w1, c2) after 1 × 1 convolution (c 2). The 1-by-1 convolution does not change the length h1 and the width w1 of the feature map, but converts the number of output channels c1 of the feature map in the previous layer into the number of input channels c2 of the feature map in the next layer. For example, the number of output channels of the L-th layer is selected from two different channel numbers of 128 and 256. The number of input channels of the L +1 th layer is selected from 96 and 192. The adaptation layer for the channel number transformation needs to store the parameter weight matrices for four different channel lengths 128 to 96,128 to 192, 256 to 96, 256 to 192, respectively. Wherein, the parameter refers to the parameter of 1 × 1 convolution. The variability of the number of channels is also satisfied while preserving the single-level parameter sharing property. Because 1-by-1 convolution is utilized, any output channel number and input channel number can be directly selected, the calculation amount can be reduced, and the searching speed and efficiency are improved.
example two
In another embodiment, as shown in fig. 5, there is provided a search space constructing apparatus 100 for a super network, including:
an output channel number obtaining module 110, configured to obtain a number of output channels of an lth-layer feature map, where L is greater than or equal to 1;
A channel number conversion module 120, configured to input a plurality of output channel numbers of the L-th layer feature map to an adaptive layer for channel number conversion, so as to obtain a plurality of input channel numbers of the L + 1-th layer feature map;
And the search space construction module 130 is configured to construct a search space of the super network according to the number of the output channels of the L-th layer feature diagram and the number of the input channels of the L + 1-th layer feature diagram.
in one embodiment, as shown in fig. 6, there is provided a search space constructing apparatus 200 for a super network, the channel number converting module 120, including:
The first conversion sub-module 121 is configured to input a plurality of output channels of the L-th layer feature diagram into a plurality of operation layers to obtain a plurality of input channels of the L + 1-th layer feature diagram, where the plurality of output channels of the L-th layer feature diagram are the same as the plurality of input channels of the L + 1-th layer feature diagram.
in one embodiment, as shown in fig. 6, the first conversion submodule 121 includes:
a branch channel number conversion unit 1211, configured to input an mth branch input channel number of the multiple output channel numbers of the L-th layer feature map to different convolution operations to obtain multiple branch input channel numbers, where M is greater than or equal to 1;
The input channel number determining unit 1212 is configured to determine the number of all branch input channels obtained according to the number of branch output channels as the number of input channels of the L +1 th layer feature map.
In one embodiment, as shown in fig. 6, the channel number conversion module 120 includes:
and a second conversion sub-module 122, configured to input an mth branch output channel number of the plurality of output channel numbers of the L-th layer feature map into the 1 × 1 convolution layer to obtain an mth branch input channel number, where M is greater than or equal to 1, and the mth branch output channel number is different from the mth branch input channel number.
According to an embodiment of the present application, an electronic device and a readable storage medium are also provided.
fig. 7 is a block diagram of an electronic device according to an embodiment of the present application, illustrating a method for constructing a search space of a super network. Electronic devices are intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the present application that are described and/or claimed herein.
As shown in fig. 7, the electronic apparatus includes: one or more processors 701, a memory 702, and interfaces for connecting the various components, including a high-speed interface and a low-speed interface. The various components are interconnected using different buses and may be mounted on a common motherboard or in other manners as desired. The processor may process instructions for execution within the electronic device, including instructions stored in or on the memory to display Graphical information for a Graphical User Interface (GUI) on an external input/output device, such as a display device coupled to the Interface. In other embodiments, multiple processors and/or multiple buses may be used, along with multiple memories and multiple memories, as desired. Also, multiple electronic devices may be connected, with each device providing portions of the necessary operations (e.g., as a server array, a group of blade servers, or a multi-processor system). In fig. 7, one processor 701 is taken as an example.
the memory 702 is a non-transitory computer readable storage medium as provided herein. The memory stores instructions executable by the at least one processor to cause the at least one processor to perform a search space construction method for a hyper-network provided by the present application. The non-transitory computer-readable storage medium of the present application stores computer instructions for causing a computer to execute a search space construction method of a hyper network provided by the present application.
The memory 702 serves as a non-transitory computer readable storage medium, and can be used for storing non-transitory software programs, non-transitory computer executable programs, and modules, such as program instructions/modules corresponding to a search space construction method of a super network in the embodiment of the present application (for example, the output channel number obtaining module 110, the channel number conversion module 120, and the search space construction module 130 shown in fig. 5). The processor 701 executes various functional applications of the server and data processing by running non-transitory software programs, instructions, and modules stored in the memory 702, that is, implements a search space construction method of a hyper-network in the above method embodiments.
The memory 702 may include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application program required for at least one function; the storage data area may store data created from use of an electronic device constructed according to a search space of a kind of hyper network, and the like. Further, the memory 702 may include high speed random access memory, and may also include non-transitory memory, such as at least one magnetic disk storage device, flash memory device, or other non-transitory solid state storage device. In some embodiments, memory 702 optionally includes memory located remotely from processor 701, which may be networked to a search space-building electronic device of a extranet. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.
The electronic device of the search space construction method of the super network may further include: an input device 703 and an output device 704. The processor 701, the memory 702, the input device 703 and the output device 704 may be connected by a bus or other means, and fig. 7 illustrates an example of a connection by a bus.
the input device 703 may receive input numeric or character information and generate key signal inputs related to user settings and function controls of an electronic device constructed with a search space of a super network, such as a touch screen, keypad, mouse, track pad, touch pad, pointer, one or more mouse buttons, track ball, joystick, or other input device. The output devices 704 may include a display device, auxiliary lighting devices (e.g., LEDs), and tactile feedback devices (e.g., vibrating motors), among others. The display device may include, but is not limited to, a Liquid Crystal Display (LCD) such as a Cr7 star display 7, a Light Emitting Diode (LED) display, and a plasma display. In some implementations, the display device can be a touch screen.
various implementations of the systems and techniques described here can be realized in digital electronic circuitry, Integrated circuitry, Application Specific Integrated Circuits (ASICs), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include: implemented in one or more computer programs that are executable and/or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, receiving data and instructions from, and transmitting data and instructions to, a storage system, at least one input device, and at least one output device.
These computer programs (also known as programs, software applications, or code) include machine instructions for a programmable processor, and may be implemented using high-level procedural and/or object-oriented programming languages, and/or assembly/machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and/or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and/or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal used to provide machine instructions and/or data to a programmable processor.
To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having: a display device (e.g., a CRT (Cathode Ray Tube) or LCD (liquid crystal display) monitor) for displaying information to a user; and a keyboard and a pointing device (e.g., a mouse or a trackball) by which a user can provide input to the computer. Other kinds of devices may also be used to provide for interaction with a user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form, including acoustic, speech, or tactile input.
The systems and techniques described here can be implemented in a computing system that includes a back-end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front-end component (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: local Area Networks (LANs), Wide Area Networks (WANs), and the internet.
the computer system may include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.
According to the technical scheme of the embodiment of the application, the bottleneck that the traditional super-network search space based on unit structure stacking cannot search the size of the deep neuron network is broken through. Through the adaptive hierarchical connection design of channel number conversion, parameter sharing under the scene of supporting the size change of the deep neural network structure is realized, and efficient automatic searching of the model structure can be realized under the constraint conditions of any model size and model speed.
it should be understood that various forms of the flows shown above may be used, with steps reordered, added, or deleted. For example, the steps described in the present application may be executed in parallel, sequentially, or in different orders, and the present invention is not limited thereto as long as the desired results of the technical solutions disclosed in the present application can be achieved.
the above-described embodiments should not be construed as limiting the scope of the present application. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may be made in accordance with design requirements and other factors. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims (10)
1. A search space construction method of a hyper network is characterized by comprising the following steps:
acquiring a plurality of output channels of the L-th layer characteristic diagram, wherein L is greater than or equal to 1;
Inputting a plurality of output channel numbers of the L-th layer feature diagram to an adaptive layer for channel number conversion to obtain a plurality of input channel numbers of an L + 1-th layer feature diagram;
And constructing a search space of the super network according to the number of the output channels of the L-th layer characteristic diagram and the number of the input channels of the L + 1-th layer characteristic diagram.
2. The method of claim 1, wherein inputting the plurality of output channels of the L-th layer feature map to an adaptation layer for channel number transformation to obtain a plurality of input channels of an L + 1-th layer feature map comprises:
inputting a plurality of output channel numbers of the L-th layer feature diagram into a plurality of operation layers to obtain a plurality of input channel numbers of the L + 1-th layer feature diagram;
Wherein the number of output channels of the L-th layer feature diagram is the same as the number of input channels of the L + 1-th layer feature diagram.
3. the method of claim 2, wherein inputting the number of output channels of the L-th layer feature map into a plurality of operation layers to obtain a number of input channels of the L + 1-th layer feature map comprises:
inputting the Mth branch input channel number in the plurality of output channel numbers of the L-th layer characteristic diagram to different convolution operations to obtain a plurality of branch input channel numbers, wherein M is greater than or equal to 1;
And determining the number of all branch input channels obtained according to the number of all branch output channels as a plurality of input channels of the L + 1-th layer characteristic diagram.
4. the method of claim 1, wherein inputting the plurality of output channels of the L-th layer feature map to an adaptation layer for channel number transformation to obtain a plurality of input channels of an L + 1-th layer feature map comprises:
inputting the Mth branch output channel number in the plurality of output channel numbers of the L-th layer feature diagram into a 1 x 1 convolution layer to obtain the Mth branch input channel number, wherein M is greater than or equal to 1;
Wherein the Mth branch output channel number is different from the Mth branch input channel number.
5. a search space construction apparatus for a super network, comprising:
the output channel number acquisition module is used for acquiring a plurality of output channel numbers of the L-th layer feature diagram, wherein L is greater than or equal to 1;
The channel number conversion module is used for inputting a plurality of output channel numbers of the L-th layer characteristic diagram to an adaptive layer for channel number conversion to obtain a plurality of input channel numbers of the L + 1-th layer characteristic diagram;
and the search space construction module is used for constructing a search space of the super network according to the number of the output channels of the L-th layer characteristic diagram and the number of the input channels of the L + 1-th layer characteristic diagram.
6. the apparatus of claim 5, wherein the channel number conversion module comprises:
the first conversion submodule is configured to input the number of output channels of the L-th layer feature map into multiple operation layers to obtain the number of input channels of the L + 1-th layer feature map, where the number of output channels of the L-th layer feature map is the same as the number of input channels of the L + 1-th layer feature map.
7. The apparatus of claim 6, wherein the first conversion submodule comprises:
The branch channel number conversion unit is used for inputting the Mth branch input channel number in the output channel numbers of the L-th layer characteristic diagram into different convolution operations to obtain a plurality of branch input channel numbers, wherein M is greater than or equal to 1;
and the input channel number determining unit is used for determining the number of all branch input channels obtained according to the number of branch output channels as a plurality of input channels of the L + 1-th layer characteristic diagram.
8. The apparatus of claim 5, wherein the channel number conversion module comprises:
and the second conversion submodule is used for inputting the Mth branch output channel number in the output channel numbers of the L-th layer characteristic diagram into the 1 x 1 convolution layer to obtain the Mth branch input channel number, wherein M is greater than or equal to 1, and the Mth branch output channel number is different from the Mth branch input channel number.
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 1-4.
10. a non-transitory computer readable storage medium having stored thereon computer instructions for causing the computer to perform the method of any one of claims 1-4.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910860577.2A CN110569972A (en) | 2019-09-11 | 2019-09-11 | search space construction method and device of hyper network and electronic equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910860577.2A CN110569972A (en) | 2019-09-11 | 2019-09-11 | search space construction method and device of hyper network and electronic equipment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110569972A true CN110569972A (en) | 2019-12-13 |
Family
ID=68779397
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910860577.2A Pending CN110569972A (en) | 2019-09-11 | 2019-09-11 | search space construction method and device of hyper network and electronic equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110569972A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111401517A (en) * | 2020-02-21 | 2020-07-10 | 华为技术有限公司 | Method and device for searching perception network structure |
| CN111582454A (en) * | 2020-05-09 | 2020-08-25 | 北京百度网讯科技有限公司 | Method and device for generating neural network model |
| CN111680600A (en) * | 2020-05-29 | 2020-09-18 | 北京百度网讯科技有限公司 | Face recognition model processing method, device, equipment and storage medium |
| CN112116090A (en) * | 2020-09-28 | 2020-12-22 | 腾讯科技（深圳）有限公司 | Neural network structure searching method and device, computer equipment and storage medium |
| WO2021258964A1 (en) * | 2020-06-22 | 2021-12-30 | 华为技术有限公司 | Neural network architecture search method, apparatus and system |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110197258A (en) * | 2019-05-29 | 2019-09-03 | 北京市商汤科技开发有限公司 | Neural network searching method, image processing method and device, equipment and medium |
-
2019
- 2019-09-11 CN CN201910860577.2A patent/CN110569972A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110197258A (en) * | 2019-05-29 | 2019-09-03 | 北京市商汤科技开发有限公司 | Neural network searching method, image processing method and device, equipment and medium |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111401517A (en) * | 2020-02-21 | 2020-07-10 | 华为技术有限公司 | Method and device for searching perception network structure |
| CN111401517B (en) * | 2020-02-21 | 2023-11-03 | 华为技术有限公司 | Method and device for searching perceived network structure |
| CN111582454A (en) * | 2020-05-09 | 2020-08-25 | 北京百度网讯科技有限公司 | Method and device for generating neural network model |
| CN111582454B (en) * | 2020-05-09 | 2023-08-25 | 北京百度网讯科技有限公司 | Method and device for generating neural network model |
| CN111680600A (en) * | 2020-05-29 | 2020-09-18 | 北京百度网讯科技有限公司 | Face recognition model processing method, device, equipment and storage medium |
| CN111680600B (en) * | 2020-05-29 | 2023-09-01 | 北京百度网讯科技有限公司 | Face recognition model processing method, device, equipment and storage medium |
| WO2021258964A1 (en) * | 2020-06-22 | 2021-12-30 | 华为技术有限公司 | Neural network architecture search method, apparatus and system |
| CN113902088A (en) * | 2020-06-22 | 2022-01-07 | 华为技术有限公司 | Method, device and system for searching neural network structure |
| CN112116090A (en) * | 2020-09-28 | 2020-12-22 | 腾讯科技（深圳）有限公司 | Neural network structure searching method and device, computer equipment and storage medium |
| CN112116090B (en) * | 2020-09-28 | 2022-08-30 | 腾讯科技（深圳）有限公司 | Neural network structure searching method and device, computer equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110569972A (en) | search space construction method and device of hyper network and electronic equipment | |
| CN111667057B (en) | Method and apparatus for searching model structures | |
| US11574146B2 (en) | Method for updating parameter of model, distributed training system and electric device | |
| CN110826696B (en) | Super-network search space construction method and device and electronic equipment | |
| CN111275190B (en) | Compression method and device of neural network model, image processing method and processor | |
| CN111582454B (en) | Method and device for generating neural network model | |
| CN110580520A (en) | model structure sampling device based on hyper-network and electronic equipment | |
| CN110766089A (en) | Model structure sampling method and device of hyper network and electronic equipment | |
| CN111553844B (en) | Method and device for updating point cloud | |
| CN111563593B (en) | Training method and device for neural network model | |
| CN110569969A (en) | Network model structure sampling method and device and electronic equipment | |
| CN111539224B (en) | Pruning method and device of semantic understanding model, electronic equipment and storage medium | |
| CN115880132B (en) | Graphics processor, matrix multiplication task processing method, device and storage medium | |
| CN111652354B (en) | Method, apparatus, device and storage medium for training super network | |
| CN111882035A (en) | Super network searching method, device, equipment and medium based on convolution kernel | |
| CN110852449A (en) | Model migration method and electronic device | |
| CN111783951B (en) | Model acquisition method, device, equipment and storage medium based on super network | |
| CN110598629B (en) | Super-network search space construction method and device and electronic equipment | |
| CN111325332A (en) | Convolutional neural network processing method and device | |
| CN111680597A (en) | Face recognition model processing method, device, equipment and storage medium | |
| CN110569973A (en) | Network structure searching method and device and electronic equipment | |
| EP3872704A2 (en) | Header model for instance segmentation, instance segmentation model, image segmentation method and apparatus | |
| CN112580723B (en) | Multi-model fusion method, device, electronic equipment and storage medium | |
| CN111553169B (en) | Pruning method and device of semantic understanding model, electronic equipment and storage medium | |
| CN111709514A (en) | Processing method and device of neural network model |
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: 20191213 |
|
| RJ01 | Rejection of invention patent application after publication |