CN111582475A - Data processing method and device based on automatic lightweight neural network - Google Patents
Data processing method and device based on automatic lightweight neural network Download PDFInfo
- Publication number
- CN111582475A CN111582475A CN202010351374.3A CN202010351374A CN111582475A CN 111582475 A CN111582475 A CN 111582475A CN 202010351374 A CN202010351374 A CN 202010351374A CN 111582475 A CN111582475 A CN 111582475A
- Authority
- CN
- China
- Prior art keywords
- data
- neural network
- block
- lightweight neural
- mth
- 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 80
- 238000003672 processing method Methods 0.000 title claims abstract description 24
- 238000000034 method Methods 0.000 claims abstract description 36
- 238000012545 processing Methods 0.000 claims abstract description 28
- 238000001514 detection method Methods 0.000 claims abstract description 13
- 230000011218 segmentation Effects 0.000 claims abstract description 9
- 210000002569 neuron Anatomy 0.000 claims description 37
- 238000012549 training Methods 0.000 claims description 16
- 238000012795 verification Methods 0.000 claims description 13
- 230000006870 function Effects 0.000 claims description 12
- 238000013507 mapping Methods 0.000 claims description 10
- 238000011176 pooling Methods 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 abstract description 12
- 230000008859 change Effects 0.000 abstract description 3
- 238000010586 diagram Methods 0.000 description 16
- 238000004590 computer program Methods 0.000 description 7
- 230000015654 memory Effects 0.000 description 7
- 238000003860 storage Methods 0.000 description 5
- 230000009286 beneficial effect Effects 0.000 description 3
- 238000004422 calculation algorithm Methods 0.000 description 3
- 238000013461 design Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000002787 reinforcement Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000006403 short-term memory Effects 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
- G06N3/08—Learning methods
-
- 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
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 invention relates to a data processing method and a device based on an automatic lightweight neural network, comprising the following steps: acquiring data to be processed; inputting the data to be processed into a pre-established automatic lightweight neural network, and acquiring a processing result of the data to be processed output by the pre-established automatic lightweight neural network; the automatic lightweight neural network in the technical scheme provided by the invention has abundant topological connection forms of the neural network, so that the searched network is more diversified, and various convolutions are matched for use, so that the network is lighter, the receptive field of the network is more flexible, and the identification, detection and segmentation of objects or scenes with larger scale change in the data processing process are facilitated.
Description
Technical Field
The invention relates to the field of computer vision, in particular to a data processing method and device based on an automatic lightweight neural network.
Background
The construction of large deep neural networks usually requires strong expert knowledge and generally consumes a great deal of time and effort of researchers, especially in the field of remote sensing image processing. Meanwhile, these artificially designed networks are usually high in computational complexity and large in memory overhead, and also pose a great challenge to deployment on edge computing devices. Neural Architecture Search (NAS) is an automatic design method specially for a deep neural network, the method can effectively reduce the degree of artificial participation, and the deep neural network is automatically built under the machine view angle.
The current NAS research mainly uses Reinforcement Learning (RL) as an optimization algorithm to select the quality of a neural network structure in a search space (search space). NAS based on reinforcement learning generally uses a Long Short Term Memory (LSTM) as a Controller (Controller), the Controller combines each structure in a search space, and feeds back the performance of the combined model as a Reward (Reward) to the Controller. Since the process is discontinuous, the process is generally optimized by Policy Gradient (Policy Gradient), and after the PG algorithm converges, the final neural network model structure can be derived by the controller, and then the network is finely tuned.
However, the network topology connection form derived by the RL-based NAS is relatively single, and the model has high computational complexity and is not suitable for large-size and multi-scale remote sensing images. Meanwhile, the RL optimization cost is also very large, generally, this method needs to perform parallel operation on several hundred computer Graphics Processing Units (GPUs) for about four weeks, and the search process is also unstable.
Disclosure of Invention
In view of the shortcomings of the prior art, the invention aims to provide a processing method suitable for identifying, detecting and segmenting an object or a scene with large scale change.
The purpose of the invention is realized by adopting the following technical scheme:
in a method of data processing based on an automated lightweight neural network, the improvement comprising:
acquiring data to be processed;
and inputting the data to be processed into a pre-established automatic lightweight neural network, and acquiring a processing result of the data to be processed output by the pre-established automatic lightweight neural network.
Preferably, the data to be processed is remote sensing data.
Preferably, the processing result includes: data recognition results, data detection results and data segmentation results.
Preferably, the process of establishing the pre-established automated lightweight neural network includes:
step 2, setting the number M of blocks of the pre-established automatic lightweight neural network, and numbering the blocks, wherein the 1 st Block is a stem layer;
step 3, setting the number of neurons in each of the blocks from the 2 nd to the Mth as N, and numbering the neurons in each of the blocks from the 2 nd to the Mth from 1 to N;
step 4, selecting a random graph which is depended by each Block from the 2 nd Block to the Mth Block;
step 5, generating a topological relation among the neurons of each Block in the 2 nd to Mth blocks based on the random graphs depended on by the 2 nd to Mth blocks, wherein the neurons of each Block in the 2 nd to Mth blocks meet the condition that the neurons with smaller sequence numbers point to the neurons with larger sequence numbers;
step 6, setting a search space of the neurons of each Block from the 2 nd Block to the Mth Block;
step 7, stacking the 1 st to Mth blocks from small to large according to the numbers, inserting all the blocks from the 2 nd to the Mth to the 1 st after each Block into a pooling layer, introducing a head network corresponding to the processing method after the Mth Block, and finally forming an automatic lightweight neural network initial model;
step 8, the training data is sent into the automatic lightweight neural network initial model in a minipatch form, and learning is carried out in a supervision mode;
and 9, setting the total verification times, verifying the performance index of the trained automatic lightweight neural network initial model by using the verification data after each T times of training, and selecting the automatic lightweight neural network initial model with the maximum performance index as the pre-established automatic lightweight neural network.
Further, the step 4 comprises:
and each of the 2 nd to Mth blocks randomly selects a random graph from a BA random graph, an ER random graph and a WS random graph as a random graph depended by the Block, wherein the probability of selecting the BA random graph, the ER random graph and the WS random graph is the same.
Further, the step 6 comprises:
the neurons in each of the 2 nd to mth blocks select mapping functions belonging to themselves at {3 × 3DWConv,5 × 5DWConv,7 × 7DWConv,9 × 9DWConv }, wherein the probabilities of sampling 3 × 3DWConv,5 × 5DWConv,7 × 7DWConv, and 9 × 9DWConv are all the same, DWConv being a depth separable convolution.
Further, when the processing method is a data identification method, the performance index of the initial model of the automatic lightweight neural network is an OA performance index, when the processing method is a data detection method, the performance index of the initial model of the automatic lightweight neural network is an mAP performance index, and when the processing method is a data segmentation method, the performance index of the initial model of the automatic lightweight neural network is an mF1 performance index.
Based on the same inventive concept, the invention also provides a data processing device based on the automatic lightweight neural network, and the improvement is that the device comprises:
the acquisition module is used for acquiring data to be processed;
and the processing module is used for inputting the data to be processed into a pre-established automatic lightweight neural network and acquiring a processing result of the data to be processed output by the pre-established automatic lightweight neural network.
Compared with the closest prior art, the invention has the following beneficial effects:
according to the technical scheme provided by the invention, firstly, data to be processed is obtained, then the data to be processed is input into a pre-established automatic lightweight neural network, and a processing result of the data to be processed output by the pre-established automatic lightweight neural network is obtained, so that the method and the device can be suitable for identifying, detecting and segmenting objects or scenes with large scale change;
furthermore, the automatic lightweight neural network searches the topological relation of the blocks based on the random graph method, is beneficial to improving the diversity of the automatically searched model, enriches the connection form inside the blocks, can adjust the number of nodes in the blocks, has certain controllability when the model is deployed on edge computing equipment, samples the mapping function of each neuron in a search space formed by deep separable convolution, can make the model lighter, enables the perception field of the neural network to be more flexible due to the matching use of convolution kernels with various sizes, is beneficial to identifying objects or scenes with large scale changes in remote sensing images, and can deepen the depth of the model and improve the final performance of the model by stacking the blocks with topology and mapping.
Drawings
FIG. 1 is a flow chart of a data processing method based on an automated lightweight neural network according to the present invention;
FIG. 2 is a schematic flow chart of an automated lightweight neural network design in an embodiment of the present invention;
FIG. 3 is a schematic diagram of a topological connection relationship of a BA random graph model in the embodiment of the present invention;
FIG. 4 is a schematic diagram of topological connection relationships of an ER random graph model in an embodiment of the present invention;
FIG. 5 is a schematic diagram illustrating a topological connection relationship of a WS random graph model in an embodiment of the present invention;
FIG. 6 is a diagram illustrating the operation of the depth separable convolution according to an embodiment of the present invention;
FIG. 7 is a schematic structural diagram of a data processing apparatus based on an automated lightweight neural network according to the present invention;
Detailed Description
The following describes embodiments of the present invention in further detail with reference to the accompanying drawings.
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
In order to solve the problems that the existing NAS topology connection form based on RL is single, the generated model is high in complexity and is not suitable for remote sensing images, the invention provides a data processing method based on an automatic lightweight neural network, as shown in FIG. 1, the method comprises the following steps:
101, acquiring data to be processed;
102, inputting the data to be processed to a pre-established automatic lightweight neural network, and acquiring a processing result of the data to be processed output by the pre-established automatic lightweight neural network.
Wherein, the data to be processed is remote sensing data, and the processing result comprises: data recognition results, data detection results and data segmentation results.
Further, the process of establishing the pre-established automated lightweight neural network includes:
step 2, setting the number M of blocks of the pre-established automatic lightweight neural network, and numbering the blocks, wherein the 1 st Block is a Stem Layer (Stem Layer) which is used for changing the size of an input image, changing the width and the height of the input image into the former half, simultaneously increasing the number of channels of input data, and changing the channels from 3 channels into C channels;
step 3, setting the number of neurons in each of the blocks from the 2 nd to the Mth as N, and numbering the neurons in each of the blocks from the 2 nd to the Mth from 1 to N;
wherein, the setting of M and N is to satisfy the constraint of edge computing device hardware, such as CPU or GPU performance, memory or storage space size, etc.;
step 4, selecting a random graph which is depended by each Block from the 2 nd Block to the Mth Block;
step 5, generating a topological relation among the neurons of each Block in the 2 nd to Mth blocks based on the random graphs depended on by the 2 nd to Mth blocks, wherein the neurons of each Block in the 2 nd to Mth blocks meet the condition that the neurons with smaller sequence numbers point to the neurons with larger sequence numbers;
step 6, setting a search space of the neurons of each Block from the 2 nd Block to the Mth Block;
step 7, stacking the 1 st to Mth blocks from small to large according to numbers, inserting a pooling layer after the 2 nd to Mth-1 th blocks, introducing a head network corresponding to a processing method after the Mth Block, and finally forming an automatic lightweight neural network initial model, taking data detection processing as an example, introducing a detection head network as shown in FIG. 2;
step 8, the training data is sent into the automatic lightweight neural network initial model in a minipatch form, and learning is carried out in a supervision mode;
furthermore, in the learning process, training data sequentially pass through each Block and Pobing Layer, and the detected target coordinates and the confidence score thereof are output by the detection head network. The cross-over ratio (IoU) of the output detection box and the real label (ground route) is then calculated, and the weight in the network is finely adjusted by using a Stochastic Gradient Descent (SGD) algorithm with a strategy of warmup learning rate (learning rate) under the constraint of the local loss function.
And 9, setting the total verification times, verifying the performance index of the trained automatic lightweight neural network initial model by using the verification data after each T times of training, and selecting the automatic lightweight neural network initial model with the maximum performance index as the pre-established automatic lightweight neural network.
In a preferred embodiment of the present invention, the step 4 includes:
each of the 2 nd to mth blocks randomly selects a random graph from a BA random graph, an ER random graph and a WS random graph as a random graph depended on by the Block, wherein the probabilities of the selection of the BA random graph, the ER random graph and the WS random graph are the same, and schematic diagrams of topological connection relationships of the BA random graph, the ER random graph and the WS random graph are respectively shown in fig. 3, 4 and 5.
Further, the step 6 comprises:
the neurons in each of the 2 nd to mth blocks select mapping functions belonging to the neurons at {3 × 3DWConv,5 × 5DWConv,7 × 7DWConv,9 × 9DWConv }, wherein the probabilities of sampling 3 × 3DWConv,5 × 5DWConv,7 × 7DWConv and 9 × 9DWConv are all the same, DWConv is depth separable convolution, and a schematic diagram of the calculation process of 3 × 3DWConv convolution is shown in fig. 6.
When the processing method is a data identification method, the performance index of the initial model of the automatic lightweight neural network is an OA performance index, when the processing method is a data detection method, the performance index of the initial model of the automatic lightweight neural network is an mAP performance index, and when the processing method is a data segmentation method, the performance index of the initial model of the automatic lightweight neural network is an mF1 performance index.
In the network generation process, the topological connection relation between each neuron of each Block in the deep neural network is constructed on the basis of a classical random graph (random graph) model, the setting of artificial rules is reduced, and the connection form of the model is enriched. This example employs three random graph models: BA random graph, ER random graph, WS random graph. The number of nodes in the random graph is related to the complexity of the finally generated model, and the model is more complex when the number of nodes is larger. The number should be set in conjunction with the actual hardware conditions of the edge computing device, such as CPU or GPU performance, memory or storage size, etc. The classical random Graph model is an undirected Graph structure, while neural networks are generally viewed as a Directed Acyclic Graph (DAG) structure. Here, in the embodiment of the present application, each node in Block is numbered, and it is specified that the connection direction can only be pointed to a node with a larger sequence number by a node with a smaller sequence number, so that confusion caused by forward propagation and backward propagation due to a searched network forming a loop is effectively avoided.
Meanwhile, the embodiment provided by the invention performs operation sampling (operation sampling) on the constructed network containing the topological connection relation, and perfects Block. After the topological connection relationship is determined, each neuron of the neural network also needs to map (mapping) data flowing in the network. The mapping function of each neuron is sampled according to some probability distribution in a search space (search space) that includes all the required mappings. In consideration of a lightweight structure and a large-size multi-scale remote sensing image, the present embodiment takes depth separable convolution (DWConv) of different sizes as a search space. The deep separable convolution is a special form of grouped convolution, the number of channels of a convolution kernel is equal to that of a feature map, and the convolution kernel of each channel only performs convolution operation with the feature map of the corresponding channel and does not perform any operation with the feature maps of other channels. DWConv not only has good expression capacity for data characteristics, but also can effectively reduce Parameters (Parameters) and floating point operations (FLOPs) of the constructed neural network, and the convolution in the form is used as a search space, so that the finally derived network model is lighter, and the deployment and application on edge computing equipment are very facilitated. Meanwhile, DWConv with various sizes can be used together in a matched mode, and the receptive field (receptive field) of the neural network can be adjusted, so that the finally generated model is more robust to recognition of objects or scenes with large scale changes in the remote sensing image.
And finally, stacking the obtained network blocks with the topological connection relation and the mapping function, removing the first Block and the last Block, inserting a Pooling Layer (Pooling Layer) into the rest blocks, and changing the size of the characteristic diagram extracted by each Block. And finally, aiming at different tasks, introducing different head networks at the top end of the network to generate a final model structure, and then finely adjusting the generated network through iterative training on training data.
Based on the same inventive concept, the present invention further provides a data processing apparatus based on an automated lightweight neural network, as shown in fig. 7, the apparatus includes:
the acquisition module is used for acquiring data to be processed;
and the processing module is used for inputting the data to be processed into a pre-established automatic lightweight neural network and acquiring a processing result of the data to be processed output by the pre-established automatic lightweight neural network.
Preferably, the data to be processed is remote sensing data.
Preferably, the processing result includes: data recognition results, data detection results and data segmentation results.
Preferably, the process of establishing the pre-established automated lightweight neural network includes:
step 2, setting the number M of blocks of the pre-established automatic lightweight neural network, and numbering the blocks, wherein the 1 st Block is a stem layer;
step 3, setting the number of neurons in each of the blocks from the 2 nd to the Mth as N, and numbering the neurons in each of the blocks from the 2 nd to the Mth from 1 to N;
step 4, selecting a random graph which is depended by each Block from the 2 nd Block to the Mth Block;
step 5, generating a topological relation among the neurons of each Block in the 2 nd to Mth blocks based on the random graphs depended on by the 2 nd to Mth blocks, wherein the neurons of each Block in the 2 nd to Mth blocks meet the condition that the neurons with smaller sequence numbers point to the neurons with larger sequence numbers;
step 6, setting a search space of the neurons of each Block from the 2 nd Block to the Mth Block;
step 7, stacking the 1 st to Mth blocks from small to large according to the numbers, inserting all the blocks from the 2 nd to the Mth to the 1 st after each Block into a pooling layer, introducing a head network corresponding to the processing method after the Mth Block, and finally forming an automatic lightweight neural network initial model;
step 8, the training data is sent into the automatic lightweight neural network initial model in a minipatch form, and learning is carried out in a supervision mode;
and 9, setting the total verification times, verifying the performance index of the trained automatic lightweight neural network initial model by using the verification data after each T times of training, and selecting the automatic lightweight neural network initial model with the maximum performance index as the pre-established automatic lightweight neural network.
Further, the step 4 comprises:
and each of the 2 nd to Mth blocks randomly selects a random graph from a BA random graph, an ER random graph and a WS random graph as a random graph depended by the Block, wherein the probability of selecting the BA random graph, the ER random graph and the WS random graph is the same.
Further, the step 6 comprises:
the neurons in each of the 2 nd to mth blocks select mapping functions belonging to themselves at {3 × 3DWConv,5 × 5DWConv,7 × 7DWConv,9 × 9DWConv }, wherein the probabilities of sampling 3 × 3DWConv,5 × 5DWConv,7 × 7DWConv, and 9 × 9DWConv are all the same, DWConv being a depth separable convolution.
Further, when the processing method is a data identification method, the performance index of the initial model of the automatic lightweight neural network is an OA performance index, when the processing method is a data detection method, the performance index of the initial model of the automatic lightweight neural network is an mAP performance index, and when the processing method is a data segmentation method, the performance index of the initial model of the automatic lightweight neural network is an mF1 performance index.
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.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solutions of the present invention and not for limiting the same, and although the present invention is described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications and equivalents may be made to the embodiments of the invention without departing from the spirit and scope of the invention, which is to be covered by the claims.
Claims (8)
1. A data processing method based on an automatic lightweight neural network is characterized by comprising the following steps:
acquiring data to be processed;
and inputting the data to be processed into a pre-established automatic lightweight neural network, and acquiring a processing result of the data to be processed output by the pre-established automatic lightweight neural network.
2. The method of claim 1, wherein the data to be processed is remotely sensed data.
3. The method of claim 1, wherein the processing results comprise: data recognition results, data detection results and data segmentation results.
4. The method of claim 1, wherein the pre-established automated lightweight neural network establishment procedure comprises:
step 1, acquiring training set data and verification set data corresponding to data to be processed;
step 2, setting the number M of blocks of the pre-established automatic lightweight neural network, and numbering the blocks, wherein the 1 st Block is a stem layer;
step 3, setting the number of neurons in each of the blocks from the 2 nd to the Mth as N, and numbering the neurons in each of the blocks from the 2 nd to the Mth from 1 to N;
step 4, selecting a random graph which is depended by each Block from the 2 nd Block to the Mth Block;
step 5, generating a topological relation among the neurons of each Block in the 2 nd to Mth blocks based on the random graphs depended on by the 2 nd to Mth blocks, wherein the neurons of each Block in the 2 nd to Mth blocks meet the condition that the neurons with smaller sequence numbers point to the neurons with larger sequence numbers;
step 6, setting a search space of the neurons of each Block from the 2 nd Block to the Mth Block;
step 7, stacking the 1 st to Mth blocks from small to large according to the numbers, inserting all the blocks from the 2 nd to the Mth to the 1 st after each Block into a pooling layer, introducing a head network corresponding to the processing method after the Mth Block, and finally forming an automatic lightweight neural network initial model;
step 8, the training data is sent into the automatic lightweight neural network initial model in a minipatch form, and learning is carried out in a supervision mode;
and 9, setting the total verification times, verifying the performance index of the trained automatic lightweight neural network initial model by using the verification data after each T times of training, and selecting the automatic lightweight neural network initial model with the maximum performance index as the pre-established automatic lightweight neural network.
5. The method of claim 4, wherein step 4 comprises:
and each of the 2 nd to Mth blocks randomly selects a random graph from a BA random graph, an ER random graph and a WS random graph as a random graph depended by the Block, wherein the probability of selecting the BA random graph, the ER random graph and the WS random graph is the same.
6. The method of claim 4, wherein the step 6 comprises:
the neurons in each of the 2 nd to mth blocks select mapping functions belonging to themselves at {3 × 3DWConv,5 × 5DWConv,7 × 7DWConv,9 × 9DWConv }, wherein the probabilities of sampling 3 × 3DWConv,5 × 5DWConv,7 × 7DWConv, and 9 × 9DWConv are all the same, DWConv being a depth separable convolution.
7. The method of claim 4, wherein the performance index of the initial model of the automated lightweight neural network is an OA performance index when the processing method is a data recognition method, an mAP performance index when the processing method is a data detection method, and an mF1 performance index when the processing method is a data segmentation method.
8. A data processing apparatus based on an automated lightweight neural network, the apparatus comprising:
the acquisition module is used for acquiring data to be processed;
and the processing module is used for inputting the data to be processed into a pre-established automatic lightweight neural network and acquiring a processing result of the data to be processed output by the pre-established automatic lightweight neural network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010351374.3A CN111582475A (en) | 2020-04-28 | 2020-04-28 | Data processing method and device based on automatic lightweight neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010351374.3A CN111582475A (en) | 2020-04-28 | 2020-04-28 | Data processing method and device based on automatic lightweight neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111582475A true CN111582475A (en) | 2020-08-25 |
Family
ID=72126227
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010351374.3A Pending CN111582475A (en) | 2020-04-28 | 2020-04-28 | Data processing method and device based on automatic lightweight neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111582475A (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104820840A (en) * | 2015-05-08 | 2015-08-05 | 西安电子科技大学 | Nearest neighborhood hyper-spectral image classification method based on dictionary and band restructuring |
| CN105809185A (en) * | 2015-12-31 | 2016-07-27 | 天津大学 | High-spectrum image nonlinear demixing method based on neural network and differential search |
| CN106991397A (en) * | 2017-03-31 | 2017-07-28 | 中国科学院电子学研究所 | View-based access control model conspicuousness constrains the remote sensing images detection method of depth confidence network |
| CN109299142A (en) * | 2018-11-14 | 2019-02-01 | 中山大学 | A kind of convolutional neural networks search structure method and system based on evolution algorithm |
| CN109948029A (en) * | 2019-01-25 | 2019-06-28 | 南京邮电大学 | Based on the adaptive depth hashing image searching method of neural network |
-
2020
- 2020-04-28 CN CN202010351374.3A patent/CN111582475A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104820840A (en) * | 2015-05-08 | 2015-08-05 | 西安电子科技大学 | Nearest neighborhood hyper-spectral image classification method based on dictionary and band restructuring |
| CN105809185A (en) * | 2015-12-31 | 2016-07-27 | 天津大学 | High-spectrum image nonlinear demixing method based on neural network and differential search |
| CN106991397A (en) * | 2017-03-31 | 2017-07-28 | 中国科学院电子学研究所 | View-based access control model conspicuousness constrains the remote sensing images detection method of depth confidence network |
| CN109299142A (en) * | 2018-11-14 | 2019-02-01 | 中山大学 | A kind of convolutional neural networks search structure method and system based on evolution algorithm |
| CN109948029A (en) * | 2019-01-25 | 2019-06-28 | 南京邮电大学 | Based on the adaptive depth hashing image searching method of neural network |
Non-Patent Citations (3)
| Title |
|---|
| CHENXI LIU等: "《Auto-DeepLab: Hierarchical Neural Architecture Search for Semantic Image Segmentation》", 《2019 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR)》 * |
| SAINING XIE等: "《Exploring Randomly Wired Neural Networks for Image Recognition》", 《2019 IEEE/CVF INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV)》 * |
| SIGAI: "《NAS（神经结构搜索）综述》", 《HTTPS://ZHUANLAN.ZHIHU.COM/P/60414004》 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110298404B (en) | Target tracking method based on triple twin Hash network learning | |
| CN109919108B (en) | Remote sensing image rapid target detection method based on deep hash auxiliary network | |
| CN108009525B (en) | A kind of specific objective recognition methods over the ground of the unmanned plane based on convolutional neural networks | |
| CN110633745B (en) | Image classification training method and device based on artificial intelligence and storage medium | |
| CN113110482B (en) | Indoor environment robot exploration method and system based on priori information heuristic method | |
| CN109902192B (en) | Remote sensing image retrieval method, system, equipment and medium based on unsupervised depth regression | |
| CN108805149A (en) | A kind of winding detection method and device of visual synchronization positioning and map structuring | |
| EP3671555A1 (en) | Object shape regression using wasserstein distance | |
| CN115294407A (en) | Model compression method and system based on preview mechanism knowledge distillation | |
| CN117292283B (en) | Target identification method based on unmanned aerial vehicle | |
| CN114881225A (en) | Power transmission and transformation inspection model network structure searching method, system and storage medium | |
| CN109948457A (en) | The real time target recognitio algorithm accelerated based on convolutional neural networks and CUDA | |
| CN116189281B (en) | End-to-end human behavior classification method and system based on space-time self-adaptive fusion | |
| CN117744759A (en) | Text information identification method and device, storage medium and electronic equipment | |
| CN118334323B (en) | Insulator detection method and system based on ultraviolet image | |
| CN113762049B (en) | Content identification method, content identification device, storage medium and terminal equipment | |
| CN113762019B (en) | Training method of feature extraction network, face recognition method and device | |
| CN113888597A (en) | Target tracking identification method and system based on lightweight target tracking network | |
| CN111582475A (en) | Data processing method and device based on automatic lightweight neural network | |
| CN117437467A (en) | Model training method and device, electronic equipment and storage medium | |
| CN115457269A (en) | Semantic segmentation method based on improved DenseNAS | |
| CN115661123A (en) | Industrial product surface defect position detection method based on weak supervision target detection | |
| CN115620238A (en) | Park pedestrian attribute identification method based on multivariate information fusion | |
| CN114067155A (en) | Image classification method, device, product and storage medium based on meta learning | |
| CN114038058A (en) | Parallel human body posture detection tracking method based on posture guiding re-recognition features |
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: 20200825 |
|
| RJ01 | Rejection of invention patent application after publication |