CN112101538B - Graphic neural network hardware computing system and method based on memory computing - Google Patents
Graphic neural network hardware computing system and method based on memory computing Download PDFInfo
- Publication number
- CN112101538B CN112101538B CN202011011776.5A CN202011011776A CN112101538B CN 112101538 B CN112101538 B CN 112101538B CN 202011011776 A CN202011011776 A CN 202011011776A CN 112101538 B CN112101538 B CN 112101538B
- Authority
- CN
- China
- Prior art keywords
- calculation
- module
- memory
- data
- control unit
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 230000015654 memory Effects 0.000 title claims abstract description 69
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 34
- 238000000034 method Methods 0.000 title claims abstract description 15
- 238000004364 calculation method Methods 0.000 claims description 150
- 239000011159 matrix material Substances 0.000 claims description 47
- 230000005540 biological transmission Effects 0.000 claims description 18
- 230000004913 activation Effects 0.000 claims description 10
- 239000013598 vector Substances 0.000 claims description 9
- 230000006870 function Effects 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 6
- 238000010586 diagram Methods 0.000 description 7
- 230000008569 process Effects 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 4
- 230000009471 action Effects 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 230000001788 irregular Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012549 training Methods 0.000 description 1
- 238000013519 translation Methods 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
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Neurology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a graph neural network hardware computing system and a method based on memory computing, which relate to the technical field of memory computing hardware and comprise two memory computing modules, two MUX modules, one DMUX module, one control unit and one Relu module.
Description
Technical Field
The invention relates to the technical field of memory computing hardware, in particular to a graphic neural network hardware computing system and method based on memory computing.
Background
Over the last decade, with the rapid development of computing resources and the availability of large amounts of training data, deep learning and neural networks have rapidly emerged and evolved. Many machine learning tasks (such as object detection, machine translation, and speech recognition) that once heavily relied on manually extracted features have been thoroughly replaced today by various end-to-end deep learning neural network models (e.g., convolutional Neural Networks (CNNs), long short term memory models (LSTM), and automatic encoders).
Although the conventional neural network method has been applied to great success in extracting features of euclidean space data, data in many practical application scenarios are generated from non-euclidean space, and the performance of the conventional neural network method in processing non-euclidean space data is still unsatisfactory. For example, in electronic commerce, a Graph-based learning system can make very accurate recommendations using interactions between users and products, but the complexity of the Graph makes existing deep learning algorithms very challenging to handle. Because the graphs are irregular, each graph has a variable size out-of-order node, and each node in the graph has a different number of neighbors, which results in some important operations such as convolution that are easy to calculate on the image, but difficult for the graph. Furthermore, one core assumption of existing deep learning algorithms is that the data samples are independent of each other. However, this is not the case for graphs where each data sample (node) in the graph has edges that are related to other real data samples (nodes) in the graph, and this information can be used to capture the inter-dependencies between instances. In recent years, researchers have studied the ideas of convolutional networks, cyclic networks and depth automatic encoders, and have defined and designed the structure of the graphic neural network specially used for processing graphic data, this structure realizes the convolution operation on the topological graph by means of spectrogram theory (Spectral Graph Theory), the specific method is that the topological graph structure in the space domain is mapped into the frequency domain through Fourier transformation and convolved, and then the inverse transformation is utilized to return to the space domain, thus completing the graph convolution operation.
The traditional graph neural network calculation is completed on a traditional PC, data are stored in a magnetic disk, required data are loaded into a memory when calculation is performed, the calculation is completed by a CPU, and in the process, whether Laplacian matrix calculation or convolution calculation is performed, a large number of memory access operations are required, and the memory access speed becomes a bottleneck of calculation. In addition, since all the calculation operations and the control of the whole calculation flow are completed by the external CPU, a large amount of calculation resources are occupied, so that the CPU cannot complete other tasks. Therefore, how to increase the access speed and release the power of the external computing unit becomes an urgent problem to be solved.
Disclosure of Invention
The invention aims at: in order to solve the bottleneck problem of the conventional graph neural network calculation, the calculation system realizes hardware mapping of graph convolution operation by means of a memory calculation module, integrates calculation and storage of data into one, directly completes all operations in a storage unit, and controls the whole calculation flow by an internal control unit, thereby completely eliminating memory operation, greatly saving the calculation force of an external ALU, greatly improving the overall calculation efficiency of the system and reducing the operation power consumption.
The technical scheme adopted by the invention is as follows:
the invention relates to a graphic neural network hardware computing system based on memory computation, which comprises two memory computing modules, two MUXs, a DMUX, a control unit module and a Relu module,
the two memory calculation modules are respectively used for completing Laplace matrix calculation and neural network convolution calculation;
the two MUX modules respectively select the input data of the two memory calculation modules according to the instruction of the control unit;
the DMUX module selects the direction of outputting the data according to the instruction of the control unit, and outputs the data to the outside or transmits the data to the MUX module for the next round of calculation;
the Relu module performs activation processing on the calculation result of the memory calculation module to obtain a final result of one-layer neural network calculation, and then sends the result to the subsequent DMUX module;
the control unit is used for generating control signals of all other modules and controlling the whole calculation flow.
Further, the internal structures of the two memory calculation modules are the same, and the memory calculation modules comprise a storage unit array, a mode control module and a read-write transmission module;
the memory cell array is composed of basic memory cells and is used for storing data and calculating data;
the mode control module receives an instruction of an external control unit, activates the control unit and controls two working modes of the control unit, namely a Laplace matrix calculation mode and a convolution calculation mode;
the read-write transmission module is used for caching the input and output data of the storage unit, inputting or outputting the data to the storage unit, writing the external input data into the storage unit, inputting the external input data into the storage unit for calculation, outputting the calculation result to the subsequent module or reading the data in the storage unit outwards.
Further, the storage unit is RRAM, MRAM or Flash.
The graph neural network hardware computing method based on memory computing comprises the computing system and further comprises the following working procedures:
step 1, determining external graph data in a graph adjacency matrix form and feature vectors of graph nodes;
step 2, global reset, and initializing all modules in the system;
step 3, writing the neural network weight into a memory calculation module;
step 4, inputting the graph adjacent matrix into the system, and generating a control signal by the control unit to control the MUX module to input the input data into a read-write transmission module in a memory calculation module for Laplace matrix calculation;
step 5, the control unit generates a control signal of the memory calculation module in step 4, the control mode control module sets the storage unit to be in a Laplace matrix calculation mode, and activates the corresponding storage unit so as to perform Laplace matrix calculation and store calculation results;
step 6, inputting the characteristic vector of the graph node into a system, and controlling a MUX module to input the input data into a read-write transmission module in a memory calculation module for Laplace matrix calculation by a control unit to generate a control signal;
step 7, the control unit generates a control signal of the memory calculation module in step 6, the control mode control module sets the storage unit to be in a matrix multiplication calculation mode, so that matrix multiplication calculation is carried out on the Laplace matrix and input data, and a calculation result is output to a subsequent MUX module through the read-write transmission module;
step 8, the control unit controls the MUX module in step 7 to input the read-write transmission module of the memory calculation module for convolution calculation;
step 9, the control unit generates a control signal of the memory calculation module in step 8, the control mode control module sets the storage unit as matrix multiplication calculation, so as to perform convolution calculation of the weight matrix and input data, and output the result data to the subsequent Relu module;
step 10, the Relu module uses the Relu activation function to activate the convolution calculation result in step 9, and outputs the generated result to the subsequent DMUX module;
step 11, the control unit generates a DMUX control signal, selects and outputs a calculation result, if the calculation operation is not finished, the control unit transmits the input to the MUX module, and then jumps to step 6, otherwise, the control unit executes step 12;
and step 12, after the calculation is completed, the system outputs a final result, all buffer data are cleared, all modules in the system are reset, and the whole operation flow is finished.
In summary, by adopting the technical scheme, the invention has the beneficial effects that:
1. compared with the traditional computer architecture, the hardware computing system provided by the invention transfers all operations into the storage unit, and the whole computing flow is completed by the internal control unit, thereby completely eliminating memory access operation, greatly releasing the computing power of the external computing unit, greatly improving the overall computing efficiency of the system and reducing the operation power consumption.
Drawings
For a clearer description of the technical solutions of the embodiments of the present invention, the drawings that are needed in the embodiments will be briefly described below, it should be understood that the following drawings only illustrate some embodiments of the present invention, and should not be considered as limiting the scope, for those skilled in the art, without performing creative efforts, other related drawings may be obtained according to the drawings, where the proportional relationships of the components in the drawings in the present specification do not represent the proportional relationships in actual material selection design, and are merely schematic diagrams of structures or positions, where:
FIG. 1 is a schematic diagram of a hardware computing system architecture of the present invention;
FIG. 2 is a schematic diagram illustrating an internal structure of a memory computing module according to the present invention;
FIG. 3 is a schematic diagram of the neural network architecture of the present invention for computation;
FIG. 4 is a schematic diagram of the data flow of the present invention in practice;
fig. 5 is a flowchart of a computing method of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the particular embodiments described herein are illustrative only and are not intended to limit the invention, i.e., the embodiments described are merely some, but not all, of the embodiments of the invention. The components of the embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations.
The present invention will be described in detail with reference to the accompanying drawings.
Example 1
As shown in fig. 1, the present invention is a memory-calculation-based hardware computing system for a neural network, comprising two memory calculation modules, two muxes, a DMUX, a control unit module and a Relu module,
the two memory calculation modules are respectively used for completing Laplace matrix calculation and neural network convolution calculation;
the two MUX modules respectively select the input data of the two memory calculation modules according to the instruction of the control unit;
the DMUX module selects the direction of outputting the data according to the instruction of the control unit, and outputs the data to the outside or transmits the data to the MUX module for the next round of calculation;
the Relu module performs activation processing on the calculation result of the memory calculation module to obtain a final result of one-layer neural network calculation, and then sends the result to the subsequent DMUX module;
the control unit is used for generating control signals of all other modules and controlling the whole calculation flow.
As shown in fig. 2, the internal structures of the two memory calculation modules are the same, and the memory calculation module comprises a storage unit array, a mode control module and a read-write transmission module;
the memory cell array is composed of basic memory cells and is used for storing data and calculating data;
the mode control module receives an instruction of an external control unit, activates the control unit and controls two working modes of the control unit, namely a Laplace matrix calculation mode and a convolution calculation mode;
the read-write transmission module is used for caching the input and output data of the storage unit, inputting or outputting the data to the storage unit, writing the external input data into the storage unit, inputting the external input data into the storage unit for calculation, outputting the calculation result to the subsequent module or reading the data in the storage unit outwards.
The storage unit is RRAM, MRAM or Flash.
In this embodiment, the two memory calculation modules are respectively configured to complete laplace matrix calculation and neural network convolution calculation. Each memory calculation module has two calculation modes, wherein the Laplace memory calculation module calculates the Laplace matrix of the input image before calculation starts, and then performs matrix multiplication operation on the input data and the Laplace matrix in the operation process to generate data used as convolution; the convolution calculation module stores the weight data of the neural network before calculation starts, and then performs matrix multiplication operation on the data and the weights in the calculation process, so that a convolution operation result of the input data is finally generated.
Fig. 3 is a schematic diagram of a schematic structure of a neural network for calculation according to the present invention, where the neural network is mainly divided into two parts, i.e., a hidden layer (hidden layer) and an activation function layer (relu), the hidden layer is mainly used for performing convolution operation on output data of a previous layer and weights of the previous layer, and the activation function layer is used for performing activation processing on a convolution operation result. The input data of the graph neural network is graph information, the computation of a layer of neural network is completed through one convolution operation and one activation function operation, and the final graph information is output after the computation of all layers is completed.
FIG. 4 is a schematic diagram of a data flow in practical application of the present invention, where the hardware computing system provided by the present invention can process all data that can be represented as a graph, such as a user-project graph, a scene graph, a point cloud graph, a traffic road graph, a molecular structure graph, and so on. However, the graph is abstract data, and cannot be directly calculated, so that the adjacency matrix of the graph and the feature vector of the graph node are obtained, and then the feature vector can be input into the hardware computing system for calculation, and final graph information such as user-project relation prediction, scene generation, target recognition, action recognition, semantic segmentation and the like is obtained after calculation.
Example two
This example is a further illustration of the present invention.
The method for calculating the hardware of the neural network of the graph based on the memory calculation comprises a calculation system of the first embodiment, and further comprises the following workflow:
step 1, determining external graph data in a graph adjacency matrix form and feature vectors of graph nodes;
step 2, global reset, and initializing all modules in the system;
step 3, writing the neural network weight into a memory calculation module;
step 4, inputting the graph adjacent matrix into the system, and generating a control signal by the control unit to control the MUX module to input the input data into a read-write transmission module in a memory calculation module for Laplace matrix calculation;
step 5, the control unit generates a control signal of the memory calculation module in step 4, the control mode control module sets the storage unit to be in a Laplace matrix calculation mode, and activates the corresponding storage unit so as to perform Laplace matrix calculation and store calculation results;
step 6, inputting the characteristic vector of the graph node into a system, and controlling a MUX module to input the input data into a read-write transmission module in a memory calculation module for Laplace matrix calculation by a control unit to generate a control signal;
step 7, the control unit generates a control signal of the memory calculation module in step 6, the control mode control module sets the storage unit to be in a matrix multiplication calculation mode, so that matrix multiplication calculation is carried out on the Laplace matrix and input data, and a calculation result is output to a subsequent MUX module through the read-write transmission module;
step 8, the control unit controls the MUX module in step 7 to input the read-write transmission module of the memory calculation module for convolution calculation;
step 9, the control unit generates a control signal of the memory calculation module in step 8, the control mode control module sets the storage unit as matrix multiplication calculation, so as to perform convolution calculation of the weight matrix and input data, and output the result data to the subsequent Relu module;
step 10, the Relu module uses the Relu activation function to activate the convolution calculation result in step 9, and outputs the generated result to the subsequent DMUX module;
step 11, the control unit generates a DMUX control signal, selects and outputs a calculation result, if the calculation operation is not finished, the control unit transmits the input to the MUX module, and then jumps to step 6, otherwise, the control unit executes step 12;
and step 12, after the calculation is completed, the system outputs a final result, all buffer data are cleared, all modules in the system are reset, and the whole operation flow is finished.
As shown in fig. 5, the whole calculation process of the hardware calculation system is simply divided into 5 steps:
firstly, writing weights of a graph neural network into a memory computing module, then converting input data into a graph form, such as a user-project relationship graph, a scene graph, a point cloud relationship graph, a traffic road relationship graph, a molecular structure graph and the like, then obtaining a feature vector of a graph adjacent matrix and a graph node, inputting the data into the memory computing system for computation, and finally obtaining a final computing result, such as user-project relationship prediction, scene generation, target recognition, action recognition, semantic segmentation and the like.
The above description is only a preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that are not creatively contemplated by those skilled in the art within the technical scope of the present invention should be included in the scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope defined by the claims.
Claims (3)
1. The graphic neural network hardware computing system based on memory computation comprises two memory computing modules, two MUX modules, one DMUX module, one control unit and one Relu module, and is characterized in that:
the two memory calculation modules are respectively used for completing Laplace matrix calculation and neural network convolution calculation;
the two MUX modules respectively select the input data of the two memory calculation modules according to the instruction of the control unit;
the DMUX module selects the direction of outputting the data according to the instruction of the control unit, and outputs the data to the outside or transmits the data to the MUX module for the next round of calculation;
the Relu module performs activation processing on the calculation result of the memory calculation module to obtain a final result of one-layer neural network calculation, and then sends the result to the subsequent DMUX module;
the control unit is used for generating control signals of all other modules and controlling the whole calculation flow;
the internal structures of the two memory calculation modules are the same, and the memory calculation modules comprise a storage unit array, a mode control module and a read-write transmission module;
the memory cell array is composed of basic memory cells and is used for storing data and calculating data;
the mode control module receives an instruction of an external control unit, activates the control unit and controls two working modes of the control unit, namely a Laplace matrix calculation mode and a convolution calculation mode;
the read-write transmission module is used for caching the input and output data of the storage unit, inputting or outputting the data to the storage unit, writing the external input data into the storage unit, inputting the external input data into the storage unit for calculation, outputting the calculation result to the subsequent module or reading the data in the storage unit outwards.
2. The memory-computation-based graphics neural network hardware computing system of claim 1, wherein: the storage unit is RRAM, MRAM or Flash.
3. The method for calculating the graphic neural network hardware based on the memory calculation comprises a calculation system according to any one of claims 1-2, and further comprises the following workflow:
step 1, determining external graph data in a graph adjacency matrix form and feature vectors of graph nodes;
step 2, global reset, and initializing all modules in the system;
step 3, writing the neural network weight into a memory calculation module;
step 4, inputting the graph adjacent matrix into the system, and generating a control signal by the control unit to control the MUX module to input the input data into a read-write transmission module in a memory calculation module for Laplace matrix calculation;
step 5, the control unit generates a control signal of the memory calculation module in step 4, the control mode control module sets the storage unit to be in a Laplace matrix calculation mode, and activates the corresponding storage unit so as to perform Laplace matrix calculation and store calculation results;
step 6, inputting the characteristic vector of the graph node into a system, and controlling a MUX module to input the input data into a read-write transmission module in a memory calculation module for Laplace matrix calculation by a control unit to generate a control signal;
step 7, the control unit generates a control signal of the memory calculation module in step 6, the control mode control module sets the storage unit to be in a matrix multiplication calculation mode, so that matrix multiplication calculation is carried out on the Laplace matrix and input data, and a calculation result is output to a subsequent MUX module through the read-write transmission module;
step 8, the control unit controls the MUX module in step 7 to input the read-write transmission module of the memory calculation module for convolution calculation;
step 9, the control unit generates a control signal of the memory calculation module in step 8, the control mode control module sets the storage unit as matrix multiplication calculation, so as to perform convolution calculation of the weight matrix and input data, and output the result data to the subsequent Relu module;
step 10, the Relu module uses the Relu activation function to activate the convolution calculation result in step 9, and outputs the generated result to the subsequent DMUX module;
step 11, the control unit generates a DMUX control signal, selects and outputs a calculation result, if the calculation operation is not finished, the control unit transmits the input to the MUX module, and then jumps to step 6, otherwise, the control unit executes step 12;
and step 12, after the calculation is completed, the system outputs a final result, all buffer data are cleared, all modules in the system are reset, and the whole operation flow is finished.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011011776.5A CN112101538B (en) | 2020-09-23 | 2020-09-23 | Graphic neural network hardware computing system and method based on memory computing |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011011776.5A CN112101538B (en) | 2020-09-23 | 2020-09-23 | Graphic neural network hardware computing system and method based on memory computing |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112101538A CN112101538A (en) | 2020-12-18 |
CN112101538B true CN112101538B (en) | 2023-11-17 |
Family
ID=73756004
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202011011776.5A Active CN112101538B (en) | 2020-09-23 | 2020-09-23 | Graphic neural network hardware computing system and method based on memory computing |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112101538B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113657577B (en) * | 2021-07-21 | 2023-08-18 | 阿里巴巴达摩院(杭州)科技有限公司 | Model training method and computing system |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106210593A (en) * | 2016-08-19 | 2016-12-07 | 京东方科技集团股份有限公司 | Display control unit, display control method and display device |
CN109032781A (en) * | 2018-07-13 | 2018-12-18 | 重庆邮电大学 | A kind of FPGA parallel system of convolutional neural networks algorithm |
CN109102065A (en) * | 2018-06-28 | 2018-12-28 | 广东工业大学 | A kind of convolutional neural networks accelerator based on PSoC |
CN209231976U (en) * | 2018-12-29 | 2019-08-09 | 南京宁麒智能计算芯片研究院有限公司 | A kind of accelerator of restructural neural network algorithm |
CN111126668A (en) * | 2019-11-28 | 2020-05-08 | 中国人民解放军国防科技大学 | Spark operation time prediction method and device based on graph convolution network |
-
2020
- 2020-09-23 CN CN202011011776.5A patent/CN112101538B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106210593A (en) * | 2016-08-19 | 2016-12-07 | 京东方科技集团股份有限公司 | Display control unit, display control method and display device |
CN109102065A (en) * | 2018-06-28 | 2018-12-28 | 广东工业大学 | A kind of convolutional neural networks accelerator based on PSoC |
CN109032781A (en) * | 2018-07-13 | 2018-12-18 | 重庆邮电大学 | A kind of FPGA parallel system of convolutional neural networks algorithm |
CN209231976U (en) * | 2018-12-29 | 2019-08-09 | 南京宁麒智能计算芯片研究院有限公司 | A kind of accelerator of restructural neural network algorithm |
CN111126668A (en) * | 2019-11-28 | 2020-05-08 | 中国人民解放军国防科技大学 | Spark operation time prediction method and device based on graph convolution network |
Non-Patent Citations (1)
Title |
---|
基于GPU的深度神经网络优化方法研究;陈一鸣;《中国优秀硕士学位论文全文数据库》;I138-1193 * |
Also Published As
Publication number | Publication date |
---|---|
CN112101538A (en) | 2020-12-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2018171717A1 (en) | Automated design method and system for neural network processor | |
CN110633153A (en) | Method for realizing neural network model splitting by using multi-core processor and related product | |
WO2022027937A1 (en) | Neural network compression method, apparatus and device, and storage medium | |
CN112651438A (en) | Multi-class image classification method and device, terminal equipment and storage medium | |
CN110826708B (en) | Method for realizing neural network model splitting by using multi-core processor and related product | |
JP2020187736A (en) | Learning data generation method for classifier learning having regional features, and system thereof | |
CN112163601A (en) | Image classification method, system, computer device and storage medium | |
CN116450486B (en) | Modeling method, device, equipment and medium for nodes in multi-element heterogeneous computing system | |
TW202138999A (en) | Data dividing method and processor for convolution operation | |
CN115457492A (en) | Target detection method and device, computer equipment and storage medium | |
CN112101538B (en) | Graphic neural network hardware computing system and method based on memory computing | |
CN116401552A (en) | Classification model training method and related device | |
CN112990433B (en) | Model time consumption prediction method and device, electronic equipment and storage medium | |
CN114330686A (en) | Configurable convolution processing device and convolution calculation method | |
CN109389215B (en) | Network structure determination method and device of deep learning network | |
US20220101120A1 (en) | Interpretable visualization system for graph neural network | |
CN112200310B (en) | Intelligent processor, data processing method and storage medium | |
Chen et al. | Research on object detection algorithm based on improved Yolov5 | |
CN116797973A (en) | Data mining method and system applied to sanitation intelligent management platform | |
Ying et al. | EdgePC: Efficient deep learning analytics for point clouds on edge devices | |
Gaihua et al. | Instance segmentation convolutional neural network based on multi-scale attention mechanism | |
JP7488375B2 (en) | Method, device and computer-readable storage medium for generating neural networks | |
KR20200061154A (en) | Method and apparatus of analyzing diagram containing visual and textual information | |
CN108388886A (en) | Method, apparatus, terminal and the computer readable storage medium of image scene identification | |
CN115294361A (en) | Feature extraction method and device |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |