Hybrid prediction calculation method and device for deep neural network, and medium
Patent Information
- Application Number
- CN202411286738.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-13
- Publication Date
- 2024-12-24
Smart Images
Figure CN119180304A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of deep learning hardware acceleration, and in particular to a hybrid prediction calculation method, device and medium for deep neural networks. Background Art
[0002] With the rapid development of deep learning technology, deep neural networks (DNNs) have become a key tool for processing complex data such as images, videos, and speech. However, the computationally intensive and memory-access intensive characteristics of deep neural networks place extremely high demands on computing hardware, especially when dealing with large-scale data sets and deep network structures.
[0003] In deep neural networks, most of the calculations are in the fully connected layers and convolutional layers. Fully connected layers and convolutional layers are usually followed by a layer of activation function. Rectified linear unit (ReLU) is one of the most common activation functions, which is used to return negative input values to zero. In practical applications, a large part of the input values of ReLU are negative. If it is possible to predict in advance which elements calculated by neurons in the fully connected layers and convolutional layers will produce negative inputs, the output of the element at that position can be set to zero without calculating the corresponding elements, thus saving a lot of calculations.
[0004] Therefore, there is an urgent need for a hybrid prediction unit accelerator for deep neural networks to solve the problems of negative input prediction accuracy and the utilization efficiency of limited resources, improve computing flexibility, improve the computing performance of deep neural networks and reduce energy consumption. Summary of the invention
[0005] The embodiments of the present application provide a hybrid prediction calculation method, device and medium for deep neural networks, which are used to solve the following technical problems: the element negative input prediction of neuron calculation in the fully connected layer and convolutional layer of the existing deep neural network is inaccurate, the overall calculation amount and energy consumption are large, and the resource utilization efficiency is low.
[0006] The present application embodiment adopts the following technical solutions:
[0007] On the one hand, the embodiments of the present application provide a hybrid prediction calculation method, device and medium for a deep neural network, including: according to the weight data of neurons during deep neural network training, calculating the phase angle data corresponding to each of the neurons, and determining the proxy neurons; based on the proxy neurons, clustering the directed graph with neurons as nodes to obtain a proxy neuron set; performing feature calculation on the input data and weight data in the proxy neuron set to obtain a feature calculation result; if the feature calculation result is zero, performing a position prediction zeroing request on the remaining neurons in the proxy neuron set; performing result judgment on the prediction zeroing result corresponding to the prediction zeroing request; if the prediction zeroing result is zero, directly determining the output element as zero, and storing the output element.
[0008] The embodiment of the present application combines the neuron weight vector phase angle prediction unit with the position prediction zeroing unit to provide a higher negative input prediction accuracy during deep neural network training, thereby improving the utilization efficiency of limited resources, enhancing computing flexibility, and also improving the computing performance of the deep neural network and reducing energy consumption.
[0009] In a feasible implementation, according to the weight data of neurons during deep neural network training, the phase angle data corresponding to each of the neurons is calculated, and the proxy neuron is determined, which specifically includes: according to the training process of the deep neural network in the preset neural network model, the weight data corresponding to each layer of neurons in the deep neural network is obtained; based on the weight data corresponding to each layer of neurons, and through a preset inverse tangent function, the phase angle data corresponding to the weight data of each neuron is calculated; according to the size order of the phase angle data, each of the neurons is sorted and processed to generate a directed graph with neurons as nodes; wherein the nodes of the directed graph are constituted by edges connecting each neuron with its closest neuron; each of the nodes in the directed graph is sorted and processed by descending order of the in-degree of the nodes, and the neuron corresponding to the node with the highest in-degree is determined as the proxy neuron.
[0010] In a feasible implementation, based on the proxy neurons, a directed graph with neurons as nodes is clustered to obtain a proxy neuron set, specifically including: removing the nodes corresponding to the proxy neurons from the directed graph, and placing the proxy neurons into a preset proxy set; iteratively removing the nodes corresponding to the remaining neurons associated with the proxy neurons through the directed graph until the directed graph no longer has any nodes; and determining the nodes corresponding to the remaining neurons as cluster nodes of the proxy neurons; and generating a proxy neuron set based on the proxy set and the cluster nodes.
[0011] In a feasible implementation manner, feature calculation is performed on the input data and weight data in the proxy neuron set to obtain a feature calculation result, specifically including: feature calculation is performed on the input data and weight data corresponding to the proxy neurons in the proxy neuron set to obtain the feature calculation result related to the proxy neurons; based on the feature calculation result and through a mask signal, calculation waiting processing is performed on the remaining neurons corresponding to the clustering nodes in the proxy neuron set; wherein the clustering calculation result after the calculation waiting processing is associated with whether the feature calculation result is zero.
[0012] In a feasible implementation, if the characteristic calculation result is zero, a bit prediction zeroing request is made to the remaining neurons in the proxy neuron set, specifically including: if the characteristic calculation result is zero, the calculation result of the clustering node in the proxy neuron set is determined to be zero by requesting a zeroing signal; and a bit prediction zeroing request is made to the remaining neurons in the proxy neuron set through a preset arbitration unit.
[0013] In a feasible implementation, if the characteristic calculation result is not zero, the calculation result of the clustering node in the proxy neuron set is determined to be not zero through the request zeroing information; and the remaining neurons in the proxy neuron set are clustered with respect to the input data through the arbitration unit to obtain a clustering calculation result.
[0014] In a feasible implementation manner, the predicted zeroing result corresponding to the predicted zeroing request is subjected to result judgment, specifically including: if the feature calculation result is zero, a bit calculation is performed between the weight data and the input data under the predicted zeroing request through a bit prediction zeroing unit to obtain the predicted zeroing result; if the predicted zeroing result is not zero, a normal neural network training calculation is performed on the input data and weight data corresponding to each layer of neurons to obtain a normal calculation result.
[0015] In a feasible implementation, if the predicted zeroing result is zero, the output element is determined to be zero and the output element is stored, specifically including: if the predicted zeroing result is zero, the neural network training calculation of the input data and weight data corresponding to each layer of neurons is not performed, and the output element in the output data corresponding to the input data is determined to be zero; the output element is stored, and the normal calculation results when the predicted zeroing result is not zero and the clustering calculation results when the feature calculation result is not zero are also stored.
[0016] In a second aspect, an embodiment of the present application also provides a hybrid prediction computing device for a deep neural network, the device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor so that the at least one processor can execute a hybrid prediction computing method for a deep neural network as described in any of the above embodiments.
[0017] In a third aspect, an embodiment of the present application further provides a non-volatile computer storage medium, characterized in that the storage medium is a non-volatile computer-readable storage medium, and the non-volatile computer-readable storage medium stores at least one program, each of which includes instructions, and when the instructions are executed by a terminal, the terminal executes a hybrid prediction calculation method for a deep neural network described in any of the above embodiments.
[0018] The present application provides a hybrid prediction calculation method, device and medium for deep neural network. Compared with the prior art, the embodiments of the present application have the following beneficial technical effects:
[0019] 1. Improve computational efficiency: By calculating the phase angle data of neurons and proxy neurons, and performing feature calculation and prediction zeroing processing on the proxy neuron set, unnecessary calculations can be reduced and the computational efficiency of deep neural networks can be improved.
[0020] 2. Reduced energy consumption: Reducing the amount of computation can reduce the energy consumption of deep neural networks, which is very important for running neural networks on resource-constrained devices or deploying neural networks on a large scale.
[0021] 3. Improve prediction accuracy: By calculating the features of the proxy neuron set and setting the prediction to zero, the output of the neuron can be predicted more accurately, thereby improving the prediction accuracy of the deep neural network.
[0022] 4. Save storage space: Output elements with zero prediction results are directly determined as zero and stored, which can save storage space, especially when processing large-scale data.
[0023] 5. Improve model compression effect: It can also be used as a model compression technology to achieve model compression by reducing the number of neurons and the amount of calculation while maintaining a certain prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:
[0025] Figure 1 A flow chart of a hybrid prediction calculation method for a deep neural network provided in an embodiment of the present application;
[0026] Figure 2 An execution judgment flow chart of a hybrid prediction unit accelerator provided in an embodiment of the present application;
[0027] Figure 3 A schematic diagram of the association between an arbitration unit and a bit prediction unit provided in an embodiment of the present application;
[0028] Figure 4 A schematic diagram of the structure of a hybrid prediction computing device for a deep neural network provided in an embodiment of the present application. DETAILED DESCRIPTION
[0029] In order to enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are only part of the embodiments of this application, not all of them. Based on the embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.
[0030] The present application embodiment provides a hybrid prediction calculation method for a deep neural network, such as Figure 1 As shown, the hybrid prediction calculation method for deep neural network specifically includes steps S101-S106:
[0031] It should be noted that the present application provides a hybrid prediction unit accelerator for deep neural networks, the core of which mainly includes: a phase angle prediction unit, a bit prediction unit and an arbitration unit. Among them, the main function of the phase angle prediction unit is to collect the phase angle of each layer of neuron weight data during training. Neurons with similar phase angles are classified into a group of clusters. After sorting according to the in-degree, each group selects a proxy neuron for fast prediction. The main function of the bit prediction unit is to activate the bit prediction unit when the proxy neuron prediction result is zero. The bit prediction unit checks whether the calculation result of the characteristic bit of the input data and weight data is zero. If it is not zero, it is calculated normally; if it is zero, the calculation is skipped and the output is directly set to zero. The main function of the arbitration unit is to receive the phase angle prediction signal from the proxy neuron. Control the calculation process of neurons in each group. Send a mask signal or a zero signal to other neurons in the group.
[0032] S101. According to the weight data of neurons during deep neural network training, the phase angle data corresponding to each neuron is calculated, and the proxy neuron is determined.
[0033] Specifically, it is necessary to first obtain the weight data corresponding to each layer of neurons in the deep neural network according to the training process of the deep neural network in the preset neural network model.
[0034] Furthermore, based on the weight data corresponding to each layer of neurons, the phase angle data corresponding to the weight data of each neuron is calculated through a preset inverse tangent function.
[0035] Furthermore, each neuron is sorted according to the order of the phase angle data to generate a directed graph with neurons as nodes, wherein the nodes of the directed graph are formed by edges connecting each neuron with its closest neuron.
[0036] Furthermore, each node in the directed graph is sorted in descending order of the in-degree of the nodes, and the neuron corresponding to the node with the highest in-degree is determined as the proxy neuron.
[0037] In one embodiment, Figure 2 The execution judgment flow chart of a hybrid prediction unit accelerator provided in the embodiment of the present application is as follows: Figure 2 As shown in the figure, during the training process of the deep neural network, for each layer of neurons, its weight data is first obtained. Then, the phase angle data of the weight data is obtained by mathematical calculation. The phase angle can be calculated using mathematical methods such as the inverse tangent function. Next, the neurons are sorted according to the size of the phase angle in the phase angle data. A directed graph with neurons as nodes is generated, and the edges connect each neuron to its closest neuron. The nodes in the directed graph are sorted in descending order according to their in-degree (the number of incident edges). Nodes with high in-degree are selected as proxy neurons.
[0038] S102: Based on the proxy neurons, a directed graph with neurons as nodes is clustered to obtain a set of proxy neurons.
[0039] Specifically, the nodes corresponding to the proxy neurons are removed from the directed graph, and the proxy neurons are placed in a preset proxy set.
[0040] Furthermore, through the directed graph, the nodes corresponding to the remaining neurons associated with the proxy neuron are iteratively removed until the directed graph no longer has any nodes, and the nodes corresponding to the remaining neurons are determined as cluster nodes of the proxy neuron.
[0041] Furthermore, based on the agent set and the clustering nodes, an agent neuron set is generated.
[0042] In one embodiment, Figure 3 A schematic diagram of the association between an arbitration unit and a bit prediction unit provided in an embodiment of the present application is shown in FIG. Figure 2 as well as Figure 3 As shown, the nodes corresponding to the proxy neurons are removed from the directed graph and placed in the proxy set. At the same time, the nodes connected to the proxy neurons in the directed graph are also removed as members of the cluster where the proxy neurons are located. That is, the nodes corresponding to the remaining neurons are determined as cluster nodes of the proxy neurons. The steps are iterated until there are no more nodes in the graph. Finally, the proxy set and the cluster nodes are combined to generate a proxy neuron set.
[0043] S103, performing feature calculation on the input data and weight data in the proxy neuron set to obtain feature calculation results.
[0044] Specifically, feature calculation is performed on input data and weight data corresponding to the proxy neurons in the proxy neuron set to obtain feature calculation results related to the proxy neurons.
[0045] Further, based on the feature calculation result and through the mask signal, the remaining neurons in the proxy neuron set corresponding to the clustering node are subjected to calculation waiting processing, wherein the clustering calculation result after the calculation waiting processing is associated with whether the feature calculation result is zero.
[0046] In one embodiment, Figure 2 As shown, first, the corresponding input data and weight data are calculated by the proxy neurons in the clustered proxy neuron set to obtain the feature calculation results of the proxy neurons, and then other neurons in the cluster wait for the calculation results of the proxy neurons through mask signals, thereby subsequently forming the clustering calculation results.
[0047] S104: If the feature calculation result is zero, a position prediction zeroing request is made to the remaining neurons in the proxy neuron set.
[0048] Specifically, if the characteristic calculation result is zero, the calculation result of the cluster node in the proxy neuron set is determined to be zero through a request for zeroing signal, and then a position prediction zeroing request is made to the remaining neurons in the proxy neuron set through a preset arbitration unit.
[0049] If the characteristic calculation result is not zero, the calculation result of the clustering node in the proxy neuron set is determined to be not zero by requesting zeroing information. Then, the arbitration unit is used to perform clustering calculations on the remaining neurons in the proxy neuron set for the input data to obtain the clustering calculation results.
[0050] In one embodiment, Figure 2 as well as Figure 3 As shown, if the calculation result of the proxy neuron is not zero, the arbitration unit is notified that the calculation result in the cluster is not zero through a request to reset to zero signal; after receiving the signal, the arbitration unit notifies the remaining neurons in the cluster to calculate the corresponding input data according to the normal calculation process, thereby obtaining the cluster calculation result. If the calculation result of the proxy neuron is zero, the arbitration unit is notified that the calculation result in the cluster is zero through a request to reset to zero signal; after receiving the signal, the arbitration unit notifies the remaining neurons in the cluster to initiate a bit prediction reset request for the relevant bit prediction unit.
[0051] S105 , performing result judgment on the predicted zeroing result corresponding to the predicted zeroing request.
[0052] Specifically, if the feature calculation result is zero, a bit prediction zeroing unit is used to perform bitwise calculation between the weight data under the prediction zeroing request and the input data to obtain a prediction zeroing result.
[0053] If the predicted zeroing result is not zero, the input data and weight data corresponding to each layer of neurons are subjected to normal neural network training calculations to obtain normal calculation results.
[0054] In one embodiment, when the feature calculation result is zero, that is, if the calculation result of the proxy neuron is zero in step S104, it is also necessary to calculate the predicted zeroing result based on the predicted zeroing request in the bit prediction zeroing unit in each calculation group, and based on the corresponding input data and the bit data after the weight data is bitified. Then, if the prediction result is not zero, the corresponding neuron data calculates the corresponding data according to the normal calculation process, and the calculation unit is responsible for performing multiplication and addition operations of the neuron to obtain a normal calculation result.
[0055] S106: If the predicted zeroing result is zero, the output element is directly determined to be zero, and the output element is stored for processing.
[0056] Specifically, if the predicted zeroing result is zero, the neural network training calculation of the input data and weight data corresponding to each layer of neurons is not performed, and the output element in the output data corresponding to the input data is determined to be zero.
[0057] In one embodiment, Figure 2 As shown in FIG. 1 , if the predicted zeroing result is zero, the corresponding output element is directly set to zero and stored without performing actual calculations. This can save computing resources and improve computing efficiency.
[0058] Furthermore, if Figure 2 As shown, the output elements are stored, and the normal calculation results when the prediction zeroing result is not zero and the clustering calculation results when the feature calculation result is not zero are also stored.
[0059] In addition, the present application also provides a hybrid prediction computing device for deep neural networks, such as Figure 4 As shown, the hybrid prediction computing device for deep neural network specifically includes:
[0060] At least one processor 401. And a memory 402 in communication with the at least one processor 401. The memory 402 stores instructions that can be executed by the at least one processor 401, so that the at least one processor 401 can execute:
[0061] According to the weight data of neurons during deep neural network training, the phase angle data corresponding to each neuron is calculated, and the proxy neuron is determined;
[0062] Based on the proxy neurons, the directed graph with neurons as nodes is clustered to obtain a set of proxy neurons;
[0063] Perform feature calculation on the input data and weight data in the agent neuron set to obtain feature calculation results;
[0064] If the feature calculation result is zero, the remaining neurons in the proxy neuron set will be requested to set their positions to zero;
[0065] Perform result judgment on the prediction zeroing result corresponding to the prediction zeroing request;
[0066] If the predicted zeroing result is zero, the output element is directly determined to be zero and the output element is stored for processing.
[0067] The embodiments of the present application can reduce unnecessary calculations and improve the computational efficiency of deep neural networks by calculating the phase angle data of neurons and proxy neurons, and performing feature calculations and prediction zeroing processing on proxy neuron sets. Reducing the amount of calculation can reduce the energy consumption of deep neural networks, which is very important for running neural networks on resource-constrained devices or deploying neural networks on a large scale. By calculating the features of proxy neuron sets and performing prediction zeroing processing, the output of neurons can be predicted more accurately, thereby improving the prediction accuracy of deep neural networks. Output elements with prediction zeroing results of zero are directly determined as zero and stored, which can save storage space, especially when processing large-scale data. It is also possible to achieve model compression by reducing the number of neurons and the amount of calculation while maintaining a certain prediction performance.
[0068] Each embodiment in this application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and non-volatile computer storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.
[0069] The devices and media provided in the embodiments of the present application correspond one-to-one to the methods. Therefore, the devices and media also have similar beneficial technical effects as the corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.
[0070] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.
[0071] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0072] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.
[0073] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.
[0074] The above describes specific embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0075] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the embodiments of the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the embodiments of the present application should be included in the scope of the claims of the present application.
Claims
1. A hybrid prediction calculation method for deep neural networks, characterized in that: The method comprises: According to the weight data of neurons during deep neural network training, the phase angle data corresponding to each of the neurons is calculated, and the proxy neuron is determined; Based on the proxy neurons, a directed graph with neurons as nodes is clustered to obtain a proxy neuron set; Performing feature calculation on the input data and weight data in the proxy neuron set to obtain a feature calculation result; If the characteristic calculation result is zero, a position prediction zeroing request is made to the remaining neurons in the proxy neuron set; Perform result judgment on the prediction zeroing result corresponding to the prediction zeroing request; If the predicted zeroing result is zero, the output element is directly determined to be zero, and the output element is stored for processing.
2. A hybrid prediction calculation method for deep neural network according to claim 1, characterized in that: According to the weight data of neurons during deep neural network training, the phase angle data corresponding to each of the neurons is calculated, and the proxy neurons are determined, specifically including: According to the training process of the deep neural network in the preset neural network model, the weight data corresponding to each layer of neurons in the deep neural network is obtained; Based on the weight data corresponding to each layer of neurons, and through a preset inverse tangent function, calculate the phase angle data corresponding to each of the weight data of the neurons; According to the order of the phase angle data, each of the neurons is sorted to generate a directed graph with neurons as nodes; wherein the nodes of the directed graph are formed by edges connecting each neuron with its closest neuron; Each of the nodes in the directed graph is sorted in descending order of the in-degree of the nodes, and the neuron corresponding to the node with the highest in-degree is determined as the proxy neuron.
3. A hybrid prediction calculation method for deep neural network according to claim 1, characterized in that: Based on the proxy neurons, a directed graph with neurons as nodes is clustered to obtain a proxy neuron set, which specifically includes: Removing the node corresponding to the proxy neuron from the directed graph, and placing the proxy neuron into a preset proxy set; Iteratively removing nodes corresponding to other neurons associated with the proxy neuron through the directed graph until the directed graph no longer has any nodes; and determining the nodes corresponding to the other neurons as cluster nodes of the proxy neuron; Based on the agent set and the clustering nodes, an agent neuron set is generated.
4. A hybrid prediction calculation method for deep neural network according to claim 1, characterized in that: Performing feature calculation on the input data and weight data in the proxy neuron set to obtain feature calculation results, specifically including: Performing feature calculation on the input data and weight data corresponding to the proxy neurons in the proxy neuron set to obtain the feature calculation result related to the proxy neurons; Based on the feature calculation result and through the mask signal, the remaining neurons in the proxy neuron set corresponding to the clustering node are subjected to calculation waiting processing; wherein the clustering calculation result after the calculation waiting processing is associated with whether the feature calculation result is zero.
5. The hybrid prediction calculation method for deep neural network according to claim 1, characterized in that: If the characteristic calculation result is zero, the remaining neurons in the proxy neuron set are subjected to a position prediction zeroing request, specifically including: If the characteristic calculation result is zero, the calculation result of the clustering node in the proxy neuron set is determined to be zero by requesting a zeroing signal; Through a preset arbitration unit, a bit prediction zeroing request is made to the remaining neurons in the proxy neuron set.
6. A hybrid prediction calculation method for deep neural network according to claim 5, characterized in that: If the characteristic calculation result is not zero, the calculation result of the clustering node in the proxy neuron set is determined to be not zero through the request zeroing information; The arbitration unit performs clustering calculation on the input data for the remaining neurons in the proxy neuron set to obtain a clustering calculation result.
7. A hybrid prediction calculation method for deep neural network according to claim 1, characterized in that: The step of determining the predicted zeroing result corresponding to the predicted zeroing request specifically includes: If the characteristic calculation result is zero, a bit prediction zeroing unit is used to perform bit calculation between the weight data under the prediction zeroing request and the input data to obtain the prediction zeroing result; If the predicted zeroing result is not zero, the input data and weight data corresponding to each layer of neurons are subjected to normal neural network training calculation to obtain a normal calculation result.
8. A hybrid prediction calculation method for deep neural network according to claim 1, characterized in that: If the predicted zeroing result is zero, the output element is determined to be zero, and the output element is stored, which specifically includes: If the prediction zeroing result is zero, the neural network training calculation of the input data and weight data corresponding to each layer of neurons is not performed, and the output element in the output data corresponding to the input data is determined to be zero; The output elements are stored, and the normal calculation results where the prediction zeroing result is not zero and the clustering calculation results where the feature calculation result is not zero are also stored.
9. A hybrid prediction computing device for deep neural networks, characterized in that: The device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor so that the at least one processor can execute a hybrid prediction calculation method for a deep neural network according to any one of claims 1-8.
10. A non-volatile computer storage medium, characterized in that: The storage medium is a non-volatile computer-readable storage medium, which stores at least one program, each of which includes instructions, and when the instructions are executed by the terminal, the terminal executes a hybrid prediction calculation method for a deep neural network according to any one of claims 1-8.