Dynamic neural network deployment method and device, storage medium and program product

Abstract

The embodiment of the invention provides a dynamic neural network deployment method and device, a storage medium and a program product, and the method comprises the steps: obtaining image input data, a plurality of dynamic convolution coefficients, and to-be-adjusted weights corresponding to the plurality of dynamic convolution coefficients respectively, and carrying out the adjustment of the to-be-adjusted weights based on a static convolution operator, and determining dynamic convolution output data of the image input data according to the image input data, the plurality of dynamic convolution coefficients and the plurality of to-be-adjusted weights so as to realize deployment of the dynamic neural network. According to the embodiment of the invention, dynamic convolution operation is carried out based on existing conventional operators such as a static convolution operator, a new dynamic convolution operator does not need to be developed, mobile terminals such as a mobile phone can also support dynamic convolution operation, and efficient deployment of the dynamic neural network is realized.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of computer technology, and in particular to a dynamic neural network deployment method, device, storage medium, and program product. Background Art

[0002] Dynamic neural network is an important technology to improve the prediction performance of deep neural network.

[0003] In the related art, dynamic neural networks can be applied to multiple backbone networks and various computer vision tasks, that is, the operation of dynamic convolution operators can be supported by the server and computer ends.

[0004] However, the inventors found that there are at least the following technical problems in the relevant technology: the hardware resources of mobile terminals such as mobile phones are limited and cannot support the use of dynamic convolution operators. How to efficiently deploy dynamic neural networks to mobile terminals is a technical problem that urgently needs to be solved. Summary of the invention

[0005] The embodiments of the present disclosure provide a dynamic neural network deployment method, device, storage medium and program product to enable mobile terminals such as mobile phones to efficiently deploy dynamic neural networks.

[0006] In a first aspect, an embodiment of the present disclosure provides a dynamic neural network deployment method, comprising:

[0007] Obtaining image input data, a plurality of dynamic convolution coefficients, and a plurality of weights to be adjusted corresponding to the dynamic convolution coefficients;

[0008] Based on the static convolution operator, the dynamic convolution output data of the image input data is determined according to the image input data, the multiple dynamic convolution coefficients and the multiple weights to be adjusted to realize the deployment of the dynamic neural network.

[0009] In a second aspect, an embodiment of the present disclosure provides a dynamic neural network deployment device, including:

[0010] An acquisition module, used to acquire image input data, a plurality of dynamic convolution coefficients and weights to be adjusted corresponding to the plurality of dynamic convolution coefficients;

[0011] A determination module is used to determine the dynamic convolution output data of the image input data based on a static convolution operator, according to the image input data, multiple dynamic convolution coefficients and multiple weights to be adjusted, so as to realize the deployment of a dynamic neural network.

[0012] In a third aspect, an embodiment of the present disclosure provides an electronic device, including: a processor and a memory;

[0013] The memory stores computer-executable instructions;

[0014] The processor executes the computer-executable instructions stored in the memory, so that the at least one processor performs the dynamic neural network deployment method as described in the first aspect and various possible designs of the first aspect.

[0015] In a fourth aspect, an embodiment of the present disclosure provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions. When a processor executes the computer-executable instructions, the dynamic neural network deployment method described in the first aspect and various possible designs of the first aspect is implemented.

[0016] In a fifth aspect, an embodiment of the present disclosure provides a computer program product, including a computer program, which, when executed by a processor, implements the dynamic neural network deployment method as described in the first aspect and various possible designs of the first aspect.

[0017] The dynamic neural network deployment method, device, storage medium and program product provided in this embodiment first obtain image input data, multiple dynamic convolution coefficients and weights to be adjusted corresponding to the multiple dynamic convolution coefficients, and then determine the dynamic convolution output data of the image input data based on the image input data, the multiple dynamic convolution coefficients and the multiple weights to be adjusted based on the static convolution operator to achieve the deployment of the dynamic neural network. This embodiment performs dynamic convolution operations based on existing conventional operators such as static convolution operators, without the need to develop new dynamic convolution operators, so that mobile terminals such as mobile phones can also support dynamic convolution operations, thereby achieving efficient deployment of dynamic neural networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0019] Figure 1 A schematic diagram of an application scenario of the dynamic neural network deployment method provided in an embodiment of the present disclosure;

[0020] Figure 2 A schematic diagram of a flow chart of a dynamic neural network deployment method provided by an embodiment of the present disclosure;

[0021] Figure 3a A schematic diagram of a dynamic neural network deployment framework provided in an embodiment of the present disclosure Figure 1 ;

[0022] Figure 3bA schematic diagram of a dynamic neural network deployment framework provided in an embodiment of the present disclosure Figure 2 ;

[0023] Figure 4a Schematic diagram of the operation flow of the dynamic neural network provided in the embodiment of the present disclosure Figure 1 ;

[0024] Figure 4b Schematic diagram of the operation flow of the dynamic neural network provided in the embodiment of the present disclosure Figure 2 ;

[0025] Figure 5 A structural block diagram of a dynamic neural network deployment device provided in an embodiment of the present disclosure;

[0026] Figure 6 A schematic diagram of the hardware structure of a dynamic neural network deployment device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the embodiments of the present disclosure clearer, the technical solution in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0028] Artificial intelligence is currently experiencing the era of deep learning, and deep neural networks are dazzling and have achieved significant performance improvements in various tasks. For computer vision tasks such as image classification, detection, and segmentation, deep convolutional networks have reached or even exceeded human levels.

[0029] Among them, dynamic neural network is an important technology to improve the prediction performance of deep neural network. The samples in the image data set in computer vision tasks have commonalities and characteristics. The dynamic neural network uses the characteristics of the samples themselves, depends on the current input samples, and dynamically adjusts the convolutional network parameters based on the attention mechanism, thereby achieving input adaptation and improving the prediction accuracy of the convolutional neural network.

[0030] In the related technology, dynamic convolutional networks have been implemented and verified in multiple backbone networks and various computer vision tasks. The relatively abundant computing resources on the server or computer side can support the operation of dynamic convolution operators.

[0031] However, although dynamic neural networks have been verified in a large number of backbone networks and various computer visions, they are still a long way from being implemented in the industry. The main reason that hinders the implementation of dynamic neural networks is that the underlying hardware of the mobile terminal (such as processing chips) does not support dynamic neural network operators. Therefore, implementing dynamic neural network operators and deploying them efficiently to different mobile terminals is an urgent problem that needs to be solved.

[0032] In order to solve the above technical problems, the inventors of the present disclosure have found that the dynamic convolution calculation formula can be split to a degree that can be supported by the hardware resources of the mobile terminal. Based on this, the present disclosure provides a dynamic neural network deployment method.

[0033] refer to Figure 1 , Figure 1 The following is a schematic diagram of the principle application scenario of the dynamic neural network deployment method provided in the embodiment of the present disclosure. Figure 1 As shown, dynamic neural networks can be applied to process computer vision tasks, which may include image classification, detection, and segmentation tasks. In the specific task processing process, the processor obtains image input data, inputs the image input data into the dynamic convolution operator, and outputs dynamic convolution output data after the dynamic convolution operation of the dynamic convolution operator to realize the deployment of the dynamic neural network. Optionally, in the dynamic convolution operation process, n dynamic convolution coefficients alpha_1, alpha_2...alpha_n can be generated based on the attention mechanism first, and the n dynamic convolution coefficients are multiplied one by one with the weights to be adjusted w_1, w_2...w_n to obtain n products, and then the n products are added to obtain the total weight w', and the total weight w' is convolved with the input data x to obtain the output data y.

[0034] refer to Figure 2 , Figure 2 A schematic diagram of a dynamic neural network deployment method provided in an embodiment of the present disclosure. The method of this embodiment can be applied in a terminal device or a server, and the dynamic neural network deployment method includes:

[0035] 201. Obtain image input data, multiple dynamic convolution coefficients, and multiple weights to be adjusted corresponding to the dynamic convolution coefficients.

[0036] In this embodiment, a neural network is a computational model that simulates the interconnection between neurons in the human brain. In a neural network, weight is an important parameter for connecting information transmission between neurons, which determines the strength of signal transmission between neurons. A dynamic neural network is a special neural network structure with the ability to dynamically adjust weights. The dynamic neural network introduces an attention mechanism and deformable convolution to enhance the model's perception of spatiotemporal information. The attention mechanism allows the model to pay more attention to important information during the convolution process, thereby improving the model's perception and accuracy. Deformable convolution allows the convolution kernel to be slightly deformed within a certain range to adapt to local changes in the data. This enables the model to better capture the temporal and spatial changes in the data, thereby improving the model's representation ability. Dynamic neural networks are widely used in computer vision tasks such as target detection and image segmentation. In target detection, dynamic neural networks can better adapt to targets of different sizes and shapes, and improve the model's ability to identify and locate targets. In image segmentation, dynamic convolutional neural networks can more accurately capture the edges and textures of different regions in the image, thereby improving the accuracy and detail of segmentation.

[0037] Exemplarily, the method of this embodiment can be applied to a head-mounted device, and the image input data can be acquired through the camera of the head-mounted device. The dynamic neural network can be deployed in the processor of the head-mounted device, and the image data can be processed by the deployed dynamic neural network to complete visual tasks such as gesture recognition tasks, semantic segmentation tasks in fully automatic scenes, and object key point (such as hand, human body) detection tasks.

[0038] Specifically, the multiple dynamic convolution coefficients and the multiple weights to be adjusted are in one-to-one correspondence. The dynamic convolution coefficient is an adjustment coefficient of the weight to be adjusted. The multiple weights to be adjusted can obtain dynamic weights under the adjustment control of the corresponding dynamic convolution coefficients.

[0039] In one embodiment of the present disclosure, the acquiring of multiple dynamic convolution coefficients may include: determining the multiple dynamic convolution coefficients respectively based on an activation function operator.

[0040] Specifically, the dynamic convolution coefficient alpha is calculated by expression (1): i .

[0041] alpha i = function(X) (1)

[0042] Among them, alpha i is the dynamic convolution coefficient corresponding to the i-th weight to be adjusted. The function() function can use common operators (Operator, OP) such as sigmoid / Fc / Relu.

[0043] 202. Based on a static convolution operator, determine the dynamic convolution output data of the image input data according to the image input data, the multiple dynamic convolution coefficients and the multiple weights to be adjusted to implement deployment of a dynamic neural network.

[0044] In this embodiment, considering that the dynamic convolution operator consumes a large amount of resources, and the inference engine of the processing chip of a mobile terminal such as a mobile phone does not directly support the method body of the dynamic convolution operator, a conventional operator such as a static convolution operator supported by the processing chip of a mobile terminal such as a mobile phone can be used to implement the calculation of the dynamic convolution operator and obtain dynamic convolution output data.

[0045] In one embodiment of the present disclosure, there are multiple ways to perform dynamic convolution on input data based on a static convolution operator to obtain corresponding output data.

[0046] In one possible implementation, consider the dynamic convolution output data C which can be calculated by expression (2).

[0047] C=(alpha0*Weight0+alpha1*Weight1+alpha2*Weight2+..+alpha n *

[0048] Weight n )*X(2)

[0049] Among them, C is the dynamic convolution output data, X is the input data, alpha i is the i-th weight to be adjusted i The corresponding dynamic convolution coefficients.

[0050] For example, Figure 3a As shown, formula (2) is split into: first, the dynamic convolution coefficient alpha is calculated by the multiplication operator MUL i and the weight to be adjusted i The product of , and then the sum of the products is calculated by the addition operator Add to obtain the dynamic weight w', and finally the dynamic weight w' and the input data X are convolved by the static convolution operator.

[0051] In another possible implementation, the determining of the dynamic convolution output data of the image input data based on the static convolution operator according to the image input data, the multiple dynamic convolution coefficients and the multiple weights to be adjusted may include: based on the static convolution operator, convolving the multiple weights to be adjusted with the image input data to obtain the first results corresponding to the multiple weights to be adjusted; multiplying the multiple first results with the corresponding dynamic convolution coefficients to obtain the second results corresponding to the multiple first results; adding the multiple second results to obtain the dynamic convolution output data corresponding to the image input data. Optionally, the multiplying the multiple first results with the corresponding dynamic convolution coefficients to obtain the second results corresponding to the multiple first results may include: based on the multiplication operator, multiplying the multiple first results with the corresponding dynamic convolution coefficients to obtain the second results corresponding to the multiple first results; the adding the multiple second results to obtain the dynamic convolution output data corresponding to the image input data may include: based on the addition operator, adding the multiple second results to obtain the dynamic convolution output data corresponding to the image input data.

[0052] Specifically, according to the data combination law, expression (2) can be split into expressions (3) to (6).

[0053] C i = Weight i * X (3)

[0054] y i = alpha i *C i (4)

[0055] C =y0+y1+y2+ ...+y i (5)

[0056] Among them, C is the dynamic convolution output data, X is the input data, alpha i is the i-th weight to be adjusted i The corresponding dynamic convolution coefficient, C i is the first result obtained by convolving the i-th weight to be adjusted with the input data, y i The second result is obtained by multiplying the dynamic convolution coefficient corresponding to the i-th weight to be adjusted by the corresponding first result.

[0057] For example, Figure 3bAs shown, the input data X can be first convolved with n weights to be adjusted through static convolution operators Conv_0, Conv_1, Conv_2, ..., Conv_n to obtain n first results c_0, c_1, c_2, ..., c_n, and the n first results are multiplied by the corresponding dynamic convolution coefficients through the multiplication operator Mul to obtain the second result, and the n second results are added through the addition operator Add to obtain the dynamic convolution output data.

[0058] Optionally, multiplying the multiple first results with the corresponding dynamic convolution coefficients respectively to obtain the second results corresponding to the multiple first results respectively, and adding the multiple second results to obtain the dynamic convolution output data corresponding to the image input data may include: based on the cumulative multiplication and addition operator, multiplying the multiple first results with the corresponding dynamic convolution coefficients respectively to obtain the second results corresponding to the multiple first results respectively, and adding the multiple second results to obtain the dynamic convolution output data corresponding to the image input data.

[0059] In one embodiment of the present disclosure, a static convolution operator can be used to sequentially complete the convolution operation of n weights to be adjusted and the input data X to obtain n first results, and then the n second results are obtained by multiplying the n first results with the corresponding dynamic convolution coefficients respectively, and finally the n second results are added to obtain dynamic convolution output data.

[0060] For example, Figure 4a As shown, at time t_0, the convolution operation of the input data x and the first weight to be adjusted w_0 is completed to obtain the first result y_0. Similarly, at time t_1, the first result y_1 is obtained, and at time t_n, the first result y_n is obtained. Then, at time t_n+1, the sum of the products of each dynamic convolution coefficient and the corresponding first result is obtained to obtain the dynamic convolution output data.

[0061] In one embodiment of the present disclosure, in order to improve the operation efficiency, the operation can be performed through multiple operators. Specifically, the static convolution operator includes multiple target static convolution operators; based on the static convolution operator, the multiple weights to be adjusted are respectively convolved with the image input data to obtain the first results corresponding to the multiple weights to be adjusted, which may include: based on the multiple target static convolution operators, the multiple weights to be adjusted are respectively convolved with the image input data in parallel to obtain the first results corresponding to the multiple weights to be adjusted.

[0062] Specifically, in one possible implementation, the number of static convolution operators currently available for operation can be first determined. If the number is less than the number of weights to be adjusted, the convolution operation can be performed in batches. For example, if the current available number of static convolution operators is 3 and the number of weights to be adjusted is 12, the weights to be adjusted can be divided into 4 batches, 3 in each batch, and 3 available static convolution operators are used to perform convolution operations on the weights to be adjusted and the input data in the batch at the same time.

[0063] In another implementable manner, based on the multiple target static convolution operators, the multiple weights to be adjusted are respectively convolved with the image input data in parallel to obtain the first results corresponding to the multiple weights to be adjusted, which may include: if the first number of the convolution operators is greater than or equal to the second number of the multiple weights to be adjusted, then based on the second number of the target static convolution operators, the multiple weights to be adjusted are respectively convolved with the image input data in parallel to obtain the first results corresponding to the multiple weights to be adjusted.

[0064] For example, Figure 4b As shown, the current available number of static convolution operators is 15, and the number n of weights to be adjusted is 12. Then, 12 of the 15 available static convolution operators can be used to simultaneously perform convolution operations on the 12 weights to be adjusted and the input data at time t_0, and then perform multiplication operations of formula (4) and addition operations of formula (5) at time t_1.

[0065] In one embodiment of the present disclosure, in order to improve the operation efficiency, when performing multiplication operations, parallel operations can be performed based on the number of multiplication operators currently available for calculation. Specifically, the multiplication of the plurality of first results with the corresponding dynamic convolution coefficients to obtain the second results corresponding to the plurality of first results may include: based on the plurality of multiplication operators, the plurality of first results are multiplied in parallel with the corresponding dynamic convolution coefficients to obtain the second results corresponding to the plurality of first results.

[0066] Exemplarily, after obtaining the first result of the convolution operation, the operator that can be used for the multiplication operation can be determined. Assuming that there are 12 first results and there are currently 4 multiplication operators that can be used for the multiplication operation, the first results can be divided into 3 batches, each batch includes 4 first results, and the multiplication operation is performed in parallel by the above 4 multiplication operators, that is, for each of the 4 first results, the first result and the dynamic convolution coefficient corresponding to the first result are multiplied simultaneously.

[0067] From the above description, it can be seen that this embodiment performs dynamic convolution operations based on existing conventional operators such as static convolution operators, without the need to develop new dynamic convolution operators. This enables mobile terminals such as mobile phones to support dynamic convolution operations and achieve efficient deployment of dynamic neural networks.

[0068] Corresponding to the dynamic neural network deployment method of the above embodiment, Figure 5 A structural block diagram of a dynamic neural network deployment device provided in an embodiment of the present disclosure. For ease of explanation, only the parts related to the embodiment of the present disclosure are shown. Figure 5 , the device includes: an acquisition module 501 and a determination module 502.

[0069] The acquisition module 501 is used to acquire image input data, a plurality of dynamic convolution coefficients and weights to be adjusted corresponding to the plurality of dynamic convolution coefficients;

[0070] The determination module 502 is used to determine the dynamic convolution output data of the image input data based on the static convolution operator, according to the image input data, the multiple dynamic convolution coefficients and the multiple weights to be adjusted, so as to realize the deployment of the dynamic neural network.

[0071] In one embodiment of the present disclosure, the determination module 502 is specifically used to: obtain image input data, multiple dynamic convolution coefficients and multiple weights to be adjusted corresponding to the dynamic convolution coefficients; based on the static convolution operator, determine the dynamic convolution output data of the image input data according to the image input data, the multiple dynamic convolution coefficients and the multiple weights to be adjusted to realize the deployment of the dynamic neural network.

[0072] In one embodiment of the present disclosure, the determination module 502 is specifically used to: based on the static convolution operator, convolve the multiple weights to be adjusted with the image input data respectively to obtain the first results corresponding to the multiple weights to be adjusted respectively; multiply the multiple first results with the corresponding dynamic convolution coefficients respectively to obtain the second results corresponding to the multiple first results respectively; add the multiple second results to obtain the dynamic convolution output data corresponding to the image input data.

[0073] In one embodiment of the present disclosure, the determination module 502 is specifically used to: based on a multiplication operator, multiply the multiple first results by the corresponding dynamic convolution coefficients respectively to obtain the second results corresponding to the multiple first results respectively; based on an addition operator, add the multiple second results to obtain the dynamic convolution output data corresponding to the image input data.

[0074] In one embodiment of the present disclosure, the static convolution operator includes multiple target static convolution operators; the determination module 502 is specifically used to: based on the multiple target static convolution operators, perform convolution operations on the multiple weights to be adjusted in parallel with the image input data to obtain first results corresponding to the multiple weights to be adjusted.

[0075] In one embodiment of the present disclosure, the determination module 502 is specifically used to: if the first number of the convolution operators is greater than or equal to the second number of the multiple weights to be adjusted, then based on the second number of the target static convolution operators, the multiple weights to be adjusted are respectively convolved with the image input data in parallel to obtain first results corresponding to the multiple weights to be adjusted.

[0076] In one embodiment of the present disclosure, the determination module 502 is specifically used to: based on multiple multiplication operators, perform multiplication operations on multiple first results in parallel with corresponding dynamic convolution coefficients respectively, to obtain second results corresponding to the multiple first results respectively.

[0077] In one embodiment of the present disclosure, the acquisition module 501 is specifically used to: determine multiple dynamic convolution coefficients respectively based on the activation function operator.

[0078] The device provided in this embodiment can be used to execute the technical solution of the above method embodiment. Its implementation principle and technical effect are similar, and this embodiment will not be repeated here.

[0079] In order to implement the above embodiment, the embodiment of the present disclosure also provides an electronic device.

[0080] refer to Figure 6 , which shows a schematic diagram of the structure of an electronic device 900 suitable for implementing the embodiment of the present disclosure, and the electronic device 900 may be a terminal device or a server. The terminal device may include but is not limited to mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, personal digital assistants (PDAs), tablet computers (Portable Android Devices, PADs), portable multimedia players (PMPs), vehicle terminals (such as vehicle navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 6 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0081] like Figure 6As shown, the electronic device 900 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 901, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 902 or a program loaded from a storage device 908 to a random access memory (RAM) 903. In the RAM 903, various programs and data required for the operation of the electronic device 900 are also stored. The processing device 901, the ROM 902, and the RAM 903 are connected to each other via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0082] Typically, the following devices may be connected to the I / O interface 905: input devices 906 including, for example, a touch screen, a touch pad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; output devices 907 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices 908 including, for example, a magnetic tape, a hard disk, etc.; and communication devices 909. The communication device 909 may allow the electronic device 900 to communicate with other devices wirelessly or by wire to exchange data. Although Figure 6 The electronic device 900 is shown with various devices, but it should be understood that it is not required to implement or possess all the devices shown. More or fewer devices may be implemented or possessed instead.

[0083] In particular, according to an embodiment of the present disclosure, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program contains program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device 909, or installed from a storage device 908, or installed from a ROM 902. When the computer program is executed by the processing device 901, the above-mentioned functions defined in the method of the embodiment of the present disclosure are executed.

[0084] It should be noted that the computer-readable medium disclosed above may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that may be used by or in combination with an instruction execution system, device or device. In the present disclosure, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which a computer-readable program code is carried. This propagated data signal may take a variety of forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the above. The computer readable signal medium may also be any computer readable medium other than a computer readable storage medium, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0085] The computer-readable medium may be included in the electronic device, or may exist independently without being incorporated into the electronic device.

[0086] The computer-readable medium carries one or more programs. When the one or more programs are executed by the electronic device, the electronic device executes the method shown in the above embodiment.

[0087] Computer program code for performing the operations of the present disclosure may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0088] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some implementations as replacements, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0089] The units involved in the embodiments described in the present disclosure may be implemented by software or hardware. The name of a unit does not limit the unit itself in some cases. For example, the first acquisition unit may also be described as a "unit for acquiring at least two Internet Protocol addresses".

[0090] The functions described above herein may be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0091] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0092] The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of disclosure involved in the present disclosure is not limited to the technical solutions formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above disclosed concept. For example, the above features are replaced with the technical features with similar functions disclosed in the present disclosure (but not limited to) by each other to form a technical solution.

[0093] In addition, although each operation is described in a specific order, this should not be understood as requiring these operations to be performed in the specific order shown or in a sequential order. Under certain circumstances, multitasking and parallel processing may be advantageous. Similarly, although some specific implementation details are included in the above discussion, these should not be interpreted as limiting the scope of the present disclosure. Some features described in the context of a separate embodiment can also be implemented in a single embodiment in combination. On the contrary, the various features described in the context of a single embodiment can also be implemented in multiple embodiments individually or in any suitable sub-combination mode.

[0094] Although the subject matter has been described in language specific to structural features and / or methodological logical actions, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. On the contrary, the specific features and actions described above are merely example forms of implementing the claims.

Claims

1. A dynamic neural network deployment method, characterized in that: include: Obtaining image input data, a plurality of dynamic convolution coefficients, and a plurality of weights to be adjusted corresponding to the dynamic convolution coefficients; Based on the static convolution operator, the dynamic convolution output data of the image input data is determined according to the image input data, the multiple dynamic convolution coefficients and the multiple weights to be adjusted to realize the deployment of the dynamic neural network.

2. The method according to claim 1, characterized in that The method of determining the dynamic convolution output data of the image input data based on the static convolution operator according to the image input data, the multiple dynamic convolution coefficients and the multiple weights to be adjusted includes: Based on the static convolution operator, convolution operations are performed on the plurality of weights to be adjusted and the image input data respectively to obtain first results corresponding to the plurality of weights to be adjusted respectively; Multiplying the first results by corresponding dynamic convolution coefficients respectively to obtain second results corresponding to the first results respectively; Add multiple second results to obtain dynamic convolution output data corresponding to the image input data.

3. The method according to claim 2, characterized in that The step of multiplying the plurality of first results by corresponding dynamic convolution coefficients to obtain second results corresponding to the plurality of first results respectively includes: Based on a multiplication operator, multiply the first results by corresponding dynamic convolution coefficients respectively to obtain second results corresponding to the first results respectively; The adding of the plurality of the second results to obtain the dynamic convolution output data corresponding to the image input data comprises: Based on the addition operator, multiple second results are added together to obtain dynamic convolution output data corresponding to the image input data.

4. The method according to claim 2, characterized in that The static convolution operator includes a plurality of target static convolution operators; based on the static convolution operator, the plurality of weights to be adjusted are respectively convolved with the image input data to obtain first results corresponding to the plurality of weights to be adjusted, including: Based on the multiple target static convolution operators, the multiple weights to be adjusted are respectively convolved with the image input data in parallel to obtain first results corresponding to the multiple weights to be adjusted.

5. The method according to claim 4, characterized in that The method of performing convolution operations on the plurality of weights to be adjusted and the image input data in parallel based on the plurality of target static convolution operators to obtain first results corresponding to the plurality of weights to be adjusted respectively includes: If the first number of the convolution operators is greater than or equal to the second number of the multiple weights to be adjusted, then based on the second number of the target static convolution operators, the multiple weights to be adjusted are convolved with the image input data in parallel to obtain first results corresponding to the multiple weights to be adjusted.

6. The method according to claim 4, characterized in that The step of multiplying the plurality of first results by corresponding dynamic convolution coefficients to obtain second results corresponding to the plurality of first results respectively includes: Based on multiple multiplication operators, the multiple first results are multiplied in parallel with the corresponding dynamic convolution coefficients to obtain second results corresponding to the multiple first results.

7. The method according to any one of claims 1 to 6, characterized in that: The obtaining of multiple dynamic convolution coefficients includes: Based on the activation function operator, multiple dynamic convolution coefficients are determined respectively.

8. A dynamic neural network deployment device, characterized in that: include: An acquisition module, used to acquire image input data, a plurality of dynamic convolution coefficients and weights to be adjusted corresponding to the plurality of dynamic convolution coefficients; A determination module is used to determine the dynamic convolution output data of the image input data based on a static convolution operator, according to the image input data, multiple dynamic convolution coefficients and multiple weights to be adjusted, so as to realize the deployment of a dynamic neural network.

9. An electronic device, characterized in that: include: Processor and memory; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory, so that the processor performs the dynamic neural network deployment method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and when the processor executes the computer-executable instructions, the dynamic neural network deployment method according to any one of claims 1 to 7 is implemented.

11. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for dynamic neural network deployment according to any one of claims 1 to 7 is implemented.

Images