Adaptive cloud-edge collaborative inference method and device based on data compression

Abstract

The application provides an adaptive cloud-edge collaborative deduction method and device based on data high compression, wherein the adaptive cloud-edge collaborative deduction method based on data high compression comprises the following steps: extracting features of to-be-transmitted data of a target service to obtain intermediate features corresponding to the to-be-transmitted data; inputting the intermediate features into an encoder to obtain target encoded data and prior information for reflecting the distribution of the intermediate feature data; sending the target encoded data and the prior information to a server; receiving a target result sent by the server; and the target result is a prediction result for the target service output by a neural network model deployed in the server. Through the above method, the feature expression capability of the original encoded data is improved based on a reparameterization convolution layer; the target encoded data is sent to the server, the server can decode accurate target decoded data based on the prior information, and model deduction is performed by using the target decoded data, so that the model deduction accuracy is ensured.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to an adaptive cloud-edge collaborative inference method and apparatus based on high data compression. Background Technology

[0002] With the rapid development of Artificial Intelligence (AI) technology, deep learning has achieved remarkable results in fields such as computer vision, natural language processing, and speech recognition. However, while the performance of AI models continues to improve, the high computational demands on their operation have become a bottleneck for deployment: when performing model inference on a terminal, the terminal's computing power is insufficient, making it difficult for the AI ​​model to run on the terminal; when performing model inference on a cloud server, the terminal needs to transmit a large amount of data to the cloud server, resulting in excessive transmission latency, which cannot meet the high latency requirements of tasks such as autonomous driving and disaster detection.

[0003] To address the issues of insufficient computing power at the terminal or excessive transmission latency at the cloud server, related technologies employ a cloud-edge collaborative model inference framework. This framework involves compressing data at the terminal, followed by decompressing the compressed data at the cloud server and then using the model for inference.

[0004] However, compressing data using a terminal can lead to excessive loss of effective information. When the cloud server uses the decompressed data for model inference, the inference accuracy will decrease, and the accuracy of the model inference cannot be guaranteed.

[0005] Therefore, how to ensure low data transmission latency while improving the model inference accuracy of cloud servers is an important issue that the industry urgently needs to address. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides an adaptive cloud-edge collaborative inference method and apparatus based on high data compression.

[0007] In a first aspect, the present invention provides an adaptive cloud-edge collaborative inference method based on high data compression, comprising:

[0008] Feature extraction is performed on the data to be transmitted for the target service to obtain intermediate features corresponding to the data to be transmitted; the intermediate features are adapted to the data transmission environment at the current moment;

[0009] The intermediate features are input into the encoder to obtain target encoded data and prior information reflecting the distribution of the intermediate feature data; the target encoded data is obtained by encoding based on the encoding layer in the encoder, and the encoding layer includes at least one reparameterized convolutional layer;

[0010] Send the target encoded data and the prior information to the server;

[0011] Receive the target result sent by the server; the target result is the prediction result for the target service output by the neural network model deployed in the server.

[0012] Optionally, the encoder further includes a quantization layer; the quantization layer is used to reduce the amount of data in the original encoded data output by the encoding layer;

[0013] The step of inputting the intermediate features into the encoder to obtain the target encoded data includes:

[0014] The intermediate features are input into the coding layer to obtain the original encoded data and the prior information output by the coding layer;

[0015] The original encoded data is input into the quantization layer to obtain the target encoded data output by the quantization layer.

[0016] Optionally, the method further includes:

[0017] Based on the intermediate features, the prior information is determined; the prior information includes the average value of each channel of the intermediate features and the standard deviation of each channel of the intermediate features.

[0018] Optionally, the step of extracting features from the data to be transmitted to obtain intermediate features corresponding to the data to be transmitted includes:

[0019] Obtain the data transmission environment information at the current moment; the transmission environment information includes bandwidth status, transmission delay limits, and the accuracy limitations of the data transmission environment at the current moment on the neural network model;

[0020] Based on the data transmission environment information, determine the maximum data transmission volume under the current data transmission environment;

[0021] Based on the maximum data transmission volume and the data volume of the data to be transmitted, feature compression parameters are determined; the feature compression parameters are used to control the data transmission volume of the data to be transmitted.

[0022] Based on the feature compression parameters and the accuracy limitations, the intermediate features are extracted.

[0023] Secondly, this invention also provides an adaptive cloud-edge collaborative inference method based on high data compression, applied to a server, comprising:

[0024] The receiving terminal sends target encoded data and prior information reflecting the distribution of intermediate feature data; the intermediate features are obtained by the terminal through feature extraction of the data to be transmitted for the target service; the intermediate features are adapted to the data transmission environment at the current moment.

[0025] The target encoded data is input into the decoder, and the target encoded data is decoded based on the prior information to obtain target decoded data; the decoder includes a decoding layer; the decoding layer includes at least one reparameterized convolutional layer;

[0026] The target decoded data is input into a neural network model for processing to obtain the target result output by the neural network model; the target result is the prediction result for the target business output by the neural network model.

[0027] The target result is sent to the terminal.

[0028] Optionally, the prior information includes the average value of each channel of the intermediate feature and the standard deviation of each channel of the intermediate feature.

[0029] Thirdly, the present invention also provides an adaptive cloud-edge collaborative inference device based on high data compression, applied to a terminal, comprising:

[0030] The feature extraction module is used to extract features from the data to be transmitted for the target service to obtain intermediate features corresponding to the data to be transmitted; the intermediate features are adapted to the data transmission environment at the current moment.

[0031] The first input module is used to input the intermediate features into the encoder to obtain target encoded data and prior information reflecting the distribution of the intermediate feature data; the target encoded data is obtained by encoding based on the encoding layer in the encoder, and the encoding layer includes at least one reparameterized convolutional layer.

[0032] The first sending module is used to send the target encoded data and the prior information to the server;

[0033] The first receiving module is used to receive the target result sent by the server; the target result is the prediction result for the target service output by the neural network model deployed in the server.

[0034] Fourthly, the present invention also provides an adaptive cloud-edge collaborative inference device based on high data compression, applied to a server, comprising:

[0035] The second receiving module is used to receive target encoded data sent by the terminal and prior information reflecting the distribution of intermediate feature data; the intermediate features are obtained by the terminal through feature extraction of the data to be transmitted for the target service; the intermediate features are adapted to the data transmission environment at the current moment.

[0036] The second input module is used to input the target encoded data into the decoder, and decode the target encoded data based on the prior information to obtain target decoded data; the decoder includes a decoding layer; the decoding layer includes at least one reparameterized convolutional layer;

[0037] The third input module is used to input the target decoded data into a neural network model for processing, and obtain the target result output by the neural network model; the target result is the prediction result of the neural network model for the target business.

[0038] The second sending module is used to send the target result to the terminal.

[0039] Fifthly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the adaptive cloud-edge collaborative inference method based on high data compression as described in the first aspect above, or implements the adaptive cloud-edge collaborative inference method based on high data compression as described in the second aspect above.

[0040] In a sixth aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the adaptive cloud-edge collaborative inference method based on high data compression as described in the first aspect above, or implements the adaptive cloud-edge collaborative inference method based on high data compression as described in the second aspect above.

[0041] In a seventh aspect, the present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the adaptive cloud-edge collaborative inference method based on high data compression as described in the first aspect above, or implements the adaptive cloud-edge collaborative inference method based on high data compression as described in the second aspect above.

[0042] The adaptive cloud-edge collaborative inference method and apparatus based on high data compression provided by this invention acquires the data to be transmitted for a target service, extracts features from the data to be transmitted to obtain intermediate features suitable for the data transmission environment at the current moment, inputs the intermediate features into an encoder, and compresses and encodes the intermediate features through at least one reparameterized convolutional layer in the encoder, avoiding the loss of effective information in the intermediate features during compression and improving the feature representation capability of the original encoded data, thereby obtaining the target encoded data and prior information reflecting the distribution of the intermediate feature data, and sending the target encoded data to a server, which can compensate for the data during the decoding process based on the prior information to accurately decode the target decoded data, since the server can decode and restore the target decoded data that is more similar to the data to be transmitted, the server inputs the target decoded data into a neural network model, improving the inference accuracy of the neural network model, and the terminal can then receive the prediction results for the target service sent by the server. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0044] Figure 1 This is one of the flowcharts of the adaptive cloud-edge collaborative inference method based on high data compression provided by the present invention;

[0045] Figure 2 This is the second flowchart of the adaptive cloud-edge collaborative inference method based on high data compression provided by the present invention;

[0046] Figure 3 This is a schematic diagram of the data transmission process using an adaptive cloud-edge collaborative inference system based on high data compression, as provided by the present invention.

[0047] Figure 4 This is one of the structural schematic diagrams of the adaptive cloud-edge collaborative inference device based on high data compression provided by the present invention;

[0048] Figure 5 This is one of the structural schematic diagrams of the adaptive cloud-edge collaborative inference device based on high data compression provided by the present invention;

[0049] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0051] To facilitate a clearer understanding of the various embodiments of this application, some relevant background knowledge will be introduced as follows.

[0052] While the performance of AI models continues to improve, the high computational load they rely on for operation has become a bottleneck for deployment. For example, the third-generation generative pre-trained Transformer (GPT-3) has a staggering 170 billion parameters, which is difficult for some cloud servers to meet its operational requirements, and it is even more difficult to deploy on various mobile devices.

[0053] To address these issues, related technologies can be categorized into terminal-based inference models and cloud-based (i.e., cloud server) inference models. Terminal-based inference models primarily focus on large model compression, employing methods such as model pruning, model distillation, and model quantization to achieve the goal of consuming fewer computing resources, less memory, less disk space, and maintaining minimal accuracy loss. However, these methods cannot fundamentally solve the problem of insufficient computing power on terminal devices.

[0054] For cloud-based inference models, the cloud has abundant computing, power, and storage resources, and is often used to solve the problem of the large computing power required by deep learning models. However, in practical applications, such as emergency disaster relief scenarios, intelligent services need to transmit a large amount of image and video data to the cloud for analysis; but the destruction of many base stations in disaster areas leads to a lack of communication resources, and the private network built in complex terrain and vegetation environments will suffer from scattering, diffraction, and shadows in wireless signal propagation.

[0055] Therefore, it is difficult for the terminal to communicate with the cloud server, the cloud architecture is difficult to adapt to the latency requirements in this scenario, and the terminal device has low computing power and limited energy. Using terminal-only simulation is also difficult to meet the task requirements. Therefore, using only the terminal and using only the cloud architecture are both difficult to meet the needs of scenarios with limited and fluctuating bandwidth resources.

[0056] The cloud-edge collaborative model inference framework can achieve overall low latency and low power consumption on the terminal, and has become a new trend in solving inference problems caused by insufficient terminal computing power or huge cloud parameters. It involves using terminal devices to collect the images to be recognized, transmitting them to a cloud server over the network, and then the cloud server returning the results to the terminal device after recognition.

[0057] However, cloud-edge collaborative methods have certain limitations in related technologies:

[0058] 1) When communication resources are severely scarce, intermediate features obtained by compression methods based on low compression ratios cannot meet the requirements of low latency transmission in the cloud; 2) Excessive compression of intermediate features will lead to excessive loss of effective information and a decrease in model inference accuracy; 3) In inference scenarios with unstable communication links, different latency requirements, and different accuracy requirements, the cloud-edge collaborative framework cannot achieve adaptive inference to adapt to the needs of dynamic bandwidth changes.

[0059] In summary, to address the problems existing in the above-mentioned technologies and to improve the model inference accuracy of cloud servers while ensuring low data transmission latency, this invention provides an adaptive cloud-edge collaborative inference method and apparatus based on high data compression.

[0060] The following is combined with Figures 1-2 The adaptive cloud-edge collaborative inference method based on high data compression provided by this invention is described in detail. Figure 1 This is one of the flowcharts of the adaptive cloud-edge collaborative inference method based on high data compression provided by the present invention. See [link / reference]. Figure 1 As shown, this method is applied to a terminal and includes steps 101-104, wherein:

[0061] Step 101: Extract features from the data to be transmitted for the target service to obtain intermediate features corresponding to the data to be transmitted; the intermediate features are adapted to the data transmission environment at the current moment.

[0062] First, it should be noted that the subject of this invention can be any terminal capable of data transmission, such as a smartphone, smartwatch, desktop computer, laptop, or any other type.

[0063] In order to improve the accuracy of the server's model inference while ensuring low data transmission latency, this embodiment first needs to acquire the data to be transmitted, which can be images, videos, audio, etc.

[0064] In practical applications, whether the neural network model is deduced by the server alone or by using cloud-edge collaboration, the problem of dynamic changes in network bandwidth will be faced. For example, in emergency rescue scenarios in dense forest areas, drones collecting data and edge servers conduct collaborative deduction. Due to the load-bearing geographical environment of the forest area, there are many communication blind spots and bandwidth resources are constantly changing. Cloud-edge collaborative deduction with unknown bandwidth changes will make it difficult to meet constraints such as latency requirements.

[0065] Therefore, after obtaining the data to be transmitted for the target service, it is necessary to extract features from the data to be transmitted to obtain intermediate features corresponding to the data to be transmitted, thereby achieving adaptive perception of network bandwidth changes and obtaining intermediate features that are adapted to the network bandwidth.

[0066] Step 102: Input the intermediate features into the encoder to obtain target encoded data and prior information reflecting the distribution of the intermediate feature data; the target encoded data is obtained by encoding based on the encoding layer in the encoder, and the encoding layer includes at least one reparameterized convolutional layer.

[0067] In this embodiment, after obtaining the intermediate features, the intermediate features need to be input into the encoder to obtain the target encoded data and prior information that reflects the distribution of the intermediate feature data.

[0068] It should be noted that the encoder includes a coding layer; in practical applications, in order to ensure the fitting ability of the coding layer without adding extra parameters, grouped convolution is used to reduce the computational load of the intermediate feature compression process, while ensuring the high feature representation capability of the original encoded data output by the coding layer.

[0069] Specifically, in one possible implementation of the present invention, the coding layer adopts a 6-layer network structure, including 3 reparameterized convolutional layers, 2 first convolutional layers and 1 second convolutional layer, wherein the second convolutional layer plays a dimensionality reduction role.

[0070] The reparameterized convolutional layer consists of a reparameterized convolution, a batch regularization layer, and an activation function in sequence; the second convolutional layer consists of a convolutional layer, a batch regularization layer, and an activation function in sequence.

[0071] The connection methods of each layer in the coding layer are as follows:

[0072] The first layer is a reparameterized convolutional layer, where the input and output channels are the same, the kernel size is 3, the stride size is 1, the padding size is 1, and the grouping parameter is 1 / 4 of the number of intermediate feature channels.

[0073] The second layer is the first convolutional layer, where the kernel size is 1 and the stride size is 1.

[0074] The third layer is a reparameterized convolutional layer, where the input and output channels are the same, the kernel size is 3, the stride size is 1, the padding size is 1, and the grouping parameter is 1 / 2 of the number of intermediate feature channels.

[0075] The fourth layer is the first convolutional layer, where the kernel size is 1 and the stride size is 1.

[0076] The fifth layer is a reparameterized convolutional layer, where the input and output channels are the same, the kernel size is 3, the stride size is 1, the padding size is 1, and the grouping parameter is the number of channels of the intermediate features.

[0077] The sixth layer is the second convolutional layer, where the kernel size is 2, the stride size is 1, and the padding size is 0. It can be understood that the purpose of the second convolutional layer is to reduce the dimensionality of the intermediate feature channels.

[0078] Step 103: Send the target encoded data and the prior information to the server.

[0079] In this embodiment, after obtaining the target encoded data output by the encoder, the target encoded data is sent to the server; accordingly, the server uses the decoder to decode the target encoded data based on the prior information carried in the target data, thereby obtaining the target decoded data.

[0080] Understandably, since prior information is used to reflect the data distribution of intermediate features, the server can decode the target encoded data with the assistance of prior information, thereby performing data compensation on the decoded data and restoring it to obtain target decoded data that is more similar to the data to be transmitted.

[0081] Step 104: Receive the target result sent by the server; the target result is the prediction result for the target service output by the neural network model deployed in the server.

[0082] In this embodiment, after the server encodes the target data to obtain the target decoded data, the server inputs the target decoded data into a neural network model for deduction to obtain the target result output by the neural network model; then the server sends the target result to the terminal.

[0083] It should be noted that since the neural network model is deployed on a server, the computing power of the server can be used to improve the inference accuracy of the neural network model.

[0084] The adaptive cloud-edge collaborative inference method based on high data compression provided by this invention obtains the data to be transmitted for the target service, extracts features from the data to be transmitted to obtain intermediate features suitable for the data transmission environment at the current moment, inputs the intermediate features into an encoder, and compresses and encodes the intermediate features through at least one reparameterized convolutional layer in the encoder, avoiding the loss of effective information in the intermediate features during compression and improving the feature representation capability of the original encoded data, thereby obtaining the target encoded data and prior information reflecting the distribution of intermediate feature data, and sending the target encoded data to the server, which can compensate for the data during the decoding process based on the prior information to accurately decode the target decoded data, since the server can decode and restore the target decoded data that is more similar to the data to be transmitted, the server inputs the target decoded data into a neural network model, improving the inference accuracy of the neural network model, and the terminal can then receive the prediction results for the target service sent by the server.

[0085] Optionally, the encoder further includes a quantization layer; the quantization layer is used to reduce the amount of data in the original encoded data output by the encoding layer;

[0086] The intermediate features are input into the encoder to obtain the target encoded data, which is specifically achieved through the following steps 1)-2):

[0087] Step 1) Input the intermediate features into the coding layer to obtain the original coded data output by the coding layer;

[0088] Step 2) Input the original encoded data into the quantization layer to obtain the target encoded data output by the quantization layer.

[0089] In this embodiment, intermediate features are input into the coding layer to obtain the original encoded data output by the coding layer, which can be specifically represented by the following formula (1):

[0090] embed = Compressor edge (feat) (1)

[0091] Where, embed represents the original encoded data, and the size of embed is n×c'×h'×w', where n represents the batch size; c' represents the channel size of the original encoded data; h' and w' represent the length and width of the feature map of the original encoded data; feature represents the intermediate feature, and the size of feature is n×c×h×w, where n represents the batch size; c represents the channel size of the intermediate feature; h and w represent the length and width of the feature map of the intermediate feature; and c'≤c, h'≤h, w'≤w.

[0092] In this embodiment, the quantization layer is used to reduce the amount of data in the original encoded data output by the encoding layer.

[0093] In other words, after inputting intermediate features into the encoder's encoding layer and obtaining the raw encoded data output by the encoding layer, the raw encoded data is represented using 32-bit floating-point data, which results in a huge amount of data being transmitted.

[0094] Therefore, a quantization layer is needed to select the lowest 1 bit to perform feature quantization on the original encoded data, thereby obtaining the target encoded data. Since the encoding layer has a high fitting capability, the high compression rate of 1-bit quantization will not cause the server-side neural network model to lose too much accuracy.

[0095] Since using 1-bit quantization can reduce the amount of original encoded data to 1 / 32 of the original, it can greatly reduce the amount of original encoded data output by the encoding layer, thereby reducing transmission latency.

[0096] Therefore, after obtaining the original encoded data output by the encoding layer, the original encoded data is input into the quantization layer to reduce the amount of data in the original encoded data and obtain the target encoded data; specifically, this can be expressed by the following formulas (2)-(3):

[0097] embed = tanh(conv(embed)) (2)

[0098]

[0099] Where embed represents the original encoded data; tanh represents the activation function; conv represents a single convolutional layer; embed b Represents the target encoded data; Sign represents the sign function.

[0100] It should be noted that Sign in the quantization layer is a non-differentiable function, which will prevent the model from being trained end-to-end. Therefore, in this embodiment, a Straight-Through Estimator (STE) is used to ensure that the gradient of the quantization process can be backpropagated, that is, the gradient of the layer after the quantization layer is directly passed back to the layer before the quantization layer.

[0101] In the above implementation, intermediate features are input into the encoding layer, which can compress the channel and spatial dimensions of the intermediate features, remove irrelevant information, and compress the intermediate features. This avoids the loss of effective information during the compression process and improves the feature representation capability of the original encoded data. Then, the original encoded data is input into the quantization layer, which can quantize the original encoded data by 1 bit, reducing the amount of data in the original encoded data to 1 / 32 of the original. Therefore, the amount of data in the original encoded data output by the encoding layer can be greatly reduced. By inputting intermediate features into the encoding and quantization layers in the encoder, high compression ratio data compression is achieved, thereby reducing data transmission latency.

[0102] Optionally, in one possible implementation of the present invention, the prior information can be determined based on the intermediate features; the prior information includes the average value of each channel of the intermediate features and the standard deviation of each channel of the intermediate features.

[0103] The average value of each channel of the intermediate feature can be calculated using the following formula (4):

[0104]

[0105] Where, μ c (feat) represents the average value of each channel of the intermediate features; nchw This represents an intermediate feature with dimensions n×c×h×w.

[0106] The standard deviation of each channel of the intermediate feature can be calculated using the following formula (5):

[0107]

[0108] Where, σ c (feat) represents the standard deviation of each channel of the intermediate feature; ε is a minimum value to ensure data stability.

[0109] In the above implementation, the prior information includes the average value and standard deviation of each channel of the intermediate features, thus reflecting the data distribution of the intermediate features. Therefore, with the assistance of prior information, the server can decode the target encoded data, thereby performing data compensation on the decoded data and restoring it to obtain target decoded data that is more similar to the data to be transmitted. The server then inputs the target decoded data into the neural network model, thereby improving the inference accuracy of the neural network model.

[0110] Optionally, since data transmission latency is related to network bandwidth, in order to make data transmission more adaptable to the network bandwidth at the current moment, in one possible implementation of this invention, a feature compression parameter is introduced to control the amount of data transmitted, thereby making it more adaptable to the network bandwidth at the current moment; specifically, this can be achieved through the following steps [1]-[4]:

[0111] Step [1]: Obtain the data transmission environment information at the current moment; the data transmission environment information includes bandwidth status, transmission delay limit, and the accuracy limit of the data transmission environment at the current moment on the neural network model;

[0112] Step [2]: Based on the data transmission environment information, determine the maximum data transmission volume under the current data transmission environment;

[0113] Step [3]: Based on the maximum data transmission volume and the data volume of the data to be transmitted, determine the feature compression parameters; the feature compression parameters are used to control the data transmission volume of the data to be transmitted.

[0114] Step [4]: ​​Based on the feature compression parameters and the precision limit, determine the intermediate feature; the intermediate feature is adapted to the data transmission environment at the current moment.

[0115] In this embodiment, the data transmission environment information at the current moment is first obtained, including the bandwidth status B, the transmission latency limit Latency' at the current moment, and the accuracy limit ACC' of the data transmission environment at the current moment on the neural network model deployed on the server;

[0116] Then, based on the bandwidth state B and the current transmission delay limit Latency', the maximum data transmission amount DS' under the current data transmission environment is determined, which can be expressed by the following formula (6):

[0117] DS'=B·Latency' (6)

[0118] After determining the maximum data transmission volume DS', it is necessary to determine the feature compression parameter S based on the maximum data transmission volume DS' under the current data transmission environment and the data volume DS of the data to be transmitted. The feature compression parameter is used to control the data transmission volume of the data to be transmitted. Specifically, it is expressed by the following formula (7):

[0119] S = DS / DS' (7)

[0120] Since the inference accuracy ACC of the neural network model in the server conforms to the relationship curve ACC=f(S) with the feature compression parameter S, the inference accuracy ACC of the neural network model when the feature compression parameter is S can be predicted by f(S).

[0121] When ACC≥ACC', the bilinear interpolation method is used to adjust the size of the feature space dimension of the data to be transmitted using the feature compression parameter S, so as to obtain intermediate features after bandwidth adaptation, making the intermediate features more suitable for the data transmission environment at the current moment, and at the same time satisfying the inference constraints of the neural network model.

[0122] It should be noted that if ACC < ACC', the cloud-edge collaborative simulation will be degraded to terminal-only simulation to ensure the simulation is completed.

[0123] In the above implementation, for environments with dynamically changing bandwidth, the amount of data transmitted is controlled by feature compression parameters, so that the intermediate features are more adapted to the data transmission environment at the current moment, thereby ensuring that the collaborative inference of the neural network model under different precision limitations is satisfied.

[0124] Figure 2 This is the second flowchart of the adaptive cloud-edge collaborative inference method based on high data compression provided by this invention. See also... Figure 2 As shown, this method is applied to a server and includes steps 201-203, wherein:

[0125] Step 201: Receive target encoded data and prior information reflecting the distribution of intermediate feature data sent by the receiving terminal; the intermediate features are obtained by the terminal through feature extraction of the data to be transmitted for the target service; the intermediate features are adapted to the data transmission environment at the current moment.

[0126] Step 202: Input the target encoded data into the decoder, and decode the target encoded data based on the prior information to obtain target decoded data; the decoder includes a decoding layer; the decoding layer includes at least one reparameterized convolutional layer.

[0127] In this embodiment, after receiving the target encoded data, it is necessary to input the target encoded data into the decoder to obtain the target decoded data.

[0128] It should be noted that the decoder includes a decoding layer; in one possible implementation of the present invention, the decoding layer adopts an 8-layer network structure, including one third convolutional layer, four reparameterized convolutional layers, and three fourth convolutional layers, wherein the third convolutional layer plays the role of dimension restoration; the reparameterized convolutional layer is composed of reparameterized convolution, batch regularization layer and activation function in sequence;

[0129] The connection methods of each layer in the decoding layer are as follows:

[0130] The first layer is the third convolutional layer, where the output channel is 4 times the input channel, the kernel size is 2, the stride size is 1, and the padding size is 0.

[0131] Specifically, the convolutional weights in the first layer use full-precision floating weights embedded with 1-bit target encoded data. b The output of the third convolutional layer is calculated and expressed by the following formula (8):

[0132] embedo ut =conv(weight,embed) b (8)

[0133] Among them, embedo ut Represents the output of the third convolutional layer; weight represents the floating weights in full precision; embed b This represents the target encoded data.

[0134] The second layer is a heavily parameterized convolutional layer, where the kernel size is 3, the stride is 1, the padding is 1, and the grouping parameter is the number of channels of the intermediate features.

[0135] The third layer is the fourth convolutional layer, which is a 1×1 convolution with a kernel size of 1 and a stride of 1.

[0136] The fourth layer is a heavily parameterized convolutional layer, where the kernel size is 3, the stride is 1, the padding is 1, and the grouping parameter is 1 / 2 of the number of intermediate feature channels.

[0137] The fifth layer is the fourth convolutional layer, which consists of a 1×1 convolution with a kernel size of 1 and a stride of 1.

[0138] The sixth layer is a heavily parameterized convolutional layer, where the kernel size is 3, the stride is 1, the padding is 1, and the grouping parameter is 1 / 4 of the number of intermediate feature channels.

[0139] The seventh layer is the fourth convolutional layer, which is a 1×1 convolution with a kernel size of 1 and a stride of 1.

[0140] The eighth layer is a heavily parameterized convolutional layer, with a kernel size of 3, a stride of 1, padding of 1, and a grouping parameter of 1.

[0141] Optionally, the prior information includes the average value of each channel of the intermediate feature and the standard deviation of each channel of the intermediate feature.

[0142] Optionally, the target encoded data can be decoded based on prior information, which can be achieved through the following steps [a]-[b]:

[0143] Step [a]: Decode the target encoded data to obtain the initial decoded data, which can be expressed by the following formula (9):

[0144] feat' = Compressorcloud(embed) b (9)

[0145] Here, 'feat' represents the initial decoded data; 'embed' represents the data embedded in the code. b This represents the target encoded data.

[0146] Step [b]: Based on prior information, perform data compensation on the initial decoded data to obtain the target decoded data; specifically, this can be expressed by the following formula (10):

[0147]

[0148] Here, "feat" represents the target decoded data; σ and μ represent prior information.

[0149] It should be noted that by using the data distribution of intermediate features as prior information, information compensation for the initial decoded data is achieved after decoding and recovery. The mean and variance of the initial feature data are aligned with the mean and variance of the intermediate features before compression. This method transmits a very small number of additional features without requiring additional training or a large amount of computation. It also reduces the learning difficulty during feature compression training, making the target decoded data more similar to the data to be transmitted, thereby effectively reducing the accuracy loss of the neural network model.

[0150] Step 203: Input the target decoded data into a neural network model for processing to obtain the target result output by the neural network model; the target result is the prediction result of the neural network model for the target service.

[0151] Step 204: Send the target result to the terminal.

[0152] The adaptive cloud-edge collaborative inference method based on high data compression provided by this invention involves a server receiving target encoded data sent by a terminal, which carries prior information reflecting the distribution of intermediate feature data. The server then inputs the target encoded data into a decoder and decodes the target decoded data based on the prior information. This allows for data compensation during the decoding process based on the prior information, thereby accurately decoding the target decoded data. Since the server can decode and reconstruct target decoded data that is more similar to the data to be transmitted, the server inputs the target decoded data into a neural network model, improving the inference accuracy of the neural network model.

[0153] Optionally, in one possible implementation of the present invention, when training the encoder (also known as the compression module), the original loss function Loss net Based on this, a loss function, Loss, is added to supervise feature consistency. feat That is, for the task y = Net(x) of the original model, its loss function is expressed by the following formula (11):

[0154] Loss net =Loss origin (Net(x),gt) (11)

[0155] Here, gt represents the actual label corresponding to x.

[0156] For the process of the compression module compressing the intermediate feature "feat" and then restoring it to the target decoded data "feat", the loss function Loss is used. feat Supervision is carried out, specifically expressed by the following formula (12):

[0157] Loss feat =Smooth L1 (feat,feat”) (12)

[0158] The loss function Loss feat It can ensure that when the difference in feature values ​​is small, the gradient value will not be too small during gradient update, and when the difference is large, gradient explosion will not occur, thus ensuring the stability of training.

[0159] The total loss function Loss after inserting the compressed module into the original network is expressed by the following formula (13):

[0160] Loss = Loss net +λ·Loss feat (13)

[0161] Where λ is the feature inhibition loss function Loss feat The weights are given by λ, where λ = 1.

[0162] Figure 3 This is a schematic diagram of the data transmission process using an adaptive cloud-edge collaborative simulation system based on high data compression provided by the present invention. The adaptive cloud-edge collaborative simulation system based on high data compression includes a terminal and a server, and specifically includes the following steps 1 to 7.

[0163] Step 1: The terminal obtains the data to be transmitted for the target service.

[0164] Step 2: The terminal uses the adapter to extract features from the data to be transmitted to obtain intermediate features;

[0165] Specifically, the terminal uses the adapter to extract features from the data to be transmitted, which can be achieved in the following ways:

[0166] (1) Obtain the data transmission environment information at the current moment; wherein, the data transmission environment information includes bandwidth status, transmission delay limit and the accuracy limit of the data transmission environment at the current moment on the neural network model;

[0167] (2) Based on the data transmission environment information, determine the maximum data transmission volume under the current data transmission environment;

[0168] (3) Determine the feature compression parameters based on the maximum data transmission volume and the data volume to be transmitted; wherein, the feature compression parameters are used to control the data transmission volume of the data to be transmitted;

[0169] (4) Based on the feature compression parameters and accuracy limitations, intermediate features are extracted; wherein, the intermediate features are adapted to the data transmission environment at the current moment.

[0170] Step 3: The terminal inputs the intermediate features into the encoder of the encoder to obtain the target encoded data output by the encoder and the prior information used to reflect the distribution of the intermediate feature data.

[0171] Step 4: The terminal sends the target encoded data and prior information to the server.

[0172] Step 5: The server inputs the target encoded data into the decoder and decodes the target encoded data based on prior information to obtain the target decoded data.

[0173] Step 6: The server inputs the target decoded data into the neural network model and obtains the target result output by the neural network model.

[0174] Step 7: The server sends the target result to the terminal.

[0175] The adaptive cloud-edge collaborative inference device based on high data compression provided by the present invention will be described below. The adaptive cloud-edge collaborative inference device based on high data compression described below can be referred to in correspondence with the adaptive cloud-edge collaborative inference method based on high data compression described above. Figure 4 This is one of the structural schematic diagrams of the adaptive cloud-edge collaborative inference device based on high data compression provided by the present invention, as shown below. Figure 4 As shown, the adaptive cloud-edge collaborative inference device 400 based on high data compression, applied to a terminal, includes: an acquisition module 401, a first input module 402, a first sending module 403, and a first receiving module 404, wherein:

[0176] The feature extraction module 401 is used to extract features from the data to be transmitted of the target service to obtain intermediate features corresponding to the data to be transmitted; the intermediate features are adapted to the data transmission environment at the current moment.

[0177] The first input module 402 is used to input the intermediate features into the encoder to obtain target encoded data and prior information reflecting the distribution of the intermediate feature data; the target encoded data is obtained by encoding based on the encoding layer in the encoder, and the encoding layer includes at least one reparameterized convolutional layer.

[0178] The first sending module 403 is used to send the target encoded data and the prior information to the server;

[0179] The first receiving module 404 is used to receive the target result sent by the server; the target result is the prediction result for the target service output by the neural network model deployed in the server.

[0180] The adaptive cloud-edge collaborative inference device based on high data compression provided by this invention acquires the data to be transmitted for the target service, extracts features from the data to be transmitted, and obtains intermediate features suitable for the data transmission environment at the current moment. The intermediate features are input into an encoder, and compressed and encoded by at least one reparameterized convolutional layer in the encoder, avoiding the loss of effective information in the intermediate features during compression and improving the feature representation capability of the original encoded data, thereby obtaining the target encoded data and prior information reflecting the distribution of the intermediate feature data. The target encoded data is sent to a server, which can compensate for the data during the decoding process based on the prior information, thereby accurately decoding the target decoded data. Since the server can decode and restore the target decoded data that is more similar to the data to be transmitted, the server inputs the target decoded data into a neural network model, improving the inference accuracy of the neural network model. The terminal can then receive the prediction results for the target service sent by the server.

[0181] Optionally, the encoder further includes a quantization layer; the quantization layer is used to reduce the amount of data in the original encoded data output by the encoding layer;

[0182] The first input module 402 is further configured to:

[0183] The intermediate features are input into the coding layer to obtain the original encoded data and the prior information output by the coding layer;

[0184] The original encoded data is input into the quantization layer to obtain the target encoded data output by the quantization layer.

[0185] Optionally, the device further includes:

[0186] A determination module is used to determine the prior information based on the intermediate features; the prior information includes the average value of each channel of the intermediate features and the standard deviation of each channel of the intermediate features.

[0187] Optionally, the acquisition module 402 is further configured to:

[0188] Obtain the data transmission environment information at the current moment; the transmission environment information includes bandwidth status, transmission delay limits, and the accuracy limitations of the data transmission environment at the current moment on the neural network model;

[0189] Based on the data transmission environment information, determine the maximum data transmission volume under the current data transmission environment;

[0190] Based on the maximum data transmission volume and the data volume of the data to be transmitted, feature compression parameters are determined; the feature compression parameters are used to control the data transmission volume of the data to be transmitted.

[0191] Based on the feature compression parameters and the accuracy limitations, the intermediate features are extracted.

[0192] Figure 5 This is the second schematic diagram of the adaptive cloud-edge collaborative inference device based on high data compression provided by the present invention, as shown below. Figure 5 As shown, the adaptive cloud-edge collaborative inference device 500 based on high data compression, applied to a server, includes: a second receiving module 501, a second input module 502, a third input module 503, and a second sending module 504, wherein:

[0193] The second receiving module 501 is used to receive target encoded data sent by the terminal and prior information reflecting the distribution of intermediate feature data; the intermediate features are obtained by the terminal through feature extraction of the data to be transmitted for the target service; the intermediate features are adapted to the data transmission environment at the current moment.

[0194] The second input module 502 is used to input the target encoded data into the decoder, and decode the target encoded data based on the prior information to obtain target decoded data; the decoder includes a decoding layer; the decoding layer includes at least one reparameterized convolutional layer;

[0195] The third input module 503 is used to input the target decoded data into a neural network model for processing to obtain the target result output by the neural network model; the target result is the prediction result for the target service output by the neural network model.

[0196] The second sending module 504 is used to send the target result to the terminal.

[0197] The adaptive cloud-edge collaborative inference device based on high data compression provided by this invention receives target encoded data sent by a terminal, which carries prior information reflecting the distribution of intermediate feature data. Then, the target encoded data is input into a decoder, and the target decoded data is decoded based on the prior information. This allows for data compensation during the decoding process based on the prior information, thereby accurately decoding the target decoded data. Since the server can decode and restore the target decoded data to be more similar to the data to be transmitted, the server inputs the target decoded data into the neural network model, improving the inference accuracy of the neural network model.

[0198] Optionally, the prior information includes the average value of each channel of the intermediate feature and the standard deviation of each channel of the intermediate feature.

[0199] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 6 As shown, the electronic device may include a processor 610, a communications interface 620, a memory 630, and a communication bus 640, wherein the processor 610, communications interface 620, and memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute an adaptive cloud-edge collaborative inference method based on high data compression. This method includes: acquiring data to be transmitted for a target service; extracting features from the data to be transmitted to obtain intermediate features corresponding to the data; adapting the intermediate features to the current data transmission environment; inputting the intermediate features into an encoder to obtain target encoded data and prior information reflecting the distribution of the intermediate feature data; the target encoded data is obtained by encoding based on an encoding layer in the encoder, the encoding layer including at least one reparameterized convolutional layer; sending the target encoded data and the prior information to a server; receiving the target result sent by the server; the target result is a prediction result for the target service output by a neural network model deployed in the server.

[0200] or,

[0201] The system receives target encoded data and prior information reflecting the distribution of intermediate feature data sent by a receiving terminal; the intermediate features are obtained by the terminal through feature extraction of the data to be transmitted for the target service; the intermediate features are adapted to the data transmission environment at the current moment; the target encoded data is input into a decoder, and the target encoded data is decoded based on the prior information to obtain target decoded data; the decoder includes a decoding layer; the decoding layer includes at least one reparameterized convolutional layer; the target decoded data is input into a neural network model for processing to obtain the target result output by the neural network model; the target result is the prediction result for the target service output by the neural network model; and the target result is sent to the terminal.

[0202] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0203] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the adaptive cloud-edge collaborative inference method based on high data compression provided by the above methods. The method includes: acquiring data to be transmitted for a target service; extracting features from the data to be transmitted to obtain intermediate features corresponding to the data to be transmitted; the intermediate features being adapted to the data transmission environment at the current moment; inputting the intermediate features into an encoder to obtain target encoded data and prior information reflecting the distribution of the intermediate feature data; the target encoded data being encoded based on an encoding layer in the encoder, the encoding layer including at least one reparameterized convolutional layer; sending the target encoded data and the prior information to a server; receiving a target result sent by the server; the target result being a prediction result for the target service output by a neural network model deployed in the server.

[0204] or,

[0205] The system receives target encoded data and prior information reflecting the distribution of intermediate feature data sent by a receiving terminal; the intermediate features are obtained by the terminal through feature extraction of the data to be transmitted for the target service; the intermediate features are adapted to the data transmission environment at the current moment; the target encoded data is input into a decoder, and the target encoded data is decoded based on the prior information to obtain target decoded data; the decoder includes a decoding layer; the decoding layer includes at least one reparameterized convolutional layer; the target decoded data is input into a neural network model for processing to obtain the target result output by the neural network model; the target result is the prediction result for the target service output by the neural network model; and the target result is sent to the terminal.

[0206] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the adaptive cloud-edge collaborative inference method based on high data compression provided by the methods described above. This method includes: acquiring data to be transmitted for a target service; extracting features from the data to be transmitted to obtain intermediate features corresponding to the data; the intermediate features being adapted to the current data transmission environment; inputting the intermediate features into an encoder to obtain target encoded data and prior information reflecting the distribution of the intermediate feature data; the target encoded data being encoded based on an encoding layer in the encoder, the encoding layer including at least one reparameterized convolutional layer; sending the target encoded data and the prior information to a server; receiving a target result sent by the server; the target result being a prediction result for the target service output by a neural network model deployed in the server.

[0207] or,

[0208] The system receives target encoded data and prior information reflecting the distribution of intermediate feature data sent by a receiving terminal; the intermediate features are obtained by the terminal through feature extraction of the data to be transmitted for the target service; the intermediate features are adapted to the data transmission environment at the current moment; the target encoded data is input into a decoder, and the target encoded data is decoded based on the prior information to obtain target decoded data; the decoder includes a decoding layer; the decoding layer includes at least one reparameterized convolutional layer; the target decoded data is input into a neural network model for processing to obtain the target result output by the neural network model; the target result is the prediction result for the target service output by the neural network model; and the target result is sent to the terminal.

[0209] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0210] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0211] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive cloud-edge collaborative inference method based on high data compression, characterized in that, Applied to terminals, including: Obtain the data transmission environment information of the target service at the current moment; the data transmission environment information includes bandwidth status, transmission latency limits, and the accuracy limitations of the data transmission environment at the current moment on the neural network model in the server; Based on the data transmission environment information, determine the maximum data transmission volume under the data transmission environment at the current moment; Based on the maximum data transmission volume and the data volume of the target service to be transmitted, feature compression parameters are determined; the feature compression parameters are used to control the data transmission volume of the data to be transmitted. Based on the feature compression parameters and the precision limitations, intermediate features corresponding to the data to be transmitted are extracted; the intermediate features are adapted to the data transmission environment at the current moment. The intermediate features are input into the encoder to obtain target encoded data and prior information reflecting the distribution of the intermediate feature data; the target encoded data is obtained by encoding based on the encoding layer in the encoder, and the encoding layer includes at least one reparameterized convolutional layer; Send the target encoded data and the prior information to the server; Receive the target result sent by the server; the target result is the prediction result for the target service output by the neural network model deployed in the server.

2. The adaptive cloud-edge collaborative inference method based on high data compression according to claim 1, characterized in that, The encoder further includes a quantization layer; the quantization layer is used to reduce the amount of data in the original encoded data output by the encoding layer; The step of inputting the intermediate features into the encoder to obtain the target encoded data includes: The intermediate features are input into the coding layer to obtain the original encoded data and the prior information output by the coding layer; The original encoded data is input into the quantization layer to obtain the target encoded data output by the quantization layer.

3. The adaptive cloud-edge collaborative inference method based on high data compression according to claim 2, characterized in that, The method further includes: Based on the intermediate features, the prior information is determined; the prior information includes the average value of each channel of the intermediate features and the standard deviation of each channel of the intermediate features.

4. An adaptive cloud-edge collaborative inference method based on high data compression, characterized in that, Applied to servers, including: The receiving terminal sends target encoded data and prior information that reflects the distribution of intermediate feature data; The target encoded data is input into the decoder, and the target encoded data is decoded based on the prior information to obtain target decoded data; the decoder includes a decoding layer; the decoding layer includes at least one reparameterized convolutional layer; The target decoded data is input into a neural network model for processing to obtain the target result output by the neural network model; the target result is the prediction result for the target business output by the neural network model. The target result is sent to the terminal; The target encoded data and the prior information are obtained by the terminal based on the following steps: acquiring the data transmission environment information of the target service at the current moment; the data transmission environment information includes bandwidth status, transmission latency limits, and the accuracy limits of the neural network model in the server under the current data transmission environment; determining the maximum data transmission volume under the current data transmission environment based on the data transmission environment information; determining feature compression parameters based on the maximum data transmission volume and the data volume of the target service's data to be transmitted; the feature compression parameters are used to control the data transmission volume of the data to be transmitted; extracting intermediate features corresponding to the data to be transmitted based on the feature compression parameters and the accuracy limits; the intermediate features are adapted to the current data transmission environment; inputting the intermediate features into the encoder to obtain the target encoded data and the prior information; the target encoded data is obtained by encoding based on the encoding layer in the encoder, and the encoding layer includes at least one reparameterized convolutional layer.

5. The adaptive cloud-edge collaborative inference method based on high data compression according to claim 4, characterized in that, The prior information includes the average value of each channel of the intermediate feature and the standard deviation of each channel of the intermediate feature.

6. An adaptive cloud-edge collaborative inference device based on high data compression, characterized in that, Applied to terminals, including: A feature extraction module is used to acquire data transmission environment information of the target service at the current moment; the data transmission environment information includes bandwidth status, transmission latency limits, and the accuracy limits of the neural network model in the server under the current data transmission environment; based on the data transmission environment information, the maximum data transmission volume under the current data transmission environment is determined; based on the maximum data transmission volume and the data volume of the target service's data to be transmitted, feature compression parameters are determined; the feature compression parameters are used to control the data transmission volume of the data to be transmitted; based on the feature compression parameters and the accuracy limits, intermediate features corresponding to the data to be transmitted are extracted; the intermediate features are adapted to the data transmission environment at the current moment; The first input module is used to input the intermediate features into the encoder to obtain target encoded data and prior information reflecting the distribution of the intermediate feature data; the target encoded data is obtained by encoding based on the encoding layer in the encoder, and the encoding layer includes at least one reparameterized convolutional layer. The first sending module is used to send the target encoded data and the prior information to the server; The first receiving module is used to receive the target result sent by the server; the target result is the prediction result for the target service output by the neural network model deployed in the server.

7. An adaptive cloud-edge collaborative inference device based on high data compression, characterized in that, Applied to servers, including: The second receiving module is used to receive target encoded data sent by the terminal and prior information reflecting the distribution of intermediate feature data; The second input module is used to input the target encoded data into the decoder, and decode the target encoded data based on the prior information to obtain target decoded data; the decoder includes a decoding layer; the decoding layer includes at least one reparameterized convolutional layer; The third input module is used to input the target decoded data into a neural network model for processing, and obtain the target result output by the neural network model; the target result is the prediction result of the neural network model for the target business. The second sending module is used to send the target result to the terminal; The target encoded data and the prior information are obtained by the terminal based on the following steps: acquiring the data transmission environment information of the target service at the current moment; the data transmission environment information includes bandwidth status, transmission latency limits, and the accuracy limits of the neural network model in the server under the current data transmission environment; determining the maximum data transmission volume under the current data transmission environment based on the data transmission environment information; determining feature compression parameters based on the maximum data transmission volume and the data volume of the target service's data to be transmitted; the feature compression parameters are used to control the data transmission volume of the data to be transmitted; extracting intermediate features corresponding to the data to be transmitted based on the feature compression parameters and the accuracy limits; the intermediate features are adapted to the current data transmission environment; inputting the intermediate features into the encoder to obtain the target encoded data and the prior information; the target encoded data is obtained by encoding based on the encoding layer in the encoder, and the encoding layer includes at least one reparameterized convolutional layer.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the adaptive cloud-edge collaborative inference method based on high data compression as described in any one of claims 1 to 3, or implements the adaptive cloud-edge collaborative inference method based on high data compression as described in any one of claims 4 to 5.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the adaptive cloud-edge collaborative inference method based on high data compression as described in any one of claims 1 to 3, or implements the adaptive cloud-edge collaborative inference method based on high data compression as described in any one of claims 4 to 5.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the adaptive cloud-edge collaborative inference method based on high data compression as described in any one of claims 1 to 3, or implements the adaptive cloud-edge collaborative inference method based on high data compression as described in any one of claims 4 to 5.

Images