Neural network operation method, device and storage medium

Abstract

The embodiments of the present disclosure provide a neural network operation method, device, and storage medium, the method including: adding a random mask layer to a first neural network according to a type of variable dimension to obtain a second neural network, the random mask layer is for performing random mask processing on a tensor input to the random mask layer; and training the second neural network using sample data. The neural network operation method, device, and storage medium according to the embodiments of the present disclosure only need to train and deploy one neural network to adapt to different input / output dimension requirements by adding a random mask layer to the neural network, and the method has low training complexity, low storage overhead, and is easy to deploy and continuously evolve online training.

Description

[Technical field]

[0001] [Cross reference] This application claims priority to a Chinese patent application filed on December 23, 2021, bearing application number 202111593613.7 and entitled "Neural network training and configuration method, apparatus and storage medium", and a Chinese patent application filed on January 13, 2022, bearing application number 202210039452.5 and entitled "Neural network operation method, apparatus and storage medium", both of which are incorporated by reference in their entireties.

[0002] The present disclosure relates to the field of communications technology, and in particular to a neural network operation method, apparatus, and storage medium. [Background technology]

[0003] When the input dimension (the input dimension varies depending on the collected data) and / or output dimension of the neural network are uncertain, it is necessary to train one neural network for every possible input dimension and / or output dimension respectively to accommodate different dimensionality and task needs.

[0004] Figure 1 is a principle diagram of the training method of dimension-variable neural network in the prior art. As shown in Figure 1, the conventional feasible solution is to train one neural network for each different input dimension, and as shown in Figure 1, the system automatically selects the corresponding neural network according to the input dimension to make predictions, and as the input dimension continues to increase (the data collected in real time goes from partial to complete), the prediction value of the neural network gradually changes from "rough" to "accurate".

[0005] However, the number of neural networks in the above technical solutions increases with the increase in the number of possible dimensional situations, the training complexity of the neural network (model) is high, and the memory overhead of the model is large, which is disadvantageous to the actual deployment and continuous evolution of online learning of neural networks. Summary of the Invention [Problem to be solved by the invention]

[0006] The embodiments of the present disclosure provide a neural network operation method, apparatus and storage medium to solve the technical problem that the number of neural networks in the prior art increases with the increase in the possible situations of dimensions. [Means for solving the problem]

[0007] In a first aspect, an embodiment of the present disclosure comprises: According to a type of variable dimension, a random mask layer is added to the first neural network to obtain a second neural network, the random mask layer is for performing a random mask process on the tensor input to the random mask layer; and training the second neural network using sample data.

[0008] In some embodiments, the type of variable dimension is: A type with variable input dimensions; A type in which the output dimension is variable; A type in which the intermediate dimension is variable.

[0009] In some embodiments, adding a random mask layer to the first neural network in response to the type of variable dimension, as described above, comprises: If the variable dimension type is a type in which the input dimension is variable, it includes adding a random mask layer before an input layer of the first neural network.

[0010] In some embodiments, adding a random mask layer to the first neural network in response to the type of variable dimension, as described above, comprises: If the variable dimension type is a type in which the output dimension is variable, it includes adding a random mask layer after the output layer of the first neural network.

[0011] In some embodiments, adding a random mask layer to the first neural network in response to the type of variable dimension, as described above, comprises: If the variable dimension type is a type in which the intermediate dimension is variable, it includes adding a random mask layer to an intermediate layer of the first neural network.

[0012] In some embodiments, the step of adding a random mask layer to an intermediate layer of the first neural network comprises: If only one intermediate dimension is variable, this involves adding one random mask layer before or after the associated intermediate layer.

[0013] In some embodiments, the step of adding a random mask layer to an intermediate layer of the neural network comprises: If multiple intermediate dimensions are variable, this involves adding one random mask layer before or after each of the associated intermediate layers, respectively.

[0014] In some embodiments, training the second neural network using sample data comprises: When the input data dimensions of each sample data are all equal to a first threshold, directly using all the sample data to train the second neural network, wherein the first threshold is the maximum value of the input data dimensions of all the sample data.

[0015] In some embodiments, training the second neural network using sample data comprises: When at least one first sample data exists, the method includes first performing an augmentation process on the first sample data, and then training the second neural network using all of the first sample data and all of the second sample data, wherein the first sample data is sample data whose input data dimension is not a first threshold, the second sample data is sample data whose input data dimension is the first threshold, and the first threshold is the maximum value of the input data dimension of all the sample data.

[0016] In some embodiments, performing the augmentation process on the first sample data includes: The method includes padding the input data of the first sample data with leading zeros.

[0017] In some embodiments, performing the augmentation process on the first sample data includes: The method includes padding lower zeros on the input data of the first sample data.

[0018] In some embodiments, training the second neural network using sample data comprises: When at least one first sample data exists, the method includes first removing the first sample data, and then using all second sample data to train the second neural network, where the first sample data is sample data whose input data dimension is not a first threshold, the second sample data is sample data whose input data dimension is the first threshold, and the first threshold is the maximum value of the input data dimension of all sample data.

[0019] In some embodiments, the step of performing a random mask process on the tensor input to the random mask layer includes: determining a tensor input to the random mask layer and a random mask tensor; Using the random mask tensor, a random mask process is performed on a tensor input to the random mask layer.

[0020] In some embodiments, the formula for performing a random mask operation on a tensor input to the random mask layer is as follows:

number

number

number

number

number

[0021] In some embodiments, the random mask tensor is a hard masking tensor.

[0022] In some embodiments, the random mask tensor is a soft masking tensor.

[0023] In some embodiments, the random mask tensor is a combination of a hard masking tensor and a soft masking tensor.

[0024] In some embodiments, after training the second neural network using sample data as described above, The method further includes replacing the random mask layer in the second neural network after it has been trained with a tensor transform layer, the tensor transform layer being for performing dimensional transformation on the tensor input to the tensor transform layer.

[0025] In some embodiments, replacing the random mask layer in the second neural network after it has been trained with a tensor transform layer as described above includes: If a random mask layer exists before the input layer of the second neural network after being trained, replacing the random mask layer before the input layer of the second neural network after being trained with a tensor padding layer, the tensor padding layer is for increasing a dimension of a tensor input to the tensor padding layer up to a first threshold, the first threshold being a maximum value of a dimension of input data of all sample data.

[0026] In some embodiments, replacing the random mask layer in the second neural network after it has been trained with a tensor transform layer as described above includes: If a random mask layer exists after the output layer of the second neural network after being trained, replacing the random mask layer after the output layer of the second neural network after being trained with a tensor cropping layer, the tensor cropping layer for reducing a dimension of a tensor input to the tensor cropping layer to a first target value.

[0027] In some embodiments, replacing the random mask layer in the second neural network after it has been trained with a tensor transform layer as described above includes: If a random mask layer exists in the intermediate layer of the second neural network after being trained, replacing the random mask layer in the intermediate layer of the second neural network after being trained with a cascaded tensor cropping and tensor padding layer, wherein the cascaded tensor cropping and tensor padding layer is for first reducing a dimension of a tensor input to the cascaded tensor cropping and tensor padding layer to a second target value and then increasing it to an original dimension.

[0028] In a second aspect, an embodiment of the present disclosure comprises: An electronic device is provided that includes a memory for storing a computer program, a transceiver for transmitting and receiving data under control of the processor, and a processor for reading the computer program in the memory and performing operations such as:

[0029] In some embodiments, the type of variable dimension is: A type with variable input dimensions; A type in which the output dimension is variable; A type in which the intermediate dimension is variable.

[0030] In some embodiments, adding a random mask layer to the first neural network in response to the type of variable dimension, as described above, comprises: If the variable dimension type is a type in which the input dimension is variable, it includes adding a random mask layer before an input layer of the first neural network.

[0031] In some embodiments, adding a random mask layer to the first neural network in response to the type of variable dimension, as described above, comprises: If the variable dimension type is a type in which the output dimension is variable, it includes adding a random mask layer after the output layer of the first neural network.

[0032] In some embodiments, adding a random mask layer to the first neural network in response to the type of variable dimension, as described above, comprises: If the variable dimension type is a type in which the intermediate dimension is variable, it includes adding a random mask layer to an intermediate layer of the first neural network.

[0033] In some embodiments, the step of adding a random mask layer to an intermediate layer of the first neural network comprises: If only one intermediate dimension is variable, this involves adding one random mask layer before or after the associated intermediate layer.

[0034] In some embodiments, the step of adding a random mask layer to an intermediate layer of the neural network comprises: If multiple intermediate dimensions are variable, this involves adding one random mask layer before or after each of the associated intermediate layers, respectively.

[0035] In some embodiments, training the second neural network using sample data comprises: When the input data dimensions of each sample data are all equal to a first threshold, directly using all the sample data to train the second neural network, wherein the first threshold is the maximum value of the input data dimensions of all the sample data.

[0036] In some embodiments, training the second neural network using sample data comprises: When at least one first sample data exists, the method includes first performing an augmentation process on the first sample data, and then training the second neural network using all of the first sample data and all of the second sample data, wherein the first sample data is sample data whose input data dimension is not a first threshold, the second sample data is sample data whose input data dimension is the first threshold, and the first threshold is the maximum value of the input data dimension of all the sample data.

[0037] In some embodiments, performing the augmentation process on the first sample data includes: The method includes padding the input data of the first sample data with leading zeros.

[0038] In some embodiments, performing the augmentation process on the first sample data includes: The method includes padding lower zeros on the input data of the first sample data.

[0039] In some embodiments, training the second neural network using sample data comprises: When at least one first sample data exists, the method includes first removing the first sample data, and then using all second sample data to train the second neural network, where the first sample data is sample data whose input data dimension is not a first threshold, the second sample data is sample data whose input data dimension is the first threshold, and the first threshold is the maximum value of the input data dimension of all sample data.

[0040] In some embodiments, the step of performing a random mask process on the tensor input to the random mask layer includes: determining a tensor input to the random mask layer and a random mask tensor; Using the random mask tensor, a random mask process is performed on a tensor input to the random mask layer.

[0041] In some embodiments, the formula for performing a random mask operation on a tensor input to the random mask layer is as follows:

number

number

number

number

number

[0042] In some embodiments, the random mask tensor is a hard masking tensor.

[0043] In some embodiments, the random mask tensor is a soft masking tensor.

[0044] In some embodiments, the random mask tensor is a combination of a hard masking tensor and a soft masking tensor.

[0045] In some embodiments, after training the second neural network using sample data as described above, The method further includes replacing the random mask layer in the second neural network after it has been trained with a tensor transform layer, the tensor transform layer being for performing dimensional transformation on the tensor input to the tensor transform layer.

[0046] In some embodiments, replacing the random mask layer in the second neural network after it has been trained with a tensor transform layer as described above includes: If a random mask layer exists before the input layer of the second neural network after being trained, replacing the random mask layer before the input layer of the second neural network after being trained with a tensor padding layer, the tensor padding layer is for increasing a dimension of a tensor input to the tensor padding layer up to a first threshold, the first threshold being a maximum value of a dimension of input data of all sample data.

[0047] In some embodiments, replacing the random mask layer in the second neural network after it has been trained with a tensor transform layer as described above includes: If a random mask layer exists after the output layer of the second neural network after being trained, replacing the random mask layer after the output layer of the second neural network after being trained with a tensor cropping layer, the tensor cropping layer for reducing a dimension of a tensor input to the tensor cropping layer to a first target value.

[0048] In some embodiments, replacing the random mask layer in the second neural network after it has been trained with a tensor transform layer as described above includes: If a random mask layer exists in the intermediate layer of the second neural network after being trained, replacing the random mask layer in the intermediate layer of the second neural network after being trained with a cascaded tensor cropping and tensor padding layer, wherein the cascaded tensor cropping and tensor padding layer is for first reducing a dimension of a tensor input to the cascaded tensor cropping and tensor padding layer to a second target value and then increasing it to an original dimension.

[0049] In a third aspect, an embodiment of the present disclosure comprises: Obtaining forecast data; and inputting the prediction data into a target neural network that has already been deployed to obtain a prediction result output by the target neural network, the target neural network including at least one tensor transformation layer, the tensor transformation layer being for performing dimensional transformation on a tensor input to the tensor transformation layer.

[0050] In some embodiments, the target neural network is the second neural network of any one of the first aspects.

[0051] In a fourth aspect, an embodiment of the present disclosure provides a method for producing a composition comprising: a memory for storing a computer program; a transceiver for transmitting and receiving data under the control of said processor; and reading the computer program in said memory; Obtaining forecast data; and and a processor for performing operations including: inputting the prediction data into a deployed target neural network to obtain a prediction result output by the target neural network, the target neural network including at least one tensor transform layer, the tensor transform layer for performing dimensional transformation on a tensor input to the tensor transform layer.

[0052] In some embodiments, the target neural network is the second neural network of any one of the first aspects.

[0053] In a fifth aspect, an embodiment of the present disclosure provides a method for producing a composition comprising: an additional module configured to add a random mask layer to the first neural network according to a type of variable dimension to obtain a second neural network, the random mask layer being for performing a random mask operation on a tensor input to the random mask layer; a training module configured to train the second neural network using sample data.

[0054] In some embodiments, the type of variable dimension is: A type with variable input dimensions; A type in which the output dimension is variable; A type in which the intermediate dimension is variable.

[0055] In some embodiments, the additional modules are specifically: The variable dimension type is configured to add a random mask layer before an input layer of the first neural network if the variable dimension type is a type in which the input dimension is variable.

[0056] In some embodiments, the additional modules are specifically: The variable dimension type is configured to add a random mask layer after an output layer of the first neural network if the variable dimension type is a type in which the output dimension is variable.

[0057] In some embodiments, the additional modules are specifically: The variable dimension type is configured to add a random mask layer to an intermediate layer of the first neural network if the variable dimension type is a type in which the intermediate dimension is variable.

[0058] In some embodiments, the additional modules are specifically: If only one intermediate dimension is variable, then it is configured to add one random mask layer before or after the associated intermediate layer.

[0059] In some embodiments, the additional modules are specifically: If multiple intermediate dimensions are variable, then it is configured to add one random mask layer before or after each of the associated intermediate layers, respectively.

[0060] In some embodiments, the training module specifically comprises: When the input data dimensions of each sample data are all equal to a first threshold, the second neural network is trained by directly using all the sample data, and the first threshold is the maximum value of the input data dimensions of all the sample data.

[0061] In some embodiments, the training module specifically comprises: When at least one first sample data exists, the method is configured to first perform an augmentation process on the first sample data, and then train the second neural network using all of the first sample data and all of the second sample data, where the first sample data is sample data whose input data dimension is not a first threshold, the second sample data is sample data whose input data dimension is the first threshold, and the first threshold is the maximum value of the input data dimension of all the sample data.

[0062] In some embodiments, the training module specifically comprises: The input data of the first sample data is configured to be padded with upper zeros.

[0063] In some embodiments, the training module specifically comprises: The input data of the first sample data is configured to be padded with lower zeros.

[0064] In some embodiments, the training module specifically comprises: When at least one first sample data exists, the method is configured to first remove the first sample data, and then use all second sample data to train the second neural network, where the first sample data is sample data whose input data dimension is not a first threshold, the second sample data is sample data whose input data dimension is the first threshold, and the first threshold is the maximum value of the input data dimension of all sample data.

[0065] In some embodiments, the training module specifically comprises: determining a tensor input to the random mask layer and a random mask tensor; and performing a random mask operation on a tensor input to the random mask layer using the random mask tensor.

[0066] In some embodiments, the formula for performing a random mask operation on a tensor input to the random mask layer is as follows:

number

number

number

number

number

[0067] In some embodiments, the random mask tensor is a hard masking tensor.

[0068] In some embodiments, the random mask tensor is a soft masking tensor.

[0069] In some embodiments, the random mask tensor is a combination of a hard masking tensor and a soft masking tensor.

[0070] In some embodiments, The method further includes a placement module configured to replace a random mask layer in the second neural network after it has been trained with a tensor transform layer, the tensor transform layer for performing dimensional transformation on a tensor input to the tensor transform layer.

[0071] In some embodiments, the placement module specifically: If a random mask layer exists before the input layer of the second neural network after training, the random mask layer before the input layer of the second neural network after training is replaced with a tensor padding layer, the tensor padding layer is for increasing a dimension of a tensor input to the tensor padding layer up to a first threshold, the first threshold being a maximum value of a dimension of input data of all sample data.

[0072] In some embodiments, the placement module specifically: If a random mask layer exists after the output layer of the second neural network after training, the random mask layer after the output layer of the second neural network after training is replaced with a tensor cropping layer, the tensor cropping layer for reducing a dimension of a tensor input to the tensor cropping layer to a first target value.

[0073] In some embodiments, the placement module specifically: If a random mask layer exists in the intermediate layer of the second neural network after being trained, the random mask layer in the intermediate layer of the second neural network after being trained is replaced with a cascaded tensor cropping and tensor padding layer, wherein the cascaded tensor cropping and tensor padding layer is for first reducing a dimension of a tensor input to the cascaded tensor cropping and tensor padding layer to a second target value and then increasing it to an original dimension.

[0074] In a sixth aspect, an embodiment of the present disclosure provides a method for producing a composition comprising: an acquisition module configured to acquire forecast data; and a processing module configured to input the prediction data into a deployed target neural network to obtain a prediction result output by the target neural network, the target neural network including at least one tensor transformation layer, the tensor transformation layer for performing dimensional transformation on a tensor input to the tensor transformation layer.

[0075] In some embodiments, the target neural network is the second neural network of any one of the fifth aspects.

[0076] In a seventh aspect, an embodiment of the present disclosure further provides a processor readable storage medium having stored thereon a computer program for causing a processor to perform the steps of the method for operating a neural network according to the first or third aspect as set forth above.

[0077] In an eighth aspect, an embodiment of the present disclosure further provides a computer readable storage medium having stored thereon a computer program for causing a computer to perform the steps of the method for operating a neural network according to the first or third aspect as described above.

[0078] In a ninth aspect, an embodiment of the present disclosure further provides a communication device readable storage medium having stored thereon a computer program for causing a communication device to execute the steps of the neural network operation method according to the first or third aspect as described above.

[0079] In a tenth aspect, an embodiment of the present disclosure further provides a chip product readable storage medium having stored thereon a computer program for causing a chip product to perform steps of the neural network operation method described in the first or third aspect as described above. Effect of the Invention

[0080] The neural network operation method, apparatus, and storage medium according to the embodiments of the present disclosure need only train and deploy one neural network to adapt to different input / output dimensional requirements by adding a random mask layer to the neural network, and the method has low training complexity, low storage overhead, and is easy to deploy and continuously evolve online training. [Brief description of the drawings]

[0081] In order to more clearly describe the embodiments of the present disclosure or the technical solutions in the prior art, the drawings necessary for the description of the embodiments or the prior art will be briefly described below. Of course, the drawings in the following description are some embodiments of the present disclosure, and those skilled in the art can further obtain other drawings based on these drawings without creative labor. [Figure 1] FIG. 1 is a principle diagram of a training method for a dimension variable neural network in the prior art; [Diagram 2] 1 is a flowchart of a method for operating a neural network according to an embodiment of the present disclosure. [Diagram 3] FIG. 1 is a block diagram of a training dimension variable neural network according to an embodiment of the present disclosure. [Figure 4] FIG. 2 is a schematic diagram of a hard masking tensor according to an embodiment of the present disclosure. [Diagram 5] FIG. 1 is a block diagram of a configuration of a dimension variable neural network according to an embodiment of the present disclosure. [Figure 6] FIG. 2 is a schematic diagram of a tensor padding and tensor cropping module according to an embodiment of the present disclosure. [Figure 7] FIG. 1 is a schematic diagram of an autoencoder neural network according to an embodiment of the present disclosure. [Figure 8] FIG. 2 is a schematic diagram of a single-symbol and dual-symbol DMRS according to an embodiment of the present disclosure. [Figure 9] FIG. 1 is a block diagram of an autoencoder based CSI compressed feedback according to an embodiment of the present disclosure. [Figure 10] 1 is a schematic diagram of a structure of an electronic device according to an embodiment of the present disclosure. [Figure 11] FIG. 2 is a schematic diagram of the structure of a neural network operation device according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0082] An artificial neural network is a simplified model established by abstracting the neuron network of the human brain from the perspective of information processing. In academia and industry, an artificial neural network, abbreviated as neural network, is a computational model that is composed of a large number of computational nodes connected to each other in a specific manner, and the model is usually a representation or approximation of an algorithm or a function mapping. In the training stage of a neural network, engineers must first determine the structure and hyperparameters of the neural network, and then feed data into the model and apply gradient backpropagation to update the weights of the trained neural network. The trained neural network is an approximation to the mapping relationship between the input data and the labels, and feeding new data into the neural network in the placement stage can obtain a predicted value. Common neural networks include fully connected forward propagation neural networks, convolutional neural networks, recurrent neural networks, etc. With the improvement of computer performance and the massive deployment of graphic processors (GPUs), neural networks are often used in the fields of computer vision, natural language processing, biomedicine, and automatic control.

[0083] Since the training and deployment of a neural network are usually oriented to a specific task, the input and output dimensions of the neural network are generally determined. A neural network oriented to A task training generally cannot be used for B task, or has poor performance when used for B task. When the input and output dimensions of a neural network are uncertain, it is necessary to train one neural network for all possible input and output dimensions respectively to meet different dimensions and task needs. For example, a system responds based on the predicted value of a deployed neural network, and the input of the neural network is data collected in real time. When the real-time requirement of the system is high, the system hopes that the neural network can make some "rough" prediction results based on some collected data, so that the system can respond in the optimal direction in advance, thereby saving the total time of responding in the optimal direction. A conventional feasible solution is to train one neural network for different input dimensions (input dimensions are different by collecting some data) respectively, as shown in FIG. 1. The system automatically selects the corresponding neural network according to the input dimension to make predictions, and as the input dimension continues to increase (the data collected in real time goes from partial to complete), the prediction value of the neural network gradually changes from "rough" to "accurate."

[0084] When the input and output dimensions of a neural network are uncertain, the prior art requires multiple neural networks to be deployed to meet different input and output dimension requirements. The number of neural networks increases with the increase in the possible situations of the dimensions, the model training complexity is high, and the model storage overhead is large, which is unfavorable to the actual deployment and continuous evolution of online learning of neural networks.

[0085] Based on the above technical problems, the embodiments of the present disclosure provide a neural network operation method, apparatus, and storage medium, which needs to train and deploy only one neural network to adapt to different input / output dimensional requirements by adding a random mask layer to the neural network, and has low training complexity, low storage overhead, and is easy to deploy and continuously evolve online training.

[0086] In order to make the objectives, technical solutions and advantages of the embodiments of the present disclosure clearer, the technical solutions of the embodiments of the present disclosure will be described below clearly and completely with reference to the drawings in the embodiments of the present disclosure, and of course, the described embodiments are only some embodiments of the present disclosure, and are not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present disclosure without creative labor belong to the scope of protection of the present disclosure.

[0087] 2 is a flowchart of a neural network operation method according to an embodiment of the present disclosure. As shown in FIG. 2, an embodiment of the present disclosure provides a neural network operation method, the execution body of which may be an electronic device, such as a terminal, a base station, a core network element, etc. The method includes steps 201 and 202, In step 201, according to a type of variable dimension, a random mask layer is added to the first neural network to obtain a second neural network, the random mask layer is for performing random mask processing on the tensor input to the random mask layer; In step 202, the second neural network is trained using sample data.

[0088] In some embodiments, the type of variable dimension is: A type with variable input dimensions; A type in which the output dimension is variable; A type in which the intermediate dimension is variable.

[0089] In some embodiments, adding a random mask layer to the first neural network in response to a type of the variable dimension comprises: If the variable dimension type is a type in which the input dimension is variable, it includes adding a random mask layer before the input layer of the first neural network.

[0090] In some embodiments, adding a random mask layer to the first neural network in response to a type of the variable dimension comprises: If the variable dimension type is a type in which the output dimension is variable, it includes adding a random mask layer after the output layer of the first neural network.

[0091] In some embodiments, adding a random mask layer to the first neural network in response to a type of the variable dimension comprises: If the variable dimension type is a type in which the intermediate dimension is variable, it includes adding a random mask layer to an intermediate layer of the first neural network.

[0092] In some embodiments, adding a random mask layer to an intermediate layer of the first neural network comprises: If only one intermediate dimension is variable, this involves adding one random mask layer before or after the associated intermediate layer.

[0093] In some embodiments, adding a random mask layer to an intermediate layer of the neural network comprises: If multiple intermediate dimensions are variable, this involves adding one random mask layer before or after each of the associated intermediate layers, respectively.

[0094] In some embodiments, training the second neural network using sample data comprises: When the input data dimensions of each sample data are all equal to a first threshold, directly using all the sample data to train the second neural network, where the first threshold is the maximum value of the input data dimensions of all the sample data.

[0095] In some embodiments, training the second neural network using sample data comprises: When at least one first sample data exists, the method includes first performing an augmentation process on the first sample data, and then training the second neural network using all the first sample data and all the second sample data, wherein the first sample data is sample data whose input data dimension is not a first threshold, the second sample data is sample data whose input data dimension is the first threshold, and the first threshold is the maximum value of the input data dimension of all the sample data.

[0096] In some embodiments, performing augmentation processing on the first sample data includes: This includes padding the input data of the first sample data with leading zeros.

[0097] In some embodiments, performing augmentation processing on the first sample data includes: The method includes padding lower zeros on the input data of the first sample data.

[0098] In some embodiments, training the second neural network using sample data comprises: If at least one first sample data exists, the method includes first removing the first sample data, and then using all second sample data to train the second neural network, where the first sample data is sample data whose input data dimension is not a first threshold, and the second sample data is sample data whose input data dimension is the first threshold, and the first threshold is the maximum value of the input data dimension of all sample data.

[0099] In some embodiments, performing a random mask operation on a tensor input to the random mask layer comprises: determining a tensor input to the random mask layer and a random mask tensor; Utilizing the random mask tensor to perform a random mask operation on a tensor input to the random mask layer.

[0100] In some embodiments, the formula for performing the random masking operation on the tensors input to the random mask layer is:

number

number

number

number

number

[0101] In some embodiments, the random mask tensor is a hard masking tensor.

[0102] In some embodiments, the random mask tensor is a soft masking tensor.

[0103] In some embodiments, the random mask tensor is a combination of a hard masking tensor and a soft masking tensor.

[0104] In some embodiments, after training the second neural network using the sample data, The method further includes replacing the random mask layer in the second neural network after it has been trained with a tensor transform layer, the tensor transform layer being for performing dimensional transformation on tensors input to the tensor transform layer.

[0105] In some embodiments, replacing the random mask layer with a tensor transform layer in the second neural network after it has been trained comprises: If a random mask layer exists before the input layer of the second neural network after training, replacing the random mask layer before the input layer of the second neural network after training with a tensor padding layer, the tensor padding layer is for increasing a dimension of a tensor input to the tensor padding layer up to a first threshold, the first threshold being a maximum value of the dimension of input data of all sample data.

[0106] In some embodiments, replacing the random mask layer with a tensor transform layer in the second neural network after it has been trained comprises: If a random mask layer is present after the output layer of the second neural network after being trained, replacing the random mask layer after the output layer of the second neural network after being trained with a tensor cropping layer, the tensor cropping layer for reducing a dimension of a tensor input to the tensor cropping layer to a first target value.

[0107] In some embodiments, replacing the random mask layer with a tensor transform layer in the second neural network after it has been trained comprises: If a random mask layer exists in the intermediate layer of the second neural network after being trained, replacing the random mask layer in the intermediate layer of the second neural network after being trained with a cascaded tensor cropping and tensor padding layer, wherein the cascaded tensor cropping and tensor padding layer is for first reducing a dimension of a tensor input to the cascaded tensor cropping and tensor padding layer to a second target value and then increasing the dimension to the original dimension.

[0108] An embodiment of the present disclosure determines the input tensor dimensions and output tensor dimensions of the neural input layer, output layer, and some intermediate layers based on the maximum dimension of the training data, the maximum dimension of the labels, and the task requirements.

[0109] 3 is a block diagram of a dimension variable neural network training according to an embodiment of the present disclosure, in the training stage, a random mask module (also called a "random mask layer") is added to the layer corresponding to the tensor that needs dimension variable to train the network, as shown in FIG. 3. Specifically, the following steps may be included: 1. If the input dimension of the neural network is variable, add a random mask module to the left of the input layer. 2. If the output dimension of the neural network is variable, add a random mask module to the right of the output layer. 3. When the output dimension of the i-th hidden layer and the input dimension of the i+1-th hidden layer in a neural network are variable, a random mask module is added between hidden layer i and hidden layer i+1. 4. When multiple dimension variable requirements exist simultaneously in the input layer, output layer, and hidden layer of the neural network, multiple random mask modules are added at the corresponding positions. 5. On the left side of the input layer, the training data may adopt only the maximum dimensional data (if non-maximum dimensional data exists, these data can be removed first), or a data augmentation process may be first performed before training to increase the low-dimensional input data to the maximum dimension by using an upper zero padding method or a lower zero padding method.

[0110] After adding the random mask module, the neural network updates the trained neural network by calculating the gradient value of the loss function with respect to the weights of the neural network. The random mask operation (random masking process) is performed on the tensor

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

[0111] 1. Hard Masking

number

number

number

number

number

number

number

number

number

number

number

number

[0112] 2. Soft-Masking For example, exponentially decaying soft masking is

number

number

number

number

number

number

number

[0113] 3.

number

[0114] When the mask tensor acts on a tensor, it is equivalent to performing a weighting operation on the tensor without changing the dimension of the tensor itself.

number

[0115] For example, for first-order tensors (vectors), hard masking

number

number

number

number

number

number

[0116] During the training process, the optimizer calculates gradient backpropagation based on the loss and updates the weight parameters of the neural network.

[0117] In addition, FIG. 3 may include the following: 1. The neural network may be, but is not limited to, a convolutional neural network, a forward propagation fully connected neural network, a recurrent neural network, and combinations thereof. 2. The loss function may be, but is not limited to, mean squared error, 1-norm, categorical cross entropy, and cosine similarity. 3. The random mask module avoids the problem of vanishing or non-differentiable gradients by customizing the gradient and treating the mask tensor as a constant tensor, among other methods. 4. The input random mask, the output random mask and the intermediate layer random mask may exist simultaneously, or only some of them may exist. The number of random masks to be used may be determined according to the specific task.

[0118] 5 is a block diagram of a dimension-variable neural network according to an embodiment of the present disclosure, in the configuration stage, as shown in FIG. 5, the dimension-variable neural network is configured by removing the random mask module from the trained neural network and adding a tensor padding module (also called "tensor padding layer") and a tensor cropping module (also called "tensor cropping layer") to the trained neural network. Specifically, the configuration may include the following steps: 1. Replace the random mask module on the left side of the input layer with a tensor padding module, and pad the input tensor up to the maximum input dimension using lower or upper zero padding scheme. 2. Replace the random mask module on the right side of the output layer with a tensor cropping module, and output the tensor according to the upper or lower cropping method. 3. Replace the random mask module between hidden layer i and hidden layer i+1 with a cascaded tensor cropping and tensor padding module.

[0119] In the whole process, only one neural network needs to be trained and deployed to adapt to different input / output dimensional requirements, avoiding the problems of training and storing multiple neural networks.

[0120] As shown in Figure 5, a variable dimension neural network is constructed based on the trained neural network by removing the random mask module and replacing it with a tensor padding and tensor cropping module. As shown in Figure 6, the replaced tensor padding and tensor cropping modules are

number

number

number

number

[0121] Output Tensor

number

number

number

[0122] It can be seen that the system only needs to deploy one neural network, the output tensor is cropped as necessary, and the input tensor is simply zero-padded and padded to meet different input and output dimensional requirements.

[0123] The following further illustrates the above method with some specific examples.

[0124] Example 1: As shown in FIG. 7, the neural network is an autoencoder used for image compression, in which the encoder module is used to compress the image, and the decoder module is used to restore the image. The input of the encoder and the output of the decoder are both

number

number

number

[0125] Example 2: As in the first embodiment, an autoencoder is used for image compression, and the input of the encoder and the output of the decoder are both

number

number

number

number

number

number

number

[0126] Example 3: In the channel estimation algorithm based on artificial intelligence, the receiving side estimates the channel coefficients on all time-frequency resource blocks based on the received value of the demodulation reference signal (DMRS), and the channel recovery algorithm is realized based on the neural network. For a single time slot (14 OFDM symbols), 8 physical resource blocks (96 frequencies), the single-symbol DMRS and dual-symbol DMRS configurations are as shown in Figure 8. The single-symbol DMRS configuration corresponds to a neural network whose input dimension is 48×1×2 (48 corresponds to frequency, 1 corresponds to time, and 2 corresponds to the real and imaginary parts of the complex number), and the dual-symbol DMRS configuration (one DMRS symbol is further configured on top of the single-symbol configuration) corresponds to a neural network whose input dimension is 48×2×2 and whose output dimension is 96×14×2. The maximum input dimension of the neural network for channel estimation is

number

number

[0127] Example 4: Similar to Example 3, for a single time slot (14 OFDM symbols) and 8 physical resource blocks (96 frequencies), the single-symbol DMRS configuration corresponds to a neural network with input dimensions of 48×1×2, and the dual-symbol DMRS configuration corresponds to a neural network with input dimensions of 48×2×2. The maximum input dimension of the neural network for channel estimation is

number

number

number

[0128] Example 5: As shown in Figure 9, the artificial intelligence-based channel state information (CSI) compression algorithm is realized by using an autoencoder, where the encoder is located on the UE side and the decoder is located on the base station side. The UE side estimates the channel H based on the CSI reference signal (RS) to calculate the precoding vector V, and the dimension size of the precoding vector V is:

number

number

number

number

number

[0129] The neural network operation method provided by the embodiments of the present disclosure only needs to train and deploy one neural network to adapt to different input / output dimensional requirements, and the method has low training complexity, low memory overhead, and is easy to deploy and continuously evolve online training.

[0130] FIG. 10 is a schematic diagram of a structure of an electronic device according to an embodiment of the present disclosure. As shown in FIG. 10, the electronic device includes: A memory 1020 for storing a computer program, a transceiver 1000 for transmitting and receiving data under the control of the processor 1010, and reading the computer program in the memory 1020, According to a type of variable dimension, a random mask layer is added to the first neural network to obtain a second neural network, the random mask layer is for performing a random mask process on the tensor input to the random mask layer; and training the second neural network using sample data.

[0131] Specifically, the transceiver 1000 is used to receive and transmit data under the control of a processor 1010 .

[0132] Here, in FIG. 10, the bus architecture may include any number of interconnected buses and bridges, specifically, various circuits of one or more processors represented by processor 1010 and memory represented by memory 1020 are linked together. The bus architecture may also link together various other circuits, such as peripheral devices, voltage regulators, power management circuits, etc., all of which are well known in the art and will not be described further herein. The bus interface provides an interface. The transceiver 1000 may be multiple elements, i.e., includes a transmitter and a receiver, providing a unit for communicating with various other devices over a transmission medium. These transmission media include transmission media such as wireless channels, wired channels, optical cables, etc. The processor 1010 manages the bus architecture and normal processing, and the memory 1020 may store data used by the processor 1010 in performing operations.

[0133] The processor 1010 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or a complex programmable logic device (CPLD), and the processor may employ a multi-core architecture.

[0134] In some embodiments, the type of variable dimension is: A type with variable input dimensions; A type in which the output dimension is variable; A type in which the intermediate dimension is variable.

[0135] In some embodiments, adding a random mask layer to the first neural network in response to the type of variable dimension, as described above, comprises: If the variable dimension type is a type in which the input dimension is variable, it includes adding a random mask layer before an input layer of the first neural network.

[0136] In some embodiments, adding a random mask layer to the first neural network in response to the type of variable dimension, as described above, comprises: If the variable dimension type is a type in which the output dimension is variable, it includes adding a random mask layer after the output layer of the first neural network.

[0137] In some embodiments, adding a random mask layer to the first neural network in response to the type of variable dimension, as described above, comprises: If the variable dimension type is a type in which the intermediate dimension is variable, it includes adding a random mask layer to an intermediate layer of the first neural network.

[0138] In some embodiments, the step of adding a random mask layer to an intermediate layer of the first neural network comprises: If only one intermediate dimension is variable, this involves adding one random mask layer before or after the associated intermediate layer.

[0139] In some embodiments, the step of adding a random mask layer to an intermediate layer of the neural network comprises: If multiple intermediate dimensions are variable, this involves adding one random mask layer before or after each of the associated intermediate layers, respectively.

[0140] In some embodiments, training the second neural network using sample data comprises: When the input data dimensions of each sample data are all equal to a first threshold, directly using all the sample data to train the second neural network, wherein the first threshold is the maximum value of the input data dimensions of all the sample data.

[0141] In some embodiments, training the second neural network using sample data comprises: When at least one first sample data exists, the method includes first performing an augmentation process on the first sample data, and then training the second neural network using all of the first sample data and all of the second sample data, wherein the first sample data is sample data whose input data dimension is not a first threshold, the second sample data is sample data whose input data dimension is the first threshold, and the first threshold is the maximum value of the input data dimension of all the sample data.

[0142] In some embodiments, performing the augmentation process on the first sample data includes: The method includes padding the input data of the first sample data with leading zeros.

[0143] In some embodiments, performing the augmentation process on the first sample data includes: The method includes padding lower zeros on the input data of the first sample data.

[0144] In some embodiments, training the second neural network using sample data comprises: When at least one first sample data exists, the method includes first removing the first sample data, and then using all second sample data to train the second neural network, where the first sample data is sample data whose input data dimension is not a first threshold, the second sample data is sample data whose input data dimension is the first threshold, and the first threshold is the maximum value of the input data dimension of all sample data.

[0145] In some embodiments, the step of performing a random mask process on the tensor input to the random mask layer includes: determining a tensor input to the random mask layer and a random mask tensor; Using the random mask tensor, a random mask process is performed on a tensor input to the random mask layer.

[0146] In some embodiments, the formula for performing a random mask operation on a tensor input to the random mask layer is as follows:

number

number

number

number

number

[0147] In some embodiments, the random mask tensor is a hard masking tensor.

[0148] In some embodiments, the random mask tensor is a soft masking tensor.

[0149] In some embodiments, the random mask tensor is a combination of a hard masking tensor and a soft masking tensor.

[0150] In some embodiments, after training the second neural network using sample data as described above, The method further includes replacing the random mask layer in the second neural network after it has been trained with a tensor transform layer, the tensor transform layer being for performing dimensional transformation on the tensor input to the tensor transform layer.

[0151] In some embodiments, replacing the random mask layer in the second neural network after it has been trained with a tensor transform layer as described above includes: If a random mask layer exists before the input layer of the second neural network after being trained, replacing the random mask layer before the input layer of the second neural network after being trained with a tensor padding layer, the tensor padding layer is for increasing a dimension of a tensor input to the tensor padding layer up to a first threshold, the first threshold being a maximum value of a dimension of input data of all sample data.

[0152] In some embodiments, replacing the random mask layer in the second neural network after it has been trained with a tensor transform layer as described above includes: If a random mask layer exists after the output layer of the second neural network after being trained, replacing the random mask layer after the output layer of the second neural network after being trained with a tensor cropping layer, the tensor cropping layer for reducing a dimension of a tensor input to the tensor cropping layer to a first target value.

[0153] In some embodiments, replacing the random mask layer in the second neural network after it has been trained with a tensor transform layer as described above includes: If a random mask layer exists in the intermediate layer of the second neural network after being trained, replacing the random mask layer in the intermediate layer of the second neural network after being trained with a cascaded tensor cropping and tensor padding layer, wherein the cascaded tensor cropping and tensor padding layer is for first reducing a dimension of a tensor input to the cascaded tensor cropping and tensor padding layer to a second target value and then increasing it to an original dimension.

[0154] Specifically, the electronic device according to the embodiment of the present disclosure can realize all the method steps realized by the method embodiment in which the execution body is an electronic device, and can achieve the same technical effects, and details of the same parts and beneficial effects of this embodiment as those of the above method embodiments will be omitted here.

[0155] An embodiment of the present disclosure comprises: Obtaining forecast data; and The present invention further provides a neural network operation method, comprising: inputting the prediction data into a deployed target neural network to obtain a prediction result output by the target neural network, the target neural network including at least one tensor transformation layer, the tensor transformation layer for performing dimension transformation on a tensor input to the tensor transformation layer.

[0156] In some embodiments, the target neural network is the second neural network of any one of the above embodiments.

[0157] The neural network operation method provided by the embodiments of the present disclosure only needs to train and deploy one neural network to adapt to different input / output dimensional requirements, and the method has low training complexity, low memory overhead, and is easy to deploy and continuously evolve online training.

[0158] An embodiment of the present disclosure comprises: a memory for storing a computer program; a transceiver for transmitting and receiving data under the control of said processor; and reading the computer program in said memory; Obtaining forecast data; and and a processor for performing operations including: inputting the prediction data into a deployed target neural network to obtain a prediction result output by the target neural network, the target neural network including at least one tensor transform layer, the tensor transform layer for performing dimensional transformation on a tensor input to the tensor transform layer.

[0159] Specifically, the transceiver is used to receive and transmit data under the control of a processor.

[0160] The bus architecture may include any number of interconnected buses and bridges, specifically, various circuits, one or more processors represented by the processor, and memory represented by the memory, are linked together. The bus architecture may also link together various other circuits, such as peripheral devices, voltage regulators, power management circuits, etc., all of which are well known in the art and will not be described further herein. The bus interface provides an interface. The transceiver may be multiple elements, i.e., includes a transmitter and a receiver, and provides a unit for communicating with various other devices over a transmission medium. These transmission media include transmission media such as wireless channels, wired channels, optical cables, etc. The processor manages the bus architecture and normal processing, and the memory may store data used by the processor to perform operations.

[0161] The processor may be a central processing unit (CPU), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or a complex programmable logic device (CPLD), and the processor may employ a multi-core architecture.

[0162] In some embodiments, the target neural network is the second neural network of any one of the above embodiments.

[0163] The electronic device provided by the embodiments of the present disclosure only needs to train and configure one neural network to adapt to different input / output dimensional requirements, and the method has low training complexity, low storage overhead, and is easy to configure and continuously evolve online training.

[0164] FIG. 11 is a schematic diagram of the structure of a neural network operation device according to an embodiment of the present disclosure. As shown in FIG. 11, the embodiment of the present disclosure includes: The present invention provides a neural network operation device, the device including: an additional module 1101 configured to add a random mask layer to a first neural network according to a type of variable dimension to obtain a second neural network, the random mask layer being for performing a random mask process on a tensor input to the random mask layer; and a training module 1102 configured to train the second neural network using sample data.

[0165] In some embodiments, the type of variable dimension is: A type with variable input dimensions; A type in which the output dimension is variable; A type in which the intermediate dimension is variable.

[0166] In some embodiments, the additional modules are specifically: The variable dimension type is configured to add a random mask layer before an input layer of the first neural network if the variable dimension type is a type in which the input dimension is variable.

[0167] In some embodiments, the additional modules are specifically: The variable dimension type is configured to add a random mask layer after an output layer of the first neural network if the variable dimension type is a type in which the output dimension is variable.

[0168] In some embodiments, the additional modules are specifically: The variable dimension type is configured to add a random mask layer to an intermediate layer of the first neural network if the variable dimension type is a type in which the intermediate dimension is variable.

[0169] In some embodiments, the additional modules are specifically: If only one intermediate dimension is variable, then it is configured to add one random mask layer before or after the associated intermediate layer.

[0170] In some embodiments, the additional modules are specifically: If multiple intermediate dimensions are variable, then it is configured to add one random mask layer before or after each of the associated intermediate layers, respectively.

[0171] In some embodiments, the training module specifically comprises: When the input data dimensions of each sample data are all equal to a first threshold, the second neural network is trained by directly using all the sample data, and the first threshold is the maximum value of the input data dimensions of all the sample data.

[0172] In some embodiments, the training module specifically comprises: When at least one first sample data exists, the method is configured to first perform an augmentation process on the first sample data, and then train the second neural network using all of the first sample data and all of the second sample data, where the first sample data is sample data whose input data dimension is not a first threshold, the second sample data is sample data whose input data dimension is the first threshold, and the first threshold is the maximum value of the input data dimension of all the sample data.

[0173] In some embodiments, the training module specifically comprises: The input data of the first sample data is configured to be padded with upper zeros.

[0174] In some embodiments, the training module specifically comprises: The input data of the first sample data is configured to be padded with lower zeros.

[0175] In some embodiments, the training module specifically comprises: When at least one first sample data exists, the method is configured to first remove the first sample data, and then use all second sample data to train the second neural network, where the first sample data is sample data whose input data dimension is not a first threshold, the second sample data is sample data whose input data dimension is the first threshold, and the first threshold is the maximum value of the input data dimension of all sample data.

[0176] In some embodiments, the training module specifically comprises: determining a tensor input to the random mask layer and a random mask tensor; and performing a random mask operation on a tensor input to the random mask layer using the random mask tensor.

[0177] In some embodiments, the formula for performing a random mask operation on a tensor input to the random mask layer is as follows:

number

number

number

number

number

[0178] In some embodiments, the random mask tensor is a hard masking tensor.

[0179] In some embodiments, the random mask tensor is a soft masking tensor.

[0180] In some embodiments, the random mask tensor is a combination of a hard masking tensor and a soft masking tensor.

[0181] In some embodiments, The method further includes a placement module configured to replace a random mask layer in the second neural network after it has been trained with a tensor transform layer, the tensor transform layer for performing dimensional transformation on a tensor input to the tensor transform layer.

[0182] In some embodiments, the placement module specifically: If a random mask layer exists before the input layer of the second neural network after training, the random mask layer before the input layer of the second neural network after training is replaced with a tensor padding layer, the tensor padding layer is for increasing a dimension of a tensor input to the tensor padding layer up to a first threshold, the first threshold being a maximum value of a dimension of input data of all sample data.

[0183] In some embodiments, the placement module specifically: If a random mask layer exists after the output layer of the second neural network after training, the random mask layer after the output layer of the second neural network after training is replaced with a tensor cropping layer, the tensor cropping layer for reducing a dimension of a tensor input to the tensor cropping layer to a first target value.

[0184] In some embodiments, the placement module specifically: If a random mask layer exists in the intermediate layer of the second neural network after being trained, the random mask layer in the intermediate layer of the second neural network after being trained is replaced with a cascaded tensor cropping and tensor padding layer, wherein the cascaded tensor cropping and tensor padding layer is for first reducing a dimension of a tensor input to the cascaded tensor cropping and tensor padding layer to a second target value and then increasing it to an original dimension.

[0185] Specifically, the neural network operation device according to the embodiment of the present disclosure can realize all the method steps realized by the embodiment of the method in which the execution body is an electronic device, and can achieve the same technical effects, and here, details of the same parts and beneficial effects of this embodiment as those of the above method embodiments will be omitted.

[0186] An embodiment of the present disclosure comprises: an acquisition module configured to acquire forecast data; and a processing module configured to input the prediction data into a deployed target neural network to obtain a prediction result output by the target neural network, where the target neural network includes at least one tensor transformation layer, the tensor transformation layer for performing dimension transformation on a tensor input to the tensor transformation layer.

[0187] In some embodiments, the target neural network is the second neural network of any one of the above embodiments.

[0188] The neural network operation device provided by the embodiment of the present disclosure only needs to train and configure one neural network to adapt to different input / output dimensional requirements, and the method has low training complexity, low memory overhead, and is easy to configure and continuously evolve online training.

[0189] In addition, the division into units in the embodiments of the present disclosure is merely a schematic and logical functional division, and other division methods may be used in actual implementation. In addition, in each embodiment of the present disclosure, each functional unit may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit. The above integrated units may be realized in the form of hardware or software functional units.

[0190] The above integrated units can be realized in the form of a software functional unit and stored in a processor-readable storage medium when sold or used as an independent product. Based on this understanding, the technical proposal of the present disclosure may be embodied in the form of a software product, either in its essence or in the part that contributes to the prior art, or in all or part of the technical means. The computer software product is stored in a storage medium and includes some instructions that cause a computer device (which may be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the method according to each embodiment of the present disclosure. The storage medium includes various media capable of storing program code, such as a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0191] In some embodiments, there is further provided a computer readable storage medium having stored thereon a computer program for causing a computer to execute the steps of the method for operating a neural network according to each of the above method embodiments.

[0192] Specifically, the computer-readable storage medium according to the embodiment of the present disclosure can realize all the method steps realized by the above-mentioned method embodiments and can achieve the same technical effects, and details of the same parts and beneficial effects of this embodiment as those of the above-mentioned method embodiments will be omitted here.

[0193] In addition, the computer-readable storage medium may be any available medium or data storage device that can be accessed by a processor, including, but not limited to, magnetic memory (e.g., flexible disk, hard disk, magnetic tape, magneto-optical disk (MO), etc.), optical memory (e.g., CD, DVD, BD, HVD, etc.), semiconductor memory (e.g., ROM, EPROM, EEPROM, non-volatile memory (NAND FLASH), solid-state drive (SSD)), etc.

[0194] In addition, the terms "first", "second", etc. in the embodiments of the present disclosure are intended to distinguish between similar objects, and are not intended to describe a specific order or sequence. The terms used in this manner are interchangeable where appropriate, so that the embodiments of the present disclosure can be implemented in an order other than that illustrated or described herein, and it should be understood that the objects distinguished by "first" and "second" are generally of the same type, and the number of objects is not limited, for example, the first object may be one or more.

[0195] In addition, the term "and / or" in the embodiments of the present disclosure describes a relation between related objects and indicates that three kinds of relations may exist. For example, A and / or B may indicate three cases: A exists alone, A and B exist simultaneously, and B exists alone. The character " / " generally indicates that the related objects before and after are in an "or" relation.

[0196] The term "plurality" in the presently disclosed embodiments means two or more than two and other quantifiers similar thereto.

[0197] The technical solutions according to the embodiments of the present disclosure are applicable to various systems, particularly 5G systems. For example, applicable systems include a global system of mobile communication (GSM) system, a code division multiple access (CDMA) system, a wideband code division multiple access (WCDMA), a general packet radio service (GPRS) system, a long term evolution (LTE) system, a LTE frequency division duplex (FDD) system, a LTE time division duplex (TDD) system, a long term evolution advanced (LTE-A) system, a universal mobile telecommunication system (UMTS), a worldwide interoperability for microwave access (WiMAX) system, and a 5G New Radio (NR) system. All of these various systems include a terminal device and a network device. The system may also include a core network portion, such as an evolved packet system (EPS) or a 5G system (5GS).

[0198] The terminal equipment according to the embodiment of the present disclosure may be a device that provides voice and / or data connectivity to a user, a handheld device with wireless connectivity, or other processing device connected to a wireless modem. In different systems, the name of the terminal equipment may be different, for example, in a 5G system, the terminal equipment may be called User Equipment (UE). The wireless terminal equipment may communicate with one or more core networks (CN) via a Radio Access Network (RAN), and the wireless terminal equipment may be a mobile terminal equipment such as a mobile phone (also called a "cellular" phone) or a computer with a mobile terminal equipment, for example, a mobile device that is portable, pocket, handheld, has a built-in computer, or is mounted on a vehicle, which exchanges language and / or data with the radio access network. For example, the devices may include a Personal Communication Service (PCS) phone, a cordless phone, a Session Initiated Protocol (SIP) phone, a Wireless Local Loop (WLL) station, a Personal Digital Assistant (PDA), and the like. A wireless terminal device may also be referred to as a system, a user unit, a subscriber station, a mobile station, a mobile, a remote station, an access point, a remote terminal, an access terminal, a user terminal, a user agent, or a user device, and is not limited to these in the embodiments of the present disclosure.

[0199] The network equipment according to the embodiment of the present disclosure may be a base station that can include multiple cells serving terminals. Depending on the specific application scenario, the base station may also be called an access point, a device that communicates with wireless terminal equipment over one or more sectors over an air interface in an access network, or other names. The network equipment may be used as a router between the wireless terminal equipment and the rest of the access network to exchange received air frames and Internet Protocol (IP) packets with each other, where the rest of the access network may include an Internet Protocol (IP) communication network. The network equipment may also coordinate attribute management of the air interface. For example, the network device according to the embodiment of the present disclosure may be a network device (BTS: Base Transceiver Station) in Global System for Mobile communications (GSM) or Code Division Multiple Access (CDMA), a network device (Node B) in Wide-band Code Division Multiple Access (WCDMA (registered trademark)), an evolved network device (eNB or e-NodeB: evolutionary Node B) in a long term evolution (LTE) system, a 5G base station (gNB) in a 5G network architecture (next generation system), a home evolved Node B (HeNB), a relay node, a home base station (femto), a pico base station (pico), or the like, and is not limited to the embodiment of the present disclosure.In some network architectures, the network equipment may include a centralized unit (CU) node and a distributed unit (DU) node, and the centralized unit and the distributed unit may be located geographically separated.

[0200] Between the network device and the terminal device, multi-input multi-output (MIMO) transmission can be performed using one or more antennas, and the MIMO transmission may be single user MIMO (SU-MIMO) or multi-user MIMO (MU-MIMO). Depending on the form and number of antenna combinations, the MIMO transmission may be 2D-MIMO, 3D-MIMO, FD-MIMO or massive-MIMO, and may be diversity transmission, precoding transmission, beam focusing transmission, etc.

[0201] As will be appreciated by those skilled in the art, the embodiments of the present disclosure may be provided as a method, a system, or a computer program product. Thus, the present disclosure may take the form of a fully hardware embodiment, a fully software embodiment, or an embodiment combining software and hardware. And, the present disclosure may take the form of a computer program product embodied in one or more computer usable storage media (including, but not limited to, magnetic disk memory, optical memory, etc.) containing computer usable program code.

[0202] The present disclosure will be described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present disclosure. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be realized by computer-executable instructions. These computer-executable instructions are provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, and the instructions executed by the processor of the computer or other programmable data processing device generate an apparatus for implementing the functions specified in one or more flows of the flowcharts and / or one or more blocks of the block diagrams.

[0203] These processor-executable instructions may be stored in a processor-readable memory that can cause a computer or other programmable data processing device to operate in a particular manner such that the instructions stored in the processor-readable memory cause an article of manufacture including instruction means that implements a function or functions specified in one or more flows in the flowcharts and / or one or more blocks in the block diagrams.

[0204] These processor-executable instructions may be loaded into a computer or other programmable data processing device to cause the computer or other programmable device to perform a series of operational steps to generate a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more flows in the flowcharts and / or one or more blocks in the block diagrams.

[0205] Of course, those skilled in the art can make various changes and modifications to the present disclosure without departing from the spirit and scope of the present disclosure. Thus, if these changes and modifications of the present disclosure are included within the scope of the claims of the present disclosure and their equivalents, the present disclosure is also intended to include these changes and modifications. [Explanation of symbols]

[0206] 1000 Transmitters 1010 Processor 1020 Memory 1101 Additional Module 1102 Training Module

Claims

1. According to the type of the variable dimension, a random mask layer is provided in the first neural network to obtain a second neural network, the random mask layer being for performing a random mask process on the tensor input to the random mask layer; training said second neural network using sample data.

2. The type of the variable dimension is: A type with variable input dimensions; A type in which the output dimension is variable; A type in which the intermediate dimension is variable.

2. A method for operating a neural network as claimed in claim 1.

3. According to the type of variable dimension, providing a random mask layer in the first neural network includes: When the variable dimension type is a type in which the input dimension is variable, the method further comprises providing a random mask layer before an input layer of the first neural network.

3. A method for operating a neural network as claimed in claim 2.

4. According to the type of variable dimension, providing a random mask layer in the first neural network includes: When the variable dimension type is a type in which the output dimension is variable, the method further comprises providing a random mask layer after the output layer of the first neural network.

3. A method for operating a neural network as claimed in claim 2.

5. According to the type of variable dimension, providing a random mask layer in the first neural network includes: When the variable dimension type is a type in which an intermediate dimension is variable, the method includes providing a random mask layer in an intermediate layer of the first neural network.

3. A method for operating a neural network as claimed in claim 2.

6. The provision of a random mask layer in the intermediate layer of the first neural network includes: If only one intermediate dimension is variable, the method includes providing one random mask layer before or after the associated intermediate layer.

6. A method for operating a neural network according to claim 5.

7. The provision of a random mask layer in the intermediate layer of the first neural network includes: When multiple intermediate dimensions are variable, the method includes providing one random mask layer before or after each associated intermediate layer, respectively.

6. A method for operating a neural network according to claim 5.

8. Training the second neural network using sample data, as described above, and directly using all the sample data to train the second neural network when the dimensions of the input data of each sample data are equal to a first threshold value, wherein the first threshold value is the maximum value of the dimensions of the input data of all the sample data.

2. A method for operating a neural network as claimed in claim 1.

9. Training the second neural network using sample data, as described above, When at least one first sample data exists, the method includes first performing a reinforcement process on the first sample data, and then training the second neural network using all of the first sample data and all of the second sample data, wherein the first sample data is sample data whose input data dimension is not equal to a first threshold, the second sample data is sample data whose input data dimension is equal to the first threshold, and the first threshold is a maximum value of the input data dimension of all the sample data.

2. A method for operating a neural network as claimed in claim 1.

10. The above-mentioned, performing the reinforcement process on the first sample data, and performing upper zero padding on the input data of the first sample data.

10. A method for operating a neural network according to claim 9.

11. The above-mentioned, performing the reinforcement process on the first sample data, and performing lower zero padding on the input data of the first sample data.

10. A method for operating a neural network according to claim 9.

12. Training the second neural network using sample data, as described above, When at least one first sample data exists, the first sample data is first removed, and then all second sample data are used to train the second neural network, wherein the first sample data is sample data whose input data dimension is not equal to a first threshold, the second sample data is sample data whose input data dimension is equal to the first threshold, and the first threshold is a maximum value of the input data dimension of all sample data.

2. A method for operating a neural network as claimed in claim 1.

13. The random mask processing for the tensor input to the random mask layer is determining a tensor input to the random mask layer and a random mask tensor; and performing a random mask process on the tensor input to the random mask layer using the random mask tensor.

2. A method for operating a neural network as claimed in claim 1.

14. The above-mentioned formula for performing random mask processing on the tensor input to the random mask layer is as follows: [0010] Where: [0025] denotes the tensor output by the random mask layer, [0030] denotes the tensor input to the random mask layer, [0045] denotes the bitwise multiplication operation on tensors, [0050] denotes a random mask tensor.

14. A method for operating a neural network according to claim 13.

15. The random mask tensor is a hard masking tensor.

14. A method for operating a neural network according to claim 13.

16. The random mask tensor is a soft masking tensor.

14. A method for operating a neural network according to claim 13.

17. The random mask tensor is a combination of a hard masking tensor and a soft masking tensor.

14. A method for operating a neural network according to claim 13.

18. After training the second neural network using the sample data as described above, The method further includes updating the random mask layer in the second neural network after training to a tensor transformation layer, the tensor transformation layer being for performing a dimension transformation on a tensor input to the tensor transformation layer.

2. A method for operating a neural network as claimed in claim 1.

19. The above-mentioned updating of the random mask layer in the second neural network after training to a tensor transform layer may include: If a random mask layer exists before the input layer of the second neural network after training, the random mask layer before the input layer of the second neural network after training is updated to a tensor padding layer, the tensor padding layer is for increasing a dimension of a tensor input to the tensor padding layer up to a first threshold, the first threshold being a maximum value of the dimension of input data of all sample data.

20. A method for operating a neural network according to claim 18.

20. The above-mentioned updating of the random mask layer in the second neural network after training to a tensor transform layer may include: and updating the random mask layer after the output layer of the second neural network after training to a tensor cropping layer when a random mask layer exists after the output layer of the second neural network after training, the tensor cropping layer being for reducing a dimension of a tensor input to the tensor cropping layer to a first target value.

20. A method for operating a neural network according to claim 18.

21. The above-mentioned updating of the random mask layer in the second neural network after training to a tensor transform layer may include: and updating the random mask layer of the intermediate layer of the trained second neural network to a cascaded tensor cropping and tensor padding layer when a random mask layer exists in the intermediate layer of the trained second neural network, the cascaded tensor cropping and tensor padding layer being for first reducing a dimension of a tensor input to the cascaded tensor cropping and tensor padding layer to a second target value and then increasing the dimension to the original dimension.

20. A method for operating a neural network according to claim 18.

22. a memory for storing a computer program; a transceiver for transmitting and receiving data under control of said processor; Reading the computer program in the memory; According to the type of the variable dimension, a random mask layer is provided in the first neural network to obtain a second neural network, the random mask layer being for performing a random mask process on the tensor input to the random mask layer; and training the second neural network using sample data.

23. The type of the variable dimension is: A type with variable input dimensions; A type in which the output dimension is variable; A type in which the intermediate dimension is variable.

23. The electronic device of claim 22.

24. According to the type of variable dimension, providing a random mask layer in the first neural network includes: When the variable dimension type is a type in which the input dimension is variable, the method further comprises providing a random mask layer before an input layer of the first neural network.

24. The electronic device of claim 23.

25. According to the type of variable dimension, providing a random mask layer in the first neural network includes: When the variable dimension type is a type in which the output dimension is variable, the method further comprises providing a random mask layer after the output layer of the first neural network.

24. The electronic device of claim 23.

26. According to the type of variable dimension, providing a random mask layer in the first neural network includes: When the variable dimension type is a type in which an intermediate dimension is variable, the method includes providing a random mask layer in an intermediate layer of the first neural network.

24. The electronic device of claim 23.

27. The provision of a random mask layer in the intermediate layer of the first neural network includes: If only one intermediate dimension is variable, the method includes providing one random mask layer before or after the associated intermediate layer.

27. The electronic device of claim 26.

28. The provision of a random mask layer in the intermediate layer of the first neural network includes: When multiple intermediate dimensions are variable, the method includes providing one random mask layer before or after each associated intermediate layer, respectively.

27. The electronic device of claim 26.

29. Training the second neural network using sample data, as described above, and directly using all the sample data to train the second neural network when the dimensions of the input data of each sample data are equal to a first threshold value, wherein the first threshold value is the maximum value of the dimensions of the input data of all the sample data.

23. The electronic device of claim 22.

30. Training the second neural network using sample data, as described above, When at least one first sample data exists, the method includes first performing a reinforcement process on the first sample data, and then training the second neural network using all of the first sample data and all of the second sample data, wherein the first sample data is sample data whose input data dimension is not equal to a first threshold, the second sample data is sample data whose input data dimension is equal to the first threshold, and the first threshold is a maximum value of the input data dimension of all the sample data.

23. The electronic device of claim 22.

31. The above-mentioned, performing the reinforcement process on the first sample data, and performing upper zero padding on the input data of the first sample data.

31. The electronic device of claim 30.

32. The above-mentioned, performing the reinforcement process on the first sample data, and performing lower zero padding on the input data of the first sample data.

31. The electronic device of claim 30.

33. Training the second neural network using sample data, as described above, When at least one first sample data exists, the first sample data is first removed, and then all second sample data are used to train the second neural network, wherein the first sample data is sample data whose input data dimension is not equal to a first threshold, the second sample data is sample data whose input data dimension is equal to the first threshold, and the first threshold is a maximum value of the input data dimension of all sample data.

23. The electronic device of claim 22.

34. The random mask processing for the tensor input to the random mask layer is determining a tensor input to the random mask layer and a random mask tensor; and performing a random mask process on the tensor input to the random mask layer using the random mask tensor.

23. The electronic device of claim 22.

35. The above-mentioned formula for performing random mask processing on the tensor input to the random mask layer is as follows: [006] Where: [0070] denotes the tensor output by the random mask layer, [0080] denotes the tensor input to the random mask layer, [0097] denotes the bitwise multiplication operation on tensors, [0010] denotes a random mask tensor.

35. The electronic device of claim 34.

36. The random mask tensor is a hard masking tensor.

35. The electronic device of claim 34.

37. The random mask tensor is a soft masking tensor.

35. The electronic device of claim 34.

38. The random mask tensor is a combination of a hard masking tensor and a soft masking tensor.

35. The electronic device of claim 34.

39. After training the second neural network using the sample data as described above, The method further includes updating the random mask layer in the second neural network after training to a tensor transformation layer, the tensor transformation layer being for performing a dimension transformation on a tensor input to the tensor transformation layer.

23. The electronic device of claim 22.

40. The above-mentioned updating of the random mask layer in the second neural network after training to a tensor transform layer may include: If a random mask layer exists before the input layer of the second neural network after training, the random mask layer before the input layer of the second neural network after training is updated to a tensor padding layer, the tensor padding layer is for increasing a dimension of a tensor input to the tensor padding layer up to a first threshold, the first threshold being a maximum value of the dimension of input data of all sample data.

40. The electronic device of claim 39.

41. The above-mentioned updating of the random mask layer in the second neural network after training to a tensor transform layer may include: and updating the random mask layer after the output layer of the second neural network after training to a tensor cropping layer when a random mask layer exists after the output layer of the second neural network after training, the tensor cropping layer being for reducing a dimension of a tensor input to the tensor cropping layer to a first target value.

40. The electronic device of claim 39.

42. The above-mentioned updating of the random mask layer in the second neural network after training to a tensor transform layer may include: and updating the random mask layer of the intermediate layer of the trained second neural network to a cascaded tensor cropping and tensor padding layer when a random mask layer exists in the intermediate layer of the trained second neural network, the cascaded tensor cropping and tensor padding layer being for first reducing a dimension of a tensor input to the cascaded tensor cropping and tensor padding layer to a second target value and then increasing the dimension to the original dimension.

40. The electronic device of claim 39.

43. Obtaining forecast data; and inputting the prediction data into a pre-configured target neural network to obtain a prediction result output by the target neural network, the target neural network including at least one tensor transformation layer, the tensor transformation layer for performing dimensional transformation on a tensor input to the tensor transformation layer.

44. The target neural network is a second neural network according to any one of claims 1 to 21.

44. A method of operating a neural network as claimed in claim 43.

45. a memory for storing a computer program; a transceiver for transmitting and receiving data under control of said processor; Reading the computer program in the memory; Obtaining forecast data; and and inputting the prediction data into a deployed target neural network to obtain a prediction result output by the target neural network, the target neural network including at least one tensor transform layer, the tensor transform layer for performing a dimensional transformation on a tensor input to the tensor transform layer.

46. The target neural network is a second neural network according to any one of claims 1 to 21.

46. ​​The electronic device of claim 45.

47. an additional module configured to provide a random mask layer to the first neural network to obtain a second neural network depending on the type of variable dimension, the random mask layer being for performing a random mask operation on a tensor input to the random mask layer; a training module configured to train the second neural network using sample data.

48. The type of the variable dimension is: A type with variable input dimensions; A type in which the output dimension is variable; A type in which the intermediate dimension is variable.

48. Apparatus for operating a neural network according to claim 47.

49. Specifically, the additional module includes: The variable dimension type is configured to add a random mask layer before an input layer of the first neural network when the variable dimension type is a type in which the input dimension is variable.

49. Apparatus for operating a neural network according to claim 48.

50. Specifically, the additional module includes: When the variable dimension type is a type in which the output dimension is variable, a random mask layer is added after the output layer of the first neural network.

49. Apparatus for operating a neural network according to claim 48.

51. Specifically, the additional module includes: When the variable dimension type is a type in which an intermediate dimension is variable, a random mask layer is added to an intermediate layer of the first neural network.

49. Apparatus for operating a neural network according to claim 48.

52. Specifically, the additional module includes: If only one intermediate dimension is variable, then it is configured to add one random mask layer before or after the associated intermediate layer.

52. Apparatus for operating a neural network according to claim 51.

53. Specifically, the additional module includes: When multiple intermediate dimensions are variable, the method is configured to add one random mask layer before or after each associated intermediate layer, respectively.

52. Apparatus for operating a neural network according to claim 51.

54. The training module specifically includes: The method is configured to directly use all sample data to train the second neural network when the dimensions of the input data of each sample data are equal to a first threshold value, and the first threshold value is a maximum value of the dimensions of the input data of all sample data.

48. Apparatus for operating a neural network according to claim 47.

55. The training module specifically includes: When at least one first sample data exists, the method is configured to first perform an augmentation process on the first sample data, and then train the second neural network using all of the first sample data and all of the second sample data, the first sample data being sample data whose input data dimension is not equal to a first threshold, the second sample data being sample data whose input data dimension is equal to the first threshold, and the first threshold being a maximum value of the input data dimension of all the sample data.

48. Apparatus for operating a neural network according to claim 47.

56. The training module specifically includes: The method is characterized in that the method is configured to perform upper zero padding on the input data of the first sample data.

56. Apparatus for operating a neural network according to claim 55.

57. The training module specifically includes: The method is characterized in that the method is configured to perform lower zero padding on the input data of the first sample data.

56. Apparatus for operating a neural network according to claim 55.

58. The training module specifically includes: The method is configured to first remove the first sample data when at least one first sample data exists, and then use all second sample data to train the second neural network, the first sample data being sample data whose input data dimension is not a first threshold, the second sample data being sample data whose input data dimension is the first threshold, and the first threshold being a maximum value of the input data dimension of all sample data.

48. Apparatus for operating a neural network according to claim 47.

59. The training module specifically includes: determining a tensor input to the random mask layer and a random mask tensor; and performing a random mask process on a tensor input to the random mask layer using the random mask tensor.

48. Apparatus for operating a neural network according to claim 47.

60. The above-mentioned formula for performing random mask processing on the tensor input to the random mask layer is as follows: ##EQU00011## Where: ##EQU00012## denotes the tensor output by the random mask layer, ##EQU00013## denotes the tensor input to the random mask layer, ##EQU14## denotes the bitwise multiplication operation on tensors, ##EQU00015## denotes a random mask tensor.

60. Apparatus for operating a neural network as claimed in claim 59.

61. The random mask tensor is a hard masking tensor.

60. Apparatus for operating a neural network as claimed in claim 59.

62. The random mask tensor is a soft masking tensor.

60. Apparatus for operating a neural network as claimed in claim 59.

63. The random mask tensor is a combination of a hard masking tensor and a soft masking tensor.

60. Apparatus for operating a neural network as claimed in claim 59.

64. and a placement module configured to replace a random mask layer in the second neural network after it has been trained with a tensor transform layer, the tensor transform layer being for performing a dimensional transformation on a tensor input to the tensor transform layer.

48. Apparatus for operating a neural network according to claim 47.

65. The arrangement module specifically includes: The method is configured to replace the random mask layer before the input layer of the second neural network after training with a tensor padding layer if a random mask layer exists before the input layer of the second neural network after training, the tensor padding layer being for increasing a dimension of a tensor input to the tensor padding layer up to a first threshold, the first threshold being a maximum value of the dimension of input data of all sample data.

65. Apparatus for operating a neural network according to claim 64.

66. The arrangement module specifically includes: and replacing the random mask layer after the output layer of the second neural network after the training with a tensor cropping layer if a random mask layer exists after the output layer of the second neural network after the training, the tensor cropping layer being for reducing a dimension of a tensor input to the tensor cropping layer to a first target value.

65. Apparatus for operating a neural network according to claim 64.

67. The arrangement module specifically includes: and replacing the random mask layer in the intermediate layer of the second neural network after training with a cascade tensor cropping and tensor padding layer when a random mask layer exists in the intermediate layer of the second neural network after training, the cascade tensor cropping and tensor padding layer being configured to first reduce a dimension of a tensor input to the cascade tensor cropping and tensor padding layer to a second target value and then increase the dimension to the original dimension.

65. Apparatus for operating a neural network according to claim 64.

68. an acquisition module configured to acquire forecast data; and a processing module configured to input the prediction data into a deployed target neural network to obtain a prediction result output by the target neural network, the target neural network including at least one tensor transform layer, the tensor transform layer for performing dimensional transformation on a tensor input to the tensor transform layer.

69. The target neural network is a second neural network according to any one of claims 47 to 67.

69. Apparatus for operating a neural network as claimed in claim 68.

70. A computer-readable storage medium having stored thereon a computer program for causing a computer to execute the method for operating a neural network according to any one of claims 1 to 21 or the method for operating a neural network according to any one of claims 43 to 44.

Images