Training method of bandwidth estimation model, bandwidth estimation method and related equipment

Abstract

The invention provides a training method of a bandwidth estimation model, a bandwidth estimation method and related equipment. The training method comprises the following steps: acquiring a training data set; inputting the training feature data into a pre-processing network to obtain input feature data suitable for a large language model; inputting the input feature data into a large language model to obtain output feature data used for representing a bandwidth estimation value; inputting the output characteristic data into a post-processing network to obtain a bandwidth estimation value; and training the bandwidth estimation model based on the bandwidth estimation value and the corresponding bandwidth label value. Therefore, the available bandwidth of the link in the multimedia transmission system is estimated by utilizing the huge parameter quantity and the strong generation and reasoning capability of the large language model (LLM). By designing the bandwidth estimation model formed by combining the large language model and the pre-processing / post-processing network, the accuracy and generalization of bandwidth estimation in each complex scene can be improved.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and more specifically, to a training method for a bandwidth estimation model, a bandwidth estimation method, and related devices. Background Art

[0002] Currently, available bandwidth estimation plays an important role in many streaming media applications. For example, in on-demand systems, short video consumption relies on available bandwidth estimation to select the video clarity level that the network can carry to maximize the user experience; short video publishing relies on available bandwidth estimation to maximize the video publishing speed to reduce user waiting time and improve user creation experience; in real-time communication (RTC) systems, available bandwidth estimation is relied on to control the encoder to output the video stream with the best bit rate to provide users with low latency and high-definition call experience. Since available bandwidth estimation is the cornerstone of many streaming media transmission algorithms, how to more accurately estimate the available bandwidth in streaming media transmission systems has always been one of the core issues in the audio and video field.

[0003] In recent years, many bandwidth estimation methods have emerged, including smoothing statistics with artificially designed rules (for example, sliding exponential average, harmonic mean, and Kalman filter), heuristic algorithms combined with features, and learning methods based on deep neural networks, etc. However, different business scenarios may have different requirements for the accuracy and emphasis of bandwidth estimation, and the existing bandwidth estimation methods cannot cope with bandwidth estimation tasks in a variety of different business scenarios. Summary of the invention

[0004] Examples of the present disclosure provide a training method for a bandwidth estimation model, a bandwidth estimation method, and related devices.

[0005] According to a first aspect of the present disclosure, a training method for a bandwidth estimation model is provided, the bandwidth estimation model comprising a pre-processing network, a large language model and a post-processing network, the training method comprising: obtaining a training data set, the training data set comprising training feature data related to a bandwidth estimation task; inputting the training feature data into the pre-processing network to obtain input feature data suitable for the large language model; inputting the input feature data into the large language model to obtain output feature data for characterizing a bandwidth estimation value, wherein the large language model is a pre-trained model with an output layer removed; inputting the output feature data into the post-processing network to obtain the bandwidth estimation value; and training the bandwidth estimation model based on the bandwidth estimation value and a corresponding bandwidth label value.

[0006] Optionally, the pre-processing network is implemented by a feature processing network and a fully connected layer, wherein the step of inputting the training feature data into the pre-processing network to obtain input feature data suitable for the large language model comprises: converting the training feature data into a feature vector through the feature processing network; and encoding the feature vector into a feature input tensor suitable for the large language model through the fully connected layer.

[0007] Optionally, the training feature data includes discrete feature data and sequence feature data, and the feature processing network includes an embedding processing network, a feature extraction network and a merging network, wherein the converting of the training feature data into a feature vector through the feature processing network includes: embedding the discrete feature data through the embedding processing network to convert the discrete feature data into a first feature vector; extracting features from the sequence feature data through the feature extraction network to convert the sequence feature data into a second feature vector; and merging the first feature vector and the second feature vector through the merging network to obtain a merged feature vector.

[0008] Optionally, the post-processing network is implemented by at least one fully connected layer, wherein inputting the output feature data into the post-processing network to obtain the bandwidth estimation value includes: mapping the output feature data to the bandwidth estimation value through the at least one fully connected layer.

[0009] Optionally, the training the bandwidth estimation model based on the bandwidth estimation value and the corresponding bandwidth label value includes: using a large language model fine-tuning technology to train the bandwidth estimation model based on the bandwidth estimation value and the corresponding bandwidth label value.

[0010] Optionally, the training feature data includes at least one feature data of user feature data, device feature data and network feature data related to the bandwidth estimation task.

[0011] Optionally, the training feature data includes feature data obtained under at least one condition of a target business scenario, a target device type, and a target network type; the bandwidth estimation model obtained through training is applied to a bandwidth estimation task under the at least one condition.

[0012] Optionally, the training feature data includes feature data obtained under at least one of multiple business scenarios, multiple device types and multiple network types; the bandwidth estimation model obtained through training is applied to the bandwidth estimation task under the at least one condition.

[0013] According to a second aspect of the present disclosure, a bandwidth estimation method based on a bandwidth estimation model is provided, wherein the bandwidth estimation model includes a pre-processing network, a large language model and a post-processing network, and the bandwidth estimation method includes: acquiring feature data related to a bandwidth estimation task; inputting the feature data into the pre-processing network to obtain input feature data suitable for the large language model; inputting the input feature data into the large language model to obtain output feature data for characterizing a bandwidth estimation value, wherein the large language model is a pre-trained model with an output layer removed; inputting the output feature data into the post-processing network to obtain the bandwidth estimation value; and performing the bandwidth estimation task based on the bandwidth estimation value.

[0014] Optionally, the pre-processing network is implemented by a feature processing network and a fully connected layer, wherein the feature data is input into the pre-processing network to obtain input feature data suitable for the large language model, including: converting the feature data into a feature vector through the feature processing network; encoding the feature vector into a feature input tensor suitable for the large language model through the fully connected layer.

[0015] Optionally, the feature data includes discrete feature data and sequence feature data, and the feature processing network includes an embedding processing network, a feature extraction network and a merging network, wherein the converting of the feature data into a feature vector through the feature processing network includes: embedding the discrete feature data through the embedding processing network to convert the discrete feature data into a first feature vector; extracting features from the sequence feature data through the feature extraction network to convert the sequence feature data into a second feature vector; and merging the first feature vector and the second feature vector through the merging network to obtain a merged feature vector.

[0016] Optionally, the post-processing network is implemented by at least one fully connected layer, wherein inputting the output feature data into the post-processing network to obtain the bandwidth estimation value includes: mapping the output feature data to the bandwidth estimation value through the at least one fully connected layer.

[0017] Optionally, the feature data includes at least one of user feature data, device feature data and network feature data related to the bandwidth estimation task.

[0018] Optionally, the bandwidth estimation model is obtained by training according to the training method disclosed in the present invention.

[0019] According to a third aspect of the present disclosure, a training device for a bandwidth estimation model is provided, the bandwidth estimation model includes a pre-processing network, a large language model and a post-processing network, the training device includes: a training data acquisition module, configured to: acquire a training data set, the training data set includes training feature data related to a bandwidth estimation task; a pre-processing module, configured to: input the training feature data into the pre-processing network to obtain input feature data suitable for the large language model; a bandwidth estimation module, configured to: input the input feature data into the large language model to obtain output feature data for characterizing a bandwidth estimation value, wherein the large language model is a pre-trained model with an output layer removed; a post-processing module, configured to: input the output feature data into the post-processing network to obtain the bandwidth estimation value; and a training module, configured to: train the bandwidth estimation model based on the bandwidth estimation value and the corresponding bandwidth label value.

[0020] Optionally, the pre-processing network is implemented by a feature processing network and a fully connected layer, wherein the pre-processing module is configured to: convert the training feature data into a feature vector through the feature processing network; and encode the feature vector into a feature input tensor suitable for the large language model through the fully connected layer.

[0021] Optionally, the training feature data includes discrete feature data and sequence feature data, and the feature processing network includes an embedding processing network, a feature extraction network and a merging network, wherein the pre-processing module is configured to: embed the discrete feature data through the embedding processing network to convert the discrete feature data into a first feature vector; extract features from the sequence feature data through the feature extraction network to convert the sequence feature data into a second feature vector; and merge the first feature vector and the second feature vector through the merging network to obtain a merged feature vector.

[0022] Optionally, the post-processing network is implemented by at least one fully connected layer, wherein the post-processing module is configured to: map the output feature data to the bandwidth estimation value through the at least one fully connected layer.

[0023] Optionally, the training module is configured to: utilize a large language model fine-tuning technique to train the bandwidth estimation model based on the bandwidth estimation value and the corresponding bandwidth label value.

[0024] Optionally, the training feature data includes at least one feature data of user feature data, device feature data and network feature data related to the bandwidth estimation task.

[0025] Optionally, the training feature data includes feature data obtained under at least one condition of a target business scenario, a target device type, and a target network type; the bandwidth estimation model obtained through training is applied to a bandwidth estimation task under the at least one condition.

[0026] Optionally, the training feature data includes feature data obtained under at least one of multiple business scenarios, multiple device types and multiple network types; the bandwidth estimation model obtained through training is applied to the bandwidth estimation task under the at least one condition.

[0027] According to a fourth aspect of the present disclosure, a bandwidth estimation device based on a bandwidth estimation model is provided, wherein the bandwidth estimation model includes a pre-processing network, a large language model and a post-processing network, and the bandwidth estimation device includes: a feature data acquisition module, configured to: acquire feature data related to a bandwidth estimation task; a pre-processing module, configured to: input the feature data into the pre-processing network to obtain input feature data suitable for the large language model; a bandwidth estimation module, configured to: input the input feature data into the large language model to obtain output feature data for characterizing a bandwidth estimation value, wherein the large language model is a pre-trained model with an output layer removed; a post-processing module, configured to: input the output feature data into the post-processing network to obtain the bandwidth estimation value; and a task execution module, configured to: execute the bandwidth estimation task based on the bandwidth estimation value.

[0028] Optionally, the pre-processing network is implemented by a feature processing network and a fully connected layer, wherein the pre-processing module is configured to: convert the feature data into a feature vector through the feature processing network; and encode the feature vector into a feature input tensor suitable for the large language model through the fully connected layer.

[0029] Optionally, the feature data includes discrete feature data and sequence feature data, and the feature processing network includes an embedding processing network, a feature extraction network and a merging network, wherein the pre-processing module is configured to: embed the discrete feature data through the embedding processing network to convert the discrete feature data into a first feature vector; extract features from the sequence feature data through the feature extraction network to convert the sequence feature data into a second feature vector; and merge the first feature vector and the second feature vector through the merging network to obtain a merged feature vector.

[0030] Optionally, the post-processing network is implemented by at least one fully connected layer, wherein the post-processing module is configured to: map the output feature data to the bandwidth estimation value through the at least one fully connected layer.

[0031] Optionally, the feature data includes at least one of user feature data, device feature data and network feature data related to the bandwidth estimation task.

[0032] Optionally, the bandwidth estimation model is obtained by training according to the training method disclosed in the present invention.

[0033] According to a fifth aspect of the present disclosure, an electronic device is provided, comprising: one or more processors; and a memory, coupled to the one or more processors and configured to store instructions that can be executed by the one or more processors, wherein the one or more processors are configured to execute a training method or a bandwidth estimation method according to the present disclosure when executing the instructions.

[0034] According to a sixth aspect of the present disclosure, a non-transitory computer-readable storage medium for storing computer-executable instructions is provided, wherein the computer-executable instructions, when executed by one or more computer processors, cause the one or more computer processors to perform a training method or a bandwidth estimation method according to the present disclosure.

[0035] According to a seventh aspect of the present disclosure, a computer program product is provided, comprising computer instructions, which, when executed by at least one processor, implement the training method or bandwidth estimation method according to the present disclosure.

[0036] The technical solution provided by the embodiments of the present disclosure brings at least the following beneficial effects:

[0037] In the present disclosure, the large language model (LLM) is used to estimate the available bandwidth of the link in the multimedia transmission system by using its huge parameter volume and powerful generation and reasoning capabilities. By designing a bandwidth estimation model formed by combining the large language model with the pre- / post-processing network, the accuracy and generalization of bandwidth estimation in various complex scenarios can be improved.

[0038] It is to be understood that both the foregoing general description and the following detailed description are examples only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate examples according to the disclosure and, together with the description, serve to explain the principles of the disclosure.

[0040] Figure 1 is a structural diagram showing a bandwidth estimation model according to an exemplary embodiment of the present disclosure.

[0041] Figure 2 is a flowchart illustrating a training method of a bandwidth estimation model according to an exemplary embodiment of the present disclosure.

[0042] Figure 3 is a schematic diagram showing the structure of a pre-processing network according to an exemplary embodiment of the present disclosure.

[0043] Figure 4 is a structural diagram showing a post-processing network according to an exemplary embodiment of the present disclosure.

[0044] Figure 5 is a flowchart illustrating a bandwidth estimation method based on a bandwidth estimation model according to an exemplary embodiment of the present disclosure.

[0045] Figure 6 is a block diagram illustrating a training apparatus for a bandwidth estimation model according to an exemplary embodiment of the present disclosure.

[0046] Figure 7 is a block diagram illustrating a bandwidth estimation apparatus based on a bandwidth estimation model according to an exemplary embodiment of the present disclosure.

[0047] Figure 8 is a diagram illustrating a computing environment coupled with a user interface according to some embodiments of the present disclosure. DETAILED DESCRIPTION

[0048] Reference will now be made in detail to specific embodiments, examples of which are shown in the accompanying drawings. In the following detailed description, a number of non-limiting specific details are set forth to aid in understanding the subject matter presented herein. However, various alternatives may be used without departing from the scope of the claims, and the subject matter may be practiced without these specific details. For example, the subject matter presented herein may be implemented on many types of electronic devices having digital video capabilities.

[0049] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the drawings are used to distinguish objects, but not to describe any specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those shown in the drawings or described in the present disclosure.

[0050] Currently, there are many challenges in estimating available bandwidth. On the one hand, the available bandwidth of the transmission link is highly uncertain. For example, factors such as user behavior, time period changes, the introduction of new services, network congestion, equipment failure, etc. may cause bandwidth demand to fluctuate significantly in a short period of time. On the other hand, different audio and video services have different quality of service (QoS) requirements. For example, in the RTC transmission scenario, the video bit rate requirement is not high but the delay requirement is more emphasized; in the short video consumption scenario, users prefer to see high-definition content, that is, the short video consumption scenario has higher requirements for the clarity of the video content.

[0051] In the related art, the network features that have been manually screened and processed are mainly processed by methods based on deep neural networks and supervised learning, and then the bandwidth is predicted and output based on the network features that have been manually screened and processed. However, the bandwidth estimation method in the related art can only perform bandwidth estimation in certain specific business scenarios, and its bandwidth estimation effect is often poor when facing new business scenarios. In addition, the types of features used for bandwidth estimation in the related art are relatively single, usually some measurable numerical features and sequence features in the network, and cannot comprehensively consider factors such as business scenarios, device status, and time changes to perform more refined bandwidth estimation.

[0052] In order to solve the above problems existing in the related art, the present disclosure provides a training method for a bandwidth estimation model, a bandwidth estimation method and related devices, which use the huge number of parameters and powerful generation and reasoning capabilities of a large language model (LLM) to estimate the link available bandwidth in a multimedia transmission system. By designing a bandwidth estimation model formed by combining a large language model with a pre- / post-processing network, the accuracy and generalization of bandwidth estimation in various complex scenarios can be improved.

[0053] Next, we will combine Figures 1 to 8 To describe the specific implementation process of the bandwidth estimation model training method and the bandwidth estimation method according to the present disclosure.

[0054] Figure 1 FIG. 1 is a schematic diagram showing a structure of a bandwidth estimation model according to an exemplary embodiment of the present disclosure. Figure 1 As shown, the present disclosure designs a bandwidth estimation model formed by combining a large language model with a pre-processing network and a post-processing network.

[0055] Since the tasks processed by the large language model are often related to natural language tasks, the input it accepts is usually a word vector generated after the token is embedded. The input used in the bandwidth estimation task is often some time series features and discrete user features related to the network. Therefore, these original feature inputs need to be processed separately by the pre-processing network before being handed over to the large language model for subsequent processing. That is, in the present disclosure, the pre-processing network is mainly used to convert feature data into input feature data suitable for the large language model.

[0056] The large language model is mainly used to perform bandwidth estimation tasks. Specifically, it is mainly used to obtain output feature data for characterizing bandwidth estimation values ​​based on input feature data. In addition, the output of the large language model is often the output for natural language processing tasks or the output of image and speech processing tasks. For example, the output of an image task can usually be a pixel RGB value of 0 to 255; the output of a text task is generally the probability of each token. The output of the bandwidth estimation task is often numerically continuous and has a larger range of values. Therefore, it is necessary to map the output of the large language model into a bandwidth estimation value through a specific neural network. In the present disclosure, the output feature data of the large language model output is converted into a bandwidth estimation value mainly by setting a post-processing network.

[0057] Thus, in the present disclosure, the pre-processing network can be used to convert feature data into input feature data suitable for a large language model; and the large language model can also be used to perform bandwidth estimation tasks, that is, the large language model can be used to obtain output feature data for characterizing bandwidth estimation values ​​based on the input feature data. Thanks to the huge number of parameters and powerful generation and reasoning capabilities of the large language model, it can provide effective guarantees for the accuracy and generalization of bandwidth prediction; and the bandwidth estimation post-processing network can also be used to map the feature vector output by the large language model to the final result of bandwidth estimation. That is, in the present disclosure, by combining the large language model, the pre-processing network and the post-processing network, it is possible to use the same model to process bandwidth estimation tasks in different business scenarios, thereby ensuring the accuracy of bandwidth estimation in each business scenario.

[0058] Figure 2 is a flow chart showing a training method of a bandwidth estimation model according to an exemplary embodiment of the present disclosure. As mentioned above, the bandwidth estimation model may include a "pre-processing network", a "large language model" and a "post-processing network".

[0059] Reference Figure 2 In step 201, a training data set may be obtained, and the training data set may include training feature data related to the bandwidth estimation task.

[0060] According to an exemplary embodiment of the present disclosure, the above-mentioned training feature data may include at least one feature data among user feature data, device feature data and network feature data related to the bandwidth estimation task.

[0061] User characteristic data may include but are not limited to: business type, geographic location, etc.; device characteristic data may include but are not limited to: device type, signal strength, model performance score, and device power, etc.; network characteristic data may include but are not limited to: historical transmission data volume, transmission time, historical transmission rate statistics, etc.

[0062] According to an exemplary embodiment of the present disclosure, the above-mentioned training feature data may include feature data obtained under at least one of the target business scenario, target device type and target network type. At this time, the trained bandwidth estimation model can be applied to the bandwidth estimation task under the above at least one condition. In this way, in the present disclosure, by constructing a training data set under a specific scenario, a bandwidth estimation model for the bandwidth estimation task under a specific scenario can be trained. That is, in the present disclosure, the user can control the training direction of the bandwidth estimation model according to his own needs, and then obtain a personalized bandwidth estimation model that meets the user's specific needs.

[0063] According to an exemplary embodiment of the present disclosure, the above-mentioned training feature data may include feature data obtained under at least one of multiple business scenarios, multiple device types, and multiple network types. In this case, the trained bandwidth estimation model can be applied to the bandwidth estimation task under at least one of the above conditions. In this way, in the present disclosure, by constructing data sets under multiple business scenarios, multiple device types, and multiple network types, it is possible to ensure that a general bandwidth estimation model applicable to various scenarios is trained, thereby improving the generalization ability of the bandwidth estimation model in the face of new scenarios.

[0064] In addition, in the present disclosure, corresponding training data sets can be constructed in a targeted manner according to the user's emphasis on different performances. For example, if the user wants to optimize the delay performance, more data about low-delay real-time calls can be collected as training data; if the user wants high throughput, data about high-bandwidth video transmission can be collected as training data, and so on. In this way, a personalized bandwidth estimation model that focuses on the performance that the user actually wants can be obtained.

[0065] In step 202, the training feature data may be input into a pre-processing network to obtain input feature data suitable for a large language model.

[0066] Figure 3 is a schematic diagram showing the structure of a pre-processing network according to an exemplary embodiment of the present disclosure. Figure 3The pre-processing network shown is only an exemplary pre-processing network structure. According to the present disclosure, the pre-processing network can also be any other network structure that can convert training feature data into input feature data suitable for a large language model. Exemplarily, the pre-processing network can be, but is not limited to, a long short-term memory network (Long Short-Term Memory, LSTM), a transformer, and the like.

[0067] Reference Figure 3 , the pre-processing network can mainly include a feature processing network and a fully connected layer (Fully ConnectedLaye, FC). The training feature data can be input into the feature processing network, and the training feature data can be converted into a feature vector through the feature processing network. Then, the feature vector can be input into the fully connected layer, and the feature vector can be encoded into a feature input tensor suitable for a large language model through the fully connected layer.

[0068] According to an exemplary embodiment of the present disclosure, the feature input of the pre-processing network can be mainly divided into two types, namely, discrete features and sequence features. That is, in the present disclosure, the training feature data can be divided into two parts, namely, discrete features and sequence features, and the two parts can be processed separately. In this way, the input requirements of the large language model can be further met.

[0069] Discrete features generally refer to features that can distinguish categories and are not easily expressed numerically. Exemplarily, discrete features may include, but are not limited to: current application type (e.g., RTC, on-demand video downloads, etc.), current time interval (per hour), whether it is the weekend, user network type, user terminal type, etc.; sequence features generally refer to continuous numerical features that can be continuously sampled and collected. Exemplarily, sequence features may include, but are not limited to: historical transmission data volume, transmission time, connection establishment duration, historical observed transmission rate, etc. In addition, in the present disclosure, features can be organized and stored in a matrix manner, and general preprocessing for machine learning can also be performed on the features. Exemplarily, "general preprocessing for machine learning" may include, but is not limited to: normalization, removal of outliers and extreme values, etc.

[0070] It should be noted that, in the present disclosure, in addition to being divided into discrete features and sequence features, the training feature data can also be divided into other types, and the aforementioned type division method is only an exemplary description.

[0071] In addition, in the case where the feature input of the pre-processing network is divided into the aforementioned two types, the aforementioned feature processing network may include an embedding processing network, a feature extraction network, and a merging network. The discrete features may be embedded by the embedding processing network to convert the discrete features into a first feature vector. Specifically, due to their discrete characteristics, the discrete features may be embedded (Embedding) with reference to the token processing method in the natural language processing task to convert the discrete features into feature vectors.

[0072] Furthermore, the sequence features may be extracted through a "feature extraction network" to convert the sequence features into a second feature vector. Specifically, a convolutional neural network or a recurrent neural network may be used to perform deep feature extraction and encoding on the sequence features of the numerical type to convert the sequence features into a feature vector.

[0073] Furthermore, the first feature vector and the second feature vector can be merged (Concat) through a merging network, which can also be called concatenation, to obtain a merged feature vector. That is, the feature vector converted from discrete features and the feature vector converted from sequence features can be merged through a merging network to obtain a merged feature vector. Then, the merged feature vector can be encoded through a fully connected neural network to obtain the final feature input tensor suitable for large language model input.

[0074] Thus, in the present disclosure, a general feature pre-processing network can be designed for bandwidth estimation, which can process rich feature data related to the bandwidth estimation task, thereby increasing the types and quantity of training data, thereby ensuring that a bandwidth estimation model with better bandwidth estimation performance is trained.

[0075] Return to reference Figure 2 In step 203, the input feature data may be input into a large language model to obtain output feature data for characterizing the bandwidth estimation value, wherein the large language model may be a pre-trained model with the output layer removed. Exemplarily, the large language model may be a pre-trained model with the output probability layer implemented by Softmax removed. According to an exemplary embodiment of the present disclosure, the large language model may be implemented by a pre-trained model such as llama, Gemma, etc. Of course, it may also be other pre-trained large language models and large language models trained by themselves according to business needs, etc.

[0076] In step 204, the output feature data may be input into a post-processing network to obtain a bandwidth estimation value.

[0077] According to an exemplary embodiment of the present disclosure, since the function of the fully connected network is to map the high-dimensional features processed by the large language model to the final bandwidth estimation result, the post-processing network can be implemented by at least one fully connected layer for mapping the output feature data to the bandwidth estimation value. For example, Figure 4 is a schematic diagram showing the structure of a post-processing network according to an exemplary embodiment of the present disclosure. Figure 4 The post-processing network shown is only an exemplary post-processing network structure. According to the present disclosure, the post-processing network can also be any other network structure that can map the output feature data to the bandwidth estimation value. Exemplarily, the "post-processing network" can adopt but is not limited to the following structures: fully connected layer + ReLU, fully connected layer + leaky ReLU, etc.

[0078] Reference Figure 4 , the post-processing network can be implemented by two fully connected networks, namely FC1 and FC2.

[0079] In addition, the sigmoid activation function can also be used to convert the final bandwidth estimation result output by the fully connected network into a normalized value between 0 and 1 to facilitate model training and reasoning. Exemplarily, the normalized value can be multiplied by the normalization coefficient for the final bandwidth estimation result, and the normalization coefficient can be formulated according to actual needs. In addition, the normalization method can be but is not limited to: maximum and minimum value normalization method, etc.

[0080] In step 205, the bandwidth estimation model may be trained based on the bandwidth estimation value and the corresponding bandwidth label value. The "bandwidth label value" may be the average value of the available bandwidth observed in the next time period of the current time period. That is, in the present disclosure, the input features and the bandwidth label value may constitute a standard supervised learning data set for fine-tuning of a large language model.

[0081] In addition, based on the autoregressive output characteristics of the large language model, the historical bandwidth estimation results and bandwidth estimation errors can also be used as feature inputs, thereby increasing the context association of the bandwidth estimation task, thereby further improving the training effect of the bandwidth estimation model.

[0082] According to an exemplary embodiment of the present disclosure, a large language model fine-tuning technology can be used to train a bandwidth estimation model based on bandwidth estimation values ​​and corresponding bandwidth label values, that is, a training data set can be used to fine-tune the bandwidth prediction model based on supervised learning. Since the large language model has a large number of parameters, fine-tuning may require a large amount of computing resources. Therefore, in the present disclosure, a commonly used large language model fine-tuning technology (e.g., LoRA, etc.) can be used to fine-tune the bandwidth estimation model, thereby obtaining a bandwidth estimation model suitable for various application scenarios.

[0083] The present disclosure aims to use the powerful generation and reasoning capabilities of the large language model (LLM) to estimate the available bandwidth of the link in the multimedia transmission system, so as to improve the accuracy of bandwidth estimation in various complex scenarios. In addition, the feature data can also be input into a preprocessing neural network, that is, a pre-processing network, so that the complex features can be converted into an input form that can be understood by the large language model. Then, the feature vector of the bandwidth estimation can be generated by the large language model, and the output of the large language model can be parsed by the post-processing neural network to obtain the final bandwidth estimation value. In this way, the difficulty of manually customizing rules can be reduced and the full utilization of multi-dimensional features can be achieved, and the applicable scenarios of the bandwidth estimation model can be expanded, that is, the generalization ability of the bandwidth estimation model can be improved.

[0084] Figure 5 is a flow chart showing a bandwidth estimation method based on a bandwidth estimation model according to an exemplary embodiment of the present disclosure. Here, the bandwidth estimation model may have the following Figure 1 The structure described above will therefore not be described in detail here.

[0085] Reference Figure 5 In step 501, characteristic data related to the bandwidth estimation task may be obtained. Specifically, characteristic data related to the bandwidth estimation task in the previous time period of the current time period may be obtained.

[0086] According to an exemplary embodiment of the present disclosure, the above-mentioned characteristic data may include at least one characteristic data of user characteristic data, device characteristic data and network characteristic data related to the bandwidth estimation task. Specifically, as described in the previous embodiment, the user characteristic data may include but is not limited to: business type, geographic location, etc.; the device characteristic data may include but is not limited to: device type, signal strength, model performance score and device power, etc.; the network characteristic data may include but is not limited to: historical transmission data volume, transmission time, historical transmission rate statistics, etc.

[0087] In step 502, the acquired feature data may be input into a pre-processing network to obtain input feature data suitable for a large language model. Here, the pre-processing network may have, but is not limited to, Figure 3 The structure shown in the figure is not described in detail here.

[0088] According to an exemplary embodiment of the present disclosure, the above-mentioned "pre-processing network" can be implemented by a feature processing network and a fully connected layer. The feature data can be converted into a feature vector by the feature processing network. Then, the feature vector can be encoded into a feature input tensor suitable for a large language model by a fully connected layer.

[0089] According to an exemplary embodiment of the present disclosure, the feature data may include discrete feature data and sequence feature data. As mentioned above, discrete features generally refer to features that can distinguish categories and are not easily represented by numerical values; sequence features generally refer to continuous numerical features that can be continuously sampled and collected. The feature processing network may include an embedding processing network, a feature extraction network, and a merging network.

[0090] The discrete feature data may be embedded through an embedding processing network to convert the discrete feature data into a first feature vector. The sequence feature data may also be feature extracted through a feature extraction network to convert the sequence feature data into a second feature vector. Specifically, a convolutional neural network or a recurrent neural network may be used to perform deep feature extraction and encoding of sequence features of numerical type to convert the sequence features into feature vectors. Next, the first feature vector and the second feature vector may be merged through a merging network to obtain a merged feature vector.

[0091] In step 503, the input feature data may be input into a large language model to obtain output feature data for characterizing the bandwidth estimation value, wherein the large language model may be a pre-trained model with the output layer removed. Exemplarily, as described above, the large language model may be a pre-trained model with the output probability layer implemented by Softmax removed.

[0092] In step 504, the output feature data may be input into a post-processing network to obtain a bandwidth estimation value. Specifically, the output feature data may be input into a post-processing network to obtain a bandwidth estimation value for the current time period, that is, the available bandwidth for the current time period. Here, the post-processing network may have, but is not limited to, Figure 4 The structure shown in the figure is not described in detail here.

[0093] According to an exemplary embodiment of the present disclosure, the above-mentioned post-processing network can be implemented by at least one fully connected layer. Exemplarily, at least one fully connected layer can be two fully connected layers, namely FC1 and FC2. The above-mentioned output feature data can be mapped to a bandwidth estimation value through at least one fully connected layer. Specifically, the function of the fully connected network can be to map the high-dimensional features processed by the large language model to the final bandwidth estimation result. Next, the sigmoid activation function can also be used to convert the final bandwidth estimation result into a normalized value between 0 and 1.

[0094] In step 505, a bandwidth estimation task may be performed based on the bandwidth estimation value. Exemplarily, the bandwidth estimation task may include but is not limited to: a short video consumption task, a short video publishing task, an RTC real-time call task, and the like.

[0095] Figure 66 is a block diagram showing a training apparatus 600 for a bandwidth estimation model according to an exemplary embodiment of the present disclosure. The bandwidth estimation model may include a pre-processing network, a large language model, and a post-processing network.

[0096] Reference Figure 6 The training device 600 of the bandwidth estimation model may include a training data acquisition module 601, a pre-processing module 602, a bandwidth estimation module 603, a post-processing module 604 and a training module 605.

[0097] The training data acquisition module 601 may acquire a training data set, where the training data set may include training feature data related to the bandwidth estimation task.

[0098] According to an exemplary embodiment of the present disclosure, the above-mentioned training feature data may include at least one feature data among user feature data, device feature data and network feature data related to the bandwidth estimation task.

[0099] User characteristic data may include but are not limited to: business type, geographic location, etc.; device characteristic data may include but are not limited to: device type, signal strength, model performance score, and device power, etc.; network characteristic data may include but are not limited to: historical transmission data volume, transmission time, historical transmission rate statistics, etc.

[0100] According to an exemplary embodiment of the present disclosure, the above-mentioned training feature data may include feature data obtained under at least one of the target business scenario, target device type and target network type. At this time, the trained bandwidth estimation model can be applied to the bandwidth estimation task under the above at least one condition. In this way, in the present disclosure, by constructing a training data set under a specific scenario, a bandwidth estimation model for the bandwidth estimation task under a specific scenario can be trained. That is, in the present disclosure, the user can control the training direction of the bandwidth estimation model according to his own needs, and then obtain a personalized bandwidth estimation model that meets the user's specific needs.

[0101] According to an exemplary embodiment of the present disclosure, the above-mentioned training feature data may include feature data obtained under at least one of multiple business scenarios, multiple device types, and multiple network types. In this case, the trained bandwidth estimation model can be applied to the bandwidth estimation task under at least one of the above conditions. In this way, in the present disclosure, by constructing data sets under multiple business scenarios, multiple device types, and multiple network types, it is possible to ensure that a general bandwidth estimation model applicable to various scenarios is trained, thereby improving the generalization ability of the bandwidth estimation model in the face of new scenarios.

[0102] The pre-processing module 602 can input the training feature data into the pre-processing network to obtain input feature data suitable for a large language model.

[0103] According to an exemplary embodiment of the present disclosure, the pre-processing network may mainly include a feature processing network and a fully connected layer. The pre-processing module 602 may input the training feature data into the feature processing network, and convert the training feature data into a feature vector through the feature processing network. Then, the pre-processing module 602 may input the feature vector into the fully connected layer, and encode the feature vector into a feature input tensor suitable for a large language model through the fully connected layer.

[0104] According to an exemplary embodiment of the present disclosure, the feature input of the pre-processing network can be mainly divided into two types, namely, discrete features and sequence features. In addition, the aforementioned feature processing network can include an embedding processing network, a feature extraction network and a merging network.

[0105] The pre-processing module 602 can embed the discrete features through the embedding processing network to convert the discrete features into a first feature vector. Specifically, due to its discrete characteristics, the discrete features can be embedded with reference to the token processing method in the natural language processing task to convert the discrete features into a feature vector.

[0106] Furthermore, the pre-processing module 602 can also extract features from the sequence features through a feature extraction network to convert the sequence features into a second feature vector. Specifically, a convolutional neural network or a recurrent neural network can be used to perform deep feature extraction and encoding on the sequence features of the numerical type to convert the sequence features into a feature vector.

[0107] In addition, the pre-processing module 602 can also merge the first feature vector and the second feature vector through a merging network to obtain a merged feature vector. That is, the feature vector converted from the discrete feature and the feature vector converted from the sequence feature can be merged through the merging network to obtain a merged feature vector. Then, the merged feature vector can be encoded through a fully connected neural network to obtain a final feature input tensor suitable for large language model input.

[0108] The bandwidth estimation module 603 may input input feature data into a large language model to obtain output feature data for characterizing a bandwidth estimation value, wherein the large language model may be a pre-trained model with an output layer removed.

[0109] The post-processing module 604 may input the output feature data into a post-processing network to obtain a bandwidth estimation value.

[0110] According to an exemplary embodiment of the present disclosure, the post-processing network may be implemented by at least one fully connected layer, and the post-processing module 604 may map the output feature data to a bandwidth estimation value through at least one fully connected layer.

[0111] The training module 605 can train the bandwidth estimation model based on the bandwidth estimation value and the corresponding bandwidth label value. The "bandwidth label value" can be the average value of the available bandwidth in the next time period of the current time period observed. That is, in the present disclosure, the input features and the bandwidth label value can form a standard supervised learning data set for fine-tuning of a large language model.

[0112] According to an exemplary embodiment of the present disclosure, the training module 605 may utilize a large language model fine-tuning technique to train the bandwidth estimation model based on the bandwidth estimation value and the corresponding bandwidth label value.

[0113] Figure 7 7 is a block diagram of a bandwidth estimation device 700 showing a bandwidth estimation model according to an exemplary embodiment of the present disclosure. The bandwidth estimation model may include a pre-processing network, a large language model, and a post-processing network.

[0114] Reference Figure 7 The bandwidth estimation device 700 of the bandwidth estimation model may include a feature data acquisition module 701, a pre-processing module 702, a bandwidth estimation module 703, a post-processing module 704 and a task execution module 705.

[0115] The characteristic data acquisition module 701 can acquire characteristic data related to the bandwidth estimation task. Specifically, characteristic data related to the bandwidth estimation task in the previous time period of the current time period can be acquired.

[0116] According to an exemplary embodiment of the present disclosure, the above-mentioned characteristic data may include at least one characteristic data of user characteristic data, device characteristic data and network characteristic data related to the bandwidth estimation task. Specifically, as described in the previous embodiment, the user characteristic data may include but is not limited to: business type, geographic location, etc.; the device characteristic data may include but is not limited to: device type, signal strength, model performance score and device power, etc.; the network characteristic data may include but is not limited to: historical transmission data volume, transmission time, historical transmission rate statistics, etc.

[0117] The pre-processing module 702 can input the acquired feature data into the pre-processing network to obtain input feature data suitable for a large language model.

[0118] According to an exemplary embodiment of the present disclosure, the pre-processing network is implemented by a feature processing network and a fully connected layer. The pre-processing module 702 converts feature data into feature vectors through the feature processing network, and encodes the feature vectors into feature input tensors suitable for a large language model through the fully connected layer.

[0119] According to an exemplary embodiment of the present disclosure, the feature data includes discrete feature data and sequence feature data, and the feature processing network includes an embedding processing network, a feature extraction network, and a merging network.

[0120] The pre-processing module 702 may perform embedding processing on the discrete feature data through an embedding processing network to convert the discrete feature data into a first feature vector. Then, the pre-processing module 702 may perform feature extraction on the sequence feature data through a feature extraction network to convert the sequence feature data into a second feature vector. Next, the pre-processing module 702 may merge the first feature vector and the second feature vector through a merging network to obtain a merged feature vector.

[0121] The bandwidth estimation module 703 may input the input feature data into the large language model to obtain output feature data for characterizing the bandwidth estimation value, wherein the large language model may be a pre-trained model with the output layer removed. Exemplarily, as described above, the large language model may be a pre-trained model with the output probability layer implemented by Softmax removed.

[0122] The post-processing module 704 may input the output feature data into a post-processing network to obtain a bandwidth estimation value. Specifically, the output feature data may be input into a post-processing network to obtain a bandwidth estimation value of the current time period, that is, the available bandwidth of the current time period.

[0123] According to an exemplary embodiment of the present disclosure, the post-processing network is implemented by at least one fully connected layer. The post-processing module 704 maps the output feature data to a bandwidth estimation value through at least one fully connected layer.

[0124] The task execution module 705 may execute a bandwidth estimation task based on the bandwidth estimation value. For example, the bandwidth estimation task may include but is not limited to: a short video consumption task, a short video publishing task, an RTC real-time call task, and the like.

[0125] According to an exemplary embodiment of the present disclosure, the bandwidth estimation model may be trained by the training method according to the present disclosure.

[0126] Figure 8 A computing environment 810 is shown coupled to a user interface 850. The computing environment 810 may be part of a data processing server. The computing environment 810 includes a processor 820, a memory 830, and an input / output (I / O) interface 840.

[0127] The processor 820 generally controls the overall operation of the computing environment 810, such as operations associated with display, data acquisition, data communication, and image processing. The processor 820 may include one or more processors for executing instructions to perform all or some steps in the above method. In addition, the processor 820 may include one or more modules that facilitate the interaction between the processor 820 and other components. The processor may be a central processing unit (CPU), a microprocessor, a single-chip microcomputer, a graphics processing unit (GPU), etc.

[0128] The memory 830 is configured to store various types of data to support the operation of the computing environment 810. The memory 830 may include predetermined software 832. Examples of such data include instructions for any application or method operating on the computing environment 810, video data sets, image data, etc. The memory 830 may be implemented by using any type of volatile or non-volatile memory device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0129] The I / O interface 840 provides an interface between the processor 820 and peripheral interface modules (e.g., keyboard, click wheel, buttons, etc.). The buttons may include, but are not limited to, a home button, a start scan button, and a stop scan button. The I / O interface 840 may be coupled to an encoder and a decoder.

[0130] In an embodiment, an electronic device is provided, comprising: one or more processors; and a memory coupled to the one or more processors and configured to store instructions executable by the one or more processors, wherein the one or more processors are configured to execute the above-mentioned training method or the above-mentioned bandwidth estimation method when executing the instructions.

[0131] In an embodiment, a non-transitory computer-readable storage medium for storing computer-executable instructions is also provided. When the computer-executable instructions are executed by one or more computer processors, the one or more computer processors perform the above training method or the above bandwidth estimation method.

[0132] In an embodiment, a computing device is also provided, the computing device comprising: one or more processors (e.g., processor 820); and a non-transitory computer-readable storage medium or memory 830 in which a plurality of programs that can be executed by the one or more processors are stored, wherein the one or more processors are configured to perform the above method when executing the plurality of programs. The non-transitory computer-readable storage medium can be, for example, a ROM, a random access memory (RAM), a CD-ROM, a tape, a floppy disk, an optical data storage device, etc.

[0133] In an embodiment, a computer program product including a plurality of programs in, for example, a memory 830 is also provided, which may be executed by a processor 820 in a computing environment 810 for performing the above method. For example, the computer program product may include a non-transitory computer-readable storage medium.

[0134] In an embodiment, the computing environment 810 may be implemented by one or more ASICs, DSPs, digital signal processing devices (DSPDs), programmable logic devices (PLDs), FPGAs, GPUs, controllers, microcontrollers, microprocessors, or other electronic components for performing the above methods.

[0135] The description of the present disclosure has been presented for purposes of illustration and is not intended to be exhaustive or limited to the present disclosure. Many modifications, variations, and alternative embodiments will be apparent to one of ordinary skill in the art having the benefit of the teachings presented in the foregoing description and the associated drawings.

[0136] Unless otherwise specifically stated, the order of steps of the method according to the present disclosure is intended to be illustrative only, and the steps of the method according to the present disclosure are not limited to the order specifically described above, but can be changed according to actual conditions. In addition, at least one of the steps of the method according to the present disclosure can be adjusted, combined or deleted according to actual needs.

[0137] The examples are chosen and described in order to explain the principles of the present disclosure and to enable other persons skilled in the art to understand the various embodiments of the present disclosure and to best utilize the basic principles and various embodiments with various modifications suitable for the specific use contemplated. Therefore, it will be understood that the scope of the present disclosure is not limited to the specific examples of the disclosed embodiments, and modifications and other embodiments are intended to be included within the scope of the present disclosure.

Claims

1. A method for training a bandwidth estimation model, characterized in that: The bandwidth estimation model includes a pre-processing network, a large language model and a post-processing network, and the training method includes: Acquire a training data set, wherein the training data set includes training feature data related to a bandwidth estimation task; Inputting the training feature data into the pre-processing network to obtain input feature data suitable for the large language model; Inputting the input feature data into the large language model to obtain output feature data for characterizing a bandwidth estimation value, wherein the large language model is a pre-trained model with an output layer removed; Inputting the output feature data into the post-processing network to obtain the bandwidth estimation value; The bandwidth estimation model is trained based on the bandwidth estimation value and the corresponding bandwidth label value.

2. The training method according to claim 1, characterized in that: The pre-processing network is implemented by a feature processing network and a fully connected layer. The step of inputting the training feature data into the pre-processing network to obtain input feature data suitable for the large language model includes: Converting the training feature data into a feature vector through the feature processing network; The feature vector is encoded into a feature input tensor suitable for the large language model through the fully connected layer.

3. The training method according to claim 2, characterized in that: The training feature data includes discrete feature data and sequence feature data, the feature processing network includes an embedding processing network, a feature extraction network and a merging network, The step of converting the training feature data into a feature vector through the feature processing network includes: Performing embedding processing on the discrete feature data through the embedding processing network to convert the discrete feature data into a first feature vector; Extracting features from the sequence feature data through the feature extraction network, and converting the sequence feature data into a second feature vector; The first feature vector and the second feature vector are merged through the merging network to obtain a merged feature vector.

4. The training method according to claim 1, characterized in that: The post-processing network is implemented by at least one fully connected layer. The step of inputting the output feature data into the post-processing network to obtain the bandwidth estimation value comprises: The output feature data is mapped to the bandwidth estimate value through the at least one fully connected layer.

5. The training method according to claim 1, characterized in that: The training of the bandwidth estimation model based on the bandwidth estimation value and the corresponding bandwidth label value includes: The bandwidth estimation model is trained based on the bandwidth estimation value and the corresponding bandwidth label value by using a large language model fine-tuning technology.

6. The training method according to claim 1, characterized in that: The training feature data includes at least one feature data of user feature data, device feature data and network feature data related to the bandwidth estimation task.

7. The training method according to claim 6, characterized in that: The training feature data includes feature data acquired under at least one condition of a target business scenario, a target device type, and a target network type; The bandwidth estimation model obtained through training is applied to the bandwidth estimation task under the at least one condition.

8. The training method according to claim 6, characterized in that: The training feature data includes feature data obtained under at least one of multiple business scenarios, multiple device types, and multiple network types; The bandwidth estimation model obtained through training is applied to the bandwidth estimation task under the at least one condition.

9. A bandwidth estimation method based on a bandwidth estimation model, characterized in that: The bandwidth estimation model includes a pre-processing network, a large language model and a post-processing network, and the bandwidth estimation method includes: Acquiring characteristic data related to bandwidth estimation tasks; Inputting the feature data into the pre-processing network to obtain input feature data suitable for the large language model; Inputting the input feature data into the large language model to obtain output feature data for characterizing a bandwidth estimation value, wherein the large language model is a pre-trained model with an output layer removed; Inputting the output feature data into the post-processing network to obtain the bandwidth estimation value; The bandwidth estimation task is performed based on the bandwidth estimation value.

10. The bandwidth estimation method according to claim 9, characterized in that: The pre-processing network is implemented by a feature processing network and a fully connected layer. The step of inputting the feature data into the pre-processing network to obtain input feature data suitable for the large language model includes: Converting the feature data into a feature vector through the feature processing network; The feature vector is encoded into a feature input tensor suitable for the large language model through the fully connected layer.

11. The bandwidth estimation method according to claim 10, characterized in that: The feature data includes discrete feature data and sequence feature data, and the feature processing network includes an embedding processing network, a feature extraction network and a merging network. The step of converting the feature data into a feature vector through the feature processing network includes: Performing embedding processing on the discrete feature data through the embedding processing network to convert the discrete feature data into a first feature vector; Extracting features from the sequence feature data through the feature extraction network, and converting the sequence feature data into a second feature vector; The first feature vector and the second feature vector are merged through the merging network to obtain a merged feature vector.

12. The bandwidth estimation method according to claim 9, characterized in that: The post-processing network is implemented by at least one fully connected layer. The step of inputting the output feature data into the post-processing network to obtain the bandwidth estimation value comprises: The output feature data is mapped to the bandwidth estimate value through the at least one fully connected layer.

13. The bandwidth estimation method according to claim 9, characterized in that: The characteristic data includes at least one characteristic data of user characteristic data, device characteristic data and network characteristic data related to the bandwidth estimation task.

14. The bandwidth estimation method according to claim 9, characterized in that: The bandwidth estimation model is obtained by training using the training method described in any one of claims 1-8.

15. A training device for a bandwidth estimation model, characterized in that: The bandwidth estimation model includes a pre-processing network, a large language model and a post-processing network, and the training device includes: A training data acquisition module is configured to: acquire a training data set, wherein the training data set includes training feature data related to the bandwidth estimation task; A pre-processing module is configured to: input the training feature data into the pre-processing network to obtain input feature data suitable for the large language model; The bandwidth estimation module is configured to: input the input feature data into the large language model to obtain output feature data for characterizing a bandwidth estimation value, wherein the large language model is a pre-trained model with an output layer removed; A post-processing module is configured to: input the output feature data into the post-processing network to obtain the bandwidth estimation value; The training module is configured to: train the bandwidth estimation model based on the bandwidth estimation value and the corresponding bandwidth label value.

16. A bandwidth estimation device based on a bandwidth estimation model, characterized in that: The bandwidth estimation model includes a pre-processing network, a large language model and a post-processing network, and the bandwidth estimation device includes: The feature data acquisition module is configured to: acquire feature data related to the bandwidth estimation task; A pre-processing module is configured to: input the feature data into the pre-processing network to obtain input feature data suitable for the large language model; The bandwidth estimation module is configured to: input the input feature data into the large language model to obtain output feature data for characterizing a bandwidth estimation value, wherein the large language model is a pre-trained model with an output layer removed; A post-processing module is configured to: input the output feature data into the post-processing network to obtain the bandwidth estimation value; The task execution module is configured to: execute the bandwidth estimation task based on the bandwidth estimation value.

17. An electronic device comprising: one or more processors; as well as a memory coupled to the one or more processors and configured to store instructions executable by the one or more processors, Wherein, when executing the instructions, the one or more processors are configured to perform the training method according to any one of claims 1-8 or the bandwidth estimation method according to any one of claims 9-14.

18. A non-transitory computer-readable storage medium for storing computer-executable instructions, which, when executed by one or more computer processors, cause the one or more computer processors to perform the training method according to any one of claims 1-8 or the bandwidth estimation method according to any one of claims 9-14.

19. A computer program product comprising computer instructions, characterized in that When the computer instructions are executed by at least one processor, the training method according to any one of claims 1-8 or the bandwidth estimation method according to any one of claims 9-14 is implemented.

Images