Media information processing method, apparatus, device, and storage medium

Abstract

The present disclosure relates to a media information processing method, device and equipment and storage medium. The method comprises inputting an initial feature set into a media feature determination network to obtain an intermediate media feature vector corresponding to a training sample. The media feature determination network comprises a media feature screening module composed of gate vectors used to perform media feature determination operations. Each gate element in the gate vector is determined based on a preset parameter composed of an attenuation parameter and an adjustable gating parameter. A target function loss is determined based on a task processing label and a predicted task processing result and the adjustable gating parameter. The initial feature set is screened based on the target gate vector in the trained media feature determination network to obtain a target media feature. The target gate element values corresponding to the features other than the target media feature in the target gate vector are all first preset values. In this way, the number of features is reduced while the accuracy of the model is ensured.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of computer, and particularly relates to a media information processing method and device, equipment and storage medium. BACKGROUND

[0002] With the rapid development of deep learning, deep neural networks have made great achievements in the fields of search and recommendation. In order to improve the model accuracy, various deep neural network models need a large number of user and commodity features as input for prediction. However, more and more input features not only increase the model calculation amount, but also increase the consumption of computer storage resources.

[0003] In the related art, in the field of picture processing, the weight of each input picture feature is determined first, and then the input features with a weight greater than a preset weight threshold are selected as target picture features to reduce the number of input picture features. Although this method reduces the model calculation amount and storage resource occupation, the size of the determined weight cannot accurately represent the importance of the target picture feature, thereby reducing the model accuracy to some extent. SUMMARY

[0004] The present disclosure provides a media information processing method, device, equipment and storage medium to at least solve at least one problem in the related art. The technical solutions of the present disclosure are as follows:

[0005] According to a first aspect of an embodiment of the present disclosure, a media information processing method is provided, comprising:

[0006] obtaining a first training sample for a target task, the first training sample comprising an initial feature set of media information and a corresponding task processing label;

[0007] inputting the initial feature set into a media feature determination network for processing to obtain an intermediate media feature vector corresponding to the first training sample; the media feature determination network comprises a feature screening module composed of a gate vector for performing a media feature determination operation, each gate element in the gate vector being determined based on a preset parameter composed of a decay parameter and a gate adjustable parameter, and the dimension of the gate vector being the same as the vector dimension corresponding to the initial feature set;

[0008] inputting the intermediate media feature vector into a task processing network for task processing to obtain a predicted task processing result corresponding to the first training sample;

[0009] determining a target function loss based on the task processing label and the predicted task processing result, and the gate adjustable parameter, and training model parameters of the media feature determination network and the task processing network based on the target function loss to obtain a trained media feature determination network.

[0010] The initial feature set is filtered based on a target gate vector in the trained media feature determination network to obtain a target media feature; and a preset value is used for a target gate element value corresponding to a feature other than the target media feature in the target gate vector.

[0011] In an optional implementation, when the preset parameter meets a preset decay condition and the gating adjustable parameter in the preset parameter is zero, a function value of the preset parameter is zero.

[0012] In an optional implementation, the preset parameter includes a first function item and a second function item; and the first function item and the second function item are respectively composed of a decay parameter and a gating adjustable parameter.

[0013] When the gating adjustable parameter in the preset parameter is zero, a derivative of the preset parameter is positively correlated with the decay parameter.

[0014] In an optional implementation, the inputting of the initial feature set into the media feature determination network to obtain an intermediate media feature vector corresponding to the training sample includes:

[0015] The initial feature set is input into the media feature determination network for mapping processing to obtain an initial media feature vector.

[0016] Based on the initial media feature vector and a gate vector in the feature filtering module, an intermediate media feature vector corresponding to the training sample is determined.

[0017] In an optional implementation, the determining of the target function loss based on the task processing label, the predicted task processing result, and the gating adjustable parameter includes:

[0018] An initial loss function is determined based on the task processing label and the predicted task processing result.

[0019] A function adjustment item is determined based on a gating adjustable parameter corresponding to each gate element in the gate vector.

[0020] A target function loss is determined based on the initial loss function and the function adjustment item.

[0021] In an optional implementation, the filtering of the initial feature set based on the gate vector in the trained media feature determination network to obtain a target media feature includes:

[0022] A target gate element is filtered from the gate vector in the trained media feature determination network, and an element value of the target gate element is zero.

[0023] The target media features are obtained by removing the features corresponding to the target gating element from the initial feature set.

[0024] In an alternative implementation, the method further includes:

[0025] Based on the target media characteristics, the task processing network is adjusted to obtain the adjusted task processing network.

[0026] The first training sample is processed based on the target media features to generate a second training sample;

[0027] The adjusted task processing network is trained based on the second training sample to obtain the trained task processing network.

[0028] In an alternative implementation, the method further includes:

[0029] Based on the target media features, extract the target media features corresponding to the target media information;

[0030] The target media features are input into the trained task processing network for task processing to obtain task processing results, which are used to indicate whether to execute the corresponding target task.

[0031] According to a second aspect of the present disclosure, a media information processing apparatus is provided, comprising:

[0032] The acquisition module is configured to acquire a first training sample for the target task, the first training sample including an initial feature set of media information and corresponding task processing labels;

[0033] The first processing module is configured to process the initial feature set into a media feature determination network to obtain an intermediate media feature vector corresponding to the training samples. The media feature determination network includes a media feature filtering module composed of gate vectors for performing media feature determination operations. Each gate element in the gate vector is determined based on preset parameters composed of attenuation parameters and gate adjustable parameters. The dimension of the gate vector is the same as the dimension of the vector corresponding to the initial feature set. The intermediate media feature vector is associated with the gate vector.

[0034] The second processing module is configured to input the intermediate media feature vector into the task processing network for task processing, and obtain the prediction task processing result corresponding to the training sample.

[0035] determining a target function loss based on the task processing label, the predicted task processing result, and the gating adjustable parameter, training model parameters of the media feature determination network and the task processing network based on the target function loss to obtain a trained media feature determination network;

[0036] performing feature screening on the initial feature set based on a target gate vector in the trained media feature determination network to obtain a target media feature corresponding to the media information; target gating element values corresponding to features other than the target media feature in the target gate vector are all first preset values, and a difference between a target gating element corresponding to the target media feature in the target gate vector and a second preset value is less than or equal to a preset threshold.

[0037] In an optional implementation, when the gating adjustable parameter in the preset parameter is zero, a function value of the preset parameter is the first preset value if the decay parameter in the preset parameter satisfies a preset decay condition.

[0038] In an optional implementation, the preset parameter includes a first function item and a second function item; the first function item and the second function item are both composed of a decay parameter and a gating adjustable parameter.

[0039] When the gating adjustable parameter in the preset parameter is zero, a derivative of the preset parameter is positively correlated with the decay parameter.

[0040] In an optional implementation, the first processing module is specifically configured to perform:

[0041] inputting the initial feature set into a media feature determination network for mapping processing to obtain an initial media feature vector;

[0042] determining an intermediate media feature vector corresponding to the first training sample based on the initial media feature vector and a gate vector in the media feature screening module.

[0043] In an optional implementation, the determining module is specifically configured to perform:

[0044] determining an initial loss based on the task processing label and the predicted task processing result;

[0045] determining a loss adjustment value based on the gating adjustable parameter corresponding to each gating element in the gate vector;

[0046] determining a target loss based on the initial loss and the loss adjustment value.

[0047] In an optional implementation, the third processing module is specifically configured to perform:

[0048] screening a target gate element from the target gate vector in the trained media feature determination network, an element value of the target gate element being the first preset value;

[0049] eliminating the feature corresponding to the target gate element from the initial feature set to obtain a target media feature.

[0050] In an optional implementation, the apparatus further includes:

[0051] performing network adjustment on a task processing network based on the target media feature to obtain an adjusted task processing network;

[0052] processing the first training sample based on the target media feature to generate a second training sample;

[0053] training the adjusted task processing network based on the second training sample to obtain a trained task processing network.

[0054] In an optional implementation, the apparatus further includes:

[0055] extracting a to-be-processed media feature corresponding to the target media information based on the target media feature;

[0056] inputting the to-be-processed media feature into the trained task processing network for task processing to obtain a task processing result, the task processing result being used to indicate whether to perform a corresponding target task.

[0057] According to a third aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device can perform the media information processing method according to any one of the above embodiments.

[0058] According to a fourth aspect of the embodiments of the present disclosure, an electronic device is provided, including:

[0059] a processor;

[0060] a memory for storing instructions executable by the processor;

[0061] The processor is configured to execute the instructions to implement the media information processing method according to any one of the above embodiments.

[0062] According to a fifth aspect of the embodiments of the present disclosure, a computer program product is provided, the computer program product including a computer program, when the computer program is executed by a processor, the media information processing method provided in any one of the above embodiments is implemented.

[0063] The technical solutions provided by the embodiments of the present disclosure bring at least the following beneficial effects:

[0064] The embodiments of the present disclosure input the initial feature set into the media feature determination network for processing to obtain an intermediate media feature vector corresponding to the training sample; the media feature determination network includes a media feature screening module composed of a gate vector used to perform a media feature determination operation, and each gate element in the gate vector is determined based on a preset parameter composed of an attenuation parameter and an adjustable gating parameter; the intermediate media feature vector is input into the task processing network for task processing to obtain a predicted task processing result corresponding to the training sample; a target function loss is determined based on the task processing label, the predicted task processing result, and the adjustable gating parameter, and the target function loss is used for training, and the target gate vector in the trained media feature determination network is used to perform feature screening on the initial feature set to obtain a target media feature, the target gate element values corresponding to the features other than the target media feature in the target gate vector are all first preset values, and the difference between the target gate element corresponding to the target media feature in the target gate vector and a second preset value is less than or equal to a preset threshold. In this way, the gate vector is determined based on the attenuation parameter and the adjustable gating parameter, the feature screening module composed of the gate vector is applied to the media feature determination network, and the target function loss determined based on the task processing label, the predicted task processing result, and the adjustable gating parameter is used to train the task processing network. In the model training process, the attenuation parameter is gradually attenuated, the adjustable gating parameter is trained together with the model parameter, and the target gate element values corresponding to the target media feature in the trained target gate vector are all first preset values. In this way, since the target gate element values of the to-be-screened vectors in the trained target gate vector are all first preset values, it is not necessary to use the gate value to represent the weight order of each feature, so that the trained target gate vector is more consistent with the contribution degree of the corresponding feature to the model, thereby reducing the number of features while ensuring the accuracy of the model.

[0065] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0066] The accompanying drawings incorporated in the specification and forming a part of the present disclosure illustrate embodiments consistent with the present disclosure and serve to explain the principles of the present disclosure, and do not constitute an improper limitation on the present disclosure.

[0067] Figure 1 is an architecture diagram of a system applying a media information processing method according to an exemplary embodiment.

[0068] Figure 2 is a flowchart of a media information processing method according to an exemplary embodiment.

[0069] Figure 3 is a flowchart of a media information processing method according to an exemplary embodiment.

[0070] Figure 4 is a flowchart of a media information processing method according to an exemplary embodiment.

[0071] Figure 5 is a flowchart of a media information processing method according to an exemplary embodiment.

[0072] Figure 6 is a block diagram of a media information processing apparatus according to an exemplary embodiment.

[0073] Figure 7 is a block diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0074] In order to make the ordinary person skilled in the art better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be described clearly and completely below in conjunction with the drawings.

[0075] It should be noted that the terms "first", "second", and the like in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily have to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The implementation described in the following exemplary embodiments does not represent all implementations consistent with the present disclosure. Rather, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0076] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for display, analyzed data, etc.) involved in the present disclosure are all information and data authorized by the user or authorized by all parties.

[0077] Figure 1 is an architecture diagram of a system applying a media information processing method according to an exemplary embodiment, see Figure 1 The architecture diagram can include a terminal 10 and a server 20.

[0078] The terminal 10 can be, but is not limited to, a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart wearable device, a digital assistant, an augmented reality device, a virtual reality device, and the like. The terminal 10 can also include software such as an application program running in the terminal 10.

[0079] The server 20 can provide a media information processing service for the terminal 10. The server 20 can be, but is not limited to, a standalone server, a server cluster composed of multiple physical servers, or a distributed system. The server 20 can also be a cloud server providing one or more of cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, intermediate services, domain name services, security services, and big data and artificial intelligence platforms. The terminal 10 and the server 20 can be directly or indirectly connected through wired or wireless communication. The embodiments of the present disclosure are not limited in this regard.

[0080] It should be noted that the architecture diagram of the system applying the media information processing method of the present disclosure is not limited to this, and can also include more or fewer devices. Figure 1 The embodiments of the present disclosure are not limited in this regard.

[0081] The media information processing method provided by the embodiments of the present disclosure can be executed by a media information processing device, which can be integrated in a terminal or a server device in the form of hardware or software. The media information processing device can be implemented by a server or a terminal alone, or can be implemented by a terminal and a server in cooperation.

[0082] Figure 2 FIG. 1 is a flowchart of a media information processing method according to an exemplary embodiment. As shown in FIG. 1, the media information processing method can be applied to an electronic device, and the electronic device is taken as an example of the server in the above-mentioned implementation environment diagram, and includes the following steps. Figure 2

[0083] In step S201, a first training sample for a target task is obtained. The first training sample includes an initial feature set of media information and a corresponding task processing label.

[0084] Optionally, the target task can be a task related to a business scenario. For example, in a recommendation scenario, the target task can include a recommendation task. Specifically, for different processing objects, the target task can include video recommendation, commodity recommendation, video recall, commodity fine arrangement, etc. For example, in a computer vision scenario, the target task can be a recognition task, a classification task, etc. Specifically, for different processing objects, the target task can include an image recognition task, a face recognition task, an image classification task, a video classification task, a music classification task, etc.​

[0085] The first training sample is a sample for training a deep neural network which is matched with a target task. The first training sample can be obtained from a storage location such as a local database, a cloud, etc., and can include an initial feature set of media information and a corresponding task processing label. The media information can be resource information such as pictures, videos, texts, goods, etc., and media resource associated information. For different target tasks, the initial feature set of media information and the corresponding task processing label can be different.

[0086] For example, for a goods recommendation task, the initial feature set of media information can include goods features and user features, where the goods features can include a goods price of a recommended goods, ratings of all users who have purchased the goods, etc., and the user features can include user attribute features (such as gender, age, etc.) of a target user, past shopping records, etc. Correspondingly, the task processing label of the goods recommendation task can include at least one of whether the target user purchases, a user purchase probability, etc. For a video recommendation task, the initial feature set of media information can include resource features and user features, where the resource features can include at least one of picture features, video features, text features, etc., of a recommended video, and the user features can include user attribute features (such as gender, age, etc.) of a target user, past shopping records, and historical operation features (such as click, like, forward, comment, etc.) of the target user. Correspondingly, the task processing label of the video recommendation task can include at least one of whether the target user clicks, a user click probability, etc. For a picture recognition task, the media information can include picture features of a picture to be recognized, and the task processing label can include whether the picture to be recognized includes a target object, a category to which the target object in the picture to be recognized belongs, etc.

[0087] In step S202, the initial feature set is input into a media feature determination network for processing to obtain an intermediate media feature vector corresponding to the training sample.

[0088] The media feature determination network includes a media feature screening module composed of a gate vector for performing a media feature determination operation. The media feature screening module can be inserted between the input end of the initial feature set and the task processing network.

[0089] Each gate element in the gate vector is determined based on a preset parameter composed of an attenuation parameter and a gate adjustable parameter. The dimension of the gate vector is the same as the vector dimension corresponding to the initial feature set, and the intermediate media feature vector is associated with the gate vector. In the model training process, the gate adjustable parameter is a learnable parameter, which is used as a gate to perform feature screening processing in the neural network. The attenuation parameter is a non-learnable parameter, but the attenuation parameter can be gradually attenuated in the training process.

[0090] Optionally, in a case where the decay parameter in the preset parameter satisfies a preset decay condition, if the gating adjustable parameter in the preset parameter is zero, the function value of the preset parameter is a first preset value.

[0091] In an optional embodiment, the preset parameter can include a first function item composed of the decay parameter and the gating adjustable parameter. In the network training process, in a case where the decay parameter in the first function item satisfies a preset decay condition, if the gating adjustable parameter in the first function item is zero, the function value of the first function item is a first preset value, which can be a constant.

[0092] For example, taking the first preset value as zero, the preset parameter g ∈ (x) can be represented by the following expression:

[0093]

[0094] wherein, is the first function item, x is the gating adjustable parameter, and ∈ is the decay parameter.

[0095] Suppose that there are N features in each initial feature set, x is a vector containing N elements, and ∈ is a number, so that g ∈ (x) is also a vector containing N gating elements. For example, at the beginning of training, the values of x and ∈ can be initialized, for example, x can be initialized as 1.0 and ∈ can be initialized as 0.1. During the training process, ∈ can be decayed by a certain decay coefficient every certain number of steps, for example, ∈ is multiplied by a decay coefficient of, for example, 0.98 every 1000 steps to obtain the decayed ∈, and the decayed ∈ is used to replace the previous ∈ and the learnable parameter x to calculate the preset parameter g ∈ (x).

[0096] When ∈ is small enough, the preset parameter g ∈ (x) in the above formula (1) can satisfy the following condition:

[0097]

[0098] With the training going on, the model gradually converges, and when the decay parameter in the preset parameter meets the preset decay condition, if the gating adjustable parameter in the preset parameter is zero, the function value of the preset parameter is strictly the first preset value (i.e. zero). When the gating adjustable parameter in the preset parameter is non-zero, the difference between the parameter value of the preset parameter and the second preset value is less than or equal to the preset threshold. For example, the second preset value here can be 1, that is, if the gating adjustable parameter in the preset parameter is non-zero, the parameter value of the preset parameter is distributed around the second preset value (i.e. 1). Since there is a certain gap between the second preset value and the first preset value (i.e. zero), the function value of the preset parameter is strictly divided into zero and non-zero, or even strictly divided into zero and the value close to 1 in the training, which significantly narrows the gap between the gating element values corresponding to each feature, facilitating better distinguishing between the features to be retained and the features to be removed.

[0099] It should be understood that the preset parameter g ∈ The expression of (x) is not limited to the above expression (1), but can also be appropriately transformed on the basis of the above expression (1), such as multiplying by a coefficient, etc., and can also be other smooth derivable even functions that meet the above condition (2). Correspondingly, in the case of transforming the above expression (1), the above condition (2) is also adaptively adjusted.

[0100] Each gating element in the above gate vector is determined based on the preset parameter composed of the decay parameter and the gating adjustable parameter. In the early stage of model training, each value of the gate can be continuously taken between 0 and 1, and is derivable. With the training going on, the values of the gate gradually polarize: part of the gate values become the first preset value (e.g. 0), and part of the gate values become the number close to the second preset value (e.g. 1). After the training is completed, we can keep the part of the features with the gate values close to the second preset value as the selected features, and remove the features with the gate values of the first preset value. Multiply the remaining gate values into the media feature vector or the weight of the first fully connected layer of the network, so that the accuracy of the model does not suddenly decrease when the unimportant features are cut off, which is beneficial to hot start.

[0101] In another optional implementation, since each feature is irreplaceable, if a feature is wrongly learned as the first preset value (e.g., 0), it cannot be compensated by other features. To solve this problem, the preset parameter includes a first parameter term and a second parameter term. The first parameter term and the second parameter term are respectively composed of a decay parameter and a gating adjustable parameter. During network training, when the decay parameter in the preset parameter satisfies a preset decay condition, if the gating adjustable parameter in the preset parameter is zero, the function value of the preset parameter is the first preset value. Optionally, taking the first preset value of zero as an example, if the gating adjustable parameter in the first parameter term is zero, the parameter value of the first parameter term is zero; if the gating adjustable parameter in the second parameter term is zero, the parameter value of the second parameter term is zero.

[0102] When the gating adjustable parameter in the preset parameter is zero, the derivative of the preset parameter is positively correlated with the decay parameter. That is, the preset parameter will decay with the decay parameter. Specifically, at the beginning of network training, the value of the decay parameter is large, and since the derivative of the preset parameter is positively correlated with the decay parameter, the derivative value of the preset parameter at the gating adjustable parameter being zero is a relatively large value, so the gating adjustable parameter can be optimized by gradient descent and the like, and has good noise resistance. As the network training proceeds, the model gradually converges, and the derivative value of the preset parameter at the gating adjustable parameter being zero also decreases, and eventually the gating adjustable parameter can be well divided into zero and non-zero cases.

[0103] For example only, the preset parameter g ∈ (x) can be represented by the following expression:

[0104]

[0105] wherein, is the first parameter term, is the second parameter term, x is the gating adjustable parameter, ∈ is the decay parameter, and a is a hyperparameter.

[0106] Suppose that there are N features in each initial feature set, x is a vector containing N elements, ∈ is a number, and thus g ∈ (x) is also a vector containing N gating elements. For example, at the beginning of training, the values of x and ∈ can be initialized, for example, x can be initialized to 1.0 and ∈ can be initialized to 0.1. During training, ∈ can be decayed by a certain decay coefficient every certain number of steps, for example, ∈ is multiplied by a decay coefficient of, for example, 0.98 every 1000 steps to obtain a decayed ∈, and the decayed ∈ is used to replace the previous ∈ and the learnable parameter x to calculate the preset parameter g ∈ (x).

[0107] The preset parameter g in the above expression (3) ∈ (x) also satisfies the above condition (2), that is, as the training progresses, the model gradually converges, and when the decay parameter in the preset parameter satisfies the preset decay condition, if the gating adjustable parameter in the preset parameter is zero, the function value of the preset parameter is strictly the first preset value. When the gating adjustable parameter in the preset parameter is nonzero, the function value of the preset parameter satisfies the preset difference with the second preset value. The second preset value here can be 1, that is, when the gating adjustable parameter in the preset parameter is nonzero, the parameter value of the preset parameter is distributed around the second preset value (that is, 1). Since there is a certain gap between the second preset value and zero, the function value of the preset parameter is strictly divided into zero and nonzero, or even strictly divided into the first preset value (that is, zero) and the approximate second preset value (that is, 1) in the training, which significantly narrows the gap between the gating element values corresponding to each feature, facilitating better distinguishing between features that need to be retained and features that need to be removed.

[0108] The preset parameter g in the above expression (3) ∈ (x) is a piecewise function, but it has smooth continuity and derivability, and its derivative is as follows:

[0109]

[0110] From the derivative, it can be seen that at the origin (that is, x=0), the value of the derivative is The value of the derivative will decay with the decay of ∈. Thus, at the beginning of training, when the value of ∈ is relatively large, the model has a relatively large derivative value at x=0, thereby having good noise resistance; as the training progresses, the model gradually converges, and the derivative value at x=0 also decreases, and eventually the gating value can be well divided into zero and nonzero.

[0111] Although the preset parameter g in the above expression (3) ∈ (x) takes negative numbers, and after the training is completed, the negative sign can also be absorbed into the media feature vector or the weight of the first fully connected layer.

[0112] The above embodiment provides a preset parameter including a first function term and a second function term, which can remedy the features whose gates are learned as the first preset value (that is, zero) by mistake, ensure the robustness of the algorithm, and improve the noise resistance.

[0113] In an optional implementation, the processing of the initial feature set in the media feature determination network to obtain the intermediate media feature vector corresponding to the first training sample includes:

[0114] Mapping processing of the initial feature set in the media feature determination network to obtain the initial media feature vector;

[0115] Based on the initial media feature vector and the gate vector in the feature screening module, an intermediate media feature vector corresponding to the first training sample is determined.

[0116] Optionally, the initial feature set can be input into the media feature determination network, and the initial feature set is mapped and processed by using the embedding layer in the feature processing network to vectorize the initial feature set and obtain an initial media feature vector. Then, multiplication is performed based on the initial media feature vector and the gate vector in the feature screening module, that is, each feature value in the initial media feature vector e is multiplied by the corresponding gate element value in the gate vector g(x), and g(x)e obtained by the multiplication is taken as the intermediate media feature vector corresponding to the training sample.

[0117] In step S203, the intermediate media feature vector is input into the task processing network for task processing to obtain a predicted task processing result corresponding to the training sample.

[0118] Optionally, the above-mentioned intermediate media feature vector is taken as the input of the task processing network and is input into the task processing network for task processing to obtain a predicted task processing result corresponding to the training sample. Taking a video recommendation task as an example, the predicted task processing result corresponding to the training sample can be the probability of the user clicking the recommended video.

[0119] In an optional implementation, the task processing network can be a pre-trained initial task processing network. For example, the initial task processing network can be pre-trained by using the training sample until the training converges, and the pre-trained initial task processing network is taken as the task processing network. In this way, the task processing network used is a pre-trained network, and thus the efficiency of the subsequent training network and the accuracy of the task processing network can be improved.

[0120] In step S204, based on the task processing label, the predicted task processing result, and the gate adjustable parameter, a target function loss is determined, and based on the target function loss, the model parameters of the media feature determination network and the task processing network are trained to obtain a trained media feature determination network.

[0121] In an optional implementation, the determination of the target function loss based on the task processing label, the predicted task processing result, and the gate adjustable parameter includes:

[0122] In step S301, an initial loss is determined based on the task processing label and the predicted task processing result.

[0123] In step S302, a loss adjustment value is determined based on the gate adjustable parameter corresponding to each gate element in the gate vector.

[0124] In step S303, the target loss is determined based on the initial loss and the loss adjustment value.

[0125] Optionally, an initial loss is determined based on the task processing label and the predicted task processing result. This initial loss may include, but is not limited to, the cross-entropy loss function, the exponential loss function, etc. During training, the values ​​of the adjustable gating parameters corresponding to each gating element in the gating vector are continuously learned and trained together with the task processing model. The loss adjustment value is obtained by considering the adjustable gating parameters corresponding to each gating element in the gating vector, and the target loss function is calculated in combination with the initial loss corresponding to the network structure.

[0126] During model training, the model parameters of the media feature determination network and the task processing network can be trained together based on this objective function loss until the training termination condition is met, resulting in the trained media feature determination network. This termination condition can be minimizing the objective loss, the decay parameter value reaching a preset decay threshold, or the number of training iterations reaching a threshold, etc.

[0127] In one specific implementation, the task processing network can be represented as a function. Assume the original task processing network is y = f(w; e), where w are the structural parameters of the network model, e are all media feature vectors (i.e., the features initially input into the model), and y is the model's output; where e = [e1, e2, ..., e...]. N ] indicates that e is composed of N features.

[0128] When optimizing, use the following objective function:

[0129]

[0130] Where w represents all the learnable parameters of the network, and x represents the gating adjustable parameters; the function f(w; g(x)e)) represents the task processing network, which is a function of the media feature vector and all other learnable parameters of the network; g(x) represents the gate, and e represents the media feature vector. Labels that indicate authenticity The original loss function of the model is represented by ||x||1; ||x||1 represents the L1 norm of the x vector, which is the sum of the absolute values ​​of all elements of the x vector.

[0131] For optimizing w, the optimization methods used without feature selection can be employed. For optimizing x, methods such as proximal SGD can be used.

[0132] In step S205, the initial feature set is feature-screened based on the target gate vector in the trained media feature determination network to obtain target media features corresponding to the media information; the target gate elements corresponding to features other than the target media features in the target gate vector are all first preset values, and the difference between the target gate element corresponding to the target media feature in the target gate vector and the second preset value is less than or equal to a preset threshold.

[0133] In an optional implementation, the feature screening of the initial feature set based on the gate vector in the trained media feature determination network to obtain target media features includes:

[0134] The target gate element is screened out from the target gate vector in the trained media feature determination network, and the element value of the target gate element is a first preset value.

[0135] The feature corresponding to the target gate element is removed from the initial feature set to obtain the target media feature.

[0136] Optionally, after obtaining the trained media feature determination network, the target gate element with an element value of zero in the target gate vector is extracted, the feature at the position corresponding to the target gate element is deleted from the initial feature set, and the upper feature subset is reserved as the target media feature. For example, the first preset value is zero, for the initial feature set {feature a, feature b, feature c, feature d, feature e, feature f}, the trained target gate vector is {0.98, 0, 0, 0.97, 0, 1.0}, and since the gate element values of feature b, feature c and feature e are the first preset value, i.e. zero, these features need to be removed, then after the feature screening process, the determined target media features are {feature a, feature d, feature f}.

[0137] The information processing method of the present disclosure is described below taking a recommendation scenario as an example:

[0138] In the recommendation scenario, it is usually necessary to decide whether to recommend a given product to a user according to the features of the user and the product. The features of the user include the gender, age, past shopping records, etc. of the user, and the features of the product include the price, ratings of all users who have purchased the product, etc. These features can be represented in the form of a media feature vector as input of a recommendation network, and the media feature vector and the structure parameters of the network are trained together. The output of the recommendation network is whether a certain user will click on a certain product. After training with a large amount of data, a converged model is obtained. Then, feature screening can be performed based on this.

[0139] In one pre-trained recommendation network, the gate vector defined by the above expression (1) or expression (3) is inserted at the input end of the recommendation network, and then all gate parameters and model structure parameters and media feature vector parameters are trained. After this step of training converges, part of the gate values are strictly the first preset value (for example, zero), and part of the gate values are distributed around the second preset value (for example, 1). The features corresponding to the gates with strictly zero values are unnecessary features, and the features corresponding to the gates with values distributed around 1 are important features screened out. Taking the recommendation network as an example of a deep learning recommendation model, the deep learning recommendation module can be composed of 13 continuous features and 26 discrete features, and after the 13 continuous features pass through a 4-layer MLP, they can be output as a 16-dimensional vector and combined with the 26 discrete features to form an initial media feature vector. The gate vector can be inserted between the initial media feature vector and the input end of the recommendation network, and each feature in the initial media feature vector is multiplied by the gate value of the gate vector. If a certain gate element value is the first preset value, it means that the feature corresponding to the gate element will be removed.

[0140] After the target media features are determined, the recommendation network is structurally adjusted to obtain an adjusted recommendation network, which is fully trained in a hot start form and then put into use.

[0141] The embodiment of the present disclosure inputs the initial feature set into a media feature determination network for processing to obtain an intermediate media feature vector corresponding to a training sample; the media feature determination network includes a media feature screening module composed of a gate vector used for performing a media feature determination operation, and each gated element in the gate vector is determined based on a preset parameter composed of a decay parameter and a gated adjustable parameter; the intermediate media feature vector is input into a task processing network for task processing to obtain a predicted task processing result corresponding to the training sample; a target function loss is determined based on a task processing label and the predicted task processing result and the gated adjustable parameter, and the target function loss is used for training, and a target gate vector in the trained media feature determination network is used for feature screening on the initial feature set to obtain a target media feature, a target gated element value corresponding to a feature other than the target media feature in the target gate vector is a first preset value, and a difference between a target gated element corresponding to the target media feature in the target gate vector and a second preset value is less than or equal to a preset threshold. In this way, the gate vector is determined based on the decay parameter and the gated adjustable parameter, the feature screening module composed of the gate vector is applied to the media feature determination network, and the target function loss determined based on the task processing label, the predicted task processing result and the gated adjustable parameter is used for training the task processing network. In the model training process, the decay parameter is gradually decayed, the gated adjustable parameter is trained together with the model parameter, and the target gated element value corresponding to the target media feature in the trained target gate vector is the first preset value, so that the target gated element value of the to-be-screened vector in the trained target gate vector is the preset value, and the gate value is not used to represent the weight of each feature for sorting, so that the trained target gate vector is more consistent with the contribution of the corresponding feature to the model, thereby reducing the number of features while ensuring the accuracy of the model.

[0142] In an optional implementation, the method further includes:

[0143] In step S401, the task processing network is adjusted based on the target media feature to obtain an adjusted task processing network;

[0144] In step S402, the first training sample is processed based on the target media feature to generate a second training sample;

[0145] In step S403, the adjusted task processing network is trained based on the second training sample to obtain a trained task processing network.

[0146] Optionally, after the target media features are determined, the task processing network input can be reduced in power format based on the number of target media features, to obtain an adjusted task processing network, which is a sub-network of the original task processing network. Then, the first training samples are processed for feature screening based on the target media features, to generate second training samples, and the adjusted task processing network is trained again based on the second training samples, to obtain a trained task processing network suitable for hot start online. In this way, the task processing network obtained by training is adjusted and trained again based on the determined target media features, so that the task processing efficiency of the task processing network is higher, and the task processing effect is improved.

[0147] In an optional embodiment, the method further comprises:

[0148] In step S501, based on the target media features, the target media information corresponding to the to-be-processed media features is extracted.

[0149] In step S502, the to-be-processed media features are input into the trained task processing network for task processing, to obtain a task processing result, which is used to indicate whether to execute the corresponding target task.

[0150] Optionally, in response to the task processing request, the target media information can be obtained, and based on the determined target media features, the to-be-processed media features corresponding to the target media information are extracted, which have the same media features as the target media features; the to-be-processed media features are input into the trained task processing network for task processing, to obtain a task processing result used to indicate whether to execute the corresponding target task. If the task processing result indicates to execute the task processing, the corresponding task is executed, otherwise, the corresponding task is not executed. In this way, the target media information is screened based on the target media features, and the determined to-be-processed media features are a good feature subset representing the characteristics of the target media information and suitable for the target task. In this way, the to-be-processed media features are processed by the trained task processing network, to obtain a task processing result used to indicate whether to execute the corresponding target task, so that the task processing operation is quickly responded, and the accuracy and precision of the task processing result are improved.

[0151] Figure 6 is a block diagram of a media information processing device according to an exemplary embodiment. Referring to Figure 6 The device comprises:

[0152] The obtaining module 610 is configured to execute the first training samples for the target task, and the first training samples comprise an initial feature set of media information and a corresponding task processing label.

[0153] The first processing module 620 is configured to input the initial feature set into a media feature determination network for processing to obtain an intermediate media feature vector corresponding to the training sample; the media feature determination network comprises a media feature screening module composed of a gate vector used for performing a media feature determination operation, each gate element in the gate vector is determined based on a preset parameter composed of an attenuation parameter and an adjustable gating parameter, a dimension of the gate vector is the same as a vector dimension corresponding to the initial feature set, and the intermediate media feature vector is associated with the gate vector;

[0154] The second processing module 630 is configured to input the intermediate media feature vector into a task processing network for task processing to obtain a predicted task processing result corresponding to the training sample.

[0155] The determination module 640 is configured to determine a target function loss based on the task processing label, the predicted task processing result, and the adjustable gating parameter, train model parameters of the media feature determination network and the task processing network based on the target function loss, and obtain a trained media feature determination network.

[0156] The third processing module 650 is configured to perform feature screening on the initial feature set based on a target gate vector in the trained media feature determination network to obtain a target media feature; target gate element values corresponding to features other than the target media feature in the target gate vector are all first preset values, and a difference between a target gate element corresponding to the target media feature in the target gate vector and a second preset value is less than or equal to a preset threshold.

[0157] In an optional implementation, when the attenuation parameter in the preset parameter satisfies a preset attenuation condition, if the adjustable gating parameter in the preset parameter is zero, a function value of the preset parameter is the first preset value.

[0158] In an optional implementation, the preset parameter comprises a first function item and a second function item; the first function item and the second function item are both composed of the attenuation parameter and the adjustable gating parameter.

[0159] When the adjustable gating parameter in the preset parameter is zero, a derivative of the preset parameter is positively correlated with the attenuation parameter.

[0160] In an optional implementation, the first processing module is specifically configured to perform:

[0161] map the initial feature set into the media feature determination network for mapping processing to obtain an initial media feature vector;

[0162] determine an intermediate media feature vector corresponding to the training sample based on the initial media feature vector and the gate vector in the feature screening module.

[0163] In an optional implementation, the determining module is specifically configured to perform:

[0164] determine an initial loss based on the task processing label and the predicted task processing result;

[0165] determine a loss adjustment value based on a gate adjustable parameter corresponding to each gate element in the gate vector;

[0166] determine a target loss based on the initial loss and the loss adjustment value.

[0167] In an optional implementation, the third processing module is specifically configured to perform:

[0168] screen a target gate element from the target gate vector in the trained media feature determination network, an element value of the target gate element being the first preset value;

[0169] remove a feature corresponding to the target gate element from the initial feature set to obtain a target media feature.

[0170] In an optional implementation, the apparatus further includes:

[0171] perform network adjustment on a task processing network based on the target media feature to obtain an adjusted task processing network;

[0172] perform processing on the first training sample based on the target media feature to generate a second training sample;

[0173] train the adjusted task processing network based on the second training sample to obtain a trained task processing network.

[0174] In an optional implementation, the apparatus further includes:

[0175] extract a to-be-processed media feature corresponding to target media information based on the target media feature;

[0176] input the to-be-processed media feature into the trained task processing network for task processing to obtain a task processing result, the task processing result being used to indicate whether to perform a corresponding target task.

[0177] As to the method in the foregoing embodiments, the specific manners of each step have been described in the foregoing method embodiments in detail, and thus will not be described in detail here.

[0178] Figure 7is a block diagram of an electronic device according to an example embodiment. Referring to Figure 7 The electronic device includes a processor; a memory for storing processor-executable instructions; wherein the processor is configured to implement the steps of any of the above-described media information processing methods when executing the instructions stored on the memory.

[0179] The electronic device can be a terminal, a server or similar computing device. In the following, the electronic device is taken as an example of a server, Figure 7 is a block diagram of an electronic device for media information processing according to an example embodiment. The electronic device 1000 can vary greatly due to different configurations or capabilities, and can include one or more Central Processing Units (CPUs) 1010 (the processor 1010 can include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA processing device), a memory 1030 for storing data, one or more storage media 1020 (such as one or more mass storage devices) for storing applications 1023 or data 1022. Among them, the memory 1030 and the storage medium 1020 can be temporary storage or persistent storage. The program stored in the storage medium 1020 can include one or more modules, each of which can include a series of instruction operations in the electronic device. Further, the central processing unit 1010 can be configured to communicate with the storage medium 1020 and execute a series of instruction operations in the storage medium 1020 on the electronic device 1000.

[0180] The electronic device 1000 can also include one or more power supplies 1060, one or more wired or wireless network interfaces 1050, one or more input / output interfaces 1040, and / or one or more operating systems 1021, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc.

[0181] The input / output interface 1040 can be used to receive or send data via a network. The above-mentioned network can include a wireless network provided by a communication provider of the electronic device 1000. In one example, the input / output interface 1040 includes a network adapter (Network Interface Controller, NIC) that can be connected to other network devices through a base station so as to communicate with the Internet. In an example embodiment, the input / output interface 1040 can be a radio frequency (Radio Frequency, RF) module for communicating with the Internet through a wireless manner.

[0182] Those of ordinary skill in the art can understand that,Figure 7 The structure shown is merely schematic, and does not limit the structure of the electronic device described above. For example, the electronic device 1000 can further include more or fewer components than those shown in FIG. 10, or have a different configuration from that shown in FIG. 10. Figure 7 Figure 7 The structure shown is merely schematic, and does not limit the structure of the electronic device described above. For example, the electronic device 1000 can further include more or fewer components than those shown in FIG. 10, or have a different configuration from that shown in FIG. 10.

[0183] In an example embodiment, a computer readable storage medium including instructions, such as a memory including instructions, is also provided, which can be executed by a processor of the electronic device 1000 to complete the method described above. Optionally, the computer readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0184] In an example embodiment, a computer storage medium is also provided, which, when instructions in the computer storage medium are executed by a processor of an electronic device, enables the electronic device to perform the steps of the method provided in any of the embodiments described above.

[0185] In an example embodiment, a computer program product is also provided, which includes a computer program / instructions, which, when executed by a processor, implements the method provided in any of the embodiments described above. Optionally, the computer program is stored in a computer readable storage medium. The processor of the electronic device reads the computer program from the computer readable storage medium, and the processor executes the computer program, so that the electronic device performs the method provided in any of the embodiments described above.

[0186] ​Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, storage, databases, or other media in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0187] Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the features disclosed herein. This application is intended to cover any variations, uses, or adaptations of the disclosure that are deemed to fall within the general principles of the disclosure and include commonly known or customary practice in the art. The specification and examples are to be considered exemplary only, with the true scope and spirit of the disclosure being indicated by the following claims.

[0188] It should be understood that the present disclosure is not limited to the precise structures herein described and illustrated in the drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the present disclosure is limited only by the claims that follow.

Claims

1. A media information processing method, characterized in that, include: Obtain a first training sample for the target task, the first training sample including an initial feature set of media information and corresponding task processing labels; The initial feature set of the media information includes resource features, which include at least one of the following: image features of the video to be recommended, video features of the video to be recommended, or text features of the video to be recommended; The initial feature set is input into a media feature determination network for processing to obtain an intermediate media feature vector corresponding to the first training sample. The media feature determination network includes a media feature filtering module composed of gate vectors for performing feature determination operations. Each gate element in the gate vector is determined based on preset parameters composed of attenuation parameters and gate adjustable parameters. The dimension of the gate vector is the same as the dimension of the vector corresponding to the initial feature set. The intermediate media feature vector is associated with the gate vector. The intermediate media feature vector is input into the task processing network for task processing to obtain the prediction task processing result corresponding to the first training sample. Based on the task processing label, the predicted task processing result, and the gated adjustable parameters, a target loss is determined, and the model parameters of the media feature determination network and the task processing network are trained based on the target loss to obtain the trained media feature determination network. Based on the trained media features, the target gate vector in the network is used to filter the initial feature set to obtain the target media features corresponding to the media information; the target gate element values ​​corresponding to the features other than the target media features in the target gate vector are all first preset values, and the difference between the target gate element corresponding to the target media features in the target gate vector and the second preset value is less than or equal to a preset threshold.

2. The method according to claim 1, characterized in that, If the attenuation parameter in the preset parameters meets the preset attenuation condition, and the gate adjustable parameter in the preset parameters is zero, the function value of the preset parameters is the first preset value.

3. The method according to claim 1, characterized in that, The preset parameter items include a first parameter item and a second parameter item; both the first parameter item and the second parameter item are respectively composed of an attenuation parameter and a gate adjustable parameter; When the gate adjustable parameter in the preset parameter item is zero, the derivative of the preset parameter item is positively correlated with the attenuation parameter.

4. The method according to any one of claims 1-3, characterized in that, The step of inputting the initial feature set into the media feature determination network for processing to obtain the intermediate media feature vector corresponding to the first training sample includes: The initial feature set is input into the media feature determination network for mapping processing to obtain the initial media feature vector; Based on the initial media feature vector and the gate vector in the media feature filtering module, the intermediate media feature vector corresponding to the first training sample is determined.

5. The method according to any one of claims 1-3, characterized in that, The determination of the target media loss based on the task processing label, the predicted task processing result, and the gated adjustable parameters includes: Based on the task processing label and the predicted task processing result, determine the initial loss; Based on the adjustable gating parameters corresponding to each gating element in the gate vector, the loss adjustment value is determined. The target loss is determined based on the initial loss and the parameter adjustment value.

6. The method according to any one of claims 1-3, characterized in that, The step of determining the target media features based on the target gate vector in the trained media feature determination network to perform feature filtering on the initial feature set to obtain target media features includes: Target gate elements are selected from the target gate vectors in the trained media feature determination network, and the element value of the target gate element is the first preset value; The target media features are obtained by removing the features corresponding to the target gating element from the initial feature set.

7. The method according to any one of claims 1-3, characterized in that, The method further includes: Based on the target media characteristics, the task processing network is adjusted to obtain the adjusted task processing network. The first training sample is processed based on the target media features to generate a second training sample; The adjusted task processing network is trained based on the second training sample to obtain the trained task processing network.

8. The method according to claim 7, characterized in that, The method further includes: Based on the target media features, extract the media features to be processed corresponding to the target media information; The media features to be processed are input into the trained task processing network for task processing to obtain task processing results, which include indicators for whether to execute the corresponding target task.

9. A media information processing device, characterized in that, include: The acquisition module is configured to acquire a first training sample for a target task. The first training sample includes an initial feature set of media information and a corresponding task processing label. The initial feature set of media information includes resource features, which include at least one of the following: image features of the video to be recommended, video features of the video to be recommended, or text features of the video to be recommended. The first processing module is configured to process the initial feature set into a media feature determination network to obtain an intermediate media feature vector corresponding to the training samples. The media feature determination network includes a media feature filtering module composed of gate vectors for performing media feature determination operations. Each gate element in the gate vector is determined based on preset parameters composed of attenuation parameters and gate adjustable parameters. The dimension of the gate vector is the same as the dimension of the vector corresponding to the initial feature set. The intermediate media feature vector is associated with the gate vector. The second processing module is configured to input the intermediate media feature vector into the task processing network for task processing, and obtain the prediction task processing result corresponding to the training sample. The determination module is configured to perform an objective function loss based on the task processing label, the predicted task processing result, and the gated adjustable parameters, and to train the model parameters of the media feature determination network and the task processing network based on the objective function loss, so as to obtain the trained media feature determination network. The third processing module is configured to perform feature filtering on the initial feature set based on the target gate vector in the trained media feature determination network to obtain the target media feature corresponding to the media information; the target gate element values ​​corresponding to the features other than the target media feature in the target gate vector are all first preset values, and the difference between the target gate element corresponding to the target media feature in the target gate vector and the second preset value is less than or equal to a preset threshold.

10. The apparatus according to claim 9, characterized in that, If the attenuation parameter in the preset parameters meets the preset attenuation condition, and the gate adjustable parameter in the preset parameters is zero, the function value of the preset parameters is the first preset value.

11. The apparatus according to claim 9, characterized in that, The preset parameters include a first function term and a second function term; both the first function term and the second function term are respectively composed of an attenuation parameter and a gated adjustable parameter; When the gate adjustable parameter in the preset parameters is zero, the derivative of the preset parameters is positively correlated with the attenuation parameter.

12. The apparatus according to any one of claims 9-11, characterized in that, The first processing module is specifically configured to execute: The initial feature set is input into the media feature determination network for mapping processing to obtain the initial media feature vector; Based on the initial media feature vector and the gate vector in the media feature filtering module, the intermediate media feature vector corresponding to the first training sample is determined.

13. The apparatus according to any one of claims 9-11, characterized in that, The determination module is specifically configured to execute: Based on the task processing label and the predicted task processing result, determine the initial loss; Based on the adjustable gating parameters corresponding to each gating element in the gate vector, the loss adjustment value is determined. The target loss is determined based on the initial loss and the loss adjustment value.

14. The apparatus according to any one of claims 9-11, characterized in that, The third processing module is specifically configured to execute: Target gate elements are selected from the target gate vectors in the trained media feature determination network, and the element value of the target gate element is the first preset value; The target media features are obtained by removing the features corresponding to the target gating element from the initial feature set.

15. The apparatus according to any one of claims 9-11, characterized in that, The device further includes: Based on the target media characteristics, the task processing network is adjusted to obtain the adjusted task processing network. The first training sample is processed based on the target media features to generate a second training sample; The adjusted task processing network is trained based on the second training sample to obtain the trained task processing network.

16. The apparatus according to claim 15, characterized in that, The device further includes: Based on the target media features, extract the media features to be processed corresponding to the target media information; The media features to be processed are input into the trained task processing network for task processing to obtain task processing results, which are used to indicate whether to execute the corresponding target task.

17. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the media information processing method as described in any one of claims 1 to 8.

18. A computer-readable storage medium, wherein when instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the media information processing method as claimed in any one of claims 1 to 8.

19. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the media information processing method according to any one of claims 1 to 8.

Images