Model training method of AI switch platform in creative content field

Abstract

The invention discloses a model training method for an AI switch platform in the field of creative content, and the method comprises the steps: 1, adding a training platform on the basis of an original AI switch platform infrastructure, so as to achieve the customization of a model; 2, generating a hybrid model through a training platform to realize a personalized AI model solution; comprising the following steps: dynamically selecting an independent or combined model to realize a training task according to the characteristics of subtasks, and taking the combined model as the training task by monitoring the use frequency of a specific combined model in a production task; extracting data elements of each model in the combined model, combining the data elements of the plurality of independent models item by item, and rearranging the data elements according to the use frequency; the cross data elements ranked in the front are extracted to serve as new data elements forming a new hybrid model, and finally the new hybrid model generated after training is loaded to a model library; by means of model customization development, it is ensured that the service scene is met.

Description

Technical Field

[0001] The present invention relates to model training, and in particular to a model training method for an AI switch platform in the field of creative content. Background Art

[0002] In the original business model, technology is the basis of product innovation, and products are the specific implementation form. With the rise of AI, innovative business models start from application scenarios, build an ecosystem with the platform as the core, and achieve this goal through functional technologies.

[0003] In the creative content industry, application scenarios include graphic images, video clips, smart editing, one-click filming and many other design scenarios. We build an AI switch platform specifically for creators in the fields of video animation, graphic design, fashion design, animation and games. It involves the integration of multiple elastic intelligent agents to achieve full-process services from user needs, creative tools to model training and model deployment. By integrating multimodal models and enhancing scalability, we provide users with powerful AI application support.

[0004] However, while AIGC empowers the creative content industry and brings more and more innovative application scenarios, common data update and model iteration methods still have limitations and cannot fully meet the needs. Summary of the invention

[0005] In view of the above technical problems, the purpose of the present invention is to provide a model training method for an AI switch platform in the field of creative content, and ensure that it conforms to business scenarios through model customization and development.

[0006] In order to achieve the above purpose, the technical solution adopted by the present invention is: a model training method of an AI switch platform in the field of creative content, specifically the following steps:

[0007] Step 1: Add a training platform to the existing AI switch platform infrastructure to achieve model customization;

[0008] Step 2: Generate a hybrid model through the training platform to achieve a personalized AI model solution.

[0009] As a preferred method of the present invention, step 1 specifically includes: adding a training platform to the AI ​​switch platform layer of the infrastructure, and adding a model customization service to the AI ​​resource service layer of the infrastructure.

[0010] As a preferred embodiment of the present invention, the infrastructure includes an application scenario layer, an AI switch platform layer, an AI resource service layer, and a computing power base layer;

[0011] The application scenario is to obtain application requirements, decompose them into multiple subtasks according to the application requirements, and then customize the solution.

[0012] The AI ​​switch platform uses AI capabilities to inherit existing resources and generate new AI resources, including an AI capability library and a model library.

[0013] AI resource services refer to AI resources and AI application services, including AI material libraries and AI tool plug-ins;

[0014] The computing power base is the cloud-based elastic computing pool.

[0015] As a preferred embodiment of the present invention, application scenarios include general scenarios and industry scenarios. General scenarios include video creation and graphic design, and industry scenarios include workwear and clothing design.

[0016] As a preferred embodiment of the present invention, the AI ​​capability library includes multimodal analysis, factor extraction, material retrieval, multimedia analysis and auxiliary generation, and the model library is an expandable set of special models, and the special models include text-based graph models and text-based video models.

[0017] As a preferred embodiment of the present invention, the AI ​​material library includes pictures, templates and design resources generated by artificial intelligence; the AI ​​tool plug-in includes AI-assisted design tools and AI video intelligent editing tools.

[0018] As a preferred embodiment of the present invention, step 2 specifically includes:

[0019] Step 2.1: Training tasks: The production task is composed of several specific subtasks in the application scenario, and independent or combined models are dynamically selected according to the characteristics of the subtasks to achieve it; by monitoring the usage frequency of specific combined models in the production task, that is, assuming that the production task is T, the subtask is t, the independent model is m, T = t1 + t2 ... + ti, and the specific subtask ti is implemented by one or more specific mi. If the usage frequency of the specific combined model is greater than 20%, the combined model is used as a training task;

[0020] Step 2.2: Training process: First, extract the data elements of each model in the combined model, which are the key features of the model; then, combine the data elements of multiple independent models one by one and rearrange them according to the frequency of use; finally, extract the cross data elements with the highest ranking as the new data elements constituting the new hybrid model;

[0021] Assume that subtask t is implemented by independent models mi and mj, and the data element of mi is An, and the data element of mj is Bn, then the recombined data element Zn = ∑AnBn, and extract the cross data element En with the highest frequency of use in Zn;

[0022] Step 2.3: Use the trained model: Load the new hybrid model generated after training into the model library.

[0023] Compared with the prior art, the present invention has the following technical effects:

[0024] By using the model training method of the AI ​​switch platform in the field of creative content disclosed in the present invention and through model customization and development, it can be ensured that it conforms to the business scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Schematic diagram of the basic architecture of the AI ​​switch platform of the present invention.

[0026] Figure 2 The figure is a flow chart of the model training method of the AI ​​switch platform of the present invention.

[0027] Figure 3 This is a schematic diagram of the present invention loading a new hybrid model generated after training into a model library. DETAILED DESCRIPTION

[0028] In order to explain the technical content, structural features, achieved objectives and effects of the technical solution in detail, the following is a detailed description in conjunction with specific embodiments and accompanying drawings.

[0029] This embodiment discloses a model training method for an AI switch platform in the field of creative content, comprising the following steps:

[0030] Step 1: Add a training platform to the existing AI switch platform infrastructure to achieve model customization;

[0031] Step 2: Generate a hybrid model through the training platform to achieve a personalized AI model solution.

[0032] like Figure 1 As shown in the figure, the infrastructure includes the application scenario layer, AI switch platform layer, AI resource service layer and computing power base layer.

[0033] The application scenario is to obtain application requirements, then decompose them into multiple subtasks according to the application requirements, and then customize solutions. General scenarios include video creation and graphic design, and industry scenarios include workwear / clothing design.

[0034] The AI ​​switch platform uses AI capabilities to inherit existing resources and generate new AI resources, including an AI capability library and a model library. The AI ​​capability library includes multimodal analysis, feature extraction, material retrieval, multimedia analysis, auxiliary generation, etc. The model library is an expandable collection of dedicated models. The dedicated models have powerful multimodal capabilities and can process various types of data, including text-based graph models, text-based video models, etc.

[0035] AI resource services refer to AI resources and AI application services, including AI material libraries and AI tool plug-ins. The AI ​​material library refers to pictures, templates, and design resources generated by artificial intelligence. AI tool plug-ins include AI-assisted design tools, AI video intelligent editing tools, etc.

[0036] The computing power base is the cloud-based elastic computing pool.

[0037] This embodiment mainly adds a training platform at the AI ​​switch layer and a model customization service at the AI ​​resource service layer.

[0038] Preferably, the step 2 specifically includes:

[0039] Step 2.1. Training tasks: The production task is composed of several specific subtasks in the application scenario, and the independent or combined model is dynamically selected according to the characteristics of the subtasks to achieve it. By monitoring the frequency of use of specific combined models in the production task, that is, assuming that the production task is T, the subtask is t, and the independent model is m, T = t1+t2...+t i , a specific subtask t i By one or more specific m i If the usage frequency of a specific combination model is greater than 20%, the combination model will be used as a training task.

[0040] Step 2.2. Training process: Figure 2 As shown, first, the data elements of each model in the combined model are extracted, and the data elements are the key features of the model; then, the data elements of multiple independent models are combined item by item and rearranged according to the frequency of use; finally, the cross data elements with the highest ranking are extracted as the new data elements constituting the new hybrid model.

[0041] Assume that subtask t is represented by an independent model m i and m j Together, i The data element is A n , m j The data element is B n , then the recombined data element Z n =∑AnBn, extract Z n The most frequently used cross-data element E n .

[0042] Step 2.3. Use the trained model: Figure 3As shown, the new hybrid model generated after training is loaded into the model library. It should be noted that although the above embodiments have been described in this article, the patent protection scope of the present invention is not limited thereby. Therefore, based on the innovative concept of the present invention, changes and modifications made to the embodiments described herein, or equivalent structures or equivalent process transformations made using the contents of the specification and drawings of the present invention, directly or indirectly applying the above technical solutions to other related technical fields are all included in the patent protection scope of the present invention.

Claims

1. A model training method for an AI switch platform in the field of creative content, characterized in that: The following steps are involved: Step 1: Add a training platform to the existing AI switch platform infrastructure to achieve model customization; Step 2: Generate a hybrid model through the training platform to achieve a personalized AI model solution.

2. According to claim 1, a model training method for an AI switch platform in the field of creative content, characterized in that: Specifically, step 1 includes: adding a training platform to the AI ​​switch platform layer of the infrastructure, and adding a model customization service to the AI ​​resource service layer of the infrastructure.

3. According to claim 2, a model training method for an AI switch platform in the field of creative content, characterized in that: The infrastructure includes the application scenario layer, AI switch platform layer, AI resource service layer, and computing power base layer; The application scenario is to obtain application requirements, decompose them into multiple subtasks according to the application requirements, and then customize the solution. The AI ​​switch platform uses AI capabilities to inherit existing resources and generate new AI resources, including an AI capability library and a model library. AI resource services refer to AI resources and AI application services, including AI material libraries and AI tool plug-ins; The computing power base is the cloud-based elastic computing pool.

4. According to claim 3, a model training method for an AI switch platform in the field of creative content, characterized in that: Application scenarios include general scenarios and industry scenarios. General scenarios include video creation and graphic design, and industry scenarios include workwear and clothing design.

5. According to claim 3, a model training method for an AI switch platform in the field of creative content, characterized in that: The AI ​​capability library includes multimodal analysis, factor extraction, material retrieval, multimedia analysis and auxiliary generation. The model library is an expandable set of special models, and the special models include text-based image models and text-based video models.

6. According to claim 3, a model training method for an AI switch platform in the field of creative content, characterized in that: The AI ​​material library includes pictures, templates and design resources generated by artificial intelligence; the AI ​​tool plug-ins include AI-assisted design tools and AI video intelligent editing tools.

7. A model training method for an AI switch platform in the field of creative content according to claim 1, characterized in that: The step 2 specifically includes: Step 2.1: Training tasks: The production task is composed of several specific subtasks in the application scenario, and independent or combined models are dynamically selected according to the characteristics of the subtasks to achieve it; by monitoring the usage frequency of specific combined models in the production task, that is, assuming that the production task is T, the subtask is t, the independent model is m, T = t1 + t2 ... + ti, and the specific subtask ti is implemented by one or more specific mi. If the usage frequency of the specific combined model is greater than 20%, the combined model is used as a training task; Step 2.2: Training process: First, extract the data elements of each model in the combined model, which are the key features of the model; then, combine the data elements of multiple independent models one by one and rearrange them according to the frequency of use; finally, extract the cross data elements with the highest ranking as the new data elements constituting the new hybrid model; Assume that subtask t is implemented by independent models mi and mj, and the data element of mi is An, and the data element of mj is Bn, then the recombined data element Zn = ∑AnBn, and extract the cross data element En with the highest frequency of use in Zn; Step 2.3: Use the trained model: Load the new hybrid model generated after training into the model library.

Images