Resource loading method and system based on Deepseek

Abstract

The invention provides a Deepseek-based resource loading method and system, and the method comprises the steps: analyzing each to-be-loaded scene resource through a Deepseek large model, and generating a resource topological graph with a weight; the fluctuation spectrum of the network bandwidth of the client, the fragmentation degree of the GPU video memory and the user behavior data are acquired; according to the obtained fluctuation frequency spectrum of the network bandwidth, the fragmentation degree of the GPU video memory and the user behavior data, constructing an environment state vector containing space-time dynamic characteristics; coding the resource topological graph into a topological feature vector through a graph neural network; inputting the topological feature vector and the environment state vector into a pre-trained deep reinforcement learning model to obtain a resource loading strategy; and loading each scene resource according to the resource loading strategy. Resource loading can be effectively carried out in combination with a network environment and a hardware condition load.

Description

Technical Field

[0001] This application relates to the field of computer technology, and particularly to a resource loading method and system based on Deepseek. Background Art

[0002] With the development of computer technology, the efficient loading of dynamic scene resources has become a core challenge for enhancing the user experience. In a complex and ever-changing network environment and heterogeneous hardware conditions, the client needs to load a large amount of high-precision scene resources in real time. Traditional resource loading schemes rely on static priority rules or offline pre-computation strategies, which are difficult to dynamically adapt to the real-time changing hardware load and network status, easily leading to a sharp increase in loading latency, video memory overflow, or redundant resource transmission, causing problems such as frame stuttering and scene rendering breaks, seriously restricting the smoothness and immersion of high-real-time application scenarios. Summary of the Invention

[0003] The technical problem to be solved by this application is to provide a resource loading method and system based on Deepseek, which can effectively combine the network environment and hardware condition load for resource loading. The specific solutions are as follows:

[0004] The resource loading method based on Deepseek includes:

[0005] In response to a scene loading trigger instruction, parse each scene resource to be loaded through the Deepseek large model to generate a weighted resource topology map;

[0006] Obtain the fluctuation spectrum of the network bandwidth of the client, the fragmentation degree of the GPU video memory, and user behavior data;

[0007] Construct an environmental state vector containing spatio-temporal dynamic features according to the obtained fluctuation spectrum of the network bandwidth, the fragmentation degree of the GPU video memory, and user behavior data;

[0008] Encode the resource topology map into a topology feature vector through a graph neural network;

[0009] Input the topology feature vector and the environmental state vector into a pre-trained deep reinforcement learning model to obtain a resource loading strategy, where the resource loading strategy includes a Markov decision sequence of the resource loading order, a priority coefficient for each scene resource, and a pre-cached resource subset based on the predicted perspective change;

[0010] Load each scene resource according to the resource loading strategy.

[0011] For the above method, optionally, the step of parsing each scene resource to be loaded through the Deepseek large model to generate a weighted resource topology map includes:

[0012] Analyze the material sharing relationship, light dependence chain, physical collision body topological connection, and semantic association degree among each of the scene resources through the DeepSeek large model;

[0013] Construct a resource topology map based on the material sharing relationship, light dependence chain, physical collision body topological connection, and semantic association degree among each of the scene resources.

[0014] For the above method, optionally, the deep reinforcement learning model includes an encoder, a first branch network, a second branch network, and a policy fusion module;

[0015] The step of inputting the topological feature vector and the environmental state vector into the pre-trained deep reinforcement learning model to obtain a resource loading policy includes:

[0016] Input the topological feature vector and the environmental state vector into the encoder in the pre-trained deep reinforcement learning model, so that the encoder performs graph Fourier transform on the topological feature vector to extract a frequency domain subspace vector; decompose the environmental state vector into a network bandwidth fluctuation component and a hardware fragmentation component, and fuse the frequency domain subspace vector, the network bandwidth fluctuation component, and the hardware fragmentation component through a multi-head attention mechanism to obtain a fusion vector;

[0017] Input the fusion vector into the first branch network and the second branch network respectively to obtain a Markov decision sequence, a resource priority coefficient, and a pre-cache identifier output by the first branch network, and the access probability entropy of each of the scene resources output by the second branch network;

[0018] Input the Markov decision sequence, the resource priority coefficient, the pre-cache identifier, the access probability entropy of each of the scene resources, the fluctuation spectrum of the network bandwidth, and the fragmentation degree of the GPU video memory into the policy fusion module to obtain a resource loading policy.

[0019] For the above method, optionally, the step of loading each of the scene resources according to the resource loading policy includes:

[0020] Generate a state transition probability matrix according to the Markov decision sequence;

[0021] Allocate each of the scene resources to a corresponding target area according to the state transition probability matrix and the pre-cache resource subset based on the predicted perspective change, where the target area includes a pre-allocated area of the GPU video memory, a video memory swap area, or a disk cache area;

[0022] Allocate bandwidth to each of the scene resources according to the priority coefficient of each scene resource in each target area and the network bandwidth fluctuation spectrum.

[0023] For each of the scenario resources, load the scenario resources according to the bandwidth allocated for the scenario resources.

[0024] In the above method, optionally, after loading the scenario resources according to the bandwidth allocated for the scenario resources, the method further includes:

[0025] Monitor in real time the loading time consumption, GPU video memory occupancy rate, and actual fluctuation value of the network bandwidth of the currently loaded scenario resources;

[0026] Generate a dynamic performance metric based on the loading time consumption, video memory occupancy rate, and actual fluctuation value;

[0027] Compare the dynamic performance metric with a predefined resource loading quality threshold:

[0028] If the dynamic performance metric is lower than the resource loading quality threshold, then based on the fragmentation degree of the current GPU video memory, the network bandwidth fluctuation spectrum, and the user behavior data, update the environmental state vector, and re-enter the deep reinforcement learning model to generate a corrected resource loading strategy;

[0029] Adjust the loading order of the scenario resources that have not been fully loaded based on the updated priority coefficients in the corrected resource loading strategy.

[0030] A resource loading system based on Deepseek includes:

[0031] A first generation unit, configured to, in response to a scenario loading trigger instruction, parse each scenario resource to be loaded through a Deepseek large model, and generate a weighted resource topology map;

[0032] An acquisition unit, configured to acquire the network bandwidth fluctuation spectrum, the fragmentation degree of the GPU video memory, and the user behavior data of the client;

[0033] A construction unit, configured to construct an environmental state vector including spatio-temporal dynamic features according to the acquired network bandwidth fluctuation spectrum, the fragmentation degree of the GPU video memory, and the user behavior data;

[0034] An encoding unit, configured to encode the resource topology map into a topology feature vector through a graph neural network;

[0035] An execution unit, configured to input the topology feature vector and the environmental state vector into a pre-trained deep reinforcement learning model to obtain a resource loading strategy, where the resource loading strategy includes a Markov decision sequence of the resource loading order, the priority coefficients of each scenario resource, and a pre-cached resource subset based on the predicted perspective change;

[0036] A loading unit for loading each of the scenario resources according to the resource loading policy.

[0037] For the above system, optionally, the first generation unit includes:

[0038] A first execution subunit for analyzing the material sharing relationship, light dependence chain, physical collision body topological connection, and semantic association degree among each of the scenario resources through the Deepseek large model;

[0039] A construction subunit for constructing a resource topology map according to the material sharing relationship, light dependence chain, physical collision body topological connection, and semantic association degree among each of the scenario resources.

[0040] For the above system, optionally, the deep reinforcement learning model includes an encoder, a first branch network, a second branch network, and a policy fusion module;

[0041] The execution unit includes:

[0042] An encoding subunit for inputting the topology feature vector and the environmental state vector into the encoder in the pre-trained deep reinforcement learning model, so that the encoder performs graph Fourier transform on the topology feature vector to extract a frequency domain subspace vector; decomposing the environmental state vector into a network bandwidth fluctuation component and a hardware fragmentation component, and fusing the frequency domain subspace vector, the network bandwidth fluctuation component, and the hardware fragmentation component through a multi-head attention mechanism to obtain a fusion vector;

[0043] A second execution subunit for respectively inputting the fusion vector into the first branch network and the second branch network to obtain a Markov decision sequence, a resource priority coefficient, and a pre-cache identifier output by the first branch network, and the access probability entropy of each of the scenario resources output by the second branch network;

[0044] A third execution subunit for inputting the Markov decision sequence, the resource priority coefficient, the pre-cache identifier, the access probability entropy of each of the scenario resources, the fluctuation spectrum of the network bandwidth, and the degree of fragmentation of the GPU video memory into the policy fusion module to obtain a resource loading policy.

[0045] For the above system, optionally, the loading unit includes:

[0046] A generation subunit for generating a state transition probability matrix according to the Markov decision sequence;

[0047] A first allocation subunit, configured to allocate each of the scenario resources to a corresponding target area according to the state transition probability matrix and the pre-cached resource subset based on the predicted perspective change, where the target area includes a pre-allocated area of GPU video memory, a video memory swap area, or a disk cache area;

[0048] A first allocation subunit, configured to allocate bandwidth to each of the scenario resources according to the priority coefficient of each scenario resource in each target area and the network bandwidth fluctuation spectrum;

[0049] A loading subunit, configured to load a scenario resource according to the bandwidth allocated to the scenario resource for each of the scenario resources;

[0050] The above system, optionally, further includes:

[0051] A monitoring unit, configured to monitor in real time the loading time, the GPU video memory occupancy rate, and the actual fluctuation value of the network bandwidth of the currently loaded scenario resources;

[0052] A second generation unit, configured to generate a dynamic performance index according to the loading time, the video memory occupancy rate, and the actual fluctuation value;

[0053] A comparison unit, configured to compare the dynamic performance index with a predefined resource loading quality threshold:

[0054] An update unit, configured to, if the dynamic performance index is lower than the resource loading quality threshold, update the environment state vector based on the fragmentation degree of the current GPU video memory, the network bandwidth fluctuation spectrum, and the user behavior data, and re-enter the deep reinforcement learning model to generate a corrected resource loading policy;

[0055] An adjustment unit, configured to adjust the loading order of the scenario resources that have not been completely loaded based on the updated priority coefficient in the corrected resource loading policy.

[0056] Based on the resource loading method and system provided by the present application based on DeepSeek, in which, in response to a scene loading trigger instruction, each scene resource to be loaded is parsed by the DeepSeek large model to generate a weighted resource topology map; the fluctuation spectrum of the network bandwidth of the client, the fragmentation degree of the GPU video memory, and the user behavior data are obtained; an environmental state vector including spatio-temporal dynamic characteristics is constructed according to the obtained fluctuation spectrum of the network bandwidth, the fragmentation degree of the GPU video memory, and the user behavior data; the resource topology map is encoded into a topology feature vector through a graph neural network; the topology feature vector and the environmental state vector are input into a pre-trained deep reinforcement learning model to obtain a resource loading strategy, where the resource loading strategy includes a Markov decision sequence of the resource loading order, a priority coefficient of each of the scene resources, and a pre-cached resource subset based on the predicted perspective change; each of the scene resources is loaded according to the resource loading strategy. Applying the method provided by the embodiments of the present application, by parsing the scene resources through the DeepSeek large model to generate a weighted resource topology map, combining the network bandwidth fluctuation spectrum of the client, the fragmentation degree of the GPU video memory, and the user behavior data to construct an environmental state vector, and using the graph neural network and the deep reinforcement learning model to generate a resource loading strategy, it can dynamically adapt to complex and changeable network environments and heterogeneous hardware conditions, effectively solve the problems in the prior art such as the disconnection between resource allocation and the actual hardware state, the lack of a globally optimal loading decision, and the insufficient robustness and self-adaptability, significantly improve the efficiency and smoothness of resource loading, reduce the loading delay, improve the video memory utilization rate, reduce the waste of network bandwidth, and bring a better real-time experience to users. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0058] Figure 1 It is a flowchart of the method for the resource loading method based on DeepSeek provided by the present application;

[0059] Figure 2 It is a flowchart of a process for generating a weighted resource topology map provided by the present application;

[0060] Figure 3 It is a flowchart of a process for obtaining a resource loading strategy provided by the present application;

[0061] Figure 4Schematic diagram of the resource loading system based on DeepSeek provided by this application. Detailed implementation manners

[0062] Next, the technical solutions in the embodiments of this application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of this application.

[0063] In this application, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "including one..." does not exclude the existence of additional identical elements in the process, method, article or device including the said element.

[0064] The embodiment of the present invention provides a resource loading method based on DeepSeek, which is applied to an electronic device. The method flow chart of the method is as Figure 1 shown and specifically includes:

[0065] S101: In response to a scene loading trigger instruction, parse each scene resource to be loaded through the DeepSeek large model to generate a weighted resource topology map.

[0066] S102: Obtain the fluctuation spectrum of the network bandwidth of the client, the fragmentation degree of the GPU video memory, and user behavior data.

[0067] S103: Construct an environmental state vector including spatio-temporal dynamic features according to the obtained fluctuation spectrum of the network bandwidth, the fragmentation degree of the GPU video memory, and user behavior data.

[0068] S104: Encode the resource topology map into a topology feature vector through a graph neural network.

[0069] S105: Input the topology feature vector and the environmental state vector into a pre-trained deep reinforcement learning model to obtain a resource loading policy, where the resource loading policy includes a Markov decision sequence of the resource loading order, the priority coefficient of each scene resource, and a pre-cached resource subset based on the predicted perspective change.

[0070] S106: Load each scene resource according to the resource loading policy.

[0071] Applying the method provided by the embodiments of the present application, a weighted resource topology map is generated by parsing scene resources through the Deepseek large model. An environmental state vector is constructed by combining the client network bandwidth fluctuation spectrum, the degree of GPU video memory fragmentation, and user behavior data. A resource loading policy is generated by using a graph neural network and a deep reinforcement learning model, which can dynamically adapt to complex and changeable network environments and heterogeneous hardware conditions, effectively solve problems in the prior art such as the disconnection between resource allocation and the actual hardware state, the lack of a globally optimal loading decision, and insufficient robustness and adaptability, significantly improve the efficiency and smoothness of resource loading, reduce the loading delay, improve the video memory utilization rate, reduce network bandwidth waste, and bring a better real-time experience to users.

[0072] In an embodiment provided by the present application, based on the above solution, optionally, the process of parsing each scene resource to be loaded through the Deepseek large model to generate a weighted resource topology map is as Figure 2 shown and includes:

[0073] S201: Analyze the material sharing relationship, light dependence chain, physical collision body topology connection, and semantic association degree among each of the scene resources through the Deepseek large model.

[0074] S202: Construct a resource topology map according to the material sharing relationship, light dependence chain, physical collision body topology connection, and semantic association degree among each of the scene resources.

[0075] In an embodiment provided by the present application, based on the above solution, optionally, the deep reinforcement learning model includes an encoder, a first branch network, a second branch network, and a policy fusion module;

[0076] The process of inputting the topology feature vector and the environmental state vector into a pre-trained deep reinforcement learning model to obtain a resource loading policy is as Figure 3 shown and includes:

[0077] S301: Input the topology feature vector and the environmental state vector into the encoder in the pre-trained deep reinforcement learning model, so that the encoder performs graph Fourier transform on the topology feature vector to extract a frequency domain subspace vector; decompose the environmental state vector into a network bandwidth fluctuation component and a hardware fragmentation component, and fuse the frequency domain subspace vector, the network bandwidth fluctuation component, and the hardware fragmentation component through a multi-head attention mechanism to obtain a fusion vector.

[0078] S302: Input the fusion vectors into the first branch network and the second branch network respectively to obtain the Markov decision sequence, resource priority coefficient, and pre-caching identifier output by the first branch network, and the access probability entropy of each of the scenario resources output by the second branch network.

[0079] S303: Input the Markov decision sequence, the resource priority coefficient, the pre-caching identifier, the access probability entropy of each of the scenario resources, the fluctuation spectrum of the network bandwidth, and the fragmentation degree of the GPU video memory into the policy fusion module to obtain a resource loading policy.

[0080] In an embodiment provided by the present application, based on the above solution, optionally, the loading of each of the scenario resources according to the resource loading policy includes:

[0081] Generate a state transition probability matrix according to the Markov decision sequence;

[0082] Allocate each of the scenario resources to a corresponding target area according to the state transition probability matrix and the pre-cached resource subset based on the change in the prediction perspective, where the target area includes a pre-allocated area of the GPU video memory, a video memory swap area, or a disk cache area;

[0083] Allocate bandwidth to each of the scenario resources according to the priority coefficient of each scenario resource in each target area and the network bandwidth fluctuation spectrum.

[0084] For each of the scenario resources, load the scenario resource according to the bandwidth allocated to the scenario resource.

[0085] In an embodiment provided by the present application, based on the above solution, optionally, after loading the scenario resource according to the bandwidth allocated to the scenario resource, it further includes:

[0086] Monitor in real time the loading time, GPU video memory occupancy rate, and actual fluctuation value of the network bandwidth of the currently loaded scenario resources;

[0087] Generate a dynamic performance index according to the loading time, video memory occupancy rate, and actual fluctuation value;

[0088] Compare the dynamic performance index with a predefined resource loading quality threshold:

[0089] If the dynamic performance index is lower than the resource loading quality threshold, update the environmental state vector based on the fragmentation degree of the current GPU video memory, the network bandwidth fluctuation spectrum, and user behavior data, and re-input it into the deep reinforcement learning model to generate a corrected resource loading policy;

[0090] Adjust the loading order of the scene resources that have not been fully loaded based on the updated priority coefficient in the corrected resource loading policy.

[0091] See Figure 4 , which is a schematic structural diagram of a resource loading method system based on Deepseek provided by an embodiment of the present application. The system includes:

[0092] The first generation unit 401 is used to parse each scene resource to be loaded through the Deepseek large model in response to a scene loading trigger instruction, and generate a weighted resource topology map;

[0093] The acquisition unit 402 is used to acquire the fluctuation spectrum of the network bandwidth of the client, the fragmentation degree of the GPU video memory, and the user behavior data;

[0094] The construction unit 403 is used to construct an environmental state vector containing spatio-temporal dynamic features according to the acquired fluctuation spectrum of the network bandwidth, the fragmentation degree of the GPU video memory, and the user behavior data;

[0095] The encoding unit 404 is used to encode the resource topology map into a topology feature vector through a graph neural network;

[0096] The execution unit 405 is used to input the topology feature vector and the environmental state vector into a pre-trained deep reinforcement learning model to obtain a resource loading policy, where the resource loading policy includes a Markov decision sequence of the resource loading order, the priority coefficient of each scene resource, and a pre-cached resource subset based on the change of the prediction perspective;

[0097] The loading unit 406 is used to load each scene resource according to the resource loading policy.

[0098] In an embodiment provided by the present application, based on the above solution, optionally, the first generation unit includes:

[0099] The first execution subunit is used to analyze the material sharing relationship, light dependence chain, physical collision body topology connection, and semantic association degree between each scene resource through the Deepseek large model;

[0100] The construction subunit is used to construct a resource topology map according to the material sharing relationship, light dependence chain, physical collision body topology connection, and semantic association degree between each scene resource.

[0101] In an embodiment provided by the present application, based on the above solution, optionally, the deep reinforcement learning model includes an encoder, a first branch network, a second branch network, and a policy fusion module;

[0102] The execution unit includes:

[0103] An encoding subunit, configured to input the topological feature vector and the environmental state vector into an encoder in a pre-trained deep reinforcement learning model, so that the encoder performs graph Fourier transform on the topological feature vector to extract a frequency domain subspace vector; decompose the environmental state vector into a network bandwidth fluctuation component and a hardware fragmentation component, and fuse the frequency domain subspace vector, the network bandwidth fluctuation component, and the hardware fragmentation component through a multi-head attention mechanism to obtain a fusion vector;

[0104] A second execution subunit, configured to input the fusion vector into the first branch network and the second branch network respectively, to obtain a Markov decision sequence, a resource priority coefficient, and a pre-cache identifier output by the first branch network, and an access probability entropy of each of the scenario resources output by the second branch network;

[0105] A third execution subunit, configured to input the Markov decision sequence, the resource priority coefficient, the pre-cache identifier, the access probability entropy of each of the scenario resources, the fluctuation spectrum of the network bandwidth, and the fragmentation degree of the GPU video memory into the policy fusion module to obtain a resource loading policy.

[0106] In an embodiment provided by the present application, based on the above solution, optionally, the loading unit includes:

[0107] A generation subunit, configured to generate a state transition probability matrix according to the Markov decision sequence;

[0108] A first allocation subunit, configured to allocate each of the scenario resources to a corresponding target area according to the state transition probability matrix and the pre-cache resource subset based on the predicted perspective change, where the target area includes a GPU video memory pre-allocation area, a video memory swap area, or a disk cache area;

[0109] A first allocation subunit, configured to allocate bandwidth to each of the scenario resources according to the priority coefficient of each scenario resource in each target area and the network bandwidth fluctuation spectrum;

[0110] A loading subunit, configured to load scenario resources for each of the scenario resources according to the bandwidth allocated to the scenario resources.

[0111] In an embodiment provided by the present application, based on the above solution, optionally, it further includes:

[0112] A monitoring unit, configured to monitor in real time the loading time consumption, the GPU video memory occupancy rate, and the actual fluctuation value of the network bandwidth of the currently loaded scenario resources;

[0113] A second generation unit, configured to generate a dynamic performance metric according to the loading time, the video memory occupancy rate, and the actual fluctuation value;

[0114] A comparison unit, configured to compare the dynamic performance metric with a predefined resource loading quality threshold:

[0115] An update unit, configured to, if the dynamic performance metric is lower than the resource loading quality threshold, update the environmental state vector based on the fragmentation degree of the current GPU video memory, the network bandwidth fluctuation spectrum, and the user behavior data, and re-enter the deep reinforcement learning model to generate a corrected resource loading policy;

[0116] An adjustment unit, configured to adjust the loading order of the resources in the uncompleted loading scenario based on the updated priority coefficient in the corrected resource loading policy.

[0117] It should be noted that the embodiments in this specification are all described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other.

[0118] Finally, it should also be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

[0119] For the convenience of description, when describing the above system, it is divided into various units according to functions for description. Of course, when implementing the present application, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0120] From the description of the above embodiments, those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments of the present application.

[0121] The above has introduced in detail the resource loading method based on Deepseek provided by this application. Specific examples are used in this article to elaborate on the principle and implementation manner of this application. The description of the above embodiments is only used to help understand the method and its core idea of this application; at the same time, for those of ordinary skill in the art, according to the idea of this application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to this application.

Claims

1. A resource loading method based on Deepseek, characterized in that Including: In response to the scene loading trigger instruction, parse each scene resource to be loaded through the DeepSeek large model, and generate a weighted resource topology graph; Obtain the fluctuation spectrum of the client's network bandwidth, the fragmentation degree of the GPU video memory, and the user behavior data; Construct an environmental state vector containing spatio-temporal dynamic features according to the obtained fluctuation spectrum of the network bandwidth, the fragmentation degree of the GPU video memory, and the user behavior data; Encode the resource topology graph into a topology feature vector through a graph neural network; Input the topology feature vector and the environmental state vector into a pre-trained deep reinforcement learning model to obtain a resource loading strategy, where the resource loading strategy includes a Markov decision sequence of the resource loading order, a priority coefficient of each scene resource, and a pre-cache resource subset based on the predicted perspective change; Load each scene resource according to the resource loading strategy.

2. The method according to claim 1, wherein The process of parsing each scene resource to be loaded through the DeepSeek large model and generating a weighted resource topology graph includes: Analyze the material sharing relationship, light dependence chain, physical collision body topology connection, and semantic association degree among each scene resource through the DeepSeek large model; Construct a resource topology graph according to the material sharing relationship, light dependence chain, physical collision body topology connection, and semantic association degree among each scene resource.

3. The method according to claim 1, characterized in that The deep reinforcement learning model includes an encoder, a first branch network, a second branch network, and a policy fusion module; The process of inputting the topology feature vector and the environmental state vector into the pre-trained deep reinforcement learning model to obtain a resource loading strategy includes: Input the topology feature vector and the environmental state vector into the encoder in the pre-trained deep reinforcement learning model, so that the encoder performs graph Fourier transform on the topology feature vector to extract a frequency domain subspace vector; decompose the environmental state vector into a network bandwidth fluctuation component and a hardware fragmentation component, and fuse the frequency domain subspace vector, the network bandwidth fluctuation component, and the hardware fragmentation component through a multi-head attention mechanism to obtain a fusion vector; Input the fusion vector into the first branch network and the second branch network respectively to obtain the Markov decision sequence, resource priority coefficient, and pre-cache identifier output by the first branch network, and the access probability entropy of each scene resource output by the second branch network; Input the Markov decision sequence, the resource priority coefficient, the pre-cache identifier, the access probability entropy of each scene resource, the fluctuation spectrum of the network bandwidth, and the fragmentation degree of the GPU video memory into the policy fusion module to obtain a resource loading strategy.

4. The method according to claim 1, characterized in that The process of loading each scene resource according to the resource loading strategy includes: Generate a state transition probability matrix according to the Markov decision sequence; Allocate each scene resource to a corresponding target area according to the state transition probability matrix and the pre-cache resource subset based on the predicted perspective change, where the target area includes a GPU video memory pre-allocation area, a video memory swap area, or a disk cache area; Allocate bandwidth to each of the scene resources according to the priority coefficient of each scene resource in each of the target regions and the network bandwidth fluctuation spectrum; For each of the scene resources, load the scene resources according to the bandwidth allocated to the scene resources.

5. The method according to claim 4, characterized in that After loading the scene resources according to the bandwidth allocated to the scene resources, it further includes: Monitor in real time the loading time, GPU video memory occupancy rate, and actual fluctuation value of the network bandwidth of the currently loaded scene resources; Generate a dynamic performance metric based on the loading time, video memory occupancy rate, and actual fluctuation value; Compare the dynamic performance metric with a predefined resource loading quality threshold: If the dynamic performance metric is lower than the resource loading quality threshold, update the environmental state vector based on the fragmentation degree of the current GPU video memory, the network bandwidth fluctuation spectrum, and user behavior data, and re-enter the deep reinforcement learning model to generate a corrected resource loading strategy; Adjust the loading order of the uncompleted loaded scene resources based on the updated priority coefficients in the corrected resource loading strategy.

6. A resource loading system based on Deepseek, characterized in that It includes: A first generation unit for, in response to a scene loading trigger instruction, parsing each of the scene resources to be loaded through a Deepseek large model to generate a weighted resource topology map; An acquisition unit for acquiring the network bandwidth fluctuation spectrum of the client, the fragmentation degree of the GPU video memory, and user behavior data; A construction unit for constructing an environmental state vector containing spatio-temporal dynamic features according to the acquired network bandwidth fluctuation spectrum, fragmentation degree of the GPU video memory, and user behavior data; An encoding unit for encoding the resource topology map into a topology feature vector through a graph neural network; An execution unit for inputting the topology feature vector and the environmental state vector into a pre-trained deep reinforcement learning model to obtain a resource loading strategy, where the resource loading strategy includes a Markov decision sequence of the resource loading order, the priority coefficients of each of the scene resources, and a pre-cached resource subset based on the predicted perspective change; A loading unit for loading each of the scene resources according to the resource loading strategy.

7. The system according to claim 6, wherein The first generation unit includes: A first execution subunit for analyzing the material sharing relationship, light dependence chain, physical collision body topology connection, and semantic association degree among each of the scene resources through a Deepseek large model; A construction subunit for constructing a resource topology map according to the material sharing relationship, light dependence chain, physical collision body topology connection, and semantic association degree among each of the scene resources.

8. The system according to claim 6, wherein The deep reinforcement learning model includes an encoder, a first branch network, a second branch network, and a policy fusion module; The execution unit includes: An encoding subunit, configured to input the topological feature vector and the environmental state vector into an encoder in a pre-trained deep reinforcement learning model, so that the encoder performs graph Fourier transform on the topological feature vector to extract a frequency-domain subspace vector; decompose the environmental state vector into a network bandwidth fluctuation component and a hardware fragmentation component, and fuse the frequency-domain subspace vector, the network bandwidth fluctuation component, and the hardware fragmentation component through a multi-head attention mechanism to obtain a fused vector; A second execution subunit, configured to input the fused vector into the first branch network and the second branch network respectively, to obtain a Markov decision sequence, a resource priority coefficient, and a pre-caching identifier output by the first branch network, and an access probability entropy of each of the scenario resources output by the second branch network; A third execution subunit, configured to input the Markov decision sequence, the resource priority coefficient, the pre-caching identifier, the access probability entropy of each of the scenario resources, the fluctuation spectrum of the network bandwidth, and the fragmentation degree of the GPU video memory into the policy fusion module to obtain a resource loading policy.

9. The system according to claim 6, wherein The loading unit includes: A generation subunit, configured to generate a state transition probability matrix according to the Markov decision sequence; A first allocation subunit, configured to allocate each of the scenario resources to a corresponding target area according to the state transition probability matrix and the pre-cached resource subset based on the predicted perspective change, where the target area includes a pre-allocated area of the GPU video memory, a video memory swap area, or a disk cache area; A first allocation subunit, configured to allocate bandwidth to each of the scenario resources according to the priority coefficient of each scenario resource in each target area and the network bandwidth fluctuation spectrum; A loading subunit, configured to load a scenario resource for each of the scenario resources according to the bandwidth allocated to the scenario resource.

10. The system according to claim 9, wherein It further includes: A monitoring unit, configured to monitor in real time the loading time, the GPU video memory occupancy rate, and the actual fluctuation value of the network bandwidth of the currently loaded scenario resources; A second generation unit, configured to generate a dynamic performance index according to the loading time, the video memory occupancy rate, and the actual fluctuation value; A comparison unit, configured to compare the dynamic performance index with a predefined resource loading quality threshold; An update unit, configured to, if the dynamic performance index is lower than the resource loading quality threshold, update the environmental state vector based on the fragmentation degree of the current GPU video memory, the network bandwidth fluctuation spectrum, and user behavior data, and re-input the updated vector into the deep reinforcement learning model to generate a corrected resource loading policy; An adjustment unit, configured to adjust the loading order of the scenario resources that have not been fully loaded based on the updated priority coefficient in the corrected resource loading policy.

Images