Visual task generation method adopting autoregression model

Abstract

The invention discloses a visual task generation method adopting an autoregression model, and belongs to the technical field of computer vision, and the method comprises the following steps: S1, historical task modeling: arranging historical visual tasks according to a time or logic sequence, and constructing a task sequence; s2, feature extraction; s3, training an autoregression model; s4, task generation; s5, parameter optimization: optimizing the task parameters to adapt to the performance requirements of the target equipment; s6, task feasibility verification; s7, performing dynamic feedback iteration; and S8, visualization and interpretation enhancement are carried out. According to the method, the time sequence dependency relationship of the historical task sequence is modeled through the autoregression model, and a logic verification mechanism of the dependency relationship graph is combined, so that the system can generate a task chain with compatible input and output formats and a reasonable execution sequence, the model automatically avoids task sequence errors, the logic correctness of a complex visual process is ensured, the manual intervention requirement is reduced, and the efficiency is improved. And the dependency conflict rate of the generated task chain is obviously reduced.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a method for generating visual tasks using an autoregressive model. Background Art

[0002] With the rapid development of computer vision technology, the complexity of visual tasks and the diversity of scenarios continue to increase. Traditional task generation methods mostly rely on manually preset rules or single model iterations, which have problems such as task logic discontinuities, cross-task dependency conflicts, and rigid resource allocation. In existing technologies, task orchestration based on rule engines lacks the ability to model the relevance and temporal dynamics of historical tasks, making it difficult to adapt to complex multi-stage nested processes. Although end-to-end generation models can automatically output task chains, they often fail to deploy due to ignoring hardware adaptability, and the generation process lacks explainability, hindering user trust and collaborative optimization. In addition, the lack of multimodal conditional control makes it difficult for task generation to accurately match user intent, especially in cross-device and cross-scenario applications, which are prone to semantic deviations and resource overruns. Therefore, we propose a visual task generation method using an autoregressive model to solve this problem. Summary of the Invention

[0003] The purpose of the present invention is to provide a visual task generation method using an autoregressive model to solve the problems raised in the above background technology.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] A method for generating visual tasks using an autoregressive model includes the following steps:

[0006] S1. Historical task modeling: Arrange historical visual tasks in chronological or logical order, construct task sequences, and generate analyzable dependency graphs;

[0007] S2. Feature extraction: Extract the semantic features, temporal features, and metadata features of the task and construct a unified coding representation;

[0008] S3. Autoregressive model training: Train the autoregressive model to predict the features and parameters of the next task;

[0009] S4, Task Generation: Based on conditional input, generate candidate task chains and instantiate them into executable solutions;

[0010] S5. Parameter optimization: Optimize task parameters to adapt to the performance requirements of the target device;

[0011] S6, Task feasibility verification: Verify the dependency compatibility and resource feasibility of the generated task chain;

[0012] S7. Dynamic feedback iteration: Continuously optimize model performance based on user feedback and execution results;

[0013] S8. Visualization and explanation enhancement: Provide an explanatory visualization interface to support user interactive debugging.

[0014] Preferably, in said S1, the specific steps are:

[0015] S101. Task sequence construction: Sort historical tasks according to timestamps or logical dependencies to form a structured sequence dataset;

[0016] S102, Task metadata annotation: Detailed recording of task parameters, execution results, and contextual information;

[0017] S103. Dependency graph construction: A directed graph is used to represent the input-output dependency relationship between tasks, where nodes represent tasks and edges represent data flow.

[0018] Preferably, in said S2, the specific steps are:

[0019] S201, Task Semantic Encoding: Use the BERT model to extract the semantic vector of the task description text, and use the ResNet-50 model to extract the global features of the task input / output images;

[0020] S202, Temporal Feature Modeling: Use LSTM or Transformer encoders to capture temporal dependencies in task sequences;

[0021] S203, meta-feature fusion: concatenate numerical features such as task parameters and hardware configuration with semantic vectors to form a unified feature vector.

[0022] Preferably, in said S3, the specific steps are:

[0023] S301. Model Architecture Design: Using the Transformer-Decoder structure, we input the historical task feature sequence and generate the next task feature through autoregression. We also introduce causal masks to ensure that we only rely on the previous task.

[0024] S302. Training objective definition: Minimize the mean square error of task feature prediction and the cross entropy loss of task parameter classification;

[0025] S303. Course learning strategy: gradually transition from simple task chains to complex task chains to improve model stability.

[0026] Preferably, in said S4, the specific steps are:

[0027] S401, Conditional Generation Control: The user inputs a target description, which is encoded into a conditional vector through the CLIP model and injected into the autoregressive model;

[0028] S402, Sequence Decoding Strategy: Use Beam Search to generate Top-K candidate task chains, balancing diversity and rationality;

[0029] S403, task instantiation: decoding the generated feature vector into specific task parameters.

[0030] Preferably, in said S5, the specific steps are:

[0031] S501, automatic hyperparameter search: Use Bayesian optimization to search for the optimal combination in the preset space;

[0032] S502, Hardware-aware optimization: Dynamically adjust the model scale based on the device computing power to ensure real-time performance;

[0033] S503. Multi-objective trade-offs: Designing Pareto optimization objectives.

[0034] Preferably, in said S6, the specific steps are:

[0035] S601, dependency conflict detection: check the compatibility of task chain input and output formats;

[0036] S602, resource consumption estimation: predict the peak memory usage and computing time of the task chain, and intercept solutions that exceed the device capabilities;

[0037] S603, logical rationality check: the rule engine verifies the rationality of the task sequence.

[0038] Preferably, in said S7, the specific steps are:

[0039] S701, User Correction Feedback: Record the user's manual adjustments to the task chain and build an incremental training dataset;

[0040] S702, Online Reinforcement Learning: Using user satisfaction as a reward signal, fine-tune the generation strategy through the PPO algorithm;

[0041] S703, abnormal task filtering: count the task chains with low execution success rate and add them to the generated blacklist.

[0042] Preferably, in said S8, the specific steps are:

[0043] S801, Task Chain Visualization: Generate a directed graph to show task dependencies, highlighting critical paths and bottleneck nodes;

[0044] S802, explanation of the basis for the generation: Visualize the attention weights to show the basis for the model to select a task;

[0045] S803, interactive debugging: allows users to drag and drop to adjust the task order, and preview resource consumption and accuracy changes in real time.

[0046] The beneficial effects of the present invention are:

[0047] In the present invention, a visual task generation method using an autoregressive model is described. By using the autoregressive model to model the temporal dependency of historical task sequences and combining it with a logic verification mechanism of the dependency graph, the system can generate a task chain with compatible input and output formats and a reasonable execution order. The model automatically avoids task sequence errors, ensures the logical correctness of complex visual processes, reduces the need for manual intervention, and significantly reduces the dependency conflict rate of the generated task chain. It is suitable for multi-stage nested visual task scenarios and improves the execution success rate and stability of the end-to-end process.

[0048] In the present invention, a visual task generation method using an autoregressive model is described. By integrating multimodal user input conditions such as text and images, and through semantic encoding and autoregressive model joint reasoning, the generated task chain accurately matches the user's intent. The model dynamically selects adaptation algorithms and parameters based on the context. For example, for low-light scenes, a highly robust detection model is prioritized, thereby improving the semantic matching degree and user satisfaction of the task chain and meeting the customization needs of diverse scenarios.

[0049] The present invention describes a visual task generation method using an autoregressive model. This method dynamically adjusts the model size and computational strategy through a hardware perception optimization module, combined with automated hyperparameter search technology. This significantly reduces the memory usage and computational time of the task chain. The resulting lightweight task chain adapts to edge device resource constraints, supports real-time processing of high-resolution data streams, and reduces hardware costs and energy consumption while ensuring accuracy. This method is suitable for low-power deployment scenarios on mobile and embedded terminals.

[0050] In the present invention, a visual task generation method using an autoregressive model is described. Through attention weight visualization and an interactive debugging interface, users can intuitively understand the decision-making basis of the model-generated task chain and manually adjust the task sequence or parameter configuration. The system provides real-time feedback on the adjusted resource consumption and performance changes. Combined with an online learning mechanism to continuously absorb user experience, it achieves dynamic alignment between the model and human expertise, and improves the quality of task generation in complex scenarios.

[0051] In the present invention, the described method for generating visual tasks using an autoregressive model enables the system to adapt to different hardware architectures and industry scenario requirements through a dynamic feedback iteration mechanism and a multi-objective optimization strategy, supports seamless expansion from the cloud to the terminal, significantly shortens the task design cycle, reduces manpower investment and operation and maintenance costs, and promotes the large-scale implementation of visual task automation technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flowchart of a visual task generation method using an autoregressive model proposed in the present invention. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0054] Reference Figure 1 , a visual task generation method using an autoregressive model, comprising the following steps:

[0055] S1. Historical task modeling: Arrange historical visual tasks in chronological or logical order, construct task sequences, and generate analyzable dependency graphs;

[0056] S2, Feature Extraction: Extract the semantic features, temporal features, and metadata features of the task and construct a unified coding representation;

[0057] S3. Autoregressive model training: Train the autoregressive model to predict the features and parameters of the next task;

[0058] S4, Task Generation: Based on conditional input, generate candidate task chains and instantiate them into executable solutions;

[0059] S5. Parameter optimization: Optimize task parameters to adapt to the performance requirements of the target device;

[0060] S6, Task feasibility verification: Verify the dependency compatibility and resource feasibility of the generated task chain;

[0061] S7. Dynamic feedback iteration: Continuously optimize model performance based on user feedback and execution results;

[0062] S8. Visualization and explanation enhancement: Provide an explanatory visualization interface to support user interactive debugging.

[0063] In this embodiment, in S1, the specific steps are:

[0064] S101. Task sequence construction: Sort historical tasks according to timestamps or logical dependencies to form a structured sequence dataset;

[0065] S102, Task metadata annotation: Detailed recording of task parameters, execution results, and contextual information;

[0066] S103. Dependency graph construction: A directed graph is used to represent the input-output dependency relationship between tasks, where nodes represent tasks and edges represent data flow.

[0067] In this embodiment, in S2, the specific steps are:

[0068] S201, Task Semantic Encoding: Use the BERT model to extract the semantic vector of the task description text, and use the ResNet-50 model to extract the global features of the task input / output images;

[0069] S202, Temporal Feature Modeling: Use LSTM or Transformer encoders to capture temporal dependencies in task sequences;

[0070] S203, meta-feature fusion: concatenate numerical features such as task parameters and hardware configuration with semantic vectors to form a unified feature vector.

[0071] In this embodiment, in S3, the specific steps are:

[0072] S301. Model Architecture Design: Using the Transformer-Decoder structure, we input the historical task feature sequence and generate the next task feature through autoregression. We also introduce causal masks to ensure that we only rely on the previous task.

[0073] S302. Training objective definition: Minimize the mean square error of task feature prediction and the cross entropy loss of task parameter classification;

[0074] S303. Course learning strategy: gradually transition from simple task chains to complex task chains to improve model stability.

[0075] In this embodiment, in S4, the specific steps are:

[0076] S401, Conditional Generation Control: The user inputs a target description, which is encoded into a conditional vector through the CLIP model and injected into the autoregressive model;

[0077] S402, Sequence Decoding Strategy: Use Beam Search to generate Top-K candidate task chains, balancing diversity and rationality;

[0078] S403, task instantiation: decoding the generated feature vector into specific task parameters.

[0079] In this embodiment, in S5, the specific steps are:

[0080] S501, automatic hyperparameter search: Use Bayesian optimization to search for the optimal combination in the preset space;

[0081] S502, Hardware-aware optimization: Dynamically adjust the model scale based on the device computing power to ensure real-time performance;

[0082] S503. Multi-objective trade-offs: Designing Pareto optimization objectives.

[0083] In this embodiment, in S6, the specific steps are:

[0084] S601, dependency conflict detection: check the compatibility of task chain input and output formats;

[0085] S602, resource consumption estimation: predict the peak memory usage and computing time of the task chain, and intercept solutions that exceed the device capabilities;

[0086] S603, logical rationality check: the rule engine verifies the rationality of the task sequence.

[0087] In this embodiment, in S7, the specific steps are:

[0088] S701, User Correction Feedback: Record the user's manual adjustments to the task chain and build an incremental training dataset;

[0089] S702, Online Reinforcement Learning: Using user satisfaction as a reward signal, fine-tune the generation strategy through the PPO algorithm;

[0090] S703, abnormal task filtering: count the task chains with low execution success rate and add them to the generated blacklist.

[0091] In this embodiment, in S8, the specific steps are:

[0092] S801, Task Chain Visualization: Generate a directed graph to show task dependencies, highlighting critical paths and bottleneck nodes;

[0093] S802, explanation of the basis for the generation: Visualize the attention weights to show the basis for the model to select a task;

[0094] S803, interactive debugging: allows users to drag and drop to adjust the task order, and preview resource consumption and accuracy changes in real time.

[0095] In this embodiment, starting with data collection and structured processing of historical visual tasks, the task sequence is sorted by time or logical dependency and a dependency graph is constructed. The semantic features, temporal features, and metadata such as hardware configuration of the task are extracted for multimodal fusion encoding. Subsequently, an autoregressive model architecture is used for training to capture the dynamic associations between tasks and predict the parameters of subsequent tasks. In the generation phase, a candidate task chain is generated through a decoding strategy based on the text or image conditions input by the user and instantiated into a specific executable solution. The task dependency compatibility, resource consumption, and logical rationality are verified in multiple dimensions through an automated verification module. At the same time, a hardware-aware optimization strategy is introduced to dynamically adjust the model scale and computing parameters to adapt to the computing power limitations of different terminal devices. Finally, the generated task chain is displayed through a visual interface and users are allowed to interactively adjust it. The model generation strategy is continuously optimized based on real-time feedback data, forming a closed-loop system from task modeling to generation execution to feedback iteration, thereby achieving efficient, controllable, and explainable automatic generation and deployment of visual tasks.

[0096] The above is a detailed introduction to the visual task generation method using an autoregressive model provided by the present invention. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above examples is only used to help understand the method of the present invention and its core ideas. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A visual task generation method using an autoregressive model, characterized in that: The following steps are involved: S1. Historical task modeling: Arrange historical visual tasks in chronological or logical order, construct task sequences, and generate analyzable dependency graphs; S2. Feature extraction: Extract the semantic features, temporal features, and metadata features of the task and construct a unified coding representation; S3. Autoregressive model training: Train the autoregressive model to predict the features and parameters of the next task; S4, Task Generation: Based on conditional input, generate candidate task chains and instantiate them into executable solutions; S5. Parameter optimization: Optimize task parameters to adapt to the performance requirements of the target device; S6, Task feasibility verification: Verify the dependency compatibility and resource feasibility of the generated task chain; S7. Dynamic feedback iteration: Continuously optimize model performance based on user feedback and execution results; S8. Visualization and explanation enhancement: Provide an explanatory visualization interface to support user interactive debugging.

2. The method for generating visual tasks using an autoregressive model according to claim 1, wherein: In said S1, the specific steps are: S101. Task sequence construction: Sort historical tasks according to timestamps or logical dependencies to form a structured sequence dataset; S102, Task metadata annotation: Detailed recording of task parameters, execution results, and contextual information; S103. Dependency graph construction: A directed graph is used to represent the input-output dependency relationship between tasks, where nodes represent tasks and edges represent data flow.

3. The method for generating visual tasks using an autoregressive model according to claim 1, wherein: In said S2, the specific steps are: S201, Task Semantic Encoding: Use the BERT model to extract the semantic vector of the task description text, and use the ResNet-50 model to extract the global features of the task input / output images; S202, Temporal Feature Modeling: Use LSTM or Transformer encoders to capture temporal dependencies in task sequences; S203, meta-feature fusion: concatenate numerical features such as task parameters and hardware configuration with semantic vectors to form a unified feature vector.

4. The method for generating visual tasks using an autoregressive model according to claim 1, wherein: In S3, the specific steps are: S301. Model Architecture Design: Using the Transformer-Decoder structure, we input the historical task feature sequence and generate the next task feature through autoregression. We also introduce causal masks to ensure that we only rely on the previous task. S302. Training objective definition: Minimize the mean square error of task feature prediction and the cross entropy loss of task parameter classification; S303. Course learning strategy: gradually transition from simple task chains to complex task chains to improve model stability.

5. The method for generating visual tasks using an autoregressive model according to claim 1, wherein: In said S4, the specific steps are: S401, Conditional Generation Control: The user inputs a target description, which is encoded into a conditional vector through the CLIP model and injected into the autoregressive model; S402, Sequence Decoding Strategy: Use Beam Search to generate Top-K candidate task chains, balancing diversity and rationality; S403, task instantiation: decoding the generated feature vector into specific task parameters.

6. The method for generating visual tasks using an autoregressive model according to claim 1, wherein: In said S5, the specific steps are: S501, automatic hyperparameter search: Use Bayesian optimization to search for the optimal combination in the preset space; S502, hardware-aware optimization: dynamically adjust the model size based on the device computing power to ensure real-time performance; S503. Multi-objective trade-offs: Designing Pareto optimization objectives.

7. The method for generating visual tasks using an autoregressive model according to claim 1, wherein: In said S6, the specific steps are: S601, dependency conflict detection: check the compatibility of task chain input and output formats; S602, resource consumption estimation: predict the peak memory usage and computing time of the task chain, and intercept solutions that exceed the device capabilities; S603, logical rationality check: the rule engine verifies the rationality of the task sequence.

8. The method for generating visual tasks using an autoregressive model according to claim 1, wherein: In said S7, the specific steps are: S701, User Correction Feedback: Record the user's manual adjustments to the task chain and build an incremental training dataset; S702, Online Reinforcement Learning: Using user satisfaction as a reward signal, fine-tune the generation strategy through the PPO algorithm; S703, abnormal task filtering: count the task chains with low execution success rate and add them to the generated blacklist.

9. The method for generating visual tasks using an autoregressive model according to claim 1, wherein: In said S8, the specific steps are: S801, Task Chain Visualization: Generate a directed graph to show task dependencies, highlighting critical paths and bottleneck nodes; S802, explanation of the basis for the generation: Visualize the attention weights to show the basis for the model to select a task; S803, interactive debugging: allows users to drag and drop to adjust the task order, and preview resource consumption and accuracy changes in real time.

Images