Processor-Supported Model Building Without Retraining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models face limitations in design flexibility and usage restrictions due to the need for retraining when converting between different processor libraries, particularly for tree-based models like Scikit-learn decision trees.
Innovation Solution
A method for building a model using model information in a specific format, such as CSV, that is supported by the processor's library, allowing direct execution without retraining, and includes a processing module to read and verify the model using testing data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the SK-DT model is converted to run on a processor supporting Tensorflow, then the model can execute on the processor, but the model requires retraining which reduces design flexibility and increases time consumption
Solution Approach 1:
The patent extracts the decision tree model structure and parameters from the original SK-DT model and creates a copy in a standardized format that can be directly loaded by the processor's supported library, eliminating the need for retraining while preserving the model's predictive capabilities
Solution Approach 2:
The patent performs preliminary extraction and formatting of the model information before deployment, converting the model to a standardized representation that can be directly used by different processors without requiring retraining, thus saving time in the deployment phase
2Ease of operation
If the model is converted to match the processor's supported library, then the model can run directly on the processor, but the conversion process increases device complexity
Solution Approach 1:
The patent changes the representation parameters of the model from library-specific formats to a standardized format that captures the essential decision tree structure and parameters, enabling direct loading by the processor without complex conversion processes
Solution Approach 2:
The patent extracts only the essential model information (tree structure, split conditions, leaf values) from the original model, separating the core predictive logic from library-specific implementations, thereby simplifying the deployment process while maintaining model functionality
Data Source
AI summary
A method for building a model includes: reading model information from a storage device, wherein the model information describes a user model; building the model according to the model information and a library supported by a processor; reading testing data from the storage device; and performing verification upon the model according to the testing data.