Neural Network Executable Image Generation Automation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating neural network executable images are complex and require manual implementation and use of separate training and inference engines, data format conversion, and memory checking, which complicates the process of creating a lightweight neural network executable image.
Innovation Solution
An apparatus and method that automatically generate a neural network executable image by receiving user requirements, checking if the default neural network model is compatible with the target system, converting the model and training data formats, and generating a lightweight image, while also supporting data quantization and memory optimization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If separate training engine and inference engine are supported, then compatibility with different neural network execution environments is improved, but system complexity and number of conversion modules increase
Solution Approach 1:
The patent implements a universal conversion module that can handle multiple format conversions between different neural network engines (TensorFlow, PyTorch, Caffe, etc.) and various data formats (FP32, FP16, INT8, etc.). This single multi-functional module replaces the need for separate N-to-N conversion modules for each engine pair, thereby maintaining compatibility with different execution environments while reducing system complexity.
2Quantity of substance
If quantization-based data format conversion is performed, then neural network model size is reduced, but number of conversion modules increases proportionally with number of format types
Solution Approach 1:
The patent implements a unified conversion module that dynamically adjusts conversion parameters based on the target data format requirements. The module can perform quantization to various precision levels (FP32, FP16, INT8, INT4) and apply different quantization strategies (uniform, non-uniform, asymmetric, symmetric) by changing operational parameters rather than requiring separate modules for each format type. This reduces the number of conversion modules while maintaining the ability to produce compact quantized models.
3Adaptability or versatility
If manual implementation of engine support, data conversion, and memory checking is required, then flexibility and customization are improved, but ease of operation deteriorates
Solution Approach 1:
The patent implements an automated generation system that performs engine compatibility checking, format conversion, quantization, and memory optimization automatically based on user-specified requirements. The system self-configures the conversion process by detecting the target engine type and automatically selecting appropriate conversion parameters and quantization strategies, eliminating the need for manual implementation while preserving flexibility through configurable options.
4Adaptability or versatility
If comprehensive format conversion support is provided, then adaptability to different data formats is improved, but device complexity increases
Solution Approach 1:
The patent introduces a standardized intermediate representation format that serves as a common denominator between different neural network engines and data formats. The conversion system first transforms various input formats (TF, PT, CA models and FP32/FP16/INT8/INT4 data) into this intermediate format, then converts from the intermediate format to the target format. This intermediary approach enables comprehensive format support while simplifying the conversion system architecture compared to implementing direct conversion between all possible format pairs.
Data Source
AI summary
An apparatus and method for generating a neural network executable image are disclosed. The apparatus receives user requirements including a default neural network model and training result data for generating a neural network executable image required by a user, checks whether the default neural network model included in the user requirements is capable of being supported in a target system in which the neural network executable image is to be installed, converts the default neural network model into a neural network model executable in the target system, converts the training result data by reconfiguring the data format set of the training result data, and generates a neural network executable image by combining the converted neural network model and the converted training result data.