Deep Learning Model Conversion Paths for Hardware Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models are often limited to specific AI inference accelerator chips due to differences in hardware and frameworks, making it difficult to adapt them across various hardware terminals.
Innovation Solution
A method and apparatus that utilize a conversion path table to convert an original deep learning model into an intermediate model, and then to a target model suitable for the specific hardware, by querying model and hardware information to determine a target conversion path, allowing for multiple intermediate models in the conversion process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If deep learning models are deployed on AI inference accelerator chips, then inference performance is improved, but the model is limited to specific hardware frameworks and cannot run on other hardware terminals
Solution Approach 1:
The patent introduces an intermediate model as a mediator between the original deep learning model and the target hardware. The conversion system transforms the original model into an intermediate model that serves as a universal interface, which can then be adapted to different hardware platforms. This intermediary layer enables the original model to run on various hardware terminals without requiring hardware-specific modifications.
Solution Approach 2:
The patent creates a universal conversion system that can handle multiple hardware frameworks and platforms through a single intermediate model representation. The system provides multi-functionality by supporting conversion to different hardware targets (GPUs, FPGAs, CPUs) from a single original model, eliminating the need for separate conversion paths for each hardware platform.
2Adaptability or versatility
If deep learning models are converted to different hardware frameworks, then hardware adaptability is improved, but the conversion process becomes complex and time-consuming
Solution Approach 1:
The patent segments the complex conversion process into distinct stages: converting the original model to an intermediate model, and then converting the intermediate model to the target hardware framework. This segmentation divides the complex task into manageable steps, where each stage has specific conversion rules and can be optimized independently, reducing the overall complexity of the conversion process.
Solution Approach 2:
The patent performs preliminary conversion to an intermediate model format that standardizes the model representation before final conversion to target hardware. This preliminary action prepares the model in advance with hardware-agnostic operations, so that subsequent conversions to different hardware frameworks become simpler and faster, as the heavy lifting is already done in the intermediate representation.
3Adaptability or versatility
If multiple conversion paths are supported for different hardware frameworks, then hardware versatility is improved, but the conversion time and processing overhead increase
Solution Approach 1:
The patent performs the time-consuming conversion to the intermediate model format once as a preliminary action, regardless of the target hardware framework. By standardizing the model representation in advance, subsequent conversions to different hardware frameworks become much faster, as they only need to transform from the intermediate format rather than from the original model format each time.
Solution Approach 2:
The patent creates a universal intermediate model representation that serves as a common foundation for multiple hardware frameworks. This universal format allows the system to support multiple frameworks (TensorFlow, PyTorch, ONNX, etc.) through a single intermediate representation, enabling rapid conversion to any target framework without requiring separate conversion paths from the original model for each framework.
Data Source
AI summary
The present disclosure discloses a method and an apparatus for adapting a deep learning model, an electronic device and a medium, which relates to technology fields of artificial intelligence, deep learning, and cloud computing. The specific implementation plan is: obtaining model information of an original deep learning model and hardware information of a target hardware to be adapted; querying a conversion path table according to the model information and the hardware information to obtain a matched target conversion path; and converting, according to the target conversion path, the original deep learning model to an intermediate deep learning model in the conversion path, and converting the intermediate deep learning model to the target deep learning model. Therefore, the deep learning model conversion is performed based on the model conversion path determined by the model information of the original deep learning model and the hardware information of the target hardware, which realizes converting any type of original deep learning model into the target deep learning model adapted to any target hardware, and solves the problem that the deep learning model is difficult to be applied to different hardware terminals.