Network Structure Optimization for MLU Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing detection network SSD experiences high resource overhead and reduced detection rate due to redundant data transformation operation layers, such as Permute, Reshape, and Flatten, which are not always necessary in MLU chips.
Innovation Solution
A network structure processing method that optimizes the network structure by obtaining an optimization instruction and applying it to delete redundant layers, thereby reducing system resource overhead and improving detection rate.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If redundant data transformation operation layers (Permute, Reshape, Flatten) are included in the detection network SSD, then the network structure maintains completeness for general-purpose processing, but system resource overhead increases and detection rate decreases
Solution Approach 1:
The patent implements dynamic network structure optimization by introducing an optimization instruction mechanism that adapts the network architecture based on specific chip types and application scenarios. The system dynamically identifies and removes redundant layers (Permute, Reshape, Flatten) when processing on MLU chips, transforming the static network structure into a dynamic one that adjusts to hardware characteristics, thereby improving detection rate while maintaining necessary structural completeness.
Solution Approach 2:
The optimization method applies local quality improvement by selectively removing only the redundant data transformation layers from specific positions in the network structure, while preserving other essential layers. This localized optimization targets specific problematic layers (Permute, Reshape, Flatten) without affecting the overall network architecture, reducing resource overhead in specific areas while maintaining global functionality.
2Adaptability or versatility
If redundant data transformation operation layers are included in the detection network SSD, then the network can handle diverse data formats, but device complexity increases
Solution Approach 1:
The system employs dynamic structure adjustment where the network complexity adapts based on the target chip type. For MLU chips, redundant layers are removed to simplify the structure, while the optimization mechanism ensures data format compatibility is maintained through appropriate layer retention for other chip types. This dynamic approach resolves the contradiction between versatility and complexity.
Solution Approach 2:
The patent extracts and removes the redundant data transformation layers (Permute, Reshape, Flatten) from the network structure when processing on MLU chips. By taking out these specific problematic components that cause unnecessary complexity, the system reduces device complexity while maintaining data format compatibility through the optimized network configuration.
3Productivity
If optimization operations are applied to reduce network complexity, then resource overhead decreases, but network structure completeness may be compromised
Solution Approach 1:
The optimization process applies local quality control by selectively removing only redundant layers (Permute, Reshape, Flatten) while preserving essential network layers. This localized optimization ensures resource efficiency improvement without compromising overall network structure completeness, as only specific non-essential layers are targeted for removal based on chip-type-specific requirements.
Solution Approach 2:
The system implements feedback control through the optimization instruction mechanism that monitors chip type and application requirements. The feedback loop determines which layers to remove or retain, ensuring that optimization operations improve resource efficiency while maintaining necessary structural completeness for the specific application scenario, preventing over-optimization that would compromise functionality.
Data Source
AI summary
The disclosure relates to a network structure processing method and device and a related product, including: obtaining, by a computer device, an optimization instruction; and executing a corresponding optimization processing operation on the network structure according to the optimization level in the optimization instruction to obtain an optimized network structure. According to the processing method of the network structure, the resource cost can be reduced, and the detection rate of the network structure on images is improved.