CTU Block Partitioning in Neural Network Model Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large neural networks require significant storage and computational resources, making them challenging to deploy on devices with limited processing power and memory, such as mobile devices, due to their large size and complexity.
Innovation Solution
The development of techniques for neural network model compression and decompression, including coding tree unit block partitioning, 3D pyramid structure-based coding, and unification-based coding, which allow for efficient encoding and decoding of neural network models, reducing their size and computational requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If neural network models are deployed on resource-constrained devices, then performance improvement is achieved, but storage and computational resource requirements become excessive
Solution Approach 1:
The patent segments the neural network model into multiple blocks and applies different quantization precision levels to different blocks. This allows the model to maintain high performance in critical blocks while reducing storage requirements in less critical blocks, effectively resolving the contradiction between performance and storage resources.
Solution Approach 2:
The patent implements local quality by assigning different precision levels (e.g., high precision for some blocks, low precision for others) to different parts of the neural network model. This selective approach ensures that performance-critical regions maintain high quality while non-critical regions consume fewer storage resources, thereby resolving the technical contradiction.
2Reliability
If neural network models are deployed on resource-constrained devices, then performance improvement is achieved, but computational load becomes excessive
Solution Approach 1:
The patent divides the neural network into multiple blocks and applies different quantization strategies to each block. This segmentation allows computationally intensive operations to be concentrated in fewer high-precision blocks while other blocks use lower precision, reducing overall computational load while maintaining performance.
Solution Approach 2:
The patent changes the precision parameter (quantization level) of different blocks within the neural network model. By adjusting these parameters, the system can reduce computational load in less critical blocks while maintaining sufficient precision in performance-critical blocks, thus resolving the contradiction between performance and computational load.
3Reliability
If uniform high precision is applied to all blocks, then performance is maintained, but storage and computational resources are inefficiently used
Solution Approach 1:
The patent applies local quality by assigning different precision levels to different blocks based on their importance to overall model performance. This avoids the waste of applying uniform high precision to all blocks, thereby improving resource efficiency while maintaining necessary performance levels in each block.
Solution Approach 2:
The patent introduces parameter changes by varying the quantization precision across different blocks rather than using a uniform setting. This allows the system to optimize resource efficiency by adjusting parameters according to the specific requirements of each block, resolving the contradiction between performance maintenance and resource efficiency.
Data Source
AI summary
A method of neural network decoding includes receiving a first syntax element in a model parameter set from a bitstream of a compressed neural network representation (NNR) of a neural network. The first syntax element indicates whether a coding tree unit (CTU) block partitioning is enabled for a tensor in an NNR aggregate unit. The method also includes reconstructing the tensor in the NNR aggregate unit based on the first syntax element.