Deep Learning Model Channel Pruning for Edge Detection Speed
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models for head and shoulder detection face challenges in cloud-based deployment due to high costs, large bandwidth requirements, and poor privacy, while terminal-based models suffer from low computational speed and large size.
Innovation Solution
A method and apparatus for compressing deep learning models by pruning weights in units of channels, calculating L1-norms, and retraining the model, which reduces parameter redundancy and sends the compressed model to terminal devices for storage and execution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a deep learning model is run on a cloud, then detection accuracy is maintained, but cost increases and bandwidth pressure increases
Solution Approach 1:
The patent extracts and removes redundant parameters from the deep learning model through pruning techniques. Specifically, it prunes convolutional layer parameters and fully connected layer parameters to reduce model size while maintaining detection accuracy, thereby reducing the bandwidth required for model deployment and inference on terminal devices
Solution Approach 2:
The patent changes the parameter structure of the deep learning model by applying L1-norm based pruning to transform the original dense parameter structure into a sparse structure. This parameter transformation reduces the overall model size and bandwidth requirements while preserving essential detection capabilities
2Reliability
If a deep learning model is run on a terminal device, then privacy is improved and bandwidth pressure is reduced, but model size is too large and computational speed is low
Solution Approach 1:
The patent extracts and removes redundant computational operations from the deep learning model through systematic pruning of convolutional and fully connected layers. This extraction of unnecessary parameters reduces model size and computational load, enabling faster inference on terminal devices with limited resources
Solution Approach 2:
The patent segments the pruning process into distinct phases: first pruning convolutional layer parameters, then pruning fully connected layer parameters, and finally performing retraining. This segmented approach allows for controlled model compression while maintaining detection accuracy and improving computational efficiency
3Volume of stationary object
If model parameters are pruned to reduce size, then storage space is reduced, but model complexity increases
Solution Approach 1:
The patent incorporates feedback mechanisms through iterative retraining after each pruning phase. The model is pruned, evaluated, and then retrained on the training dataset to restore performance. This feedback loop ensures that pruning operations do not excessively increase model complexity while achieving storage space reduction
Data Source
AI summary
Embodiments of the present disclosure disclose a method and apparatus for compressing a deep learning model. An embodiment of the method includes: acquiring a to-be-compressed deep learning model; pruning each layer of weights of the to-be-compressed deep learning model in units of channels to obtain a compressed deep learning model; and sending the compressed deep learning model to a terminal device, so that the terminal device stores the compressed deep learning model. By pruning each layer of weights of the deep learning model in units of channels, the parameter redundancy of the deep learning model is effectively reduced, thereby improving the computational speed of the deep learning model and maintaining the model accuracy.