Feature Map Size Detection for Neural Network Memory Allocation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network models face inefficiencies in detecting image sizes of feature maps across network layers, hindering effective memory allocation and processing efficiency.
Innovation Solution
A method involving a computer device that obtains a pre-trained image processing model, processes input images, establishes a stack based on network layers, detects feature map sizes, and iteratively deletes layers to optimize memory usage, using modules like acquisition, establishing, processing, detection, and iteration modules.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If traditional methods are used to detect image sizes of feature maps in neural network models, then memory space can be reserved in advance, but the detection process is inefficient and time-consuming
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing the relationship between input image sizes and output feature map sizes in a lookup table before actual image processing. When an image size detection is needed, the system directly queries the pre-computed table rather than performing complex calculations in real-time, significantly reducing detection time while maintaining accuracy
Solution Approach 2:
The patent creates a simplified copy of the neural network's size transformation logic in the form of a lookup table that mirrors the behavior of the actual network layers. This copy stores pre-computed size relationships without requiring the full computational overhead of the original network, enabling fast size detection through simple table queries
2Measurement precision
If all network layers are processed to detect feature map sizes, then complete size information is obtained, but memory usage and processing complexity increase
Solution Approach 1:
The patent extracts only the essential size transformation information from the neural network layers and stores it in a compact lookup table. Instead of processing and storing all intermediate feature maps, it extracts and stores only the size relationships, significantly reducing memory consumption while preserving the ability to accurately determine feature map sizes
Solution Approach 2:
The patent performs preliminary computation of size relationships and stores them in advance in a lookup table. This pre-computation phase separates the heavy calculation work from the actual image processing, allowing the detection phase to use minimal memory by simply querying pre-stored size information rather than computing it on-the-fly
Data Source
AI summary
A method for detecting image sizes using a computer device is provided. In the method, the computer device obtains a pre-trained image processing model and an input image, establishes a stack according to a number of network layers of the pre-trained image processing model. Output feature maps are obtained by processing the input image using the pre-trained image processing model. Image sizes of the output feature maps are detected. The image sizes are written into the stack in sequence, and the network layers deleted iteratively in an order from large to small according to an execution sequence of the network layers of the pre-trained image processing model until the stack is full. By performing the method, output sizes of the input image can be obtained, and a detection efficiency of image sizes can be improved.