Multidimensional Data Generation via Transposition and 1x1 Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The SE block in neural networks reduces processing speed due to a large overhead caused by the H×W times copy process in the Scale layer, especially when the number of elements in the C dimension is large, leading to slow performance.
Innovation Solution
A multidimensional data generation device and method that transforms first multidimensional data with C elements in the channel dimension and 1 elements in other dimensions into second data with one dimension being C, increases the channel dimension elements from 1 to N using a convolution layer with 1×1 filters, performs transposition to restore the original dimensionality, and generates data with C elements in the channel dimension and predetermined elements in other dimensions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the H×W times copy process is used in the Scale layer to generate multidimensional data, then the accuracy of the SE block is maintained, but the processing speed significantly decreases due to large overhead
Solution Approach 1:
The patent performs preliminary transformation of the input multidimensional data into a transposed format before the Scale layer, so that the data is already in the required H×W×C structure. This eliminates the need for the H×W times copy process during actual execution, thereby maintaining accuracy while significantly improving processing speed.
Solution Approach 2:
The patent changes the dimensionality arrangement of the input data by performing transposition operations. Instead of processing 1×1×C data through repeated copying, the data is transformed into H×W×C format through dimensional reorganization, allowing direct generation of the required output structure without repeated memory copy operations.
2Quantity of substance
If the number of elements in the C dimension is increased to improve channel information processing, then the processing overhead in the copy process becomes enormous, leading to slower performance
Solution Approach 1:
The patent performs preliminary transposition of the input data to match the required output dimensionality before entering the SE block. This preliminary arrangement ensures that when C dimension elements need to be expanded to H×W×C format, the operation can be performed efficiently through matrix transformation rather than element-by-element copying, thus reducing processing time regardless of the number of C dimension elements.
3Measurement precision
If the SE block structure is used to improve CNN accuracy, then the device complexity and processing overhead increase due to the required H×W times copy process
Solution Approach 1:
The patent performs preliminary transposition of input data to the required H×W×C format before the SE block processing. This preliminary action simplifies the Scale layer operation from repeated copying to a single transformation step, reducing device complexity and processing overhead while maintaining the accuracy-improving functionality of the SE block.
Solution Approach 2:
The patent changes the dimensional arrangement of data through transposition operations, transforming the problem from one requiring repeated copying in the original dimensionality to one that can be solved through efficient matrix transformation in the transposed dimensionality, thereby reducing overall system complexity.
Data Source
AI summary
The transforming means 72 transforms first multidimensional data in which the number of elements of dimension of channel is C and the number of elements of each dimension other than the dimension of channel is 1 into second multidimension of a predetermined form. The channel dimension element number increase means 73 generates third multidimensional data in which the number of elements of the dimension of channel is increased from 1 to N, by performing a convolution layer process with a filter size of 1×1. The transposition means 74 performs transposition on the third multidimensional data so that the number of elements of the dimension of channel becomes C. The generation means 75 generates multidimensional data in which the number of elements of the dimension of channel is C and the number of elements of each dimension other than the dimension of channel is predetermined number of elements.