Displaced Feature Map Convolution for Lightweight DNN Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep neural networks (DNNs) face inefficiencies due to multiple convolution operations, particularly in lightweight structures like MobileNet and ShuffleNet, which reduce operation efficiency.
Innovation Solution
A method involving displacement of input feature maps along axes using preset displacement parameters, followed by a single 1×1 convolution kernel, to aggregate information on both channel and spatial dimensions, thereby reducing the need for multiple convolutions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If Depthwise Convolution is used to reduce calculation complexity, then computational cost is reduced, but operation efficiency decreases due to multiple convolution operations
Solution Approach 1:
The patent merges the depthwise convolution and pointwise convolution operations into a single unified operation. Instead of performing separate depthwise convolution followed by pointwise convolution, the invention combines these steps into one integrated convolutional layer that achieves both spatial filtering and channel-wise aggregation simultaneously, thereby reducing the total number of operations while maintaining computational efficiency
Solution Approach 2:
The patent creates a universal convolutional operation that performs multiple functions simultaneously. The single convolutional layer implements both the depthwise spatial filtering function and the pointwise channel aggregation function that were previously separate operations, making the convolutional layer multi-functional and reducing overall computational steps
2Loss of information
If multiple convolution operations are used in DWConv, then information aggregation is achieved, but operation efficiency is reduced
Solution Approach 1:
The patent combines the information aggregation function previously distributed across multiple separate convolution operations into a single unified convolutional layer. This merged operation maintains comprehensive information aggregation across both spatial and channel dimensions while reducing the total number of computational steps required
Data Source
Figure 1~2
Figure 3a
Figure 3b
AI summary
A deep neural network operation method and apparatus are provided. The method comprises: obtaining an input feature map of a network layer; displacing respectively, according to a preset displacement parameter, each of channels of the input feature map of the network layer along axes, to obtain a displaced feature map, wherein the preset displacement parameter comprises displacement amounts of the channel in the axes; and performing convolution operation on the displaced feature map with a 1×1 convolution kernel to obtain an output feature map of the network layer. The operation efficiency of the DNN can be improved through the above method.