Convolution operation method and apparatus, computer device, and computer-readable storage medium
View Patent ↗Embodiments of the present application provide a method, an apparatus, a computer device for convolution operation and a computer readable storage medium. The method includes: obtaining input data of a network layer in a convolutional neural work; extracting each time a plurality of data points from the input data according to a preset step size; mapping the plurality of data points extracted each time to the same position at different depth levels of three-dimensional data to obtain rearranged data; and performing convolution operation on the rearranged data with a convolution kernel of a preset size to obtain a convolution result. Through the present solution, the operation efficiency of the convolutional neural network can be improved.
1. A method for convolution operation, performed by a computer device having a convolution kernel of a preset size, comprising:
obtaining input data of a network layer in a convolutional neural work, wherein the input data is three-dimensional data having a size represented as W×H×I, wherein I is a depth of the input data, and W×H is a data size at each depth level, wherein a size of a convolution kernel for the input data is larger than the preset size;
extracting each time a plurality of data points from the input data according to a preset step size;
mapping the plurality of data points extracted each time to the same position at different depth levels of three-dimensional data to obtain rearranged data such that a size of the input data at each depth level is reduced without changing original volume of the input data, wherein a depth of the rearranged data is larger than a depth of the input data, and a data size of the rearranged data at each depth level is smaller than the data size of the input data at each depth level; the data size of the rearranged data at each depth allows that convolution operation can be performed on the rearranged data by using the convolution kernel of the preset size; and
performing convolution operation on the rearranged data with the convolution kernel of the preset size to obtain a convolution result.
2. The method of claim 1 , wherein, before extracting each time a plurality of data points from the input data according to a preset step size, the method further comprises:
dividing the input data in a depth direction to obtain a plurality of slices;
the operation of extracting each time a plurality of data points from the input data according to a preset step size comprises:
for each slice, extracting each time data points from depth levels of the slice according to the preset step size, to obtain the plurality of data points;
the operation of mapping the plurality of data points extracted each time to the same position at different depth levels in three-dimensional data to obtain rearranged data comprises:
mapping the plurality of data points extracted each time from each slice to the same position at different depth levels in the three-dimensional data to respectively obtain to-be-merged data corresponding to each slice; and
arranging a plurality of to-be-merged data in the depth direction to obtain the rearranged data.
3. The method of claim 1 , wherein extracting each time a plurality of data points from the input data according to a preset step size comprises:
extracting each time, from each depth level of the input data, a plurality of data points according to the preset step size respectively;
the operation of mapping the plurality of data points extracted each time to the same position at different depth levels in three-dimensional data to obtain rearranged data comprises:
mapping the plurality of data points extracted each time from depth levels of the input data to the same position at different depth levels in the three-dimensional data to obtain a plurality of data to be merged; and
arranging a plurality of to-be-merged data in the depth direction to obtain the rearranged data.
4. The method of claim 1 , wherein, mapping the plurality of data points extracted each time to the same position at different depth levels of three-dimensional data to obtain rearranged data comprises:
arranging the plurality of data points extracted each time; and
storing the plurality of data points extracted each time, in an order as they are arranged, to the same position at different depth levels in the three-dimensional data to obtain the rearranged data.
5. An apparatus for convolution operation, the apparatus having a convolution kernel of a preset size, comprising:
an obtaining module, configure for obtaining input data of a network layer in a convolutional neural work, wherein the input data is three-dimensional data having a size represented as W×H×I, wherein I is a depth of the input data, and W×H is a data size at each depth level, wherein a size of a convolution kernel for the input data is larger than the preset size;
an extraction module, configured for extracting each time a plurality of data points from the input data according to a preset step size;
a mapping module, configured for mapping the plurality of data points extracted each time to the same position at different depth levels of three-dimensional data to obtain rearranged data such that a size of the input data at each depth level is reduced without changing original volume of the input data, wherein a depth of the rearranged data is larger than a depth of the input data, and a data size of the rearranged data at each depth level is smaller than the data size of the input data at each depth level; the data size of the rearranged data at each depth allows that convolution operation can be performed on the rearranged data by using the convolution kernel of the preset size; and
an operation module, configured for performing convolution operation on the rearranged data with the convolution kernel of the preset size to obtain a convolution result.
6. The apparatus of claim 5 , further comprising:
a division module, configured for dividing the input data in a depth direction to obtain a plurality of slices;
the extraction module is further configured for:
for each slice, extracting each time data points from depth levels of the slice according to the preset step size, to obtain the plurality of data points;
the mapping module is further configured for:
mapping the plurality of data points extracted each time from each slice to the same position at different depth levels in the three-dimensional data to respectively obtain to-be-merged data corresponding to each slice; and
arranging a plurality of to-be-merged data in the depth direction to obtain the rearranged data.
7. The apparatus of claim 5 , wherein the extraction module is further configured for:
extracting each time, from each depth level of the input data, a plurality of data points according to the preset step size respectively;
the mapping module is further configured for:
mapping the plurality of data points extracted each time from depth levels of the input data to the same position at different depth levels in the three-dimensional data to obtain a plurality of data to be merged; and
arranging a plurality of to-be-merged data in the depth direction to obtain the rearranged data.
8. The apparatus of claim 5 , wherein the extraction module is further configured for:
arranging the plurality of data points extracted each time; and
storing the plurality of data points extracted each time, in an order as they are arranged, to the same position at different depth levels in the three-dimensional data to obtain the rearranged data.
9. A non-transitory computer readable storage medium having executable codes stored thereon which, when executed, performs the method for convolution operation of claim 1 .
10. A computer device, comprising:
a non-transitory computer readable storage medium configured for storing executable codes; and
a processor configured for executing the executable code stored in the computer readable storage medium to perform the method for convolution operation of claim 1 .