Method and apparatus with dilated convolution
A neural network apparatus includes one or more processors configured to acquire an input feature map and trained weights, generate a plurality of sub-feature maps by splitting the input feature map based on a dilation rate, generate a plurality of intermediate feature maps by performing a convolution operation between the plurality of sub-feature maps and the trained weights, and generate a dilated output feature map by merging the plurality of intermediate feature maps based on the dilation rate.
1. A neural network apparatus, comprising:
one or more processors configured to:
acquire an input feature map and trained weights,
generate a plurality of sub-feature maps by splitting the input feature map based on a dilation rate,
generate a plurality of intermediate feature maps by performing a convolution operation between the plurality of sub-feature maps and the trained weights, and
generate a dilated output feature map by merging the plurality of intermediate feature maps based on the dilation rate,
wherein, for the generating of the sub-feature maps, the one or more processors are configured to:
compare a size of each of the plurality of sub-feature maps with a predetermined size, and
in response to a size of at least one sub-feature map of the plurality of sub-feature maps being smaller than the predetermined size, pad at least one part of edges of the at least one sub-feature map with a predetermined value such that the size of the at least one sub-feature map becomes the predetermined size.
2. The apparatus of claim 1 , wherein neighboring pixels of each of the plurality of sub-feature maps are pixels differently spaced, with respect to the input feature map, by a number resulting from subtracting 1 from the dilation rate.
3. The apparatus of claim 1 , wherein, each of the plurality of sub-feature maps includes pixels from the input feature map separated in the input feature map according to the dilation rate.
4. The apparatus of claim 1 , wherein, for the padding of the at least one part of edges of the at least one sub-feature map, the one or more processors are configured to add either one or both of a row or column of pixels of the predetermined value to the at least one sub-feature map.
5. The apparatus of claim 1 , wherein neighboring pixels of each of the plurality of intermediate feature maps are pixels spaced by a number resulting from subtracting 1 from the dilation rate in the output feature map.
6. The apparatus of claim 1 , wherein the one or more processors comprise:
a splitter configured to perform the splitting of the input feature map into the plurality of sub-feature maps based on the dilation rate, and
a merger configured to perform the merging of the plurality of intermediate feature maps based on the dilation rate to generate the output feature map.
7. The apparatus of claim 6 , wherein, for the generating of the sub-feature maps, the one or more processors are configured to:
compare the dilation rate with a threshold value, and
in response to the dilation rate exceeding the threshold value, update the dilation rate by dividing the dilation rate by another predetermined value, and split the input feature map into the plurality of sub-feature maps based on the other predetermined value.
8. The apparatus of claim 7 ,
wherein the sub-feature maps are first sub-feature maps, and
wherein the one or more processors are configured to:
compare the updated dilation rate with the threshold value, and
in response to the updated dilation rate exceeding the threshold value, update the updated dilation rate by dividing the updated dilation rate by the other predetermined value, and split the plurality of first sub-feature maps into a plurality of second-sub feature maps based on the other predetermined value.
9. The apparatus of claim 6 , wherein, for the generating of the sub-feature maps, the one or more processors are configured to:
compare the dilation rate with a threshold value, and
in response to the dilation rate exceeding the threshold value, update the dilation rate by dividing the dilation rate by a predetermined value, and merge the plurality of intermediate feature maps based on the predetermined value to generate the output feature map.
10. The apparatus of claim 1 , wherein the one or more processors comprise a convolution operator configured to perform the convolution operation between the plurality of sub-feature maps and the trained weights.
11. The apparatus of claim 1 , wherein the one or more processors are configured to acquire a dilated weight from a memory and restore the dilated weight to the trained weights based on the dilation rate.
12. The apparatus of claim 1 , wherein:
the input feature map corresponds to an input image, and
the one or more processors are configured to perform an image recognition for the input image based on the output feature map.
13. A processor-implemented neural network method, the method comprising:
acquiring an input feature map and trained weights;
generate a plurality of sub-feature maps by splitting the input feature map based on a dilation rate;
generate a plurality of intermediate feature maps by performing a convolution operation between the plurality of sub-feature maps and the trained weights; and
generate a dilated output feature map by merging the plurality of intermediate feature maps based on the dilation rate,
wherein the generating of the plurality of sub-feature maps further comprises:
comparing a size of each of the plurality of sub-feature maps with a predetermined size; and
in response to a size of at least one sub-feature map of the plurality of sub-feature maps being smaller than the predetermined size, padding at least one part of edges of the at least one sub-feature map with a predetermined value such that the size of the at least one sub-feature map becomes the predetermined size.
14. The method of claim 13 , wherein neighboring pixels of each of the plurality of sub-feature maps are pixels spaced by a number resulting from subtracting 1 from the dilation rate in the input feature map.
15. The method of claim 13 , wherein neighboring pixels of each of the plurality of intermediate feature maps are pixels spaced by a number resulting from subtracting 1 from the dilation rate in the output feature map.
16. The method of claim 13 , wherein the generating of the plurality of sub-feature maps comprises:
comparing the dilation rate with a threshold value; and
in response to the dilation rate exceeding the threshold value, updating the dilation rate by dividing the dilation rate by another predetermined value, and splitting the input feature map into the plurality of sub-feature maps based on the other predetermined value.
17. The method of claim 13 , wherein the generating of the plurality of sub-feature maps comprises:
comparing the dilation rate with a threshold value, and
in response to the dilation rate exceeding the threshold value, updating the dilation rate by dividing the dilation rate by another predetermined value, and merging the plurality of intermediate feature maps based on the other predetermined value to generate the output feature map.
18. The method of claim 13 , wherein the acquiring of the input feature map and the trained weights comprises:
acquiring the input feature map and a dilated weight from a memory;
restoring the dilated weight to the trained weights based on the dilation rate.
19. A non-transitory computer-readable storage medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform the method of claim 13 .