Convolutional neural network with building blocks
An apparatus that may include a neural network processor, the neural network processor comprises multiple building blocks. Each of the at least some of the building blocks may include, may consist or may consist essentially of a channel split unit, a convolution unit, a concatenation unit, and a shuffle unit.
1. An apparatus that comprises a neural network processor, the neural network processor comprises multiple building blocks;
wherein each of the multiple building blocks comprises:
a channel split unit, a convolution unit, a concatenation unit, and a shuffle unit;
wherein the channel split unit is executed by the neural network processor to receive information related to a group of channels, send information related to a first sub-group of the group of channels to the convolution unit, and send information related to a second sub-group of the group of channels to the concatenation unit;
wherein the convolution unit is executed by the neural network processor to perform at least one convolution operation on the information related to the first sub-group of the group of channels to provide convolution results; wherein the convolution results are associated with the first sub-group of the group of channels;
wherein the concatenation unit is executed by the neural network processor to concatenate the convolution results and the information related to the second sub-group of the group of channels to provide concatenation results, wherein each concatenation result of the concatenation results represents a dedicated channel of the group of channels;
wherein the shuffle unit is executed by the neural network processor to shuffle the concatenation results to provide shuffled results, wherein each of the shuffled results is associated with a dedicated channel of the group of channels;
wherein the shuffle unit is executed by the neural network processor to interleave shuffled results related to the first sub-group of the group of channels with shuffled results related to the second sub-group of the group of channels; and
wherein the shuffle unit is executed by the neural network processor to maintain, within the shuffled results related to the first sub-group of the group of channels, an order of shuffled results related to the first sub-group of the group of channels.
2. The apparatus according to claim 1 wherein adjacent concatenation results are mapped to spaced apart shuffled results.
3. The apparatus according to claim 1 wherein the shuffle unit is executed by the neural network processor to maintain, within the shuffled results related to the second sub-group of channels, an order of shuffled results related to the second sub-group of channels.
4. The apparatus according to claim 1 wherein the group of channels comprises a first number (N) of channels, wherein N is an even positive integer, wherein the first sub-group of channels comprises a first half of channels and wherein the second sub-group of channels comprises a second half of channels, wherein N/2 is a half of N.
5. The apparatus according to claim 1 wherein the convolution unit is a five by five convolution unit.
6. The apparatus according to claim 1 wherein the convolution unit is not a one by one convolution units.
7. The apparatus according to claim 1 wherein each of the multiple building blocks consists essentially of the channel split unit, the convolution unit, the concatenation unit, and the shuffle unit.
8. The apparatus according to claim 1 wherein each of the multiple building blocks consists of the channel split unit, the convolution unit, the concatenation unit, and the shuffle unit.
9. The apparatus according to claim 1 wherein each of the multiple building blocks consists of the channel split unit, the convolution unit, the concatenation unit, and the shuffle unit.
10. The apparatus according to claim 1 comprises at least one convolution unit that is coupled between a pair of building blocks.
11. The apparatus according to claim 1 wherein at least one convolution unit does not include one by one convolution subunits.
12. The apparatus according to claim 1 wherein at least one convolution unit comprises of five by five convolution subunits.
13. A non-transitory computer readable medium that stores instructions, the instructions are executed by a neural network processor to perform a method comprising:
receiving information related to a group of channels;
sending information related to a first sub-group of the group of channels to a convolution unit;
sending information related to a second sub-group of the group of channels to a concatenation unit;
performing, by the convolution unit, at least one convolution operation on the information related to the first sub-group of the group of channels to provide convolution results; wherein the convolution results are associated with the first sub-group of the group of channels;
concatenating, by the concatenation unit, the convolution results and the information related to the second sub-group of the group of channels to provide concatenation results, wherein each concatenation result of the concatenation results represents a dedicated channel of the group of channels; and
shuffling, by a shuffle unit, the concatenation results to provide shuffled results,
wherein each of the shuffled results is associated with a dedicated channel of the group of channels;
wherein the shuffle unit is executed by the neural network processor to interleave shuffled results related to the first sub-group of the group of channels with shuffled results related to the second sub-group of the group of channels; and
wherein the shuffle unit is executed by the neural network processor to maintain, within the shuffled results related to the first sub-group of the group of channels, an order of shuffled results related to the first sub-group of the group of channels.
14. A method executed by a neural network processor, comprising:
receiving information related to a group of channels;
sending information related to a first sub-group of the group of channels to a convolution unit;
sending information related to a second sub-group of the group of channels to a concatenation unit;
performing, by the convolution unit, at least one convolution operation on the information related to the first sub-group of the group of channels to provide convolution results; wherein the convolution results are associated with the first sub-group of the group of channels;
concatenating, by the concatenation unit, the convolution results and the information related to the second sub-group of the group of channels to provide concatenation results, wherein each concatenation result of the concatenation results represents a dedicated channel of the group of channels; and
shuffling, by a shuffle unit, the concatenation results to provide shuffled results, wherein each of the shuffled results is associated with a dedicated channel of the group of channels;
wherein the shuffle unit is executed by the neural network processor to interleave shuffled results related to the first sub-group of the group of channels with shuffled results related to the second sub-group of the group of channels; and
wherein the shuffle unit is executed by the neural network processor to maintain, within the shuffled results related to the first sub-group of the group of channels, an order of shuffled results related to the first sub-group of the group of channels.