Multi-size convolutional layer
Improved convolutional layers for neural networks can obtain an input feature map comprising groups of channels. Each group of channels can include one or more channels having a predetermined size. The predetermined sizes can differ between the groups. The convolutional layer can generate, for each one of the groups of channels, an output channel. Generation of the output channel can include resizing the channels in the remaining groups of channels to match the predetermined size of the each one of the groups of channels. Generation can further include combining the channels in the each one of the groups with the resized channels and applying the combined channels to a convolutional sub-layer to generate the output channel.
1. A processing unit comprising:
a task manager; and
one or more cores configured by the task manager to generate a neural network output from a neural network input, generation of the neural network output comprising:
generate an output feature map comprising a first output channel and a second output channel using an input feature map comprising a first input channel and a second input channel, generation of the output feature map comprising:
generate, by up-sampling the first input channel, a third input channel having a size of the second input channel;
generate, by down-sampling the second input channel, a fourth input channel having a size of the first input channel;
convolve a first input including the third input channel and the second input channel with a first kernel to generate the first output channel; and
convolve a second input including the fourth input channel and the first input channel with a second kernel to generate the second output channel.
2. The processing unit of claim 1 , wherein:
generation of the neural network output further comprises:
obtain an initial feature map;
generate, by down-sampling the initial feature map, a down-sampled initial feature map;
convolve the initial feature map by a third kernel to generate a second channel group comprising the second input channel; and
convolve the down-sampled initial feature map by a fourth kernel to generate a first channel group comprising the first input channel.
3. The processing unit of claim 1 , wherein the down-sampling comprises at least one of convolution, sampling, max pooling, or averaging pooling.
4. The processing unit of claim 1 , wherein:
generation of the output feature map further comprises combining of the first output channel and the second output channel to generate the output feature map.
5. The processing unit of claim 1 , wherein:
the input feature map comprises groups of channels, each group of channels including multiple channels having a predetermined size, the predetermined sizes differing between the groups.
6. The processing unit of claim 5 , wherein:
the input feature map comprises 2, 4, 8, 16, or 32 groups of channels.
7. The processing unit of claim 5 , wherein:
the predetermined sizes differ by powers of four or more.
8. A method for generating an output feature map including a first output channel and a second output channel from an input feature map comprising a first input channel and a second input channel, using a convolutional layer of a convolutional neural network, the method comprising:
generating, by up-sampling the first input channel, a third input channel having a size of the second input channel;
generating, by down-sampling the second input channel, a fourth input channel having a size of the first input channel;
convolving a first input including the third input channel and the second input channel with a first kernel to generate the first output channel; and
convolving a second input including the fourth input channel and the first input channel with a second kernel to generate the second output channel.
9. The method of claim 8 , further comprising:
obtaining an initial feature map;
generating, by down-sampling the initial feature map, a down-sampled initial feature map;
convolving the initial feature map by a third kernel to generate a second channel group comprising the second input channel; and
convolving the down-sampled initial feature map by a fourth kernel to generate a first channel group comprising the first input channel.
10. The method of claim 8 , wherein the down-sampling comprises at least one of convolution, sampling, max pooling, or averaging pooling.
11. The method of claim 8 , wherein:
generation of the output feature map further comprises combining the first output channel and the second output channel to generate the output feature map.
12. The method of claim 8 , wherein:
the first input feature map comprises groups of channels, each group of channels including multiple channels having a predetermined size, the predetermined sizes differing between the groups.
13. The method of claim 12 , wherein:
the first input feature map comprises 2, 4, 8, 16, or 32 groups of channels.
14. The method of claim 12 , wherein:
the predetermined sizes differ by powers of four or more.
15. A device comprising:
a host unit; and
a processing unit configurable by the host unit to:
generate a neural network output from a neural network input using one or more cores of the neural processing unit, generation of the neural network output comprising:
generate an output feature map comprising a first output channel and a second output channel using an input feature map comprising a first input channel and a second input channel, generation of the output feature map comprising:
generate, by up-sampling the first input channel, a third input channel having a size of the second input channel;
generate, by down-sampling the second input channel, a fourth input channel having a size of the first input channel;
convolve a first input including the third input channel and the second input channel with a first kernel to generate the first output channel; and
convolve a second input including the fourth input channel and the first input channel with a second kernel to generate the second output channel.
16. The device of claim 15 , wherein:
generation of the neural network output further comprises:
obtain an initial feature map;
generate, by down-sampling the initial feature map, a down-sampled initial feature map;
convolve the initial feature map by a third kernel to generate a second channel group comprising the second input channel; and
convolve the down-sampled initial feature map by a fourth kernel to generate a first channel group comprising the first input channel.
17. The device of claim 15 , wherein the down-sampling comprises at least one of convolution, sampling, max pooling, or averaging pooling.
18. The device of claim 15 , wherein:
generation of the output feature map further comprises combining of the first output channel and the second output channel to generate the output feature map.
19. The device of claim 15 , wherein:
the input feature map comprises groups of channels, each group of channels including multiple channels having a predetermined size, the predetermined sizes differing between the groups.
20. The device of claim 19 , wherein:
the input feature map comprises 2, 4, 8, 16, or 32 groups of channels.
21. The device of claim 19 , wherein:
the predetermined sizes differ by powers of four or more.