Method and apparatus for partitioning neural network data
A method and apparatus for scheduling a neural network operation. The method includes receiving data on a layer of a neural network, generating partitions to be assigned to cores by dividing the data, generating tiles by dividing the partitions, and scheduling an operation order of the tiles based on whether the data are shared between the cores.
1. A processor-implemented method, comprising:
dividing an input feature map or weight, respectively of a layer of a neural network, into partitions to be assigned to cores in different ways depending on whether an output of a previous layer is forwardable;
generating tiles by dividing each one of the partitions; and
scheduling an operation order of the tiles based on whether values of the input feature map or the weight are shared between the cores.
2. The method of claim 1 , wherein dividing of the input feature map or weigh comprises generating the partitions based on a partitioning policy of a layer previous to the layer.
3. The method of claim 1 , wherein dividing of the input feature map or weigh comprises:
generating the partitions based on a partitioning policy of the previous layer in response to the output of the previous layer being forwardable; and
generating the partitions by comparing a size of the input feature map of the layer to a size of the weight of the layer in response to the output of the previous layer being unforwardable.
4. The method of claim 3 , wherein the generating of the partitions by comparing the size of the input feature map of the layer to the size of the weight of the layer in response to the output of the previous layer being unforwardable comprises generating the partitions by comparing a loss caused by a memory size to a loss caused by unbalance in response to the input feature map or the weight being not uniformly divided and assigned to the cores.
5. The method of claim 1 , wherein the generating of the tiles comprises generating the tiles based on a partitioning policy of a layer previous to the layer, a partitioning policy of the layer, or a tile division policy of the previous layer.
6. The method of claim 5 , wherein the generating of the tiles comprises:
generating the tiles based on a partitioning policy of a previous layer, a partitioning policy of the layer, and a tile division policy of the previous layer in response to an output of the previous layer being forwardable; and
generating the tiles by comparing a size of the input feature map of the layer to a size of the weight of the layer in response to the output of the previous layer being unforwardable.
7. The method of claim 1 , wherein the generating of the partitions comprises dividing the data in a height or width direction of the data.
8. The method of claim 1 , wherein the scheduling comprises changing the operation order of the tiles based on whether operation results of the tiles are shared between the cores.
9. The method of claim 8 , wherein the changing of the operation order of the tiles comprises changing the operation order so as to operate one tile included in the tiles with priority in response to an operation result of the one tile being shared between a core to which the one tile is assigned and another core.
10. An apparatus, comprising:
a processor configured to:
dividing an input feature map or a weight, respectively of a layer of a neural network, into partitions to be assigned to cores in different ways depending on whether an output of a previous layer is forwardable;
generate tiles by dividing each one of the partitions; and
schedule an operation order of the tiles based on whether values of the input feature map or the weight are shared between the cores.
11. The apparatus of claim 10 , wherein, for the generation of the partitions, the processor is further configured to generate the partitions based on a partitioning policy of a layer previous to the layer.
12. The apparatus of claim 10 , wherein, for the generation of the partitions, the processor is further configured to generate the tiles based on a partitioning policy of a layer previous to the layer, a partitioning policy of the layer, or a tile division policy of the previous layer.
13. The apparatus of claim 10 , wherein the processor is further configured to divide the data in a height or width direction of the data.
14. The apparatus of claim 10 , wherein the processor is further configured to change the operation order of the tiles based on whether operation results of the tiles are shared between the cores.
15. The apparatus of claim 14 , wherein the processor is further configured to change the operation order so as to operate one tile included in the tiles with priority in response to an operation result of the one tile being shared between a core to which the one tile is assigned and another core.
16. The apparatus of claim 11 , wherein, for the generation of the partitions, the processor is further configured to:
generate the partitions based on the partitioning policy of the previous layer in response to an output of the previous layer being forwardable; and
generate the partitions by comparing a size of the input feature map of the layer to a size of the weight of the layer in response to the output of the previous layer being unforwardable.
17. The apparatus of claim 12 , wherein, for the generation of the partitions, the processor is further configured to:
generate the tiles based on the partitioning policy of the previous layer, the partitioning policy of the layer, and the tile division policy of the previous layer in response to an output of the previous layer being forwardable; and
generate the tiles by comparing a size of the input feature map of the layer to a size of the weight of the layer in response to the output of the previous layer being unforwardable.
18. An apparatus, comprising:
a receiver configured to receive data on a layer of a neural network; and
a processor configured to:
generate partitions to be assigned to cores by dividing the data;
generate tiles by dividing the partitions; and
schedule an operation order of the tiles based on whether the data are shared between the cores,
wherein, for the generation of the partitions, the processor is further configured to:
generate the partitions based on a partitioning policy of a previous layer in response to an output of the previous layer being forwardable; and
generate the partitions by comparing a size of an input feature map of the layer to a size of a weight of the layer in response to the output of the previous layer being unforwardable.
19. The apparatus of claim 18 , wherein, for the generation of the partitions, the processor is further configured to generate the partitions by comparing a loss caused by a memory size to a loss caused by unbalance in response to the input feature map or the weight being not uniformly divided and assigned to the cores.
20. An apparatus, comprising:
a receiver configured to receive data on a layer of a neural network; and
a processor configured to:
generate partitions to be assigned to cores by dividing the data;
generate tiles by dividing the partitions; and
schedule an operation order of the tiles based on whether the data are shared between the cores,
wherein, for the generation of the partitions, the processor is further configured to:
generate the tiles based on a partitioning policy of a previous layer, a partitioning policy of the layer, and a tile division policy of the previous layer in response to an output of the previous layer being forwardable; and
generate the tiles by comparing a size of an input feature map of the layer to a size of a weight of the layer in response to the output of the previous layer being unforwardable.