Parameter optimization device, parameter optimization method, and parameter optimization program

1 . A parameter optimization device that optimizes input CNN structure information and outputs optimized CNN structure information, the parameter optimization device comprising:

a memory storing a software component; and

one or more processors configured to execute the instructions to:

extract stride and dilation parameter information for each convolution layer from the input CNN structure information; and

change the stride and dilation parameter information of the convolution layer,

wherein, until there are no more pairs of changeable stride and dilation parameters for all convolution layers in the CNN structure, the one or more processors execute the instructions to repeatedly execute a process for extracting the parameter information, and a process for changing the parameter information.

2 . The parameter optimization device according to claim 1 , wherein:

the one or more processors execute the instructions to execute a process for extracting the parameter information for each layer in the convolution layer in the CNN structure while switching the layer from a deep layer to a shallow layer, and

the one or more processors execute a process for changing the parameter information for each layer in the convolution layer in the CNN structure while switching the layer from a deep layer to a shallow layer.

3 . The parameter optimization device according to claim 1 , wherein when the greatest common divisor of a stride value and a dilation value for a certain convolution layer is greater than 1, the one or more processors execute the instructions to change the stride value for the certain convolution layer to a value obtained by dividing the stride value by the greatest common divisor, the dilation value for the certain convolution layer to a value obtained by dividing the dilation value by the greatest common divisor, and a stride value for the convolution layer that is one-level shallower than the certain convolution layer to a value obtained by multiplying the stride value by the greatest common divisor.

4 . The parameter optimization device according to claim 1 , wherein when both a stride value and a dilation value for a certain layer are 2, the one or more processors execute the instructions to change the stride value for the certain convolution layer to a value obtained by dividing the stride value by 2, the dilation value for the certain convolution layer to a value obtained by dividing the dilation value by 2, and a stride value for the convolution layer that is one-level shallower than the certain layer to a value obtained by multiplying the stride value by 2.

5 . The parameter optimization device according to claim 1 , the one or more processors further execute the instructions to:

determine whether a modification of a shortcut process due to a result of changing the stride and dilation is required or not in the case where the CNN structure includes the shortcut process; and

modify the shortcut process when it is determined that the modification of the shortcut process is required.

6 . The parameter optimization device according to claim 5 , wherein the one or more processors further execute the instructions to determine that the modification of the shortcut process is required, when the change of the stride occurs in two convolution layers across an addition process in the shortcut process.

7 . The parameter optimization device according to claim 5 , wherein the one or more processors execute the instructions to:

introduce a thinning process with the stride whose value is equal to the stride value after the change, or a 1×1 convolution process such that the stride has the same value as the stride value after the change and the weights are represented by a unit matrix, when the shortcut process before the modification does not change input value; and

change the stride of the 1×1 convolution process to a value multiplied by the stride value after the change, when the shortcut process before the modification includes the 1×1 convolution process.

8 . A parameter optimization method, implemented by a processor, for optimizing input CNN structure information and outputting optimized CNN structure information, the parameter optimization method comprising:

extracting stride and dilation parameter information for each convolution layer from the input CNN structure information; and

changing the stride and dilation parameter information of the convolution layer,

wherein until there are no more pairs of changeable stride and dilation parameters for all convolution layers in the CNN structure, a process for extracting the parameter information and a process for changing the parameter information are executed.

9 . The parameter optimization method according to claim 8 , further comprising:

determining whether a modification of a shortcut process due to a result of changing the stride and dilation is required or not in the case where the CNN structure includes the shortcut process; and

when determining that the modification of the shortcut process is required, modifying the shortcut process.

10 . A non-transitory computer readable recording medium storing a parameter optimization program for optimizing input CNN structure information and outputting optimized CNN structure information, wherein the program causes a processor to execute:

a process of extracting stride and dilation parameter information for each convolution layer from the input CNN structure information; and

a process of changing the stride and dilation parameter information of the convolution layer,

wherein until there are no more pairs of changeable stride and dilation parameters for all convolution layers in the CNN structure, the program causes the processor to execute a process for extracting the parameter information and a process for changing the parameter information.

11 . The recording medium according to claim 10 , wherein the program causes the processor to further execute:

a process of determining whether a modification of a shortcut process due to a result of changing the stride and dilation is required or not in the case where the CNN structure includes the shortcut process; and

a process of modifying the shortcut process when determining that the modification of the shortcut process is required.