Neural network processing apparatus, neural network processing method, and neural network processing program
A CNN processing apparatus ( 1 ) includes an input buffer ( 10 ) configured to store an input signal A given to a CNN, a weight buffer ( 11 ) configured to store weights U, a convolutional operation unit ( 12 ) configured to perform a convolutional operation including a product-sum operation of the input signal A and the weights U, a storage unit 16 configured to store a table ( 160 ) which is configured to associate an input and an output of conversion-quantization processing with each other, wherein the input is an operation result of the convolutional operation, and the output is a result of the conversion-quantization processing of converting the input value based on a predetermined condition and quantizing the converted value by reducing a bit accuracy of the converted data, and a processing unit ( 14 ) configured to acquire the output of the conversion-quantization processing corresponding to the operation result by the operation unit by referring to the table ( 160 ).
1. A neural network processing apparatus comprising:
a first memory configured to store an image data given to a neural network;
a second memory configured to store weights of the neural network;
an operation unit configured to perform a convolutional operation of the neural network including a product-sum operation of the image data and the weights;
a third memory configured to store a table which is configured to associate an input and an output of a conversion-quantization process with each other; and
a processing unit configured to acquire the output of the conversion-quantization process corresponding to an operation result of the convolutional operation by the operation unit;
wherein the conversion-quantization process includes:
a conversion process that converts an input value based on a predetermined condition to obtain a converted data having a bit precision, and
a quantization process that quantizes the converted data by reducing the bit precision of the converted data, and
the processing unit is configured to perform the conversion process and the quantization process together and the output of the conversion and quantization process is acquired by referring to the table.
2. The neural network processing apparatus according to claim 1 , wherein
the table associates a plurality of input sections obtained by dividing the input of the conversion-quantization process into a plurality of continuous sections and the output of the conversion-quantization process with each other, and
the processing unit comprises:
an input determination unit configured to compare the operation result of the convolutional operation by the operation unit with the plurality of input sections and determine an input section including the operation result; and
an output acquisition unit configured to acquire the output of the conversion-quantization process corresponding to the input section according to a determination result of the input determination unit by referring to the table.
3. The neural network processing apparatus according to claim 1 , wherein
the table associates a plurality of thresholds set in advance for the input of the conversion-quantization process and the output of the conversion-quantization process with each other, and
the processing unit comprises:
a first threshold processing unit configured to compare the operation result of the convolutional operation by the operation unit with the plurality of thresholds and output a threshold corresponding to the operation result; and
an output acquisition unit configured to acquire the output of the conversion-quantization process corresponding to the threshold output by the first threshold processing unit by referring to the table.
4. The neural network processing apparatus according to claim 2 , wherein
the table associates division information for identifying an input section in which the output of the conversion-quantization process monotonously increases and an input section in which the output of the conversion-quantization process monotonously decreases, a plurality of thresholds set in advance for the input of the conversion-quantization process, and the output of the conversion-quantization process corresponding to each of the plurality of thresholds with each other,
the input determination unit determines, based on the division information, the input section of the conversion-quantization process to which the operation result of the convolutional operation by the operation unit belongs, and
the processing unit comprises:
a second threshold processing unit configured to compare the operation result of the convolutional operation by the operation unit with the plurality of thresholds in the input section determined by the input determination unit and output a threshold corresponding to the operation result; and
an output acquisition unit configured to acquire the output of the conversion-quantization process corresponding to the threshold output by the second threshold processing unit by referring to the table.
5. The neural network processing apparatus according to claim 1 , wherein
the neural network is a multilayer neural network including at least one intermediate layer.
6. The neural network processing apparatus according to claim 1 , wherein
processing of converting the operation result of the convolutional operation by the operation unit based on the predetermined condition, which is included in the conversion-quantization process, includes at least one of decision of the operation result by an activation function and normalization of the operation result.
7. A neural network processing method comprising:
a first step of storing an image data given to a neural network in a first memory;
a second step of storing weights of the neural network in a second memory;
a third step of performing a convolutional operation of the neural network including a product-sum operation of the image data and the weights;
a fourth step of storing, in a third memory, a table which is configured to associate an input and an output of a conversion-quantization process with each other; and
a fifth step of acquiring the output of the conversion-quantization process corresponding to an operation result of the convolutional operation in the third step by referring to the table,
wherein the conversion-quantization process includes:
a conversion process that converts an input value based on a predetermined condition, and
a quantization process that quantizes the converted data by reducing a bit precision of the converted data, and
the fifth step of executing the conversion process and the quantization process together and the output of the conversion and quantization process is acquired by referring to the table.
8. A non-transitory computer-readable storage medium storing a neural network processing program configured to cause a computer to execute:
a first step of storing an image data given to a neural network in a first memory;
a second step of storing weights of the neural network in a second memory;
a third step of performing a convolutional operation of the neural network including a product-sum operation of the image data and the weights;
a fourth step of storing, in a third memory, a table which is configured to associate an input and an output of conversion-quantization process with each other; and
a fifth step of acquiring the output of the conversion-quantization process corresponding to an operation result of the convolutional operation in the third step by referring to the table,
wherein the conversion-quantization process comprising:
a conversion process that converts an input value based on a predetermined condition, and
a quantization process that quantizes the converted data by reducing a bit precision of the converted data, and
the fifth step of executing the conversion process and the quantization process together and the output of the conversion and quantization process is acquired by referring to the table.