Information processing device and information processing method

A method for generating data for executing, in a semiconductor integrated circuit, calculations according to a neural network trained using a deep learning framework, the method comprising:

• executing, by one or more processors, the calculations according to the neural network based on first parameters of the trained neural network expressed in floating-point and a floating-point forward processing algorithm to obtain a first result;
• converting, by the one or more processors, the first parameters into second parameters expressed in fixed-point;
• executing, by the one or more processors, the calculations according to the neural network based on the second parameters of the trained neural network expressed in fixed-point and a fixed-point forward processing algorithm to obtain a second result;
• and
• generating, by the one or more processors and based on the trained neural network, data including the second parameters, the data being used when the semiconductor integrated circuit executes the calculations according to the neural network,
• wherein the data is in a data format that is executable by the semiconductor integrated circuit, and
• wherein the deep learning framework is configured to perform training of one or more neural networks using a define-by-run configuration.

A device for generating data for executing, in a semiconductor integrated circuit, calculations according to a neural network trained using a deep learning framework, the device comprising:

• one or more processors;
• and
• one or more memories storing one or more program instructions that, when executed by the one or more processors, cause the one or more processors to:
• execute the calculations according to the neural network based on first parameters of the trained neural network expressed in floating-point and a floating-point forward processing algorithm to obtain a first result;
• convert the first parameters into second parameters expressed in fixed-point;
• execute the calculations according to the neural network based on the second parameters of the trained neural network expressed in fixed-point and a fixed-point forward processing algorithm to obtain a second result;
• and
• generate, based on the trained neural network, data including the second parameters, the data being used when the semiconductor integrated circuit executes the calculations according to the neural network,
• wherein the data is in a data format that is executable by the semiconductor integrated circuit, and
• wherein the deep learning framework is configured to perform training of one or more neural networks using a define-by-run configuration.

A non-transitory computer readable medium storing one or more program instructions that perform a method when executed by one or more processors, the method comprising:

• executing, after a neural network is trained using a deep learning framework, calculations according to the trained neural network based on first parameters of the trained neural network expressed in floating-point and a floating-point forward processing algorithm to obtain a first result;
• converting the first parameters into second parameters expressed in fixed-point;
• executing the calculations according to the trained neural network based on the second parameters of the trained neural network expressed in fixed-point and a fixed-point forward processing algorithm to obtain a second result;
• and
• generating, based on the trained neural network, data including the second parameters, the data used when a semiconductor integrated circuit executes the calculations according to the trained neural network,
• wherein the data is in a data format that is executable by the semiconductor integrated circuit, and
• wherein the deep learning framework is configured to perform training of one or more neural networks using a define-by-run configuration.