Method of optimizing neural network model that is pre-trained, method of providing a graphical user interface related to optimizing neural network model, and neural network model processing system performing the same

A method of optimizing a neural network model, the method comprising:

• receiving original model information about a first neural network model that is pre-trained and includes a plurality of original layers;
• generating a second neural network model and compressed model information about the second neural network model by performing a compression on the first neural network model, the second neural network model including a plurality of compressed layers;
• displaying a first graphical representation on a graphical user interface (GUI), the first graphical representation including a network structure of the plurality of compressed layers;
• receiving, through the GUI, a first user input with respect to a first compressed layer among the plurality of compressed layers;
• displaying a second graphical representation on the GUI based on the first user input, the second graphical representation including a comparison of a characteristic of a first original layer among the plurality of original layers and a characteristic of the first compressed layer, the first original layer corresponding to the first compressed layer;
• receiving, through the GUI, a second user input for changing a setting of a second compressed layer among the plurality of compressed layers;
• updating a characteristic of the second compressed layer based on the second user input;
• displaying a third graphical representation on the GUI based on the second user input, the third graphical representation including a comparison of a characteristic of a second original layer among the plurality of original layers and the updated characteristic of the second compressed layer, the second original layer corresponding to the second compressed layer;
• generating a plurality of score values for the plurality of compressed layers;
• displaying a fourth graphical representation on the GUI, the fourth graphical representation including the plurality of compressed layers that are displayed according to different schemes based on the plurality of score values;
• and
• displaying a fifth graphical representation on the GUI, the fifth graphical representation including the plurality of compressed layers at least one of which is changeable based on the plurality of score values,
• wherein the first graphical representation includes a plurality of layer boxes corresponding to the plurality of compressed layers, and
• wherein the displaying the first graphical representation includes:
• displaying, on the GUI that displays the first graphical representation, a first menu including indicators of two or more from a signal-to-quantization-noise power ratio (SQNR), a latency, a power consumption, and a utilization amount;
• and
• receiving, via the GUI, a third user input to select one of the indicators from the first menu, and grouping the plurality of compressed layers according to a reference value associated with the selected indicator,
• wherein the method further comprises:
• displaying, based on a result of grouping, a sixth graphical representation on the GUI, the sixth graphical representation including a plurality of layer group boxes corresponding to a plurality of compressed layer groups, a compressed layer group of the plurality of compressed layer groups including two or more compressed layers that are grouped according to the reference value associated with the selected indicator, wherein
• the plurality of layer group boxes are representations in a reduced form in which a number of the plurality of layer group boxes is less than a number of compressed layers that are grouped into the plurality of layer group boxes, and compressed layers that are grouped are not individually represented, and
• based on a second user input, via the GUI, to select a reduced form of a representation of one of the plurality of layer group boxes, individual representations of two or more compressed layers grouped into a compressed layer group corresponding to the selected reduced form of the representation are displayed in an extended form, and
• wherein, in the sixth graphical representation, a compressed layer that does not satisfy a predetermined criterion of a performance for the selected indicator is not grouped into any compressed layer group and is individually represented in a layer box separate from the plurality of layer group boxes.

A method of providing a graphical user interface (GUI) related to optimizing a neural network model, the method comprising:

• receiving first model information about a first neural network model that is pre-trained, the first neural network model including a plurality of original layers;
• generating a second neural network model and second model information about the second neural network model by performing a data processing on the first neural network model, the second neural network model including a plurality of compressed layers;
• and
• providing a graphical user interface (GUI), which displays a graphical representation including at least a part of the first model information and at least a part of the second model information in comparison with each other,
• wherein the providing the GUI includes:
• displaying a first graphical representation on the GUI, the first graphical representation including a network structure of the plurality of compressed layers and a plurality of layer boxes corresponding to the plurality of compressed layers;
• displaying, on the GUI that displays the first graphical representation, a first menu including indicators of two or more from a signal-to-quantization-noise power ratio (SQNR), a latency, a power consumption, and a utilization amount;
• receiving, via the GUI, a first user input to select one of the indicators from the first menu, and grouping the plurality of compressed layers according to a reference value associated with the selected indicator;
• and
• displaying, based on a result of grouping, a second graphical representation on the GUI, the second graphical representation including a plurality of layer group boxes corresponding to a plurality of compressed layer groups, a compressed layer group of the plurality of compressed layer groups including two or more compressed layers that are grouped according to the reference value associated with the selected indicator, wherein
• the plurality of layer group boxes are representations in a reduced form in which a number of the plurality of layer group boxes is less than a number of compressed layers that are grouped into the plurality of layer group boxes, and compressed layers that are grouped are not individually represented, and
• based on a second user input, via the GUI, to select a reduced form of a representation of one of the plurality of layer group boxes, individual representations of two or more compressed layers grouped into a compressed layer group corresponding to the selected reduced form of the representation are displayed in an extended form, and
• wherein, in the second graphical representation, a compressed layer that does not satisfy a predetermined criterion of a performance for the selected indicator is not grouped into any compressed layer group and is individually represented in a layer box separate from the plurality of layer group boxes.