Electronic apparatus and method for controlling thereof
An electronic apparatus, including a memory configured to store weight data used for computation of a neural network model; and a processor configured to: identify, from among weight values included in the weight data, at least one weight value having a size less than or equal to a threshold value, quantize remaining weight values other than the identified at least one weight value to obtain first quantized data including quantized values corresponding to the remaining weight values, identify, from among the quantized values, a quantized value closest to a predetermined value, obtain second quantized data including a quantized value corresponding to the at least one weight value based on the quantized value closest to the predetermined value, and store the first quantized data and the second quantized data in the memory.
1 . An electronic apparatus comprising:
a memory configured to store weight data used for computation of a neural network model; and
a processor configured to:
identify, from among weight values included in the weight data, at least one weight value having a size less than or equal to a threshold value,
quantize remaining weight values other than the identified at least one weight value to obtain first quantized data including quantized values corresponding to the remaining weight values,
identify, from among the quantized values, a quantized value closest to a predetermined value,
obtain second quantized data including a quantized value corresponding to the at least one weight value based on the quantized value closest to the predetermined value,
store the first quantized data and the second quantized data in the memory,
quantize the neural network model by modifying the at least one weight value based on the second quantized data and modifying the remaining weight values based on the first quantized data to obtain a quantized neural network model, and
obtain output data by providing input data to the quantized neural network model.
2 . The electronic apparatus of claim 1 , wherein the processor is further configured to identify the quantized value closest to the predetermined value as the quantized value corresponding to the at least one weight value.
3 . The electronic apparatus of claim 1 , wherein the processor is further configured to:
based on the quantization of the remaining weight values, obtain the first quantized data including a plurality of scaling factors and bit values corresponding to the remaining weight values,
obtain a plurality of computational values based on computation using the plurality of scaling factors,
identify, from among the plurality of computational values, a computational value closest to the predetermined value, and
obtain the second quantized data based on the identified computational value.
4 . The electronic apparatus of claim 3 , wherein the processor is further configured to identify the computational value closest to the predetermined value as the quantized value corresponding to the at least one weight value.
5 . The electronic apparatus of claim 3 , wherein the processor is further configured to:
obtain the plurality of computational values by adding values obtained by multiplying each of the plurality of scaling factors by +1 or −1, and
identify the computational value closest to the predetermined value among the plurality of computational values.
6 . The electronic apparatus of claim 3 , wherein the processor is further configured to:
identify a first plurality of computations which output the computational value closest to the predetermined value among a second plurality of computations using the plurality of scaling factors,
identify, from among the first plurality of computations, a computation which outputs a computational value having a same code as a code of the at least one weight value, and
obtain the second quantized data based on the identified computation.
7 . The electronic apparatus of claim 3 , wherein the processor is further configured to:
identify codes of the plurality of scaling factors included in a computation, among a plurality of computations using the plurality of scaling factors, to output the computational value closest to the predetermined value, and
obtain the second quantized data based on the codes.
8 . The electronic apparatus of claim 1 , wherein the processor is further configured to:
receive a user input to set a pruning rate, and
identify the at least one weight value based on sizes of the weight values.
9 . The electronic apparatus of claim 1 , wherein the predetermined value is zero.
10 . A method for controlling an electronic apparatus, the method comprising:
identifying, from among weight values included in weight data, at least one weight value having a size less than or equal to a threshold value;
quantizing remaining weight values other than the identified at least one weight value to obtain first quantized data including quantized values corresponding to the remaining weight values;
identifying, from among the quantized values, a quantized value closest to a predetermined value;
obtaining second quantized data including a quantized value corresponding to the at least one weight value based on the quantized value closest to the predetermined value; and
storing the first quantized data and the second quantized data,
quantizing the neural network model by modifying the at least one weight value based on the second quantized data and modifying the remaining weight values based on the first quantized data to obtain a quantized neural network model, and
obtaining output data by providing input data to the quantized neural network model.
11 . The method of claim 10 , wherein the obtaining the second quantized data comprises identifying the quantized value closest to the predetermined value as the quantized value corresponding to the at least one weight value.
12 . The method of claim 10 , wherein the obtaining the first quantized data comprises, based on the quantization of the remaining weight values, obtaining the first quantized data including a plurality of scaling factors and bit values corresponding to the remaining weight values, and
wherein the obtaining the second quantized data comprises:
obtaining a plurality of computational values based on computation using the plurality of scaling factors;
identifying, from among the plurality of computational values, a computational value closest to the predetermined value; and
obtaining the second quantized data based on the identified computational value.
13 . The method of claim 12 , wherein the obtaining the second quantized data comprises identifying the computational value closest to the predetermined value as the quantized value corresponding to the at least one weight value.
14 . The method of claim 12 , wherein the identifying the computational value comprises:
obtaining the plurality of computational values by adding values obtained by multiplying each of the plurality of scaling factors by +1 or −1; and
identifying the computational value closest to the predetermined value among the plurality of computational values.
15 . The method of claim 12 , wherein the obtaining the second quantized data comprises:
identifying a first plurality of computations which output the computational value closest to the predetermined value among a second plurality of computations using the plurality of scaling factors;
identifying, from among the first plurality of computations, a computation which outputs a computational value having a same code as a code of the at least one weight value; and
obtaining the second quantized data based on the identified computation.
16 . A method for controlling an electronic apparatus, the method comprising:
obtaining a weight values included in weight data of a neural network model;
selecting, from among the weight values, a first weight value to be pruned from the weight values;
quantizing remaining weight values other than the first weight value to obtain quantized weight values corresponding to the remaining weight values;
identifying, from among the quantized weight values, a smallest quantized weight value;
based on the smallest quantized weight value, determining a first quantized weight value corresponding to the first weight value;
storing the quantized weight values and the first quantized weight value;
quantize the neural network model by modifying the at least one weight value based on the second quantized data and modifying the remaining weight values based on the first quantized data to obtain a quantized neural network model; and
performing a computation using the quantized neural network model.
17 . The method of claim 16 , wherein the smallest quantized weight value is determined as the first quantized weight value.
18 . The method of claim 16 , wherein the first weight value is selected based on a result of a comparison between the first weight value and a threshold value.
19 . The method of claim 18 , wherein the threshold value is set based on a user input.
20 . The method of claim 19 , wherein the threshold value is determined based on a pruning rate set based on the user input.
21 . The electronic device of claim 1 , further comprising normalizing the plurality of weight values based on fisher values corresponding to the plurality of weight values,
wherein the quantized values are obtained based on the normalized plurality of weight values.