Microsoft MVP성태의 닷넷 이야기
닷넷: 2247. C# - tensorflow 연동 (MNIST 예제) [링크 복사], [링크+제목 복사],
조회: 9362
글쓴 사람
정성태 (techsharer at outlook.com)
홈페이지
첨부 파일
(연관된 글이 1개 있습니다.)

C# - tensorflow 연동 (MNIST 예제)

요즘 접하기 쉬운 예제로 MNIST 손글씨 인식을 C#에서 tensorflow와 연동해 만들어 보겠습니다. 여기서 중요한 것은, Model을 구해야 하는 것인데요 ^^ 그 부분은 그냥 파이썬 환경에서 자유롭게 코딩해 구하기만 하면 됩니다.

예를 들어, 아래의 MNIST 예제는 my_mnist_model.keras 파일로 모델을 저장하고 있습니다.

// 케라스 창시자에게 배우는 딥러닝
// https://github.com/gilbutITbook/080315/blob/main/chapter02_mathematical-building-blocks.ipynb

import setuptools.dist
from tensorflow.keras.datasets import mnist
(train_images, train_labels), (test_images, test_labels) = mnist.load_data()

from tensorflow import keras
from tensorflow.keras import layers

model = keras.Sequential([
    layers.Dense(512, activation='relu'),
    layers.Dense(10, activation='softmax')
    ])

model.compile(optimizer='rmsprop', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

train_images = train_images.reshape((60000, 28 * 28))
train_images = train_images.astype('float32') / 255

test_images = test_images.reshape((10000, 28 * 28))
test_images = test_images.astype('float32') / 255

model.fit(train_images, train_labels, epochs=5, batch_size=128)

test_loss, test_acc = model.evaluate(test_images, test_labels)
print(f'{test_acc}')

# https://www.tensorflow.org/tutorials/keras/save_and_load?hl=ko
model.save('my_mnist_model.keras')

my_mnist_model.keras 파일의 크기는 3MB 정도 됩니다. 이렇게 구한 Model 파일은 C# 프로젝트에 추가/배포해, 실행 시 C#에서 Python.NET을 이용해 저 Model 파일을 로드해 활용할 것입니다.




자, 그럼 본격적으로 위에서 만든 MNIST 필기체 인식 Model을 C#에서 Python과 연동해 볼까요? ^^

이를 위해, 모델을 이용한 predict 코드를 호출하는 파이썬 코드를 다음과 같이 만들어 줍니다.

# mnist_predict.py

import setuptools.dist
import tensorflow as tf
import numpy as np

model = tf.keras.models.load_model('my_mnist_model.keras')

def predict(img):
    imgs = np.expand_dims(img, axis=0)
    predictions = model.predict(imgs, verbose=0)
    predict_number = np.argmax(predictions[0])
    return (predict_number.item(), predictions[0][predict_number].item())

위의 predict 함수는 model.predict 호출 시 해당 이미지로 판정되는 숫자와 그 확률을 반환합니다.

그럼, 이제 Python.NET을 이용한 C# 코드에서는 이를 호출하는 코드만 다음과 같이 작성해 주면 됩니다.

using Python.Runtime;

namespace ConsoleApp3;

internal class Program
{
    static void Main(string[] args)
    {
        Runtime.PythonDLL = @".\python\python312.dll";

        PythonEngine.Initialize();

        using (_ = Py.GIL())
        {
            DisableTensorflowLog();

            dynamic npModule = Py.Import("numpy");

            {
                dynamic sys = Py.Import("sys");
                string dirPath = Path.GetDirectoryName(typeof(Program).Assembly.Location) ?? Environment.CurrentDirectory;
                sys.path.append(dirPath);
            }

            float[]? testImgArray = // ... 28x28 크기의 이미지 데이터 ...;
            dynamic npArray = npModule.array(testImgArray);

            {
                var pyFile = Py.Import(Path.GetFileNameWithoutExtension("mnist_predict"));

                dynamic results = pyFile.InvokeMethod("predict", npArray);

                int expected = results[0];
                double percentage = results[1];

                Console.WriteLine($"{expected}: {percentage:P0}");
            }
        }

        PythonEngine.Shutdown();
    }
}

만약 testImgArray에 7과 비슷한 숫자의 이미지를 담고 있는 28x28 크기의 버퍼가 있다면 위의 프로그램을 실행 시 "7: 100%"와 유사한 출력이 나옵니다.

(첨부 파일은 이 글의 예제 코드를 포함합니다.)

만약 지난 글에 설명한 대로 CopyToOutputDirectory 설정을 했다면, 위의 예제를 실행했을 때 "C:\temp\ConsoleApp3\net8.0" 디렉터리에 출력이 모였을 것입니다. 해당 출력 파일만 다른 컴퓨터에 그대로 복사하면 (당연히 별도의 파이썬 설치 없이) 정상적으로 실행까지 됩니다.

한 가지 문제점이라면, 위의 경우 net8.0 출력에 있는 전체 바이너리의 크기가 (python + tensorflow까지 포함하므로) 1.6GB 정도, 압축하면 480MB 정도 됩니다. 만약 대상 컴퓨터에 파이썬 tensorflow 환경이 설치돼 있다면 이 용량을 없앨 수 있지만 그렇지 않은 경우라면... 뭔가 있어 보이는 ^^ 응용 프로그램의 크기를 자랑합니다.




참고로, 위의 코드를 Windows 10+ 환경에서 Python 3.12.0 버전으로 실행하면 load_model 시에 다음과 같은 오류가 발생합니다.

Traceback (most recent call last):
  File "C:\temp\ConsoleApp3\net8.0\python\test.py", line 36, in <module>
    model2 = tf.keras.models.load_model('my_mnist_model.keras')
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\temp\ConsoleApp3\net8.0\python\Lib\site-packages\keras\src\saving\saving_api.py", line 176, in load_model
    return saving_lib.load_model(
           ^^^^^^^^^^^^^^^^^^^^^^
  File "C:\temp\ConsoleApp3\net8.0\python\Lib\site-packages\keras\src\saving\saving_lib.py", line 152, in load_model
    return _load_model_from_fileobj(
           ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\temp\ConsoleApp3\net8.0\python\Lib\site-packages\keras\src\saving\saving_lib.py", line 207, in _load_model_from_fileobj
    _raise_loading_failure(error_msgs)
  File "C:\temp\ConsoleApp3\net8.0\python\Lib\site-packages\keras\src\saving\saving_lib.py", line 295, in _raise_loading_failure
    raise ValueError(msg)
ValueError: A total of 1 objects could not be loaded. Example error message for object <keras.src.optimizers.adam.Adam object at 0x000001CBB0BCFBF0>:

The shape of the target variable and the shape of the target value in `variable.assign(value)` must match. variable.shape=(10,), Received: value.shape=(512, 10). Target variable: <KerasVariable shape=(10,), dtype=float32, path=adam/dense_1_bias_momentum>

List of objects that could not be loaded:
[<keras.src.optimizers.adam.Adam object at 0x000001CBB0BCFBF0>]))

3.12.2 이상의 버전에서 하면 오류가 발생하지 않습니다.




[이 글에 대해서 여러분들과 의견을 공유하고 싶습니다. 틀리거나 미흡한 부분 또는 의문 사항이 있으시면 언제든 댓글 남겨주십시오.]

[연관 글]






[최초 등록일: ]
[최종 수정일: 4/24/2024]

Creative Commons License
이 저작물은 크리에이티브 커먼즈 코리아 저작자표시-비영리-변경금지 2.0 대한민국 라이센스에 따라 이용하실 수 있습니다.
by SeongTae Jeong, mailto:techsharer at outlook.com

비밀번호

댓글 작성자
 




1  2  3  4  [5]  6  7  8  9  10  11  12  13  14  15  ...
NoWriterDateCnt.TitleFile(s)
13818정성태11/15/20245256Windows: 272. Windows 11 24H2 - sudo 추가
13817정성태11/14/20244915Linux: 106. eBPF / bpf2go - (BPF_MAP_TYPE_HASH) Map을 이용한 전역 변수 구현
13816정성태11/14/20245380닷넷: 2312. C#, C++ - Windows / Linux 환경의 Thread Name 설정파일 다운로드1
13815정성태11/13/20244800Linux: 105. eBPF - bpf2go에서 전역 변수 설정 방법
13814정성태11/13/20245277닷넷: 2311. C# - Windows / Linux 환경에서 Native Thread ID 가져오기파일 다운로드1
13813정성태11/12/20245014닷넷: 2310. .NET의 Rune 타입과 emoji 표현파일 다운로드1
13812정성태11/11/20245245오류 유형: 933. Active Directory - The forest functional level is not supported.
13811정성태11/11/20244844Linux: 104. Linux - COLUMNS 환경변수가 언제나 80으로 설정되는 환경
13810정성태11/10/20245374Linux: 103. eBPF (bpf2go) - Tracepoint를 이용한 트레이스 (BPF_PROG_TYPE_TRACEPOINT)
13809정성태11/10/20245250Windows: 271. 윈도우 서버 2025 마이그레이션
13808정성태11/9/20245247오류 유형: 932. Linux - 커널 업그레이드 후 "error: bad shim signature" 오류 발생
13807정성태11/9/20244983Linux: 102. Linux - 커널 이미지 파일 서명 (Ubuntu 환경)
13806정성태11/8/20244891Windows: 270. 어댑터 상세 정보(Network Connection Details) 창의 내용이 비어 있는 경우
13805정성태11/8/20244732오류 유형: 931. Active Directory의 adprep 또는 복제가 안 되는 경우
13804정성태11/7/20245359Linux: 101. eBPF 함수의 인자를 다루는 방법
13803정성태11/7/20245309닷넷: 2309. C# - .NET Core에서 바뀐 DateTime.Ticks의 정밀도
13802정성태11/6/20245691Windows: 269. GetSystemTimeAsFileTime과 GetSystemTimePreciseAsFileTime의 차이점파일 다운로드1
13801정성태11/5/20245480Linux: 100. eBPF의 2가지 방식 - libbcc와 libbpf(CO-RE)
13800정성태11/3/20246314닷넷: 2308. C# - ICU 라이브러리를 활용한 문자열의 대소문자 변환 [2]파일 다운로드1
13799정성태11/2/20244891개발 환경 구성: 732. 모바일 웹 브라우저에서 유니코드 문자가 표시되지 않는 경우
13798정성태11/2/20245490개발 환경 구성: 731. 유니코드 - 출력 예시 및 폰트 찾기
13797정성태11/1/20245484C/C++: 185. C++ - 문자열의 대소문자를 변환하는 transform + std::tolower/toupper 방식의 문제점파일 다운로드1
13796정성태10/31/20245368C/C++: 184. C++ - ICU dll을 이용하는 예제 코드 (Windows)파일 다운로드1
13795정성태10/31/20245163Windows: 268. Windows - 리눅스 환경처럼 공백으로 끝나는 프롬프트 만들기
13794정성태10/30/20245261닷넷: 2307. C# - 윈도우에서 한글(및 유니코드)을 포함한 콘솔 프로그램을 컴파일 및 실행하는 방법
13793정성태10/28/20245125C/C++: 183. C++ - 윈도우에서 한글(및 유니코드)을 포함한 콘솔 프로그램을 컴파일 및 실행하는 방법
1  2  3  4  [5]  6  7  8  9  10  11  12  13  14  15  ...