3.MNIST 딥러닝 모델 예제

MNIST

모듈 임포트

import tensorflow as tf

from tensorflow.keras.datasets.mnist import load_data

from tensorflow.keras.models import Sequential

from tensorflow.keras.layers import Dense, Input, Flatten

from tensorflow.keras.utils import to_categorical, plot_model

import numpy as np

import tensorflow as tf

from keras.models import load_model

from sklearn.model_selection import train_test_split

import matplotlib.pyplot as plt

plt.style.use('seaborn-white')

데이터 로드 및 전처리

• MNIST 데이터셋을 로드
• Train Data 중, 30%를 검증 데이터(validation data)로 사용

#고정된 난수 생성을 위해 tf.random.set_seed(111)를 설정합니다.

tf.random.set_seed(111)

(x_train_full, y_train_full), (x_test, y_test) =load_data(path='mnist.npz')

x_train, x_val, y_train, y_val = train_test_split(x_train_full, y_train_full,

                                                  test_size=0.3,

   

                                               random_state=111)

 

#mnist.npz 파일로부터 훈련, 테스트 데이터셋을 불러옵니다.

num_x_train=(x_train.shape[0])

num_x_val=(x_val.shape[0])

num_x_test=(x_test.shape[0])

 

num_sample=5

random_idxs=np.random.randint(60000, size=num_sample)

plt.figure(figsize=(15,3))

for i , idx in enumerate(random_idxs):

  img=x_train_full[idx,:]

  label= y_train_full[idx]

  plt.subplot(1,len(random_idxs),i+1)

  plt.imshow(img)

  plt.title('Index:{},Label:{}'.format(idx,label))

#훈련, 검증, 테스트 데이터셋의 픽셀 값을 255로 나누어 정규화합니다.

x_train=x_train/255.

x_val = x_val/255.

x_test=x_test/255.

 

 

모델 구성(Sequential)

#신경망 구성

model = Sequential([Input(shape=(28,28),name='input'),

                    Flatten(input_shape=[28,28], name='flatten'),

                    Dense(100,activation='relu', name='dense1'),

                    Dense(64, activation='relu', name='dense2'),

                    Dense(32, activation='relu', name='dense3'),

                    Dense(10, activation='softmax', name='output')])

#원핫인코딩

y_train = to_categorical(y_train)

y_val = to_categorical(y_val)

y_test = to_categorical(y_test)

 

모델 컴파일 및 학습

model.compile(loss='categorical_crossentropy',

              optimizer='sgd',

              metrics=['accuracy'])

history = model.fit(x_train, y_train, epochs=50, batch_size=128, validation_data=(x_val, y_val))

history_dict=history.history

loss= history_dict['loss']

val_loss= history_dict['val_loss']

epochs=range(1,len(loss)+1)

fig=plt.figure(figsize=(12,5))

ax1=fig.add_subplot(1,2,1)

ax1.plot(epochs, loss, color='b', label='train_loss')

ax1.plot(epochs, val_loss, color='r', label='val_loss')

ax1.set_xlabel('Epochs')

ax1.set_ylabel('Loss')

ax1.grid()

ax1.legend()

accuracy=history_dict['accuracy']

val_accuracy=history_dict['val_accuracy']

ax2=fig.add_subplot(1,2,2)

ax2.plot(epochs, accuracy, color='b', label='train_accuracy')

ax2.plot(epochs, val_accuracy, color='r', label='val_accuracy')

ax2.set_xlabel('Epochs')

ax2.set_ylabel('accuracy')

ax2.grid()

ax2.legend()

모델 평가 및 예측

 

model.evaluate(x_test, y_test)

pred_ys=model.predict(x_test)

print(pred_ys.shape)

np.set_printoptions(precision=7)

print(pred_ys[0])

arg_pred_y = np.argmax(pred_ys, axis=1)

plt.imshow(x_test[11])

plt.title('Predicted label:{}'.format(arg_pred_y[11]))

plt.show()

 

• 혼동 행렬 (Confusion Matrix)

from sklearn.metrics import classification_report, confusion_matrix

import seaborn as sns

sns.set(style='white')

plt.figure(figsize=(8,8))

cm = confusion_matrix(np.argmax(y_test, axis=1), np.argmax(pred_ys, axis=-1))

sns.heatmap(cm, annot=True, fmt='d', cmap='Blues')

plt.xlabel('Predicted Label')

plt.ylabel('True Lable')

plt.show()