인공지능 알고리즘 분석 방법 - jupyter 활용

이 자료는 한국인공지능협회에서 진행한 2021년 사업주 직업능력 개발훈련으로 BASIC AI(32시간)-박성주 강사의 온라인 교육내용을 참고한 것입니다. 

 

MNIST 데이터셋 불러오기¶

In [1]:

import tensorflow as tf
from tensorflow.keras import datasets, layers, models
from keras.utils.vis_utils import plot_model

# 1. MNIST 데이터셋 불러오기
(train_images, train_labels), (test_images, test_labels) = datasets.mnist.load_data()

 

데이터 전처리하기¶

In [2]:

# 2. 데이터 전처리하기
train_images = train_images.reshape((60000, 28, 28, 1))
test_images = test_images.reshape((10000, 28, 28, 1))
train_images, test_images = train_images / 255.0, test_images / 255.0

 

합성곱 신경망 구성하기¶

In [3]:

# 3. 합성곱 신경망 구성하기
model = models.Sequential()
model.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=(28, 28, 1)))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))

model.summary()

 

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 conv2d (Conv2D)             (None, 26, 26, 32)        320       
                                                                 
 max_pooling2d (MaxPooling2D  (None, 13, 13, 32)       0         
 )                                                               
                                                                 
 conv2d_1 (Conv2D)           (None, 11, 11, 64)        18496     
                                                                 
 max_pooling2d_1 (MaxPooling  (None, 5, 5, 64)         0         
 2D)                                                             
                                                                 
 conv2d_2 (Conv2D)           (None, 3, 3, 64)          36928     
                                                                 
 max_pooling2d_2 (MaxPooling  (None, 1, 1, 64)         0         
 2D)                                                             
                                                                 
=================================================================
Total params: 55,744
Trainable params: 55,744
Non-trainable params: 0
_________________________________________________________________

 

Dense 층 추가하기¶

In [4]:

# 4. Dense 층 추가하기
model.add(layers.Flatten())
model.add(layers.Dense(64, activation='relu'))
model.add(layers.Dense(10, activation='softmax'))

model.summary()

 

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 conv2d (Conv2D)             (None, 26, 26, 32)        320       
                                                                 
 max_pooling2d (MaxPooling2D  (None, 13, 13, 32)       0         
 )                                                               
                                                                 
 conv2d_1 (Conv2D)           (None, 11, 11, 64)        18496     
                                                                 
 max_pooling2d_1 (MaxPooling  (None, 5, 5, 64)         0         
 2D)                                                             
                                                                 
 conv2d_2 (Conv2D)           (None, 3, 3, 64)          36928     
                                                                 
 max_pooling2d_2 (MaxPooling  (None, 1, 1, 64)         0         
 2D)                                                             
                                                                 
 flatten (Flatten)           (None, 64)                0         
                                                                 
 dense (Dense)               (None, 64)                4160      
                                                                 
 dense_1 (Dense)             (None, 10)                650       
                                                                 
=================================================================
Total params: 60,554
Trainable params: 60,554
Non-trainable params: 0
_________________________________________________________________

In [5]:

# 모델확인
plot_model(model, show_shapes=True, dpi=80)

Out[5]:

 

모델 컴파일하기¶

In [6]:

# 5. 모델 컴파일하기
model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=["accuracy"])

 

훈련하기¶

In [7]:

# 6. 훈련하기
history = model.fit(train_images, train_labels, batch_size=16, epochs=10, verbose=1, validation_data=(test_images,test_labels))

 

Epoch 1/10
3750/3750 [==============================] - 22s 6ms/step - loss: 0.1819 - accuracy: 0.9430 - val_loss: 0.0587 - val_accuracy: 0.9824
Epoch 2/10
3750/3750 [==============================] - 22s 6ms/step - loss: 0.0666 - accuracy: 0.9789 - val_loss: 0.0550 - val_accuracy: 0.9823
Epoch 3/10
3750/3750 [==============================] - 22s 6ms/step - loss: 0.0471 - accuracy: 0.9851 - val_loss: 0.0556 - val_accuracy: 0.9830
Epoch 4/10
3750/3750 [==============================] - 22s 6ms/step - loss: 0.0366 - accuracy: 0.9885 - val_loss: 0.0622 - val_accuracy: 0.9825
Epoch 5/10
3750/3750 [==============================] - 21s 6ms/step - loss: 0.0290 - accuracy: 0.9905 - val_loss: 0.0417 - val_accuracy: 0.9871
Epoch 6/10
3750/3750 [==============================] - 21s 6ms/step - loss: 0.0247 - accuracy: 0.9916 - val_loss: 0.0491 - val_accuracy: 0.9858
Epoch 7/10
3750/3750 [==============================] - 21s 6ms/step - loss: 0.0201 - accuracy: 0.9936 - val_loss: 0.0480 - val_accuracy: 0.9875
Epoch 8/10
3750/3750 [==============================] - 22s 6ms/step - loss: 0.0189 - accuracy: 0.9937 - val_loss: 0.0802 - val_accuracy: 0.9818
Epoch 9/10
3750/3750 [==============================] - 22s 6ms/step - loss: 0.0172 - accuracy: 0.9944 - val_loss: 0.0497 - val_accuracy: 0.9876
Epoch 10/10
3750/3750 [==============================] - 22s 6ms/step - loss: 0.0154 - accuracy: 0.9949 - val_loss: 0.0544 - val_accuracy: 0.9868

In [8]:

print(history.history.keys())

 

dict_keys(['loss', 'accuracy', 'val_loss', 'val_accuracy'])

In [9]:

import matplotlib.pyplot as plt

plt.plot(history.history['accuracy'])
plt.plot(history.history['val_accuracy'])
plt.title('model accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'test'], loc='upper left')
plt.show()
# summarize history for loss
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('model loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'test'], loc='upper left')
plt.show()

 

 

 

모델 평가하기¶

In [10]:

# 7. 모델 평가하기
loss, acc = model.evaluate(test_images, test_labels, verbose=2)

 

313/313 - 1s - loss: 0.0544 - accuracy: 0.9868 - 913ms/epoch - 3ms/step

In [11]:

print('테스트 정확도 : ', acc)

 

테스트 정확도 :  0.9868000149726868

 

test_images: 숫자로 실제 평가 테스트¶

In [12]:

plt.imshow(test_images[587],cmap='Greys')
plt.show()

 

In [13]:

print(test_labels[587])

 

6

 

결과 예측하기¶

In [14]:

result= model.predict(test_images)

In [15]:

print(result[587])

 

[1.1671032e-09 1.7788943e-12 2.1471147e-09 3.5404605e-12 2.2394228e-07
 1.0474565e-09 9.9999976e-01 4.9988434e-12 3.5337108e-10 1.5825743e-11]

In [16]:

import numpy as np
result_num = np.argmax(result[587])
print("예측 결과값 = %d" % (result_num))

 

예측 결과값 = 6

 

실제 데이터를 직접 처리해서 결과를 예측해보자¶

In [17]:

# 이미지를 가져오는 라이브러리 호출
# 2.7 이상붙처는 내부에 설치되어 있음. pip install pillow
from PIL import Image, ImageOps

In [18]:

image = Image.open('./test_5.jpg')

In [19]:

# 정규화
size = (28, 28)
image = ImageOps.fit(image, size, Image.ANTIALIAS)
image_array = np.asarray(image)
normal_image_array = (image_array.astype(np.float32) / 127.0) -1

data = normal_image_array
real_data = data.reshape(-1, 28, 28, 1)

In [20]:

plt.imshow(data, cmap='Greys')
plt.show()

 

In [21]:

real_result = model.predict(real_data)
print(real_result)

 

[[5.5553396e-06 1.2178321e-02 2.7646674e-02 3.3765911e-05 2.2704608e-03
  9.5192426e-01 4.4796267e-03 7.8509147e-05 9.0447837e-04 4.7833423e-04]]

In [22]:

# 실제 데이터 검증
real_num = np.argmax(real_result)
print("포토샵의 숫자 예측 결과값 = %d" % (real_num))

 

포토샵의 숫자 예측 결과값 = 5

In [ ]: