21.03.05. 딥러닝 - 동물 이미지 인식, 패션 이미지 분류

개, 고양이, 말 이미지 인식 - 개, 고양이를 학습했던 모델을 그대로 활용

 

오늘 수업 실습에 앞서서 먼저 어제 진행했던 파일을 복제하여 사용한다. 복제하는 방법은 그림으로 첨부.

그리고 동물이미지를 구하기 위해 아래의 캐글 사이트를 접속한다.

www.kaggle.com/alessiocorrado99/animals10

Animals-10

폴더명이 이태리어로 되어있다보니 '말' 폴더명이 horse가 아니고 cavallo로 되어있다. 폴더를 찾아서 열면 다음과 같이 이름이 제각기(물론 규칙이 있겠지만)로 되어있으니 사용 편의성을 위해 파일을 전체선택(ctrl+A)해서 f2를 누르고 horse를 입력하여 한 번에 이름을 바꿔주도록 한다.

사진명 변경 전 → 변경 후

이 기능을 몰랐어서... 감탄하고 있었는데...

신기한 사람이 되었다.

 

import os

base_dir = "./data/dogs_vs_cats_small2_2"

train_dir = os.path.join(base_dir, "train")
test_dir = os.path.join(base_dir, "test")
validation_dir = os.path.join(base_dir, "validation")

train_cats_dir = os.path.join(train_dir, "cats")
train_dogs_dir = os.path.join(train_dir, "dogs")
train_horses_dir = os.path.join(train_dir, "horses")

test_cats_dir = os.path.join(test_dir, "cats")
test_dogs_dir = os.path.join(test_dir, "dogs")

validation_cats_dir = os.path.join(validation_dir, "cats")
validation_dogs_dir = os.path.join(validation_dir, "dogs")
validation_horses_dir = os.path.join(validation_dir, "horses")

# 각 폴더에 있는 파일의 개수를 출력
print("훈련용 고양이 데이터 개수 : ", len(os.listdir(train_cats_dir)))
print("훈련용 개 데이터 개수 : ", len(os.listdir(train_dogs_dir)))
print("훈련용 말 데이터 개수 : ", len(os.listdir(train_horses_dir)))

print("테스트용 고양이 데이터 개수 : ", len(os.listdir(test_cats_dir)))
print("테스트용 개 데이터 개수 : ", len(os.listdir(test_dogs_dir)))

print("검증용 고양이 데이터 개수 : ", len(os.listdir(validation_cats_dir)))
print("검증용 개 데이터 개수 : ", len(os.listdir(validation_dogs_dir)))
print("검증용 말 데이터 개수 : ", len(os.listdir(validation_horses_dir)))

out :

훈련용 고양이 데이터 개수 :  1000
훈련용 개 데이터 개수 :  1000
훈련용 말 데이터 개수 :  1000
테스트용 고양이 데이터 개수 :  11
테스트용 개 데이터 개수 :  11
검증용 고양이 데이터 개수 :  500
검증용 개 데이터 개수 :  500
검증용 말 데이터 개수 :  500

 

horses까지 추가되어 총 3개가 되었으니 이진분류가 아닌 다중분류에 속한다. 고로, 더이상 binary가 아니라 categorical이다.

# 이미지 처리용 라이브러리
from tensorflow.keras.preprocessing.image import ImageDataGenerator # 전처리

# 0~1 사이값으로 픽셀값을 변환
train_datagen = ImageDataGenerator(rescale=1./255)
test_datagen = ImageDataGenerator(rescale=1./255)

# 이미지 전처리
# 폴더에 있는 이미지를 전처리
train_generator = train_datagen.flow_from_directory(
    # 폴더명
    train_dir,
    # 이미지 크기를 동일한 크기로 변환
    target_size = (150, 150),
    # 한 번에 전처리할 이미지의 수
    batch_size = 20,
    # 라벨링 : binary 이진라벨링, categorical 다중라벨링
    # 라벨링 방법 : 폴더명의 첫 문자의 알파벳으로 0부터 부여
    class_mode = "categorical"
)

test_generator = test_datagen.flow_from_directory(
    # 폴더명
    validation_dir,
    # 이미지 크기를 동일한 크기로 변환
    target_size = (150, 150),
    # 한 번에 전처리할 이미지의 수
    batch_size = 20,
    # 라벨링 : binary 이진라벨링, categorical 다중라벨링
    # 라벨링 방법 : 폴더명의 첫 문자의 알파벳으로 0부터 부여
    class_mode = "categorical"
)

out :

Found 3000 images belonging to 3 classes.

Found 1500 images belonging to 3 classes.

print(train_generator.class_indices)
print(test_generator.class_indices)

out :

{'cats': 0, 'dogs': 1, 'horses': 2}

{'cats': 0, 'dogs': 1, 'horses': 2}

 

activation도 sigmoid가 아닌 softmax로 변경

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout, Flatten
from tensorflow.keras.layers import Conv2D, MaxPooling2D

model1 = Sequential()

model1.add(Conv2D(filters=32, kernel_size=(3, 3), input_shape=(150, 150, 3), padding="same", activation="relu"))
model1.add(MaxPooling2D(pool_size=(2, 2)))
model1.add(Conv2D(filters=64, kernel_size=(3, 3), input_shape=(150, 150, 3), padding="same", activation="relu"))
model1.add(MaxPooling2D(pool_size=(2, 2)))
model1.add(Conv2D(filters=128, kernel_size=(3, 3), input_shape=(150, 150, 3), padding="same", activation="relu"))
model1.add(MaxPooling2D(pool_size=(2, 2)))
model1.add(Conv2D(filters=256, kernel_size=(3, 3), input_shape=(150, 150, 3), padding="same", activation="relu"))
model1.add(MaxPooling2D(pool_size=(2, 2)))

model1.add(Flatten())

model1.add(Dense(units=512, activation="relu"))
model1.add(Dropout(0.5))
model1.add(Dense(units=3, activation="softmax"))

model1.summary()

out :

Model: "sequential" _________________________________________________________________ Layer (type)                 Output Shape              Param #    ================================================================= conv2d (Conv2D)              (None, 150, 150, 32)      896        _________________________________________________________________ max_pooling2d (MaxPooling2D) (None, 75, 75, 32)        0          _________________________________________________________________ conv2d_1 (Conv2D)            (None, 75, 75, 64)        18496      _________________________________________________________________ max_pooling2d_1 (MaxPooling2 (None, 37, 37, 64)        0          _________________________________________________________________ conv2d_2 (Conv2D)            (None, 37, 37, 128)       73856      _________________________________________________________________ max_pooling2d_2 (MaxPooling2 (None, 18, 18, 128)       0          _________________________________________________________________ conv2d_3 (Conv2D)            (None, 18, 18, 256)       295168     _________________________________________________________________ max_pooling2d_3 (MaxPooling2 (None, 9, 9, 256)         0          _________________________________________________________________ flatten (Flatten)            (None, 20736)             0          _________________________________________________________________ dense (Dense)                (None, 512)               10617344   _________________________________________________________________ dropout (Dropout)            (None, 512)               0          _________________________________________________________________ dense_1 (Dense)              (None, 3)                 1539       ================================================================= Total params: 11,007,299 Trainable params: 11,007,299 Non-trainable params: 0 _________________________________________________________________

 

loss도 마찬가지. categorical_crossentropy로 바꾸어준다.

model1.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["acc"])

# generator : 제너레이터를 설정
h1 = model1.fit_generator(generator=train_generator, epochs=30, validation_data=test_generator)

out :

WARNING:tensorflow:From <ipython-input-7-1bb476a2e5e2>:2: Model.fit_generator (from tensorflow.python.keras.engine.training) is deprecated and will be removed in a future version.
Instructions for updating:
Please use Model.fit, which supports generators.
WARNING:tensorflow:sample_weight modes were coerced from
  ...
    to  
  ['...']
WARNING:tensorflow:sample_weight modes were coerced from
  ...
    to  
  ['...']
Train for 150 steps, validate for 75 steps
Epoch 1/30
150/150 [==============================] - 11s 75ms/step - loss: 0.9633 - acc: 0.4940 - val_loss: 0.8663 - val_acc: 0.6153
Epoch 2/30
150/150 [==============================] - 8s 51ms/step - loss: 0.7579 - acc: 0.6187 - val_loss: 0.6845 - val_acc: 0.6680
Epoch 3/30
150/150 [==============================] - 8s 51ms/step - loss: 0.6645 - acc: 0.6757 - val_loss: 0.6437 - val_acc: 0.7060
Epoch 4/30
150/150 [==============================] - 8s 51ms/step - loss: 0.6045 - acc: 0.7153 - val_loss: 0.6338 - val_acc: 0.7127
Epoch 5/30
150/150 [==============================] - 8s 51ms/step - loss: 0.5231 - acc: 0.7633 - val_loss: 0.6068 - val_acc: 0.7247
Epoch 6/30
150/150 [==============================] - 8s 51ms/step - loss: 0.4727 - acc: 0.7883 - val_loss: 0.7154 - val_acc: 0.6720
Epoch 7/30
150/150 [==============================] - 8s 52ms/step - loss: 0.4041 - acc: 0.8280 - val_loss: 0.6553 - val_acc: 0.7500
Epoch 8/30
150/150 [==============================] - 8s 52ms/step - loss: 0.3323 - acc: 0.8630 - val_loss: 0.6299 - val_acc: 0.7460
Epoch 9/30
150/150 [==============================] - 8s 52ms/step - loss: 0.2645 - acc: 0.8943 - val_loss: 0.7078 - val_acc: 0.7653
Epoch 10/30
150/150 [==============================] - 8s 52ms/step - loss: 0.2265 - acc: 0.9123 - val_loss: 0.7124 - val_acc: 0.7480
Epoch 11/30
150/150 [==============================] - 8s 52ms/step - loss: 0.1641 - acc: 0.9337 - val_loss: 0.7786 - val_acc: 0.7567
Epoch 12/30
150/150 [==============================] - 8s 52ms/step - loss: 0.1481 - acc: 0.9433 - val_loss: 0.8521 - val_acc: 0.7453
Epoch 13/30
150/150 [==============================] - 8s 52ms/step - loss: 0.1203 - acc: 0.9553 - val_loss: 0.9640 - val_acc: 0.7680
Epoch 14/30
150/150 [==============================] - 8s 52ms/step - loss: 0.1053 - acc: 0.9630 - val_loss: 0.9989 - val_acc: 0.7107
Epoch 15/30
150/150 [==============================] - 8s 52ms/step - loss: 0.0816 - acc: 0.9717 - val_loss: 1.1470 - val_acc: 0.7427
Epoch 16/30
150/150 [==============================] - 8s 51ms/step - loss: 0.0587 - acc: 0.9783 - val_loss: 1.1178 - val_acc: 0.7487
Epoch 17/30
150/150 [==============================] - 8s 52ms/step - loss: 0.0525 - acc: 0.9787 - val_loss: 1.3174 - val_acc: 0.7480
Epoch 18/30
150/150 [==============================] - 8s 52ms/step - loss: 0.0571 - acc: 0.9800 - val_loss: 1.2667 - val_acc: 0.7487
Epoch 19/30
150/150 [==============================] - 8s 52ms/step - loss: 0.0488 - acc: 0.9783 - val_loss: 1.5896 - val_acc: 0.7373
Epoch 20/30
150/150 [==============================] - 8s 52ms/step - loss: 0.0561 - acc: 0.9817 - val_loss: 1.4970 - val_acc: 0.7280
Epoch 21/30
150/150 [==============================] - 8s 52ms/step - loss: 0.0267 - acc: 0.9923 - val_loss: 1.5404 - val_acc: 0.7487
Epoch 22/30
150/150 [==============================] - 8s 52ms/step - loss: 0.0483 - acc: 0.9857 - val_loss: 1.5844 - val_acc: 0.7633
Epoch 23/30
150/150 [==============================] - 8s 52ms/step - loss: 0.0288 - acc: 0.9920 - val_loss: 1.6528 - val_acc: 0.7573
Epoch 24/30
150/150 [==============================] - 8s 52ms/step - loss: 0.0094 - acc: 0.9977 - val_loss: 1.8498 - val_acc: 0.7553
Epoch 25/30
150/150 [==============================] - 8s 52ms/step - loss: 0.0821 - acc: 0.9760 - val_loss: 1.7805 - val_acc: 0.7307
Epoch 26/30
150/150 [==============================] - 8s 51ms/step - loss: 0.0360 - acc: 0.9880 - val_loss: 1.6722 - val_acc: 0.7527
Epoch 27/30
150/150 [==============================] - 8s 51ms/step - loss: 0.0782 - acc: 0.9770 - val_loss: 1.3163 - val_acc: 0.7320
Epoch 28/30
150/150 [==============================] - 8s 52ms/step - loss: 0.0357 - acc: 0.9883 - val_loss: 1.4310 - val_acc: 0.7627
Epoch 29/30
150/150 [==============================] - 8s 52ms/step - loss: 0.0059 - acc: 0.9983 - val_loss: 1.5812 - val_acc: 0.7540
Epoch 30/30
150/150 [==============================] - 8s 52ms/step - loss: 0.0201 - acc: 0.9957 - val_loss: 1.6541 - val_acc: 0.7320

import matplotlib.pyplot as plt

acc = h1.history["acc"]
val_acc = h1.history["val_acc"]

plt.plot(range(1, len(acc)+1), acc, label="train_acc")
plt.plot(range(1, len(acc)+1), val_acc, label="test_acc")
plt.legend()

out :

 

우리반 사람들도 대개 비슷한 그래프 형태를 보였다.

중간에 아이러니한 그래프가 하나 껴있다.. 원인은 test_generator에서 binary를 categorical로 바꾸지 않은 것이 이유라고 한다.

 

증식은 생략하고 전이학습으로 넘어간다.

from tensorflow.keras.applications import VGG16

# weights : 초기화할 가중치
# include_top = False : 분류기(Dense)를 빼고 Conv레이어만 사용하겠다.
#     - True로 두면 Dense층이 추가된다
# input_shape : VGG16에 넣어줄 입력 크기
conv_base = VGG16(weights="imagenet", include_top = False, input_shape=(150, 150, 3))

conv_base.summary()

out :

Model: "vgg16" _________________________________________________________________ Layer (type)                 Output Shape              Param #    ================================================================= input_1 (InputLayer)         [(None, 150, 150, 3)]     0          _________________________________________________________________ block1_conv1 (Conv2D)        (None, 150, 150, 64)      1792       _________________________________________________________________ block1_conv2 (Conv2D)        (None, 150, 150, 64)      36928      _________________________________________________________________ block1_pool (MaxPooling2D)   (None, 75, 75, 64)        0          _________________________________________________________________ block2_conv1 (Conv2D)        (None, 75, 75, 128)       73856      _________________________________________________________________ block2_conv2 (Conv2D)        (None, 75, 75, 128)       147584     _________________________________________________________________ block2_pool (MaxPooling2D)   (None, 37, 37, 128)       0          _________________________________________________________________ block3_conv1 (Conv2D)        (None, 37, 37, 256)       295168     _________________________________________________________________ block3_conv2 (Conv2D)        (None, 37, 37, 256)       590080     _________________________________________________________________ block3_conv3 (Conv2D)        (None, 37, 37, 256)       590080     _________________________________________________________________ block3_pool (MaxPooling2D)   (None, 18, 18, 256)       0          _________________________________________________________________ block4_conv1 (Conv2D)        (None, 18, 18, 512)       1180160    _________________________________________________________________ block4_conv2 (Conv2D)        (None, 18, 18, 512)       2359808    _________________________________________________________________ block4_conv3 (Conv2D)        (None, 18, 18, 512)       2359808    _________________________________________________________________ block4_pool (MaxPooling2D)   (None, 9, 9, 512)         0          _________________________________________________________________ block5_conv1 (Conv2D)        (None, 9, 9, 512)         2359808    _________________________________________________________________ block5_conv2 (Conv2D)        (None, 9, 9, 512)         2359808    _________________________________________________________________ block5_conv3 (Conv2D)        (None, 9, 9, 512)         2359808    _________________________________________________________________ block5_pool (MaxPooling2D)   (None, 4, 4, 512)         0          ================================================================= Total params: 14,714,688 Trainable params: 14,714,688 Non-trainable params: 0 _________________________________________________________________

import os
import os.path
import numpy as np
from tensorflow.keras.preprocessing.image import ImageDataGenerator

base_dir = "./data/dogs_vs_cats_small2_2"

train_dir = os.path.join(base_dir, "train")
test_dir = os.path.join(base_dir, "test")
validation_dir = os.path.join(base_dir, "validation")

dataGen = ImageDataGenerator(rescale=1./255)
batch_size = 20

# VGG16에 개고양이 이미지를 입력하고 특성을 추출하여 특성맵을 받는 함수
# 이미지가 있는 폴더명, 폴더에 있는 이미지의 수
def extract_feature(directory, sample_count):
    # VGG16으로부터 받은 특성맵을 저장할 리스트
    # 4, 4, 512 : VGG16의 출력크기
    features = np.zeros(shape=(sample_count, 4, 4, 512))
    # 각 이미지의 라벨을 저장하기 위한 리스트
    labels = np.zeros(shape=(sample_count))
    
    # VGG16으로 이미지를 전달하기 위한 제너레이터를 설정
    generator = dataGen.flow_from_directory(directory,
                                           target_size=(150, 150),
                                           batch_size=batch_size,
                                           class_mode="sparse")
    
    # VGG16호출 횟수를 카운트
    i = 0
    # 폴더에 있는 전체 이미지를 VGG16으로 batch_size 개수만큼씩 전달
    # generator로부터 변환된 이미지와 라벨을 20개씩 받는다
    for inputs_batch, labels_batch in generator:
        # VGG16에 이미지(input_batch)입력으로 주어 특성맵을 받아온다
        features_batch = conv_base.predict(inputs_batch)
        # features리스트에 20개씩 특성맵을 추가
        features[i*batch_size : (i+1)*batch_size] = features_batch
        labels[i*batch_size : (i+1)*batch_size] = labels_batch
        
        i = i+1
        
        # 폴더에 있는 이미지수(sample_count)를 (sample_count)만큼 다 처리했다면
        if i*batch_size >= sample_count:
            braek
        return features, labels

# trian, validation 폴더에 있는 이미지를 VGG16에 전송해서 특성맵을 추출
train_features, train_labels = extract_feature(train_dir, 3000)
validation_features, validation_labels = extract_feature(validation_dir, 1500)

out :

Found 3000 images belonging to 3 classes.

Found 1500 images belonging to 3 classes.

# VGG16에서 넘어온 특성맵이 3차원 데이터이므로 1차원으로 변환
train_features = np.reshape(train_features, (3000, 4*4*512))
validation_features = np.reshape(validation_features, (1500, 4*4*512))

train_features.shape, validation_features.shape

out : ((3000, 8192), (1500, 8192))

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout

model4 = Sequential()

model4.add(Dense(256, input_dim=train_features.shape[1], activation="relu"))
model4.add(Dropout(0.5))
model4.add(Dense(3, activation="softmax"))

model4.summary()

out :

Model: "sequential_1" _________________________________________________________________ Layer (type)                 Output Shape              Param #    ================================================================= dense_2 (Dense)              (None, 256)               2097408    _________________________________________________________________ dropout_1 (Dropout)          (None, 256)               0          _________________________________________________________________ dense_3 (Dense)              (None, 3)                 771        ================================================================= Total params: 2,098,179 Trainable params: 2,098,179 Non-trainable params: 0 _________________________________________________________________

model4.compile(loss="sparse_categorical_crossentropy", optimizer="RMSprop", metrics=["acc"])

h4 = model4.fit(train_features, train_labels, epochs=30, batch_size=20, validation_data=(validation_features, validation_labels))

out :

Train on 3000 samples, validate on 1500 samples Epoch 1/30 3000/3000 [==============================] - 1s 283us/sample - loss: 0.5582 - acc: 0.9957 - val_loss: 0.1333 - val_acc: 0.9907 Epoch 2/30 3000/3000 [==============================] - 1s 178us/sample - loss: 0.0497 - acc: 0.9970 - val_loss: 0.0369 - val_acc: 0.9907 Epoch 3/30 3000/3000 [==============================] - 1s 179us/sample - loss: 0.0176 - acc: 0.9980 - val_loss: 0.0269 - val_acc: 0.9927 Epoch 4/30 3000/3000 [==============================] - 1s 183us/sample - loss: 0.0159 - acc: 0.9980 - val_loss: 0.0275 - val_acc: 0.9907 Epoch 5/30 3000/3000 [==============================] - 1s 196us/sample - loss: 0.0316 - acc: 0.9967 - val_loss: 0.0481 - val_acc: 0.9900 Epoch 6/30 3000/3000 [==============================] - 1s 201us/sample - loss: 0.0098 - acc: 0.9983 - val_loss: 0.0098 - val_acc: 0.9967 Epoch 7/30 3000/3000 [==============================] - 1s 196us/sample - loss: 0.0119 - acc: 0.9983 - val_loss: 0.0241 - val_acc: 0.9960 Epoch 8/30 3000/3000 [==============================] - 1s 183us/sample - loss: 0.0126 - acc: 0.9977 - val_loss: 0.0361 - val_acc: 0.9927 Epoch 9/30 3000/3000 [==============================] - 1s 180us/sample - loss: 0.0040 - acc: 0.9983 - val_loss: 0.0087 - val_acc: 0.9960 Epoch 10/30 3000/3000 [==============================] - 1s 194us/sample - loss: 0.0050 - acc: 0.9987 - val_loss: 0.0216 - val_acc: 0.9953 Epoch 11/30 3000/3000 [==============================] - 1s 176us/sample - loss: 0.0013 - acc: 0.9993 - val_loss: 0.0232 - val_acc: 0.9953 Epoch 12/30 3000/3000 [==============================] - 1s 181us/sample - loss: 1.3413e-05 - acc: 1.0000 - val_loss: 0.0377 - val_acc: 0.9953 Epoch 13/30 3000/3000 [==============================] - 1s 214us/sample - loss: 0.0066 - acc: 0.9990 - val_loss: 0.0095 - val_acc: 0.9967 Epoch 14/30 3000/3000 [==============================] - 1s 214us/sample - loss: 2.8784e-06 - acc: 1.0000 - val_loss: 0.0166 - val_acc: 0.9947 Epoch 15/30 3000/3000 [==============================] - 1s 205us/sample - loss: 2.6414e-04 - acc: 0.9997 - val_loss: 0.0459 - val_acc: 0.9953 Epoch 16/30 3000/3000 [==============================] - 1s 192us/sample - loss: 0.0017 - acc: 0.9997 - val_loss: 0.0230 - val_acc: 0.9967 Epoch 17/30 3000/3000 [==============================] - 1s 203us/sample - loss: 0.0044 - acc: 0.9993 - val_loss: 0.0207 - val_acc: 0.9973 Epoch 18/30 3000/3000 [==============================] - 1s 208us/sample - loss: 1.4602e-05 - acc: 1.0000 - val_loss: 0.0135 - val_acc: 0.9967 Epoch 19/30 3000/3000 [==============================] - 1s 186us/sample - loss: 1.0828e-04 - acc: 1.0000 - val_loss: 0.0232 - val_acc: 0.9940 Epoch 20/30 3000/3000 [==============================] - 1s 189us/sample - loss: 1.6643e-07 - acc: 1.0000 - val_loss: 0.0181 - val_acc: 0.9953 Epoch 21/30 3000/3000 [==============================] - 1s 186us/sample - loss: 0.0016 - acc: 0.9997 - val_loss: 0.0448 - val_acc: 0.9940 Epoch 22/30 3000/3000 [==============================] - 1s 169us/sample - loss: 1.8851e-05 - acc: 1.0000 - val_loss: 0.0186 - val_acc: 0.9947 Epoch 23/30 3000/3000 [==============================] - 1s 170us/sample - loss: 0.0050 - acc: 0.9997 - val_loss: 0.0275 - val_acc: 0.9940 Epoch 24/30 3000/3000 [==============================] - 1s 170us/sample - loss: 1.3818e-06 - acc: 1.0000 - val_loss: 0.0740 - val_acc: 0.9927 Epoch 25/30 3000/3000 [==============================] - 1s 170us/sample - loss: 8.2917e-05 - acc: 1.0000 - val_loss: 0.0305 - val_acc: 0.9960 Epoch 26/30 3000/3000 [==============================] - 1s 191us/sample - loss: 0.0026 - acc: 0.9997 - val_loss: 0.0244 - val_acc: 0.9953 Epoch 27/30 3000/3000 [==============================] - 1s 177us/sample - loss: 3.7972e-04 - acc: 0.9997 - val_loss: 0.0326 - val_acc: 0.9940 Epoch 28/30 3000/3000 [==============================] - 1s 190us/sample - loss: 0.0029 - acc: 0.9997 - val_loss: 0.0204 - val_acc: 0.9960 Epoch 29/30 3000/3000 [==============================] - 1s 167us/sample - loss: 3.1072e-06 - acc: 1.0000 - val_loss: 0.0206 - val_acc: 0.9967 Epoch 30/30 3000/3000 [==============================] - 1s 173us/sample - loss: 3.9011e-07 - acc: 1.0000 - val_loss: 0.0161 - val_acc: 0.9967

import matplotlib.pyplot as plt

acc = h4.history["acc"]
val_acc = h4.history["val_acc"]

plt.plot(range(1, len(acc)+1), acc, label="train_acc")
plt.plot(range(1, len(acc)+1), val_acc, label="test_acc")
plt.legend()

out :

그래프 성능이 훨씬 좋아진 것을 확인할 수 있다.

 

 

---

패션이미지 분류

 

패션이미지 분류 학습을 위해 학습예제를 가져온다.

from tensorflow.keras.datasets import fashion_mnist

# 28x28 크기의 10가지 패션 흑백이미지 60,000개 훈련, 10,000개 테스트 데이터
(X_train, y_train), (X_test, y_test) = fashion_mnist.load_data()

X_train.shape, X_test.shape, y_train.shape, y_test.shape

out : ((60000, 28, 28), (10000, 28, 28), (60000,), (10000,))

import matplotlib.pyplot as plt

plt.imshow(X_train[0], cmap="gray")

out :

 

훈련데이터와 테스트데이터로 적당한 갯수인 1000개, 300개만 커팅해서 가져온다.

# 훈련 데이터 1000개, 테스트 데이터 300개만 자름
X_train = X_train[:1000, :]
y_train = y_train[:1000]

X_test = X_test[:300, :]
y_test = y_test[:300]

X_train.shape, X_test.shape, y_train.shape, y_test.shape

out : ((1000, 28, 28), (300, 28, 28), (1000,), (300,))

# y의 클래스가 몇 개인지 확인
import pandas as pd

pd.Series(y_train).unique()

out : array([9, 0, 3, 2, 7, 5, 1, 6, 4, 8], dtype=uint8)

# 원핫 인코딩
from tensorflow.keras.utils import to_categorical

y_train = to_categorical(y_train)
y_test = to_categorical(y_test)

y_train.shape, y_test.shape

out : ((1000, 10), (300, 10))

 

28x28 이미지를 244x244 이미지로 크기를 변경

from PIL import Image
import numpy as np

train_resize_list = []
for i in range(len(X_train)):
    # 배열로 되어있는 이미지를 이미지타입으로 바꿔서 크기를 변경
    train_resize_image = Image.fromarray(X_train[i]).resize((244,244))
    # 다시 학습데이터로 사용하기 위해 추가 (그냥은 들어가지 않으니 numpy배열로 변경)
    train_resize_list.append(np.array(train_resize_image))
    
test_resize_list = []
for i in range(len(X_test)):
    # 배열로 되어있는 이미지를 이미지타입으로 바꿔서 크기를 변경
    test_resize_image = Image.fromarray(X_test[i]).resize((244,244))
    # 다시 학습데이터로 사용하기 위해 추가 (그냥은 들어가지 않으니 numpy배열로 변경)
    test_resize_list.append(np.array(test_resize_image))
    
# list를 배열로 변경
X_train_resized = np.array(train_resize_list)
X_test_resized = np.array(test_resize_list)

X_train_resized.shape, X_test_resized.shape

out : ((1000, 244, 244), (300, 244, 244))

 

이미지 사이즈를 변경해주었으니 출력해서 비교해본다. 먼저 리사이징 전 이미지는

plt.imshow(X_train[0], cmap="gray")

픽셀이 선명하게 보인다. 리사이징해준 이미지를 확인해보면

plt.imshow(X_train_resized[0], cmap="gray")

# 색상차원 추가
X_train_resized = X_train_resized.reshape(X_train_resized.shape[0], 244, 244, 1)
X_test_resized = X_test_resized.reshape(X_test_resized.shape[0], 244, 244, 1)

X_train_resized.shape, X_test_resized.shape

out : ((1000, 244, 244, 1), (300, 244, 244, 1))

# 훈련, 테스트 데이터로 분리
from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(X_train_resized, y_train, random_state=777)

X_train.shape, X_test.shape, y_train.shape, y_test.shape

out : ((750, 244, 244, 1), (250, 244, 244, 1), (750, 10), (250, 10))

 

신경망 학습

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout
from tensorflow.keras.layers import Flatten, Conv2D, MaxPooling2D

model1 = Sequential()

# filters와 kernel_size도 units처럼 생략하고 숫자만 작성해도 ok.
model1 = Sequential()
model1.add(Conv2D(32, (3,3), input_shape=(244, 244, 1), 
                  activation="relu"))
model1.add(MaxPooling2D((2,2)))
model1.add(Conv2D(128, (3,3), activation="relu"))
model1.add(MaxPooling2D((2,2)))
model1.add(Conv2D(256, (3,3), activation="relu"))
model1.add(MaxPooling2D((2,2)))
model1.add(Conv2D(512, (3,3), activation="relu"))
model1.add(MaxPooling2D((2,2)))

model1.add(Flatten())
model1.add(Dense(512, activation='relu'))
model1.add(Dropout(0.5))
model1.add(Dense(10, activation='softmax'))

model1.summary()

out :

Model: "sequential_1" _________________________________________________________________ Layer (type)                 Output Shape              Param #    ================================================================= conv2d (Conv2D)              (None, 242, 242, 32)      320        _________________________________________________________________ max_pooling2d (MaxPooling2D) (None, 121, 121, 32)      0          _________________________________________________________________ conv2d_1 (Conv2D)            (None, 119, 119, 128)     36992      _________________________________________________________________ max_pooling2d_1 (MaxPooling2 (None, 59, 59, 128)       0          _________________________________________________________________ conv2d_2 (Conv2D)            (None, 57, 57, 256)       295168     _________________________________________________________________ max_pooling2d_2 (MaxPooling2 (None, 28, 28, 256)       0          _________________________________________________________________ conv2d_3 (Conv2D)            (None, 26, 26, 512)       1180160    _________________________________________________________________ max_pooling2d_3 (MaxPooling2 (None, 13, 13, 512)       0          _________________________________________________________________ flatten (Flatten)            (None, 86528)             0          _________________________________________________________________ dense (Dense)                (None, 512)               44302848   _________________________________________________________________ dropout (Dropout)            (None, 512)               0          _________________________________________________________________ dense_1 (Dense)              (None, 10)                5130       ================================================================= Total params: 45,820,618 Trainable params: 45,820,618 Non-trainable params: 0 _________________________________________________________________

model1.compile(loss="categorical_crossentropy",
               optimizer="adam",
               metrics = ["acc"])

h1 = model1.fit(X_train, y_train, epochs=20, batch_size=20,
           validation_data=(X_test, y_test))

out :

Train on 750 samples, validate on 250 samples Epoch 1/20 750/750 [==============================] - 6s 8ms/sample - loss: 20.3453 - acc: 0.5893 - val_loss: 0.7260 - val_acc: 0.7800 Epoch 2/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.6252 - acc: 0.7760 - val_loss: 0.6346 - val_acc: 0.8160 Epoch 3/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.4563 - acc: 0.8387 - val_loss: 0.6568 - val_acc: 0.8000 Epoch 4/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.3575 - acc: 0.8747 - val_loss: 0.6588 - val_acc: 0.8240 Epoch 5/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.3150 - acc: 0.8973 - val_loss: 0.6725 - val_acc: 0.7920 Epoch 6/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.2192 - acc: 0.9240 - val_loss: 0.7707 - val_acc: 0.8000 Epoch 7/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.2292 - acc: 0.9307 - val_loss: 0.8896 - val_acc: 0.7840 Epoch 8/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.1710 - acc: 0.9333 - val_loss: 0.9272 - val_acc: 0.7920 Epoch 9/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.1500 - acc: 0.9493 - val_loss: 1.0057 - val_acc: 0.8000 Epoch 10/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.1271 - acc: 0.9480 - val_loss: 0.8757 - val_acc: 0.7920 Epoch 11/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.0993 - acc: 0.9613 - val_loss: 1.0343 - val_acc: 0.8160 Epoch 12/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.0939 - acc: 0.9653 - val_loss: 1.0534 - val_acc: 0.8000 Epoch 13/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.0871 - acc: 0.9733 - val_loss: 0.8950 - val_acc: 0.8240 Epoch 14/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.0559 - acc: 0.9787 - val_loss: 1.0458 - val_acc: 0.8240 Epoch 15/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.0639 - acc: 0.9773 - val_loss: 1.2778 - val_acc: 0.8240 Epoch 16/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.0195 - acc: 0.9960 - val_loss: 1.5939 - val_acc: 0.8160 Epoch 17/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.0377 - acc: 0.9840 - val_loss: 1.4754 - val_acc: 0.8120 Epoch 18/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.0292 - acc: 0.9880 - val_loss: 1.4117 - val_acc: 0.8160 Epoch 19/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.0589 - acc: 0.9760 - val_loss: 1.5752 - val_acc: 0.8000 Epoch 20/20 750/750 [==============================] - 2s 3ms/sample - loss: 0.0812 - acc: 0.9747 - val_loss: 1.3597 - val_acc: 0.7920

 

import matplotlib.pyplot as plt

acc = h1.history["acc"]
val_acc = h1.history["val_acc"]

plt.plot(range(1, len(acc)+1), acc, label="train")
plt.plot(range(1, len(acc)+1), val_acc, label="train")
plt.legend()

out :

 

Xception 모델을 이용한 전이학습

from tensorflow.keras.datasets import fashion_mnist

# 28x28 크기의 10가지 패션 흑백이미지 60000개 훈련, 10000개 테스트 데이터
(X_train, y_train), (X_test, y_test) = fashion_mnist.load_data()

# 훈련 데이터 1000개, 테스트데이터 300개만 자름
X_train = X_train[:1000, :]
y_train = y_train[:1000,]

X_test = X_test[:300, :]
y_test = y_test[:300,]

# 원핫 인코딩
from tensorflow.keras.utils import to_categorical

y_train = to_categorical(y_train)
y_test = to_categorical(y_test)

y_train.shape, y_test.shape

out : (1000, 10), (300, 10))

from PIL import Image
import numpy as np

train_resize_list = []
for i in range(len(X_train)) :
    # 배열된 이미지를 이미지타입으로 변경해서 크기를 변경
    train_resize_image = Image.fromarray(X_train[i]).resize((244, 244))
    # 다시 학습데이터로 쓰기위해 배열로 변경
    train_resize_list.append(np.array(train_resize_image))
    
test_resize_list = []
for i in range(len(X_test)) :
    # 배열된 이미지를 이미지타입으로 변경해서 크기를 변경
    test_resize_image = Image.fromarray(X_test[i]).resize((244, 244))
    # 다시 학습데이터로 쓰기위해 배열로 변경
    test_resize_list.append(np.array(test_resize_image))
    

# list를 배열로 변경
X_train_resized = np.array(train_resize_list)
X_test_resized = np.array(test_resize_list)

X_train_resized.shape, X_test_resized.shape

out : ((1000, 244, 244), (300, 244, 244))

# 색상차원 추가
X_train_resized = X_train_resized.reshape(X_train_resized.shape[0], 
                                          244, 244, 1)
X_test_resized = X_test_resized.reshape(X_test_resized.shape[0], 
                                        244, 244, 1)

X_train_resized.shape, X_test_resized.shape

out : ((1000, 244, 244, 1), (300, 244, 244, 1))

# Xception에 입력으로 넣기 위해서 컬러이미지로 변경
import numpy as np

X_train_resized = np.repeat(X_train_resized, 3, axis=3) # 같은 것을 3개로 복사
X_test_resized = np.repeat(X_test_resized, 3, axis=3)

X_train_resized.shape, X_test_resized.shape

out : ((1000, 244, 244, 3), (300, 244, 244, 3))

위의 결과는 이 그림처럼 똑같은 것을 3장을 쌓은 형태이다.

from tensorflow.keras.applications import Xception

base_conv = Xception(weights="imagenet", include_top=False, input_shape=(244,244,3))

# 동결
base_conv.trainable = False


from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout, Conv2D, MaxPooling2D, Flatten

model3 = Sequential()

model3.add(base_conv)
model3.add(Flatten())

model3.add(Dense(512, activation="relu"))
model3.add(Dense(10, activation="softmax"))

model3.summary()

Model: "sequential_2" _________________________________________________________________ Layer (type)                 Output Shape              Param #    ================================================================= xception (Model)             (None, 8, 8, 2048)        20861480   _________________________________________________________________ flatten_1 (Flatten)          (None, 131072)            0          _________________________________________________________________ dense_2 (Dense)              (None, 512)               67109376   _________________________________________________________________ dense_3 (Dense)              (None, 10)                5130       ================================================================= Total params: 87,975,986 Trainable params: 67,114,506 Non-trainable params: 20,861,480 _________________________________________________________________

model3.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["acc"])
model3.fit(X_train, y_train, epochs=20, batch_size=20, validation_data=(X_test, y_test))

오류 발생.. 눙물이 굴러간다. 또르륵. 또르르륵.