Dog & Cat Classification Exercise #01

26 minute read

CNN Exercise #01 - 개와 고양이 사진 분류하기

Exercise #00에서는 Dropout 과 Data Augmentation으로 정확도를 올리는 방법을 사용해 보았습니다.
이번에는 다른 사람이 Train 시켜 놓은 훌륭한 Pre-Trained Model을 이용하여 정확도를 더 끌어올려 보겠습니다.

앞 Exercise와 마찬가지로 이번에도 Keras를 사용하도록 하겠습니다.

import keras
keras.__version__

Using TensorFlow backend.
'2.2.4'

from IPython.core.interactiveshell import InteractiveShell 
InteractiveShell.ast_node_interactivity = "all"

0. Load Pre-Trained Model

이번 Exercise에서 사용할 Pre-Trained Model은 VGG16입니다.
VGG16은 구조가 간단하고 ImageNet 데이터셋에 널리 사용되는 Conv. Net입니다.
VGG16은 조금 오래되었고 최고 수준의 성능에는 못미치며 최근의 다른 모델보다는 조금 무겁습니다.
하지만 이 모델의 구조가 이전에 보았던 것과 비슷해서 새로운 개념을 도입하지 않고 이해하기 쉽기 때문에 선택하였습니다.
CNN의 전반적인 설명과 다른 Pre-Trained Model에 대해서 알아보시려면 아래 Link의 Post를 참고하시기 바랍니다.
- https://moonlight314.github.io/study/cnn/CNN/

VGG16 Model은 아래와 같이 구성되어 있습니다.

Keras에서 VGG Model을 Load하는 것은 매우 쉽습니다.
아래와 같은 Code를 이용해서 VGG16 Model을 Load할 수 있습니다.
include_top=False은 최종 Layer인 Softmax Layer를 빼고 사용하겠다는 의미입니다.
다른 Pre-Trained Model도 비슷하지만, 마지막 Softmax Layer는 그 앞 Layer들이 추출한 Image들의 Feature를 단순히 분류하는 기능입니다.
앞 Layer들이 ImageNet의 Data들의 Feature들을 Extract하였다면, 마지막 Softmax Layer는 ImageNet을 분류하는 Classifier가 됩니다.
이는 VGG16 Model을 우리 Model에서 Feature만 Extract하고, 이미지의 분류는 우리가 정의한 Classifier를 사용하겠다는 의미입니다.

from keras.applications import VGG16

conv_base = VGG16(weights='imagenet',
                  include_top=False,
                  input_shape=(150, 150, 3))

WARNING:tensorflow:From C:\Users\Moon\Anaconda3\lib\site-packages\keras\backend\tensorflow_backend.py:74: The name tf.get_default_graph is deprecated. Please use tf.compat.v1.get_default_graph instead.

WARNING:tensorflow:From C:\Users\Moon\Anaconda3\lib\site-packages\keras\backend\tensorflow_backend.py:517: The name tf.placeholder is deprecated. Please use tf.compat.v1.placeholder instead.

WARNING:tensorflow:From C:\Users\Moon\Anaconda3\lib\site-packages\keras\backend\tensorflow_backend.py:4138: The name tf.random_uniform is deprecated. Please use tf.random.uniform instead.

WARNING:tensorflow:From C:\Users\Moon\Anaconda3\lib\site-packages\keras\backend\tensorflow_backend.py:3976: The name tf.nn.max_pool is deprecated. Please use tf.nn.max_pool2d instead.

WARNING:tensorflow:From C:\Users\Moon\Anaconda3\lib\site-packages\keras\backend\tensorflow_backend.py:174: The name tf.get_default_session is deprecated. Please use tf.compat.v1.get_default_session instead.

WARNING:tensorflow:From C:\Users\Moon\Anaconda3\lib\site-packages\keras\backend\tensorflow_backend.py:181: The name tf.ConfigProto is deprecated. Please use tf.compat.v1.ConfigProto instead.

VGG16 Model의 구조와 Parameter수는 아래와 같습니다.

conv_base.summary()

_________________________________________________________________
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
_________________________________________________________________

1. Train Model

사전 훈련된 네트워크를 사용하는 크게 두 가지 방법이 있습니다.
- 특성 추출 ( Feature Extract )
  - 이 방법은 앞서 말씀드린 Pre-Trained Model을 Feature Extractor로 사용하는 방법입니다.
  - 첫번째 사용 방법은 기존 Model에서 마지막 Softmax Layer만 없애고, 새롭게 Classifier를 추가하는 방법입니다.
  - 최종 Softmax Layer만 새롭게 학습하기 때문에 빠른 시간내에 적용할 수 있습니다.
  - 두번째 사용 방법은 Data Augmentation을 이용하여 Ovetfitting을 피하는 방법입니다.
- 미세 조정 ( Fine Tune )
  - Pre-Trained Model의 마지막 Softmax Layer만 따로 학습하는 것이 아니라, Softmax Layer를 포함한 그 직전의 Conv. Layer도 학습하는 방법입니다.
  - 주어진 문제에 조금 더 밀접하게 재사용 모델의 표현을 일부 조정하기 때문에 미세 조정이라고 부른다.

1.1. 방법 #01-01

기존 VGG Model에서 Top(Softmax)만 제외합니다. 이를 conv_base라고 부르겠습니다.
Train Data(2000개)를 conv_base Model에 넣어서 Feature들을 Extract 합니다.
conv_base에 Train Data를 Predict하여 Feature들을 모읍니다. ( 2000 x 4 x 4 x 512 )
전체 Data는 2000개 이고, VGG16의 최종 Output이 4 x 4 x 512 입니다.
이것을 (2000 x 8192 )로 Reshape한 후에, FC에 연결하여 Binary Classification 합니다.

Feature Extract

Feature Extract은 Pre-Trained Model을 사용해 새로운 Dataset에서 Feature을 뽑아내는 것을 말합니다.
뽑아낸 Feature을 사용하여 New Classifier를 Train 시킵니다.
Conv.Net은 Image Classification를 위해 크게 두 부분으로 구성됩니다.
연속된 Conv. Layer와 Pooling Layer로 시작해서 완전 연결 분류기로 구성됩니다.
첫 번째 부분을 모델의 합성곱 기반층(convolutional base)이라고 부르겠습니다.
Conv.Net의 경우 Fearue Extract는 사전에 훈련된 Model의 합성곱 기반층을 선택해 새로운 데이터를 통과시키고 그 출력으로 새로운 분류기를 훈련합니다.

Classifier만 재학습하는 이유는 분류기에서 학습한 표현은 모델이 훈련된 클래스 집합(VGG16의 경우는 ImageNet Dataset)에 특화되어 있기 때문에 새로운 Dataset에 크게 쓸모가 없기 때문입니다.
특정 합성곱 층에서 추출한 표현의 일반성(그리고 재사용성)의 수준은 모델에 있는 층의 깊이에 달려 있습니다.
모델의 하위층(Input Layer에 가까운 쪽)은 (에지, 색깔, 질감 등과 같이) 지역적이고 매우 일반적인 특성 맵을 추출합니다.
반면 상위 층은 (‘강아지 눈’이나 ‘고양이 귀’와 같이) 좀 더 추상적인 개념을 추출하는 특성이 있습니다.
만약 새로운 Dataset이 원본 모델이 훈련한 데이터셋과 많이 다르다면 전체 합성곱 기반층을 사용하는 것보다는 모델의 하위 층 몇 개만 특성 추출에 사용하는 것이 좋습니다.

이런저런 일반적인 설정을 합니다.
Folder Name을 설정하고, Feature Extract 할 Data Generator도 만듭니다.

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

base_dir = './dataset'

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

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

Train Data 2000장의 Feature Extract합니다.
Softmax Layer를 제외한 Conv. Base Model로 Predict를 하면 해당 Image에 대한 Feature가 추출되겠죠 ?
ImageDataGenerator의 사용법을 유심히 봐 둘 필요가 있습니다.
ImageDataGenerator가 Return 해주는 Data와 그 Data의 Format도 봐주시기 바랍니다.

def extract_features(directory, sample_count):
    
    features = np.zeros(shape=(sample_count, 4, 4, 512))
    labels = np.zeros(shape=(sample_count))
    
    generator = datagen.flow_from_directory(
        directory,
        target_size=(150, 150),
        batch_size=batch_size,
        class_mode='binary')

    i = 0
    
    for inputs_batch, labels_batch in generator:
        
        #print( labels_batch.shape , labels_batch[0] )
        # label은 (20,)의 1차원 ndarrdy
        
        #print( inputs_batch.shape , inputs_batch[0] )
        # inputs_batch은 (20,150,150,3) ndarrdy
        
        features_batch = conv_base.predict(inputs_batch)
        features[i * batch_size : (i + 1) * batch_size] = features_batch
        labels[i * batch_size : (i + 1) * batch_size] = labels_batch
        
        i += 1        
        if i * batch_size >= sample_count:
            # 제너레이터는 루프 안에서 무한하게 데이터를 만들어내므로 모든 이미지를 한 번씩 처리하고 나면 중지합니다
            break
    
    return features, labels

train_features, train_labels = extract_features(train_dir, 2000)

Found 2000 images belonging to 2 classes.

extract_features 가 Return한 Data의 형태를 보도록 하겠습니다.
features, labels을 Return하고 있고, features는 (2000, 4, 4, 512) , labels은 (2000,)의 형태를 갖습니다.
(4, 4, 512)는 Conv. Base의 출력이며, 각 Image의 Feature가 되겠고, 전체 Image 갯수는 2000개 이기 때문에 (2000, 4, 4, 512) Shape이 됩니다.
labels는 각 Image의 Label( Dog or Cat)을 0,1로 표현되어 있습니다.

type( train_features )
train_features.shape
train_labels.shape
train_labels[0:10]

numpy.ndarray
(2000, 4, 4, 512)
(2000,)
array([0., 1., 0., 0., 1., 1., 0., 0., 0., 1.])

Validation / Test Data도 위와 동일하게 Feature Extract를 진행하도록 하겠습니다.

validation_features, validation_labels = extract_features(validation_dir, 1000)

Found 1000 images belonging to 2 classes.

test_features, test_labels = extract_features(test_dir, 1000)

Found 1000 images belonging to 2 classes.

자~! 이제 Train Data가 모두 준비되었습니다.
Dog & Cat Classifier를 Train 시키기 위해서 Feature Array를 Reshape하도록 하겠습니다.
Feature를 1차원 Array로 쭉 편다고 생각하시면 됩니다.

train_features = np.reshape(train_features, (2000, 4 * 4 * 512))
validation_features = np.reshape(validation_features, (1000, 4 * 4 * 512))
test_features = np.reshape(test_features, (1000, 4 * 4 * 512))

적당한 Dense Layer를 만들어서 Train을 시작해 보겠습니다.
최종 Classifier만 학습하기 때문에 시간이 많이 걸리지 않습니다.

from keras import models
from keras import layers
from keras import optimizers

model = models.Sequential()
model.add(layers.Dense(256, activation='relu', input_dim=4 * 4 * 512))
model.add(layers.Dropout(0.5))
model.add(layers.Dense(1, activation='sigmoid'))

model.compile(optimizer=optimizers.RMSprop(lr=2e-5),
              loss='binary_crossentropy',
              metrics=['acc'])

history = model.fit(train_features, train_labels,
                    epochs=30,
                    batch_size=20,
                    validation_data=(validation_features, validation_labels))

WARNING:tensorflow:From C:\Users\Moon\Anaconda3\lib\site-packages\keras\backend\tensorflow_backend.py:3445: calling dropout (from tensorflow.python.ops.nn_ops) with keep_prob is deprecated and will be removed in a future version.
Instructions for updating:
Please use `rate` instead of `keep_prob`. Rate should be set to `rate = 1 - keep_prob`.
WARNING:tensorflow:From C:\Users\Moon\Anaconda3\lib\site-packages\keras\optimizers.py:790: The name tf.train.Optimizer is deprecated. Please use tf.compat.v1.train.Optimizer instead.

WARNING:tensorflow:From C:\Users\Moon\Anaconda3\lib\site-packages\tensorflow\python\ops\nn_impl.py:180: add_dispatch_support.<locals>.wrapper (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.
Instructions for updating:
Use tf.where in 2.0, which has the same broadcast rule as np.where
Train on 2000 samples, validate on 1000 samples
Epoch 1/30
2000/2000 [==============================] - 1s 723us/step - loss: 0.6165 - acc: 0.6515 - val_loss: 0.4434 - val_acc: 0.8260
Epoch 2/30
2000/2000 [==============================] - 1s 280us/step - loss: 0.4296 - acc: 0.8090 - val_loss: 0.3600 - val_acc: 0.8590
Epoch 3/30
2000/2000 [==============================] - 1s 276us/step - loss: 0.3520 - acc: 0.8420 - val_loss: 0.3238 - val_acc: 0.8720
Epoch 4/30
2000/2000 [==============================] - 1s 276us/step - loss: 0.3103 - acc: 0.8710 - val_loss: 0.3069 - val_acc: 0.8760
Epoch 5/30
2000/2000 [==============================] - 1s 283us/step - loss: 0.2899 - acc: 0.8870 - val_loss: 0.2848 - val_acc: 0.8920
Epoch 6/30
2000/2000 [==============================] - 1s 282us/step - loss: 0.2677 - acc: 0.8930 - val_loss: 0.2721 - val_acc: 0.8970
Epoch 7/30
2000/2000 [==============================] - 1s 277us/step - loss: 0.2470 - acc: 0.9025 - val_loss: 0.2645 - val_acc: 0.8960
Epoch 8/30
2000/2000 [==============================] - 1s 280us/step - loss: 0.2328 - acc: 0.9065 - val_loss: 0.2608 - val_acc: 0.8940
Epoch 9/30
2000/2000 [==============================] - 1s 280us/step - loss: 0.2253 - acc: 0.9185 - val_loss: 0.2520 - val_acc: 0.8970
Epoch 10/30
2000/2000 [==============================] - 1s 269us/step - loss: 0.2114 - acc: 0.9235 - val_loss: 0.2732 - val_acc: 0.8770
Epoch 11/30
2000/2000 [==============================] - 1s 280us/step - loss: 0.1966 - acc: 0.9290 - val_loss: 0.2469 - val_acc: 0.9020
Epoch 12/30
2000/2000 [==============================] - 1s 290us/step - loss: 0.1881 - acc: 0.9370 - val_loss: 0.2551 - val_acc: 0.8990
Epoch 13/30
2000/2000 [==============================] - 1s 274us/step - loss: 0.1792 - acc: 0.9410 - val_loss: 0.2534 - val_acc: 0.8960
Epoch 14/30
2000/2000 [==============================] - 1s 271us/step - loss: 0.1778 - acc: 0.9385 - val_loss: 0.2378 - val_acc: 0.9060
Epoch 15/30
2000/2000 [==============================] - 1s 278us/step - loss: 0.1672 - acc: 0.9410 - val_loss: 0.2442 - val_acc: 0.9040
Epoch 16/30
2000/2000 [==============================] - 1s 277us/step - loss: 0.1584 - acc: 0.9470 - val_loss: 0.2364 - val_acc: 0.9080
Epoch 17/30
2000/2000 [==============================] - 1s 276us/step - loss: 0.1494 - acc: 0.9455 - val_loss: 0.2349 - val_acc: 0.9080
Epoch 18/30
2000/2000 [==============================] - 1s 280us/step - loss: 0.1448 - acc: 0.9490 - val_loss: 0.2338 - val_acc: 0.9080
Epoch 19/30
2000/2000 [==============================] - 1s 276us/step - loss: 0.1401 - acc: 0.9525 - val_loss: 0.2347 - val_acc: 0.9060
Epoch 20/30
2000/2000 [==============================] - 1s 290us/step - loss: 0.1299 - acc: 0.9575 - val_loss: 0.2357 - val_acc: 0.9050
Epoch 21/30
2000/2000 [==============================] - 1s 278us/step - loss: 0.1241 - acc: 0.9590 - val_loss: 0.2404 - val_acc: 0.9040
Epoch 22/30
2000/2000 [==============================] - 1s 279us/step - loss: 0.1214 - acc: 0.9585 - val_loss: 0.2359 - val_acc: 0.9060
Epoch 23/30
2000/2000 [==============================] - 1s 290us/step - loss: 0.1166 - acc: 0.9625 - val_loss: 0.2371 - val_acc: 0.9050
Epoch 24/30
2000/2000 [==============================] - 1s 290us/step - loss: 0.1122 - acc: 0.9650 - val_loss: 0.2352 - val_acc: 0.9040
Epoch 25/30
2000/2000 [==============================] - 1s 286us/step - loss: 0.1052 - acc: 0.9675 - val_loss: 0.2450 - val_acc: 0.9030
Epoch 26/30
2000/2000 [==============================] - 1s 277us/step - loss: 0.1029 - acc: 0.9700 - val_loss: 0.2380 - val_acc: 0.9070
Epoch 27/30
2000/2000 [==============================] - 1s 284us/step - loss: 0.0988 - acc: 0.9675 - val_loss: 0.2487 - val_acc: 0.9030
Epoch 28/30
2000/2000 [==============================] - 1s 279us/step - loss: 0.0966 - acc: 0.9640 - val_loss: 0.2370 - val_acc: 0.9060
Epoch 29/30
2000/2000 [==============================] - 1s 276us/step - loss: 0.0899 - acc: 0.9690 - val_loss: 0.2378 - val_acc: 0.9030
Epoch 30/30
2000/2000 [==============================] - 1s 279us/step - loss: 0.0881 - acc: 0.9700 - val_loss: 0.2424 - val_acc: 0.9070

학습이 완료되었으니 결과를 한 번 보도록 하죠.

import matplotlib.pyplot as plt

acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']

epochs = range(len(acc))

plt.plot(epochs, acc, 'bo', label='Training acc')
plt.plot(epochs, val_acc, 'b', label='Validation acc')
plt.title('Training and validation accuracy')
plt.legend()

plt.figure()

plt.plot(epochs, loss, 'bo', label='Training loss')
plt.plot(epochs, val_loss, 'b', label='Validation loss')
plt.title('Training and validation loss')
plt.legend()

plt.show()

90% 정도의 검증 정확도에 도달했으나 많은 비율로 드롭아웃을 사용했음에도 불구하고 훈련이 시작하면서 거의 바로 과대적합되고 있습니다.
이를 개선하기 위해서 Data Augmentation을 사용해 보도록 하겠습니다.

1.2. 방법 #01-02

ImageDataGenerator를 사용하여 Data Augmentation을 이용해 보도록 하겠습니다.
전체 Network을 다 실행해야합니다.
방법 #01-01에 Data Augmentation을 사용하면 안될까요 ??
- Data Augmentation을 사용하려면 ImageDataGenerator를 사용해야 하고, ImageDataGenerator를 사용하려면 fit_generator를 호출해야
  합니다.
- fit_generator를 호출하면 전체 Network을 다 Train해야합니다..

새로운 Model을 정의하는데, 기존 Conv. Base에서 Dense Layer를 추가합니다.

from keras import models
from keras import layers

model = models.Sequential()
model.add(conv_base)
model.add(layers.Flatten())
model.add(layers.Dense(256, activation='relu'))
model.add(layers.Dense(1, activation='sigmoid'))

model.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
vgg16 (Model)                (None, 4, 4, 512)         14714688  
_________________________________________________________________
flatten_1 (Flatten)          (None, 8192)              0         
_________________________________________________________________
dense_3 (Dense)              (None, 256)               2097408   
_________________________________________________________________
dense_4 (Dense)              (None, 1)                 257       
=================================================================
Total params: 16,812,353
Trainable params: 16,812,353
Non-trainable params: 0
_________________________________________________________________

print('conv_base를 동결하기 전 훈련되는 가중치의 수:', 
      len(model.trainable_weights))

conv_base를 동결하기 전 훈련되는 가중치의 수: 30

Keras에서는 trainable 속성을 False로 설정하여 네트워크를 동결할 수 있습니다:

conv_base.trainable = False

print('conv_base를 동결한 후 훈련되는 가중치의 수:', 
      len(model.trainable_weights))

conv_base를 동결한 후 훈련되는 가중치의 수: 4

컴파일 단계 후에 trainable 속성을 변경하면 반드시 모델을 다시 컴파일해야 합니다. 그렇지 않으면 변경 사항이 적용되지 않습니다.

from keras.preprocessing.image import ImageDataGenerator

train_datagen = ImageDataGenerator(
      rescale=1./255,
      rotation_range=20,
      width_shift_range=0.1,
      height_shift_range=0.1,
      shear_range=0.1,
      zoom_range=0.1,
      horizontal_flip=True,
      fill_mode='nearest')

# 검증 데이터는 증식되어서는 안 됩니다!
test_datagen = ImageDataGenerator(rescale=1./255)

train_generator = train_datagen.flow_from_directory(
        # 타깃 디렉터리
        train_dir,
        # 모든 이미지의 크기를 150 × 150로 변경합니다
        target_size=(150, 150),
        batch_size=20,
        # binary_crossentropy 손실을 사용하므로 이진 레이블이 필요합니다
        class_mode='binary')

validation_generator = test_datagen.flow_from_directory(
        validation_dir,
        target_size=(150, 150),
        batch_size=20,
        class_mode='binary')

model.compile(loss='binary_crossentropy',
              optimizer=optimizers.RMSprop(lr=2e-5),
              metrics=['acc'])

history = model.fit_generator(
      train_generator,
      steps_per_epoch=100,
      epochs=30,
      validation_data=validation_generator,
      validation_steps=50,
      verbose=2)

Found 2000 images belonging to 2 classes.
Found 1000 images belonging to 2 classes.
Epoch 1/30
 - 14s - loss: 0.5165 - acc: 0.7575 - val_loss: 0.4247 - val_acc: 0.8170
Epoch 2/30
 - 13s - loss: 0.3989 - acc: 0.8325 - val_loss: 0.3414 - val_acc: 0.8620
Epoch 3/30
 - 12s - loss: 0.3424 - acc: 0.8625 - val_loss: 0.3006 - val_acc: 0.8740
Epoch 4/30
 - 13s - loss: 0.3187 - acc: 0.8645 - val_loss: 0.2791 - val_acc: 0.8880
Epoch 5/30
 - 13s - loss: 0.2934 - acc: 0.8775 - val_loss: 0.2657 - val_acc: 0.8940
Epoch 6/30
 - 13s - loss: 0.2779 - acc: 0.8885 - val_loss: 0.2561 - val_acc: 0.8990
Epoch 7/30
 - 12s - loss: 0.2802 - acc: 0.8830 - val_loss: 0.2554 - val_acc: 0.8970
Epoch 8/30
 - 12s - loss: 0.2572 - acc: 0.8970 - val_loss: 0.2518 - val_acc: 0.9000
Epoch 9/30
 - 12s - loss: 0.2544 - acc: 0.8955 - val_loss: 0.2613 - val_acc: 0.8860
Epoch 10/30
 - 12s - loss: 0.2422 - acc: 0.9035 - val_loss: 0.2398 - val_acc: 0.9050
Epoch 11/30
 - 12s - loss: 0.2439 - acc: 0.8935 - val_loss: 0.2417 - val_acc: 0.9040
Epoch 12/30
 - 12s - loss: 0.2401 - acc: 0.9035 - val_loss: 0.2561 - val_acc: 0.9000
Epoch 13/30
 - 12s - loss: 0.2354 - acc: 0.9055 - val_loss: 0.2369 - val_acc: 0.9010
Epoch 14/30
 - 13s - loss: 0.2262 - acc: 0.9125 - val_loss: 0.2358 - val_acc: 0.9080
Epoch 15/30
 - 13s - loss: 0.2311 - acc: 0.9025 - val_loss: 0.2349 - val_acc: 0.9030
Epoch 16/30
 - 12s - loss: 0.2263 - acc: 0.9090 - val_loss: 0.2406 - val_acc: 0.9080
Epoch 17/30
 - 12s - loss: 0.2114 - acc: 0.9215 - val_loss: 0.2346 - val_acc: 0.9080
Epoch 18/30
 - 12s - loss: 0.2054 - acc: 0.9200 - val_loss: 0.2367 - val_acc: 0.9080
Epoch 19/30
 - 12s - loss: 0.2038 - acc: 0.9225 - val_loss: 0.2397 - val_acc: 0.9070
Epoch 20/30
 - 13s - loss: 0.2057 - acc: 0.9190 - val_loss: 0.2425 - val_acc: 0.8990
Epoch 21/30
 - 13s - loss: 0.2065 - acc: 0.9175 - val_loss: 0.2429 - val_acc: 0.9000
Epoch 22/30
 - 12s - loss: 0.2093 - acc: 0.9115 - val_loss: 0.2361 - val_acc: 0.9070
Epoch 23/30
 - 12s - loss: 0.1982 - acc: 0.9195 - val_loss: 0.2333 - val_acc: 0.9070
Epoch 24/30
 - 12s - loss: 0.1968 - acc: 0.9130 - val_loss: 0.2515 - val_acc: 0.8950
Epoch 25/30
 - 12s - loss: 0.1952 - acc: 0.9245 - val_loss: 0.2321 - val_acc: 0.9050
Epoch 26/30
 - 12s - loss: 0.1906 - acc: 0.9250 - val_loss: 0.2694 - val_acc: 0.8890
Epoch 27/30
 - 12s - loss: 0.1899 - acc: 0.9235 - val_loss: 0.2349 - val_acc: 0.9090
Epoch 28/30
 - 12s - loss: 0.1876 - acc: 0.9265 - val_loss: 0.2399 - val_acc: 0.9060
Epoch 29/30
 - 12s - loss: 0.1777 - acc: 0.9310 - val_loss: 0.2390 - val_acc: 0.9040
Epoch 30/30
 - 12s - loss: 0.1824 - acc: 0.9260 - val_loss: 0.2358 - val_acc: 0.9060

Train이 완료되었으면 일단 저장해 놓습니다.

model.save('cats_and_dogs_small_3.h5')

acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']

epochs = range(len(acc))

plt.plot(epochs, acc, 'bo', label='Training acc')
plt.plot(epochs, val_acc, 'b', label='Validation acc')
plt.title('Training and validation accuracy')
plt.legend()

plt.figure()

plt.plot(epochs, loss, 'bo', label='Training loss')
plt.plot(epochs, val_loss, 'b', label='Validation loss')
plt.title('Training and validation loss')
plt.legend()

plt.show()

정확도가 이전과 비슷하지만 처음부터 훈련시킨 소규모 컨브넷보다 과대적합이 줄었습니다

1.3. 방법 #02 - 미세조정

Feature Extract에 사용했던 모델의 상위 층 몇 개를 동결에서 해제하고 모델에 새로 추가한 층( FC )과 함께 훈련하는 방식입니다.
주어진 문제에 조금 더 밀접하게 재사용 모델의 표현을 일부 조정하기 때문에 미세 조정이라고 부릅니다.

앞선 방법에서는 마지막 FC만 학습하고 나머지 VGG Model은 Trainable = false로 했습니다.
이번 방법에서는 기존 VGG 마지막 Conv. Layer만 동결해제 합니다.

네트워크를 미세 조정하는 단계는 다음과 같습니다:

사전에 훈련된 기반 네트워크 위에 새로운 네트워크를 추가합니다.
기반 네트워크를 동결합니다.
새로 추가한 네트워크를 훈련합니다.
기반 네트워크에서 일부 층의 동결을 해제합니다.
동결을 해제한 층과 새로 추가한 층을 함께 훈련합니다.

처음 세 단계는 특성 추출을 할 때 이미 완료했고, 네 번째 단계를 진행해 보죠. conv_base의 동결을 해제하고 개별 층을 동결하겠습니다.

conv_base.summary()

_________________________________________________________________
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: 0
Non-trainable params: 14,714,688
_________________________________________________________________

마지막 세 개의 합성곱 층을 미세조정. 즉, block4_pool까지 모든 층은 동결하고 block5_conv1, block5_conv2, block5_conv3 층은 학습하도록 하겠습니다.
왜 더 많은 층을 미세 조정하지 않을까요?
- 합성곱 기반층에 있는 하위 층들은 좀 더 일반적이고 재사용 가능한 특성들을 인코딩한다.
- 반면 상위 층은 좀 더 특화된 특성을 인코딩합니다.
- 새로운 문제에 재활용하도록 수정이 필요한 것은 구체적인 특성이므로 이들을 미세 조정하는 것이 유리합니다.
- 하위 층으로 갈수록 미세 조정에 대한 효과가 감소합니다.
- 훈련해야 할 파라미터가 많을수록 과대적합의 위험이 커집니다.
- 합성곱 기반층은 1천 5백만 개의 파라미터를 가지고 있고, 작은 데이터셋으로 전부 훈련하려고 하면 매우 위험합니다.

그러므로 이런 상황에서는 합성곱 기반층에서 최상위 두 세개의 층만 미세 조정하는 것이 좋습니다.

block5_conv1 이후의 Layer는 동결하도록 하겠습니다.

conv_base.trainable = True

set_trainable = False
for layer in conv_base.layers:
    if layer.name == 'block5_conv1':
        set_trainable = True
    if set_trainable:
        layer.trainable = True
    else:
        layer.trainable = False

Train을 시작합니다.

model.compile(loss='binary_crossentropy',
              optimizer=optimizers.RMSprop(lr=1e-5),
              metrics=['acc'])

history = model.fit_generator(
      train_generator,
      steps_per_epoch=100,
      epochs=100,
      validation_data=validation_generator,
      validation_steps=50)

Epoch 1/100
100/100 [==============================] - 14s 137ms/step - loss: 0.1813 - acc: 0.9255 - val_loss: 0.2593 - val_acc: 0.9000
Epoch 2/100
100/100 [==============================] - 13s 128ms/step - loss: 0.1533 - acc: 0.9430 - val_loss: 0.2473 - val_acc: 0.9080
Epoch 3/100
100/100 [==============================] - 13s 128ms/step - loss: 0.1373 - acc: 0.9450 - val_loss: 0.2697 - val_acc: 0.8930
Epoch 4/100
100/100 [==============================] - 13s 128ms/step - loss: 0.1190 - acc: 0.9535 - val_loss: 0.2574 - val_acc: 0.9010
Epoch 5/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0953 - acc: 0.9655 - val_loss: 0.1973 - val_acc: 0.9250
Epoch 6/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0770 - acc: 0.9705 - val_loss: 0.2241 - val_acc: 0.9260s - loss:  - ETA: 3s - loss: 0.0793 - acc: - ETA: 2s - loss: 0.0859 - acc:  - ETA: 2s - loss: 0.0853 - acc:  - ETA: 1s - loss: 0.0
Epoch 7/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0851 - acc: 0.9650 - val_loss: 0.1914 - val_acc: 0.9260
Epoch 8/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0702 - acc: 0.9730 - val_loss: 0.2129 - val_acc: 0.9240757 - acc:
Epoch 9/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0565 - acc: 0.9790 - val_loss: 0.2240 - val_acc: 0.9260
Epoch 10/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0551 - acc: 0.9825 - val_loss: 0.2592 - val_acc: 0.9190
Epoch 11/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0481 - acc: 0.9830 - val_loss: 0.2330 - val_acc: 0.9340
Epoch 12/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0453 - acc: 0.9835 - val_loss: 0.2633 - val_acc: 0.91500434 - 
Epoch 13/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0381 - acc: 0.9860 - val_loss: 0.2258 - val_acc: 0.9200
Epoch 14/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0404 - acc: 0.9860 - val_loss: 0.2146 - val_acc: 0.9330
Epoch 15/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0278 - acc: 0.9925 - val_loss: 0.3402 - val_acc: 0.9140A: 2s - loss
Epoch 16/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0328 - acc: 0.9875 - val_loss: 0.2362 - val_acc: 0.9260
Epoch 17/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0272 - acc: 0.9895 - val_loss: 0.2303 - val_acc: 0.9350
Epoch 18/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0311 - acc: 0.9900 - val_loss: 0.2423 - val_acc: 0.9300
Epoch 19/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0167 - acc: 0.9950 - val_loss: 0.2193 - val_acc: 0.9380ETA: 2
Epoch 20/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0309 - acc: 0.9915 - val_loss: 0.2480 - val_acc: 0.9310 7s - loss: 0.0265  - ETA: 6s - loss: 0.0371 - acc: 0 - ETA: 5s - loss: 0.0562 -  - ETA: 4s - loss: - ETA: 2s - loss: 0.0364 - acc:  - ETA: 2s - loss: 0 - ETA: 0s - loss: 0.0325 - acc: 0.990 - ETA: 0s - loss: 0.0323 - acc: 
Epoch 21/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0160 - acc: 0.9950 - val_loss: 0.2892 - val_acc: 0.9220
Epoch 22/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0217 - acc: 0.9935 - val_loss: 0.3709 - val_acc: 0.9100
Epoch 23/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0238 - acc: 0.9910 - val_loss: 0.2201 - val_acc: 0.9390
Epoch 24/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0167 - acc: 0.9940 - val_loss: 0.2208 - val_acc: 0.9360
Epoch 25/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0209 - acc: 0.9930 - val_loss: 0.2692 - val_acc: 0.9320 - loss: 0.0209 - 
Epoch 26/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0165 - acc: 0.9950 - val_loss: 0.2837 - val_acc: 0.9300
Epoch 27/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0135 - acc: 0.9940 - val_loss: 0.2806 - val_acc: 0.9250
Epoch 28/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0070 - acc: 0.9985 - val_loss: 0.2602 - val_acc: 0.9330
Epoch 29/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0168 - acc: 0.9935 - val_loss: 0.2746 - val_acc: 0.9260
Epoch 30/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0238 - acc: 0.9930 - val_loss: 0.3045 - val_acc: 0.9250
Epoch 31/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0135 - acc: 0.9950 - val_loss: 0.2851 - val_acc: 0.9210
Epoch 32/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0073 - acc: 0.9975 - val_loss: 0.3244 - val_acc: 0.9230 ETA: 4s - loss: 0.0041 - a
Epoch 33/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0150 - acc: 0.9955 - val_loss: 0.2884 - val_acc: 0.9350
Epoch 34/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0131 - acc: 0.9945 - val_loss: 0.3464 - val_acc: 0.9190
Epoch 35/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0088 - acc: 0.9980 - val_loss: 0.2882 - val_acc: 0.9340
Epoch 36/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0091 - acc: 0.9975 - val_loss: 0.3254 - val_acc: 0.9290
Epoch 37/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0061 - acc: 0.9985 - val_loss: 0.3295 - val_acc: 0.9310
Epoch 38/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0130 - acc: 0.9960 - val_loss: 0.3055 - val_acc: 0.93200.0087 - acc:  - ETA: 0s - loss: 0.0131 - acc: 0.996
Epoch 39/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0047 - acc: 0.9990 - val_loss: 0.2662 - val_acc: 0.9410
Epoch 40/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0106 - acc: 0.9965 - val_loss: 0.2971 - val_acc: 0.9310
Epoch 41/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0069 - acc: 0.9980 - val_loss: 0.3663 - val_acc: 0.9110
Epoch 42/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0075 - acc: 0.9980 - val_loss: 0.2408 - val_acc: 0.9390
Epoch 43/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0077 - acc: 0.9975 - val_loss: 0.3169 - val_acc: 0.92900.99
Epoch 44/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0044 - acc: 0.9990 - val_loss: 0.3122 - val_acc: 0.9260 loss: 0.0046 - acc: 
Epoch 45/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0057 - acc: 0.9985 - val_loss: 0.2893 - val_acc: 0.9360- loss: 0.0064 
Epoch 46/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0085 - acc: 0.9965 - val_loss: 0.7179 - val_acc: 0.8810
Epoch 47/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0101 - acc: 0.9970 - val_loss: 0.4235 - val_acc: 0.9240
Epoch 48/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0119 - acc: 0.9970 - val_loss: 0.2666 - val_acc: 0.9340
Epoch 49/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0024 - acc: 0.9995 - val_loss: 0.3485 - val_acc: 0.9310
Epoch 50/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0048 - acc: 0.9980 - val_loss: 0.3051 - val_acc: 0.9340
Epoch 51/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0076 - acc: 0.9975 - val_loss: 0.3027 - val_acc: 0.9240047 - acc: 0.9 - ETA: 5s - loss:
Epoch 52/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0030 - acc: 0.9990 - val_loss: 0.3273 - val_acc: 0.9260
Epoch 53/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0042 - acc: 0.9985 - val_loss: 0.3460 - val_acc: 0.9270
Epoch 54/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0018 - acc: 0.9995 - val_loss: 0.3755 - val_acc: 0.9310- ETA: 1s - loss: 0.0019 - 
Epoch 55/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0040 - acc: 0.9990 - val_loss: 0.3012 - val_acc: 0.9410
Epoch 56/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0041 - acc: 0.9985 - val_loss: 0.3481 - val_acc: 0.9320oss: 0.0044 - acc: 
Epoch 57/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0058 - acc: 0.9980 - val_loss: 0.2842 - val_acc: 0.9360
Epoch 58/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0051 - acc: 0.9980 - val_loss: 0.3516 - val_acc: 0.9360
Epoch 59/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0022 - acc: 1.0000 - val_loss: 0.3672 - val_acc: 0.9250
Epoch 60/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0043 - acc: 0.9975 - val_loss: 0.3058 - val_acc: 0.9390
Epoch 61/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0051 - acc: 0.9980 - val_loss: 0.3159 - val_acc: 0.9350
Epoch 62/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0045 - acc: 0.9990 - val_loss: 0.3732 - val_acc: 0.9300
Epoch 63/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0026 - acc: 0.9985 - val_loss: 0.3657 - val_acc: 0.9290- ETA: 0s - loss: 0.0026 - acc: 0
Epoch 64/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0015 - acc: 0.9995 - val_loss: 0.2992 - val_acc: 0.9500
Epoch 65/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0027 - acc: 0.9990 - val_loss: 0.4004 - val_acc: 0.9210
Epoch 66/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0016 - acc: 0.9995 - val_loss: 0.4355 - val_acc: 0.9290
Epoch 67/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0022 - acc: 0.9990 - val_loss: 0.3004 - val_acc: 0.9460
Epoch 68/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0033 - acc: 0.9995 - val_loss: 0.3708 - val_acc: 0.9350
Epoch 69/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0083 - acc: 0.9980 - val_loss: 0.3919 - val_acc: 0.9250
Epoch 70/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0043 - acc: 0.9990 - val_loss: 0.4616 - val_acc: 0.9170TA: 2s - lo
Epoch 71/100
100/100 [==============================] - 13s 127ms/step - loss: 0.0034 - acc: 0.9980 - val_loss: 0.3611 - val_acc: 0.9340 ETA: 2s - loss: 0. - ETA: 0s - loss: 0.0036 - acc: 0
Epoch 72/100
100/100 [==============================] - 13s 127ms/step - loss: 0.0028 - acc: 0.9990 - val_loss: 0.4096 - val_acc: 0.9210
Epoch 73/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0049 - acc: 0.9980 - val_loss: 0.3963 - val_acc: 0.9300
Epoch 74/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0037 - acc: 0.9995 - val_loss: 0.3567 - val_acc: 0.9330 1s - loss: 0.0042 -
Epoch 75/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0012 - acc: 1.0000 - val_loss: 0.3776 - val_acc: 0.9340
Epoch 76/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0034 - acc: 0.9990 - val_loss: 0.3662 - val_acc: 0.9390
Epoch 77/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0024 - acc: 0.9990 - val_loss: 0.5282 - val_acc: 0.9280
Epoch 78/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0028 - acc: 0.9995 - val_loss: 0.4118 - val_acc: 0.9350
Epoch 79/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0131 - acc: 0.9975 - val_loss: 0.4383 - val_acc: 0.9290
Epoch 80/100
100/100 [==============================] - 13s 127ms/step - loss: 0.0020 - acc: 0.9990 - val_loss: 0.4570 - val_acc: 0.9320
Epoch 81/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0029 - acc: 0.9985 - val_loss: 0.5011 - val_acc: 0.9220
Epoch 82/100
100/100 [==============================] - 13s 127ms/step - loss: 0.0025 - acc: 0.9985 - val_loss: 0.5921 - val_acc: 0.9090
Epoch 83/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0046 - acc: 0.9985 - val_loss: 0.4468 - val_acc: 0.9280
Epoch 84/100
100/100 [==============================] - 13s 128ms/step - loss: 1.3392e-04 - acc: 1.0000 - val_loss: 0.2831 - val_acc: 0.9480
Epoch 85/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0029 - acc: 0.9985 - val_loss: 0.4810 - val_acc: 0.9280
Epoch 86/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0128 - acc: 0.9975 - val_loss: 0.4014 - val_acc: 0.9330
Epoch 87/100
100/100 [==============================] - 13s 127ms/step - loss: 9.7576e-04 - acc: 0.9995 - val_loss: 0.3834 - val_acc: 0.9380
Epoch 88/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0023 - acc: 0.9990 - val_loss: 0.3717 - val_acc: 0.9320
Epoch 89/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0049 - acc: 0.9980 - val_loss: 0.4039 - val_acc: 0.9290
Epoch 90/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0029 - acc: 0.9990 - val_loss: 0.3495 - val_acc: 0.9450
Epoch 91/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0030 - acc: 0.9980 - val_loss: 0.3481 - val_acc: 0.9380
Epoch 92/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0025 - acc: 0.9985 - val_loss: 0.4316 - val_acc: 0.9260oss: 0.0026 - acc: 0
Epoch 93/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0025 - acc: 0.9985 - val_loss: 0.3484 - val_acc: 0.9380
Epoch 94/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0059 - acc: 0.9975 - val_loss: 0.3907 - val_acc: 0.9320
Epoch 95/100
100/100 [==============================] - 13s 128ms/step - loss: 7.1898e-04 - acc: 1.0000 - val_loss: 0.6235 - val_acc: 0.9080
Epoch 96/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0029 - acc: 0.9995 - val_loss: 0.4184 - val_acc: 0.9250
Epoch 97/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0024 - acc: 0.9990 - val_loss: 0.3732 - val_acc: 0.9300
Epoch 98/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0026 - acc: 0.9985 - val_loss: 0.4068 - val_acc: 0.9390
Epoch 99/100
100/100 [==============================] - 13s 128ms/step - loss: 7.8065e-04 - acc: 1.0000 - val_loss: 0.3804 - val_acc: 0.9290
Epoch 100/100
100/100 [==============================] - 13s 128ms/step - loss: 0.0018 - acc: 0.9985 - val_loss: 0.3999 - val_acc: 0.9300

Validation Set에서 정확도가 93%정도 되네요.

model.save('cats_and_dogs_small_4.h5')

자, 이제 결과를 그래프로 그려보도록 하겠습니다.

acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']

epochs = range(len(acc))

plt.plot(epochs, acc, 'bo', label='Training acc')
plt.plot(epochs, val_acc, 'b', label='Validation acc')
plt.title('Training and validation accuracy')
plt.legend()

plt.figure()

plt.plot(epochs, loss, 'bo', label='Training loss')
plt.plot(epochs, val_loss, 'b', label='Validation loss')
plt.title('Training and validation loss')
plt.legend()

plt.show()

2. Test Set으로 검증

test_generator = test_datagen.flow_from_directory(
        test_dir,
        target_size=(150, 150),
        batch_size=20,
        class_mode='binary')

test_loss, test_acc = model.evaluate_generator(test_generator, steps=50)
print('test acc:', test_acc)

Found 1000 images belonging to 2 classes.
test acc: 0.9409999907016754

Test Set에서의 정확도는 94% 까지 올라갔네요.

Share on

Twitter Facebook LinkedIn

MoonLight

Dog & Cat Classification Exercise #01

CNN Exercise #01 - 개와 고양이 사진 분류하기

0. Load Pre-Trained Model

1. Train Model

1.1. 방법 #01-01

Feature Extract

1.2. 방법 #01-02

1.3. 방법 #02 - 미세조정

2. Test Set으로 검증

Share on

You May Also Enjoy

세상에서 제일 쉬운 챗GPT 프롬프트 엔지니어링

AI를 위한 필수 수학 ( Essential Math for AI )

Learning to (Learn at Test Time): RNNs with Expressive Hidden States

실무로 통하는 클린 코드