[Image Augmentation] 이미지 데이터 증강하는 코드 작성법

데이터 제너레이터를 통해 이미지를 증강한다.

train_datagen = ImageDataGenerator(
      rotation_range=40,
      width_shift_range=0.2,
      height_shift_range=0.2,
      shear_range=0.2,
      zoom_range=0.2,
      horizontal_flip=True,
      fill_mode='nearest')

이미지 증강만 하고, 드랍아웃 없이 해본다. 오버핏팅 됨을 확인하라.

!wget --no-check-certificate \

    https://storage.googleapis.com/mledu-datasets/cats_and_dogs_filtered.zip \

    -O /tmp/cats_and_dogs_filtered.zip

  

import os

import zipfile

filename = '/tmp/cats_and_dogs_filtered.zip'

zip_ref = zipfile.ZipFile(filename, 'r')

zip_ref.extractall('/tmp/cats_and_dogs_filtered')

zip_ref.close()

 

import tensorflow as tf

from tensorflow.keras.models import Sequential

from tensorflow.keras.layers import Conv2D, MaxPool2D, Flatten, Dense

from tensorflow.keras.optimizers import RMSprop, Adam, Adagrad, Adadelta

def build_model(): 

    model = Sequential()

    model.add(Conv2D(filters=16, kernel_size=(3,3), activation='relu', input_shape=(150,150, 3) ))

    model.add(MaxPool2D(pool_size=(2,2), strides=2))

    

    model.add(Conv2D(filters=32, kernel_size=(3,3), activation='relu'))

    model.add(MaxPool2D(pool_size=(2,2), strides=2))

    model.add(Conv2D(filters=64, kernel_size=(3,3), activation='relu'))

    model.add(MaxPool2D(pool_size=(2,2), strides=2))

    model.add(Flatten())

    model.add(Dense(units=512, activation='relu'))

    model.add(Dense(units=1, activation='sigmoid'))

    model.compile(optimizer=RMSprop(learning_rate =0.001), 

    loss='binary_crossentropy', metrics=['accuracy'])

 

(또는 model.compile(optimizer='rmsprop' , 

    loss='binary_crossentropy', metrics=['accuracy'])

    return model

model = build_model()

 

from tensorflow.keras.preprocessing.image import ImageDataGenerator

 

# 학습데이터가 인공지능의 성능을 좌우하기 때문에

# 원본 이미지를 다양하게 변형시켜서 학습을 해야 성능이 좋아진다.

train_datagen = ImageDataGenerator(rescale= 1/255.0, rotation_range= 40, 

width_shift_range=0.2, height_shift_range=0.2, shear_range=0.2, zoom_range=0.2, 

horizontal_flip= True)

# 테스트 데이터는 이미지 증강을 해도 되고, 안해도 된다. 상황에 맞게 하면 된다. 

test_datagen = ImageDataGenerator(rescale=1/255.0)

 

train_generator = train_datagen.flow_from_directory('/tmp/cats_and_dogs_filtered/

cats_and_dogs_filtered/train', 

target_size=(150, 150), class_mode='binary')

Found 2000 images belonging to 2 classes.

 

test_generator = test_datagen.flow_from_directory ('/tmp/cats_and_dogs_filtered/

cats_and_dogs_filtered/validation', target_size=(150,150), class_mode = 'binary')

Found 1000 images belonging to 2 classes.

 

 

epoch_history = model.fit(train_generator, epochs=10, validation_data= test_generator)