[이미지 딥러닝 모델링] zipfile, optimizers, ImageDataGenerator

정리 1. 

import tensorflow as tf

import os

import zipfile

 

구글코랩에서 인터넷에 있는 파일을 다운로드 하는 방법

!wget --no-check-certificate \

    "https://storage.googleapis.com/laurencemoroney-blog.appspot.com/happy-or-sad.zip" \

    -O "/tmp/happy-or-sad.zip"

 

 

파이썬으로 압축파일 푸는 방법 

 

import os

import zipfile

filename = '/tmp/happy-or-sad.zip'

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

zip_ref.extractall('/tmp/happy-or-sad')

zip_ref.close()

 

 

train_happy_dir = '/tmp/happy-or-sad/happy'

train_sad_dir = '/tmp/happy-or-sad/sad'

train_happy_names = os.listdir(train_happy_dir)

train_sad_names = os.listdir(train_sad_dir)

 

## 콜백 함수를 만든다.

class myCallback(tf.keras.callbacks.Callback):

  # Your Code

    def on_epoch_end(self, epoch, logs={}) :

        if logs['accuracy'] >= 0.999 :

          print('\nReached 99% accuracy so cancelling training!')

          self.model.stop_training = True

callbacks = myCallback()

 

 

jpg, png와 같은 이미지 파일을 학습 데이터로 만든 방법 (텐서플로우의 ImageDataGenerator)

 

## 모델링을 하고, 컴파일 코드를 작성한다.

import tensorflow as tf

from tensorflow.keras.models import Sequential

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

# name에러가 뜨면 import안했구나 생각하면 됨.

model = tf.keras.models.Sequential([                        

(Conv2D(filters=16, kernel_size=(3,3), activation='relu', input_shape=(300,300, 3) )),

(MaxPool2D(pool_size=(2,2), strides=2)),

    

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

(MaxPool2D(pool_size=(2,2), strides=2)),

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

(MaxPool2D(pool_size=(2,2), strides=2)),

(Flatten()),

(Dense(units=512, activation='relu')),

(Dense(units=1, activation='sigmoid'))])

 

from tensorflow.keras.optimizers import RMSprop

model.compile(loss='binary_crossentropy',

              optimizer=RMSprop(learning_rate =0.001),

              metrics=['accuracy'])

 

from tensorflow.keras.preprocessing.image import ImageDataGenerator

train_datagen = ImageDataGenerator(rescale= 1/255.0)

train_generator = train_datagen.flow_from_directory(

        '/tmp/happy-or-sad', target_size=(300, 300), class_mode='binary' )

history = model.fit(train_generator, epochs=15, callbacks=callbacks)

정리 2.

(모델링 함수이용)

압축 풀기

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()

 

데이터 억세스할 경로를 만든다.

train_cats_dir = '/tmp/cats_and_dogs_filtered/cats_and_dogs_filtered/train/cats'

 

validation_cats_dir = '/tmp/cats_and_dogs_filtered/cats_and_dogs_filtered/validation/cats'

파일명을 찍어본다.

train_cats_names = os.listdir(train_cats_dir)

train_dogs_names = os.listdir(train_dogs_dir)

validation_cats_names = os.listdir(validation_cats_dir)

validation_dogs_names = os.listdir(validation_dogs_dir)

 

개와 고양이 8개씩 화면에 이미지 표시해보자

%matplotlib inline

 

import matplotlib.image as mpimg

import matplotlib.pyplot as plt

 

# Parameters for our graph; we'll output images in a 4x4 configuration

nrows = 4

ncols = 4

 

pic_index = 0 # Index for iterating over images

 

# Set up matplotlib fig, and size it to fit 4x4 pics

fig = plt.gcf()

fig.set_size_inches(ncols*4, nrows*4)

pic_index+=8

next_cat_pix = [os.path.join(train_cats_dir, fname) 

                for fname in train_cats_names[ pic_index-8:pic_index] 

               ]

next_dog_pix = [os.path.join(train_dogs_dir, fname) 

                for fname in train_dogs_names[ pic_index-8:pic_index]

               ]

for i, img_path in enumerate(next_cat_pix+next_dog_pix):

  # Set up subplot; subplot indices start at 1

  sp = plt.subplot(nrows, ncols, i + 1)

  sp.axis('Off') # Don't show axes (or gridlines)

  img = mpimg.imread(img_path)

  plt.imshow(img)

plt.show()

Building a Small Model from Scratch to Get to ~72% Accuracy

이미지의 사이즈를 150x150, 칼라(rgb) 로 처리하자.

import tensorflow as tf

from tensorflow.keras.models import Sequential

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

 

def build_model(): 

    model = Sequential()

    model.add(Conv2D(filters=16, kernel_size=(3,3), activation='relu', input_shape=(300,300, 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'))

    return model

 

model = build_model()

 

서머리 해보자

model.summary()

 

RMSprop 으로 컴파일한다.

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

model.compile(optimizer=RMSprop(learning_rate =0.001), loss='binary_crossentropy', metrics=['accuracy'])

 

 

Data Preprocessing

# 파일로 되어있는 이미지를 학습을 통해서 넘파이로 바꿔줘야 한다.

# 실제로 복잡한 작업ㅇ을 해야 하는 것을, 텐서플로우에서 쉽게 처리 할 수 있게 라이브러리를 제공한다.

# ImageDataGenerator

ImageDataGenerator 사용하기

from tensorflow.keras.preprocessing.image import ImageDataGenerator

# 파일로 되어 있는 이미지의 피쳐 스케일링을 한다. ==> 255.0으로 나누는 것!

 

train_datagen = ImageDataGenerator(rescale= 1/255.0)

validation_datagen = ImageDataGenerator(rescale= 1/255.0)

 

# 파일이 들어있는 디렉토리를 알려주고, 이미지 사이즈 정보도 알려주고, 분류할 정보도 알려준다. 

# target_size 파라미터는 우리가 마음대로 정해줄 수 있다. 단, 모델의 input_shape과 동일해야 한다.

# class_mode의 2개 분류는 binary, 3개 분류는 categorical로 설정

train_generator = train_datagen.flow_from_directory('/tmp/cats_and_dogs_filtered/cats_and_dogs_filtered/train', target_size=(300, 300), class_mode='binary')

# train_generator 에는 X_train, y_train이 들어있게 한다.

# y_train의 값은 폴더의 이름으로 설정된다. 

# 따라서 폴더의 이름을 알파벳순으로 정렬한 후, 0부터 차례로 숫자를 매긴다. 

validation_generator = validation_datagen.flow_from_directory('/tmp/cats_and_dogs_filtered/cats_and_dogs_filtered/validation', target_size=(300, 300), class_mode='binary')

 

# 전체 데이터수 = batch_size * steps_per_epoch 와 같다.

# 1000          = 20         * 50

 

epoch_history = model.fit(train_generator, epochs=15, steps_per_epoch= 8, validation_data= validation_generator)

Running the Model

픽사베이에서 무료 사진 다운로드 받아서, 실제로 예측해 보자

import numpy as np

from google.colab import files

from keras.preprocessing import image

uploaded=files.upload()

for fn in uploaded.keys():

 

  # predicting images

  path='/content/' + fn

  img=image.load_img(path, target_size=(300, 300))

  

  x=image.img_to_array(img)

  x=np.expand_dims(x, axis=0)

  images = np.vstack([x])

  

  classes = model.predict(images, batch_size=10)

  

  print(classes[0])

  

  if classes[0]>0.5 :

    print(fn + " is a dog")

    

  else:

    print(fn + " is a cat")

Evaluating Accuracy and Loss for the Model

training/validation accuracy 와 loss 를 차트로 시각화 한다.

model.evaluate(validation_generator)

plt.plot(epoch_history.history['accuracy'])

plt.plot(epoch_history.history['val_accuracy'])

plt.title('Model Accuracy')

plt.xlabel('Epochs')

plt.ylabel('Accuracy')

plt.legend(['Train', 'Validation'])

plt.show()

 

plt.plot(epoch_history.history['loss'])

plt.plot(epoch_history.history['val_loss'])

plt.title('Model Accuracy')

plt.xlabel('Epochs')

plt.ylabel('loss')

plt.legend(['Train', 'Validation'])

plt.show()