2020_01_22 CIFAR10

import numpy as np
import matplotlib.pyplot as plt

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import optimizers
from tensorflow.keras.layers import Dense, Flatten, Conv2D, MaxPooling2D, BatchNormalization, Input

import time

▶ 이전 CNN 코드

* CIFAR10 적용 전 CNN 복습

(raw_train_x, raw_train_y), (raw_test_x, raw_test_y) = tf.keras.datasets.mnist.load_data()

train_x = raw_train_x/255
test_x = raw_test_x/255

train_x = train_x.reshape((60000, 28, 28, 1))
test_x = test_x.reshape((10000, 28, 28, 1))
train_y = raw_train_y
test_y = raw_test_y

# 일부만 사용(시간상)
train_x = train_x[:10000] # ADDED
train_y = train_y[:10000] # ADDED

model = keras.Sequential()
model.add(Input((28,28,1)))
model.add(Conv2D(32, (3, 3)))
model.add(MaxPooling2D((2, 2)))
model.add(Conv2D(64, (3, 3)))
model.add(MaxPooling2D((2, 2)))
model.add(Flatten())
model.add(Dense(10, activation='relu'))
model.add(Dense(10, activation='relu'))
model.add(Dense(10, activation='softmax'))

model.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"])
model.summary()


model.fit(train_x, train_y, epochs=5, verbose=1, batch_size=128)


loss, acc = model.evaluate(test_x, test_y)
print("loss=",loss)
print("acc=",acc)

y_ = model.predict(test_x)
predicted = np.argmax(y_, axis=1)

print(predicted)

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 (MaxPooling2D)	(None, 5, 5, 64)	0
flatten (Flatten)	(None, 1600)	0
dense (Dense)	(None, 10)	1010
dense_1 (Dense)	(None, 10)	110
dense_2 (Dense)	(None, 10)	110

▶ CIFAR10 적용

# (raw_train_x, raw_train_y), (raw_test_x, raw_test_y) = tf.keras.datasets.mnist.load_data()
(raw_train_x, raw_train_y), (raw_test_x, raw_test_y) = tf.keras.datasets.cifar10.load_data()

print(raw_train_x.shape)
print(raw_train_y.shape)
print(raw_test_x.shape)
print(raw_test_y.shape)


train_x = raw_train_x/255
test_x = raw_test_x/255

# train_x = train_x.reshape((60000, 28, 28, 1)) # COMMENT OUT
# test_x = test_x.reshape((10000, 28, 28, 1)) # COMMENT OUT

train_y = raw_train_y
test_y = raw_test_y

→ (50000, 32, 32, 3) (50000, 1) (10000, 32, 32, 3) (10000, 1)

 

labels = ["airplane", "automobile", "bird", "cat", "deer", "dog", "frog", "horse", "ship", "truck"]

def show_sample(i):
  print(raw_train_y[i][0], labels[raw_train_y[i][0]])
  plt.imshow(raw_train_x[i])
  plt.show()

for i in [2, 10, 12, 14]:
  show_sample(i)

9 truck

4 deer

# 데이터 불러오는 시간관계로 일부만 사용한다.
train_x = train_x[:10000] # ADDED
train_y = train_y[:10000] # ADDED

model = keras.Sequential()
# model.add(Input((28,28,1)))
model.add(Input((32,32,3)))
model.add(Conv2D(32, (3, 3)))
model.add(MaxPooling2D((2, 2)))
model.add(Conv2D(64, (3, 3)))
model.add(MaxPooling2D((2, 2)))
model.add(Flatten())
model.add(Dense(10, activation='relu'))
model.add(Dense(10, activation='relu'))
model.add(Dense(10, activation='softmax'))

model.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"])
model.summary()


model.fit(train_x, train_y, epochs=5, verbose=1, batch_size=128)


loss, acc = model.evaluate(test_x, test_y)
print("loss=",loss)
print("acc=",acc)

y_ = model.predict(test_x)
predicted = np.argmax(y_, axis=1)

print(predicted)

→ model.add(Conv2D(32, (3, 3)))

   1. 첫번째 인자 : 컨볼루션 필터의 수 입니다.

   2. 두번째 인자 : 컨볼루션 커널의 (행, 열) 입니다.

Conv2D

→ model.add(MaxPooling2D((2, 2))) : 주요값만(max) 뽑아 크기가 작은 출력 영상을 만드는 함수(사소한 변화를 무시함)

   1. 첫번째 인자 : 수직

   2. 두번째 인자 : 수평

MaxPooling2D

 

Layer (type)	Output Shape	Param #
conv2d_2 (Conv2D)	(None, 30, 30, 32)	896
max_pooling2d_2 (MaxPooling2D)	(None, 15, 15, 32)	0
conv2d_3 (Conv2D)	(None, 13, 13, 64)	18496
max_pooling2d_3 (MaxPooling2D)	(None, 6, 6, 64)	0
flatten_1 (Flatten)	(None, 2304)	0
dense_3 (Dense)	(None, 10)	23050
dense_4 (Dense)	(None, 10)	110
dense_5 (Dense)	(None, 10)	110
Total params: 42,662 Trainable params: 42,662 Non-trainable params: 0

▶ 영상 이미지 증강

 * ImageDataGenerator를 사용

(raw_train_x, raw_train_y), (raw_test_x, raw_test_y) = tf.keras.datasets.cifar10.load_data()

print(raw_train_x.shape)
print(raw_train_y.shape)
print(raw_test_x.shape)
print(raw_test_y.shape)

train_x = raw_train_x/255
test_x = raw_test_x/255

train_y = raw_train_y
test_y = raw_test_y

(50000, 32, 32, 3) (50000, 1) (10000, 32, 32, 3) (10000, 1)

train_x = train_x[:10000]
train_y = train_y[:10000]

# ADDED START
from tensorflow.keras.preprocessing.image import ImageDataGenerator

datagen = ImageDataGenerator(
    rotation_range=10,       # 0 ~ 180 회전
    width_shift_range=0.1,   # 좌우로 얼마나 움직이는가
    height_shift_range=0.1,  # 상하로 이동
    fill_mode='nearest',     # 움직이다 나오는 빈공간을 채우기
    horizontal_flip=True,    # 좌우로 회전
    vertical_flip=False      # 상하로 회전
)
# ADDED END

model = keras.Sequential()
model.add(Input((32,32,3)))
model.add(Conv2D(32, (3, 3)))
model.add(MaxPooling2D((2, 2)))
model.add(Conv2D(64, (3, 3)))
model.add(MaxPooling2D((2, 2)))
model.add(Flatten())
model.add(Dense(10, activation='relu'))
model.add(Dense(10, activation='relu'))
model.add(Dense(10, activation='softmax'))

model.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"])
model.summary()

# model.fit(train_x, train_y, epochs=5, verbose=1, batch_size=128)
model.fit_generator(datagen.flow(train_x, train_y, batch_size=128), epochs=5)

loss, acc = model.evaluate(test_x, test_y)
print("loss=",loss)
print("acc=",acc)

y_ = model.predict(test_x)
predicted = np.argmax(y_, axis=1)

print(predicted)

 

Layer (type)	Output Shape	Param #
conv2d_4 (Conv2D)	(None, 30, 30, 32)	896
max_pooling2d_4 (MaxPooling2D)	(None, 15, 15, 32)	0
conv2d_5 (Conv2D)	(None, 13, 13, 64)	18496
max_pooling2d_5 (MaxPooling2D)	(None, 6, 6, 64)	0
flatten_2 (Flatten)	(None, 2304)	0
dense_6 (Dense)	(None, 10)	23050
dense_7 (Dense)	(None, 10)	110
dense_8 (Dense)	(None, 10)	110
Total params: 42,662 Trainable params: 42,662 Non-trainable params: 0

 

 

데이터 증강을 하기 전고 비교하면 loss 더 작고, acc가 높아졌다. 성능이 향상되었다.
 - 증강 전 : loss: 1.7496 - acc: 0.3587
 - 증강 후 : loss: 1.5547 - acc: 0.4406
하지만 실행 시간은 epoch당 12초에서 15초로 증가하였다.