Deep Learning(0913_day7) - CIFAR10(ImageAugmentation)

CIFAR10 Dataset 신경망으로 훈련¶

imports¶

In [1]:

import numpy as np
import tensorflow as tf
from tensorflow.keras.datasets import cifar10
import matplotlib.pyplot as plt

from tensorflow.keras.utils import to_categorical
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Flatten, Dense
from tensorflow.keras.optimizers import SGD, Adam
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Dropout, BatchNormalization, Activation, GlobalAveragePooling2D
from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping, ReduceLROnPlateau
from sklearn.model_selection import train_test_split
from tensorflow.keras.preprocessing.image import ImageDataGenerator

In [2]:

np.random.seed(42)
tf.random.set_seed(42)

데이터 적재¶

In [3]:

(train_images, train_labels), (test_images, test_labels) = cifar10.load_data()

Downloading data from https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
170500096/170498071 [==============================] - 3s 0us/step
170508288/170498071 [==============================] - 3s 0us/step

In [4]:

print(train_images.shape, train_labels.shape)
print(test_images.shape, test_labels.shape)
print(train_images[0].shape)

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

In [5]:

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

In [6]:

train_labels[:10]

Out[6]:

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

데이터 시각화¶

In [7]:

train_labels = np.int32(train_labels.flatten())

In [8]:

figure, axes = plt.subplots(figsize=(22, 10), nrows=3, ncols=8)

for i in range(3):
    for j in range(8):
        axes[i][j].imshow(train_images[(i*8)+j])
        axes[i][j].set_title(class_names[train_labels[(i*8)+j]])

데이터 전처리¶

In [9]:

train_images = np.array(train_images/255.0, dtype=np.float32)
train_labels = np.array(train_labels, dtype=np.float32)
test_images = np.array(test_images/255.0, dtype=np.float32)
test_labels = np.array(test_labels, dtype=np.float32)

train_oh_labels = to_categorical(train_labels)
test_oh_labels = to_categorical(test_labels)

train_images, valid_images, train_oh_labels, valid_oh_labels = train_test_split(train_images, train_oh_labels, test_size=0.2, random_state=42)

In [10]:

train_images.shape, valid_images.shape, train_oh_labels.shape, valid_oh_labels.shape

Out[10]:

((40000, 32, 32, 3), (10000, 32, 32, 3), (40000, 10), (10000, 10))

데이터 증강(샘플 예시)¶

In [11]:

# data_generator = ImageDataGenerator(horizontal_flip=True) # 랜덤하게 진행하기 때문에 바뀌는 경우도 있고 아닌 경우도 있음
# data_generator = ImageDataGenerator(rotation_range=180)
data_generator = ImageDataGenerator(vertical_flip=True)

image = np.expand_dims(train_images[0], axis=0)
data_generator_iterator = data_generator.flow(image)

augmented_image = next(data_generator_iterator)

image = np.squeeze(image)
augmented_image = np.squeeze(augmented_image)

fig, axes = plt.subplots(nrows=1, ncols=2, figsize=(6*2, 6))
axes[0].imshow(image)
axes[1].imshow(augmented_image)

Out[11]:

<matplotlib.image.AxesImage at 0x7f76de314c90>

In [12]:

# data_generator = ImageDataGenerator(horizontal_flip=True) 
# data_generator = ImageDataGenerator(rotation_range=180)
# data_generator = ImageDataGenerator(vertical_flip=True)
data_generator = ImageDataGenerator(horizontal_flip=True, 
                                    rotation_range=30,
                                    vertical_flip=True, 
                                    zoom_range=(0.7, 0.9))

images = train_images[0:4]

data_generator_iterator = data_generator.flow(images)

augmented_image = next(data_generator_iterator)

fig, axes = plt.subplots(nrows=1, ncols=4, figsize=(6*4, 6))

for i, image in enumerate(images):
    axes[i].imshow(image)

fig, axes = plt.subplots(nrows=1, ncols=4, figsize=(6*4, 6))

for i, aug_image in enumerate(augmented_image):
    axes[i].imshow(aug_image)

In [13]:

BATCH_SIZE = 32

train_generator = ImageDataGenerator(horizontal_flip=True)
train_generator_iterator = train_generator.flow(train_images, train_oh_labels, batch_size=BATCH_SIZE)

valid_generator = ImageDataGenerator(horizontal_flip=True)
valid_generator_iterator = valid_generator.flow(valid_images, valid_oh_labels, batch_size=BATCH_SIZE)

Global Average Pooling 적용¶

In [14]:

def build_extended_model():
    tf.keras.backend.clear_session()

    input_tensor = Input(shape=(32, 32, 3))
    x = Conv2D(filters=32, kernel_size=(3, 3), strides=1, padding="same")(input_tensor)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)
    x = Conv2D(filters=64, kernel_size=(3, 3), strides=1, padding="same")(x)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)
    x = MaxPooling2D(pool_size=(2, 2))(x)

    x = Conv2D(filters=128, kernel_size=(3, 3), strides=1, padding="same")(x)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)
    x = Conv2D(filters=128, kernel_size=(3, 3), strides=1, padding="same")(x)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)
    x = MaxPooling2D(pool_size=(2, 2))(x)
                    
    x = Conv2D(filters=256, kernel_size=(3, 3), strides=1, padding="same")(x)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)
    x = Conv2D(filters=256, kernel_size=(3, 3), strides=1, padding="same")(x)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)
    # x = MaxPooling2D(pool_size=(2, 2))(x)                 

    x = Conv2D(filters=512, kernel_size=(3, 3), strides=1, padding="same")(x)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)
    x = MaxPooling2D(pool_size=(2, 2))(x)  

    # x = Flatten()(x)      
    x = GlobalAveragePooling2D()(x)
    x = Dropout(rate=0.5)(x)                 
    x = Dense(300, activation="relu")(x)
    x = Dropout(rate=0.3)(x)
    x = Dense(100, activation="relu")(x)
    x = Dropout(rate=0.3)(x)
    output = Dense(10, activation="softmax")(x)

    model = Model(inputs=input_tensor, outputs=output)
    return model

모델 생성 및 컴파일¶

In [15]:

model = build_extended_model()
model.summary()

Model: "model"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         [(None, 32, 32, 3)]       0         
_________________________________________________________________
conv2d (Conv2D)              (None, 32, 32, 32)        896       
_________________________________________________________________
batch_normalization (BatchNo (None, 32, 32, 32)        128       
_________________________________________________________________
activation (Activation)      (None, 32, 32, 32)        0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 32, 32, 64)        18496     
_________________________________________________________________
batch_normalization_1 (Batch (None, 32, 32, 64)        256       
_________________________________________________________________
activation_1 (Activation)    (None, 32, 32, 64)        0         
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 16, 16, 64)        0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 16, 16, 128)       73856     
_________________________________________________________________
batch_normalization_2 (Batch (None, 16, 16, 128)       512       
_________________________________________________________________
activation_2 (Activation)    (None, 16, 16, 128)       0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 16, 16, 128)       147584    
_________________________________________________________________
batch_normalization_3 (Batch (None, 16, 16, 128)       512       
_________________________________________________________________
activation_3 (Activation)    (None, 16, 16, 128)       0         
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 8, 8, 128)         0         
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 8, 8, 256)         295168    
_________________________________________________________________
batch_normalization_4 (Batch (None, 8, 8, 256)         1024      
_________________________________________________________________
activation_4 (Activation)    (None, 8, 8, 256)         0         
_________________________________________________________________
conv2d_5 (Conv2D)            (None, 8, 8, 256)         590080    
_________________________________________________________________
batch_normalization_5 (Batch (None, 8, 8, 256)         1024      
_________________________________________________________________
activation_5 (Activation)    (None, 8, 8, 256)         0         
_________________________________________________________________
conv2d_6 (Conv2D)            (None, 8, 8, 512)         1180160   
_________________________________________________________________
batch_normalization_6 (Batch (None, 8, 8, 512)         2048      
_________________________________________________________________
activation_6 (Activation)    (None, 8, 8, 512)         0         
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 4, 4, 512)         0         
_________________________________________________________________
global_average_pooling2d (Gl (None, 512)               0         
_________________________________________________________________
dropout (Dropout)            (None, 512)               0         
_________________________________________________________________
dense (Dense)                (None, 300)               153900    
_________________________________________________________________
dropout_1 (Dropout)          (None, 300)               0         
_________________________________________________________________
dense_1 (Dense)              (None, 100)               30100     
_________________________________________________________________
dropout_2 (Dropout)          (None, 100)               0         
_________________________________________________________________
dense_2 (Dense)              (None, 10)                1010      
=================================================================
Total params: 2,496,754
Trainable params: 2,494,002
Non-trainable params: 2,752
_________________________________________________________________

In [16]:

model.compile(optimizer=Adam(0.001), loss="categorical_crossentropy", metrics=["accuracy"])

Callback 정의¶

In [17]:

checkpoint_cb = ModelCheckpoint("my_keras_model.h5", save_best_only=True, verbose=1)
early_stopping_cb = EarlyStopping(patience=10, restore_best_weights=True)
reducelr_cb = ReduceLROnPlateau(monitor="val_loss", factor=0.2, patience=5, mode="min", verbose=1)

모델 학습¶

In [19]:

history = model.fit(train_generator_iterator, epochs=40, validation_data=valid_generator_iterator, 
                    callbacks=[checkpoint_cb, early_stopping_cb, reducelr_cb])

Epoch 1/40
1250/1250 [==============================] - 68s 32ms/step - loss: 1.6851 - accuracy: 0.3584 - val_loss: 1.6107 - val_accuracy: 0.3910

Epoch 00001: val_loss improved from inf to 1.61065, saving model to my_keras_model.h5
Epoch 2/40
1250/1250 [==============================] - 40s 32ms/step - loss: 1.2623 - accuracy: 0.5510 - val_loss: 1.3285 - val_accuracy: 0.5221

Epoch 00002: val_loss improved from 1.61065 to 1.32849, saving model to my_keras_model.h5
Epoch 3/40
1250/1250 [==============================] - 40s 32ms/step - loss: 1.0395 - accuracy: 0.6420 - val_loss: 0.9931 - val_accuracy: 0.6446

Epoch 00003: val_loss improved from 1.32849 to 0.99310, saving model to my_keras_model.h5
Epoch 4/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.8905 - accuracy: 0.6992 - val_loss: 0.9183 - val_accuracy: 0.6835

Epoch 00004: val_loss improved from 0.99310 to 0.91830, saving model to my_keras_model.h5
Epoch 5/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.7911 - accuracy: 0.7350 - val_loss: 0.8444 - val_accuracy: 0.7119

Epoch 00005: val_loss improved from 0.91830 to 0.84442, saving model to my_keras_model.h5
Epoch 6/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.6963 - accuracy: 0.7705 - val_loss: 0.9830 - val_accuracy: 0.6945

Epoch 00006: val_loss did not improve from 0.84442
Epoch 7/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.6329 - accuracy: 0.7918 - val_loss: 0.6412 - val_accuracy: 0.7890

Epoch 00007: val_loss improved from 0.84442 to 0.64124, saving model to my_keras_model.h5
Epoch 8/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.5662 - accuracy: 0.8156 - val_loss: 0.8325 - val_accuracy: 0.7299

Epoch 00008: val_loss did not improve from 0.64124
Epoch 9/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.5148 - accuracy: 0.8334 - val_loss: 0.6691 - val_accuracy: 0.7745

Epoch 00009: val_loss did not improve from 0.64124
Epoch 10/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.4759 - accuracy: 0.8454 - val_loss: 0.5873 - val_accuracy: 0.7992

Epoch 00010: val_loss improved from 0.64124 to 0.58735, saving model to my_keras_model.h5
Epoch 11/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.4403 - accuracy: 0.8565 - val_loss: 0.5541 - val_accuracy: 0.8219

Epoch 00011: val_loss improved from 0.58735 to 0.55413, saving model to my_keras_model.h5
Epoch 12/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.4013 - accuracy: 0.8702 - val_loss: 0.5485 - val_accuracy: 0.8263

Epoch 00012: val_loss improved from 0.55413 to 0.54846, saving model to my_keras_model.h5
Epoch 13/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.3729 - accuracy: 0.8799 - val_loss: 0.4964 - val_accuracy: 0.8408

Epoch 00013: val_loss improved from 0.54846 to 0.49642, saving model to my_keras_model.h5
Epoch 14/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.3474 - accuracy: 0.8864 - val_loss: 0.6327 - val_accuracy: 0.8029

Epoch 00014: val_loss did not improve from 0.49642
Epoch 15/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.3149 - accuracy: 0.8985 - val_loss: 0.5795 - val_accuracy: 0.8246

Epoch 00015: val_loss did not improve from 0.49642
Epoch 16/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.2888 - accuracy: 0.9062 - val_loss: 0.5494 - val_accuracy: 0.8315

Epoch 00016: val_loss did not improve from 0.49642
Epoch 17/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.2675 - accuracy: 0.9140 - val_loss: 0.5777 - val_accuracy: 0.8372

Epoch 00017: val_loss did not improve from 0.49642
Epoch 18/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.2594 - accuracy: 0.9162 - val_loss: 0.5210 - val_accuracy: 0.8370

Epoch 00018: val_loss did not improve from 0.49642

Epoch 00018: ReduceLROnPlateau reducing learning rate to 0.00020000000949949026.
Epoch 19/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.1609 - accuracy: 0.9493 - val_loss: 0.4152 - val_accuracy: 0.8847

Epoch 00019: val_loss improved from 0.49642 to 0.41522, saving model to my_keras_model.h5
Epoch 20/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.1286 - accuracy: 0.9579 - val_loss: 0.4481 - val_accuracy: 0.8865

Epoch 00020: val_loss did not improve from 0.41522
Epoch 21/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.1134 - accuracy: 0.9630 - val_loss: 0.4594 - val_accuracy: 0.8803

Epoch 00021: val_loss did not improve from 0.41522
Epoch 22/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.0998 - accuracy: 0.9671 - val_loss: 0.4668 - val_accuracy: 0.8837

Epoch 00022: val_loss did not improve from 0.41522
Epoch 23/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.0988 - accuracy: 0.9688 - val_loss: 0.4689 - val_accuracy: 0.8863

Epoch 00023: val_loss did not improve from 0.41522
Epoch 24/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.0871 - accuracy: 0.9715 - val_loss: 0.5004 - val_accuracy: 0.8839

Epoch 00024: val_loss did not improve from 0.41522

Epoch 00024: ReduceLROnPlateau reducing learning rate to 4.0000001899898055e-05.
Epoch 25/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.0671 - accuracy: 0.9776 - val_loss: 0.4701 - val_accuracy: 0.8902

Epoch 00025: val_loss did not improve from 0.41522
Epoch 26/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.0632 - accuracy: 0.9800 - val_loss: 0.4864 - val_accuracy: 0.8930

Epoch 00026: val_loss did not improve from 0.41522
Epoch 27/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.0612 - accuracy: 0.9804 - val_loss: 0.4896 - val_accuracy: 0.8908

Epoch 00027: val_loss did not improve from 0.41522
Epoch 28/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.0565 - accuracy: 0.9815 - val_loss: 0.5053 - val_accuracy: 0.8919

Epoch 00028: val_loss did not improve from 0.41522
Epoch 29/40
1250/1250 [==============================] - 40s 32ms/step - loss: 0.0563 - accuracy: 0.9826 - val_loss: 0.5125 - val_accuracy: 0.8903

Epoch 00029: val_loss did not improve from 0.41522

Epoch 00029: ReduceLROnPlateau reducing learning rate to 8.000000525498762e-06.

In [20]:

plt.plot(history.history["accuracy"], label="train")
plt.plot(history.history["val_accuracy"], label="valid")
plt.legend()

Out[20]:

<matplotlib.legend.Legend at 0x7f76d0032110>

테스트 데이터에 대해 평가¶

In [22]:

model.evaluate(test_images, test_oh_labels, batch_size=BATCH_SIZE, verbose=1)

313/313 [==============================] - 4s 12ms/step - loss: 0.4121 - accuracy: 0.8841

Out[22]:

[0.4120880365371704, 0.8841000199317932]

In [ ]: