My model vs VGG16

1. 각종 모듈 호출

import os
import cv2
import numpy as np
import pandas as pd
import tensorflow as tf
import matplotlib.pyplot as plt

from sklearn.model_selection import train_test_split

from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Dense, Conv2D, BatchNormalization, Dropout, \
                                      Flatten, Activation, MaxPooling2D, GlobalAveragePooling2D
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.callbacks import ReduceLROnPlateau, EarlyStopping
from tensorflow.keras.utils import to_categorical
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.datasets import cifar10
from tensorflow.keras.applications import VGG16, ResNet50V2, Xception

---

2. 각종 함수 정의

def make_to_float(images, labels) :
    
    images = np.array(images, dtype=np.float32)
    labels = np.array(labels, dtype=np.float32)
    
    return images, labels


def ohe(labels) :
    
    labels = to_categorical(labels)
    
    return labels


def tr_val_test(train_images, train_labels, test_images, test_labels, val_rate) :
    
    tr_images, val_images, tr_labels, val_labels = \
                           train_test_split(train_images, train_labels, test_size=val_rate)
    
    return (tr_images, tr_labels), (val_images, val_labels), (test_images, test_labels)


def resize_images(images, resize) :
    
    resized_images = np.zeros((images.shape[0], resize, resize, 3))
    
    for i in range(images.shape[0]) :
        resized_image = cv2.resize(images[i], (resize, resize))
        resized_images[i] = resized_image
        
    return resized_images


def create_pretrained_model(model_name, verbose) :
    
    input_tensor = Input(shape=(IMAGE_SIZE, IMAGE_SIZE, 3))
    
    if model_name == "vgg16" :
        base_model = VGG16(input_tensor=input_tensor, include_top=False, weights="imagenet")
    elif model_name == "resnet50" :
        base_model = ResNet50V2(input_tensor=input_tensor, include_top=False, weights="imagenet")
    elif model_name == "xception" :
        base_model = Xception(input_tensor=input_tensor, include_top=False, weights="imagenet")
        
    bm_output = base_model.output
    
    x = GlobalAveragePooling2D()(bm_output)
    if model_name != 'vgg16' :
        x = Dropout(rate=0.5)(x)
    x = Dense(50, activation="relu")(x)
    output = Dense(10, activation="softmax")(x)
    
    model = Model(inputs=input_tensor, outputs=output)
    
    if verbose == True :
        model.summary()
        
    return model


def create_my_model(verbose) :
    
    input_tensor = Input(shape = (32, 32, 3))
    
    x = Conv2D(filters=64, kernel_size=(3, 3), padding="same")(input_tensor)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)
    
    num_filter = [64, 128, 128, 256, 256, 512]
    n = 1
    
    for i in num_filter :
        
        x = Conv2D(filters=i, kernel_size=(3, 3), padding="same")(x)
        x = BatchNormalization()(x)
        x = Activation("relu")(x)
        if n%2 != 0 :
            x = MaxPooling2D(pool_size=(2, 2))(x)
        n+=1
    
    x = GlobalAveragePooling2D()(x)
    x = Dropout(rate=0.5)(x)
    x = Dense(50, activation="relu")(x)
    x = Dropout(rate=0.2)(x)
    output = Dense(10, activation="softmax")(x)
    
    model = Model(inputs=input_tensor, outputs=output)
    
    if verbose==True:
        
        model.summary()
        
    return model


def compare_acc(result_1, result_2) :
    
    fig, axs = plt.subplots(figsize=(20, 5), nrows=1, ncols=2)
    
    axs[0].plot(result_1.history["accuracy"], label="tr_acc", linewidth=4)
    axs[0].plot(result_1.history["val_accuracy"], label="val_acc", linewidth=4)
    axs[0].set_title("My model")
    axs[0].set_xlabel("epochs")
    axs[0].set_ylabel("accuracy")
    axs[0].legend()
    
    axs[1].plot(result_2.history["accuracy"], label="tr_acc", linewidth=4)
    axs[1].plot(result_2.history["val_accuracy"], label="val_acc", linewidth=4)
    axs[1].set_title("Pretrained model")
    axs[1].set_xlabel("epochs")
    axs[1].set_ylabel("accuracy")
    axs[1].legend()
    

def compare_loss(result_1, result_2) :
    
    fig, axs = plt.subplots(figsize=(20, 5), nrows=1, ncols=2)
    
    axs[0].plot(result_1.history["loss"], label="tr_loss", linewidth=4)
    axs[0].plot(result_1.history["val_loss"], label="val_loss", linewidth=4)
    axs[0].set_title("My model")
    axs[0].set_xlabel("epochs")
    axs[0].set_ylabel("loss")
    axs[0].legend()
    
    axs[1].plot(result_2.history["loss"], label="tr_loss", linewidth=4)
    axs[1].plot(result_2.history["val_loss"], label="val_loss", linewidth=4)
    axs[1].set_title("Pretrained model")
    axs[1].set_xlabel("epochs")
    axs[1].set_ylabel("loss")
    axs[1].legend()

---

3. 데이터 전처리 및 shape 확인

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

train_images, train_labels = make_to_float(train_images, train_labels)
test_images, test_labels = make_to_float(test_images, test_labels)

train_labels = ohe(train_labels)
test_labels = ohe(test_labels)

(tr_images, tr_labels), (val_images, val_labels), (test_images, test_labels) = \
               tr_val_test(train_images, train_labels, test_images, test_labels, val_rate=0.15)

print("train data's shape : \n{}, {}\n".format(tr_images.shape, tr_labels.shape))
print("validation data's shape : \n{}, {}\n".format(val_images.shape, val_labels.shape))
print("test data's shape : \n{}, {}".format(test_images.shape, test_labels.shape))

---

4. ImageDataGenerator

BATCH_SIZE=64

tr_generator    = ImageDataGenerator(horizontal_flip=True, rescale=1/255.0)
val_generator   = ImageDataGenerator(rescale=1/255.0)
test_generator  = ImageDataGenerator(rescale=1/255.0)

flow_tr_gen   = tr_generator.flow(tr_images, tr_labels, batch_size=BATCH_SIZE, shuffle=True)
flow_val_gen  = val_generator.flow(val_images, val_labels, batch_size=BATCH_SIZE, shuffle=False)
flow_test_gen = test_generator.flow(test_images, test_labels, batch_size=BATCH_SIZE, shuffle=False)

---

5. callback

rlr_cb = ReduceLROnPlateau(monitor="val_loss", mode="min", factor=0.2, patience=5, verbose=True)
ely_cb = EarlyStopping(monitor="val_loss", mode="min", patience=10, verbose=True)

---

6. my model 생성, 설정, 학습

model_mine = create_my_model(verbose=True)
model_mine.compile(optimizer=Adam(learning_rate=0.001), loss="categorical_crossentropy", metrics=["accuracy"])

result_mine = model_mine.fit(flow_tr_gen, epochs=40, validation_data=flow_val_gen, callbacks=[rlr_cb, ely_cb])

---

7. VGG16을 위한 data resize

tr_images = resize_images(tr_images, resize=64)
val_images = resize_images(val_images, resize=64)
test_images = resize_images(test_images, resize=64)

print("#### 크기를 64x64로 resize한 후 shape ####\n")
print("train data's shape : \n{}, {}\n".format(tr_images.shape, tr_labels.shape))
print("validation data's shape : \n{}, {}\n".format(val_images.shape, val_labels.shape))
print("test data's shape : \n{}, {}".format(test_images.shape, test_labels.shape))

---

8. VGG16 생성, 설정, 학습

IMAGE_SIZE=64

model_vgg16 = create_pretrained_model(model_name = "vgg16", verbose = True)
model_vgg16.compile(optimizer=Adam(learning_rate=0.001), loss="categorical_crossentropy", metrics=["accuracy"])

result_vgg16 = model_vgg16.fit(flow_tr_gen, epochs=40, validation_data=flow_val_gen, callbacks=[rlr_cb, ely_cb])

---

9. accuracy 비교

compare_acc(result_mine, result_vgg16)

---

10. loss 비교

compare_loss(result_mine, result_vgg16)

---