성능 향상 using shuffle

1. 각종 모듈 호출

import numpy as np
from tensorflow.keras.datasets import cifar10
from tensorflow.keras.utils import to_categorical
from sklearn.model_selection import train_test_split
import matplotlib.pyplot as plt
from tensorflow.keras.models import Model, Sequential
from tensorflow.keras.layers import Input, Conv2D, Activation, MaxPooling2D, Flatten, Dense,\
					                                       Dropout, BatchNormalization
from tensorflow.keras.optimizers import Adam, RMSprop

---

2. 각종 함수 정의

def make_zero_to_one(images, labels) :
    
    images = np.array(images/255., dtype = np.float32)
    labels = np.array(labels, dtype = np.float32)
    
    return images, labels


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


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


def create_model():
    
    input_size = (train_images.shape[1])
    input_tensor = Input(shape = (input_size, input_size, 3)) 

    x = Conv2D(filters = 32, kernel_size = (3, 3), padding = "same")(input_tensor)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)

    x = Conv2D(filters = 32, kernel_size = (3, 3), padding = "same")(x)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)
    x = MaxPooling2D(pool_size = (2, 2))(x)

    x = Conv2D(filters = 64, kernel_size = (3, 3), padding = "same")(x)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)

    x = Conv2D(filters = 64, kernel_size = (3, 3), 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), padding = "same")(x)
    x = BatchNormalization()(x)
    x = Activation("relu")(x)

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

    x = Flatten(name = "flatten")(x)
    x = Dropout(rate = 0.5)(x)
    x = Dense(300, activation = "relu", name = "fc1")(x)
    x = Dropout(rate = 0.3)(x)
    output = Dense(10, activation = "softmax", name = "output")(x)

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


def compare_graph(result_1, result_2) :
    
    fig, axs = plt.subplots(nrows = 1, ncols = 2, figsize = (22, 6))
    
    axs[0].plot(result_1.history["val_accuracy"], label = "before_shuffle")
    axs[0].plot(result_2.history["val_accuracy"], label = "after_shuffle")
    axs[0].set_title("val_accuracy")
    axs[0].set_xlabel("epochs")
    axs[0].set_ylabel("accuracy")
    axs[0].legend()
    
    axs[1].plot(result_1.history["val_loss"], label = "before_shuffle")
    axs[1].plot(result_2.history["val_loss"], label = "after_shuffle")
    axs[1].set_title("val_loss")
    axs[1].set_xlabel("epochs")
    axs[1].set_ylabel("loss")
    axs[1].legend()
    
    plt.show()

---

3. fit 수행 시 shuffle을 적용하지 않았을 경우

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

train_images, train_labels = make_zero_to_one(train_images, train_labels)
test_images, test_labels = make_zero_to_one(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) = \
         sper_tr_val(train_images, train_labels, test_images, test_labels, validation_rate = 0.15)

model_1 = create_model()
model_1.compile(optimizer = Adam(), loss = "categorical_crossentropy", metrics = ["accuracy"])
result_1 = model_1.fit(tr_images, tr_labels, batch_size = 64, epochs = 30, shuffle = False, \
                                                       validation_data= (val_images, val_labels))
                                                       
model_1.evaluate(test_images, test_labels)

---

4. fit 수행 시 shuffle을 적용했을 경우

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

train_images, train_labels = make_zero_to_one(train_images, train_labels)
test_images, test_labels = make_zero_to_one(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) = \
                sper_tr_val(train_images, train_labels, test_images, test_labels, validation_rate = 0.15)

model_2 = create_model()
model_2.compile(optimizer = Adam(), loss = "categorical_crossentropy", metrics = ["accuracy"])
result_2 = model_2.fit(tr_images, tr_labels, batch_size = 64, epochs = 30, shuffle = True, \
                                                         validation_data= (val_images, val_labels))
                                                         
                                                         
model_2.evaluate(test_images, test_labels)

---

5. 비교

compare_graph(result_1, result_2)

---

fit 수행 시 shuffle을 적용했을 때가 비교적 성능이 좋아진 것을 확인할 수 있었다. default값은 True이지만 이미지 처리를 할 때는 shuffle에 대한 조정일 필요할 때가 있다.