1. Sequential Api 사용

텐서플로우의 NN 실행 흐름

1. 하이퍼파라미터 설정 — 학습률, 학습 에포크, 배치 크기 등
2. 데이터 파이프라인 구성 — tf.data 사용
3. 신경망 모델 구축 — tf.keras Sequential API 사용
4. 손실 함수 정의 — 교차 엔트로피
5. gradient 계산 — tf.GradientTape 사용
6. 옵티마이저 선택 — Adam optimizer
7. 성능 지표 정의 — 정확도
8. (옵션) 체크포인트 저장 설정 — 학습된 결과 저장
9. 신경망 모델 학습 및 검증

---

모델 이미지

---

0. Import Libraries

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras.utils import to_categorical
import numpy as np
import matplotlib.pyplot as plt
import os

---

1. 하이퍼파라미터 설정

learning_rate = 0.001
training_epochs = 15
batch_size = 100

# 체크포인트를 저장할 디렉터리 생성
cur_dir = os.getcwd()
ckpt_dir_name = "checkpoints"
model_dir_name = "mnist_cnn_seq"

checkpoint_dir = os.path.join(cur_dir, ckpt_dir_name, model_dir_name)
os.makedirs(checkpoint_dir, exist_ok=True)

checkpoint_prefix = os.path.join(checkpoint_dir, model_dir_name)

---

2. 데이터 파이프라인 구성

mnist = keras.datasets.mnist

# 데이터 셋 분리
(train_images, train_labels), (test_images, test_labels) = mnist.load_data()

# 255로 나눠서 0~1로 scaling
train_images = train_images.astype(np.float32) / 255.0
test_images = test_images.astype(np.float32) / 255.0

# 3차원 데이터를 4차원 데이터로 변환
# (batch, width, height) => (batch, width, height, channel)
train_images = np.expand_dims(train_images, axis=-1)
test_images = np.expand_dims(test_images, axis=-1)

# one-hot encoding
train_labels = to_categorical(train_labels, 10)
test_labels = to_categorical(test_labels, 10)

train_dataset = (
    tf.data.Dataset.from_tensor_slices((train_images, train_labels))
    .shuffle(buffer_size=100000)  # training 할 때는 셔플이 필요함
    .batch(batch_size)
)

test_dataset = (
    tf.data.Dataset.from_tensor_slices((test_images, test_labels))
    .batch(batch_size)
)

---

3. 신경망 모델 구축 (Sequential CNN)

def create_model():
    model = keras.Sequential()  # Sequential API 사용 선언

    # 벽돌을 쌓듯 레이어를 쌓아올린다.
    model.add(
        keras.layers.Conv2D(
            filters=32,
            kernel_size=3,
            activation=tf.nn.relu,
            padding="SAME",
            input_shape=(28, 28, 1),
        )
    )
    model.add(keras.layers.MaxPool2D(padding="SAME"))  # 기본 2x2, stride 2

    model.add(
        keras.layers.Conv2D(
            filters=64,
            kernel_size=3,
            activation=tf.nn.relu,
            padding="SAME",
        )
    )
    model.add(keras.layers.MaxPool2D(padding="SAME"))

    model.add(
        keras.layers.Conv2D(
            filters=128,
            kernel_size=3,
            activation=tf.nn.relu,
            padding="SAME",
        )
    )
    model.add(keras.layers.MaxPool2D(padding="SAME"))

    # fc로 들어가기 전 벡터로 펴 주는 단계
    model.add(keras.layers.Flatten())
    model.add(keras.layers.Dense(256, activation=tf.nn.relu))
    model.add(keras.layers.Dropout(0.4))
    model.add(keras.layers.Dense(10))

    return model


model = create_model()
model.summary()

모델 확인

---

4. 손실 함수 정의

def loss_fn(model, images, labels):
    # training=True : dropout 활성화
    logits = model(images, training=True)
    loss = tf.reduce_mean(
        tf.nn.softmax_cross_entropy_with_logits_v2(
            logits=logits,
            labels=labels,
        )
    )
    return loss

---

5. 그래디언트 계산

def grad(model, images, labels):
    with tf.GradientTape() as tape:
        loss = loss_fn(model, images, labels)
    return tape.gradient(loss, model.variables)

---

6. 옵티마이저 선택

optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)

---

7. 성능 지표 정의

def evaluate(model, images, labels):
    logits = model(images, training=False)
    correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(labels, 1))
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
    return accuracy

---

8. 체크포인트 저장 설정

checkpoint = tf.train.Checkpoint(cnn=model)

---

9. 신경망 모델 학습 및 검증

for epoch in range(training_epochs):
    avg_loss = 0.0
    avg_train_acc = 0.0
    avg_test_acc = 0.0
    train_step = 0
    test_step = 0

    # 학습 루프
    for images, labels in train_dataset:
        grads = grad(model, images, labels)
        optimizer.apply_gradients(zip(grads, model.variables))

        loss = loss_fn(model, images, labels)
        acc = evaluate(model, images, labels)

        avg_loss = avg_loss + loss
        avg_train_acc = avg_train_acc + acc
        train_step += 1

    avg_loss = avg_loss / float(train_step)
    avg_train_acc = avg_train_acc / float(train_step)

    # 평가 루프
    for images, labels in test_dataset:
        acc = evaluate(model, images, labels)
        avg_test_acc = avg_test_acc + acc
        test_step += 1

    avg_test_acc = avg_test_acc / float(test_step)

    print(
        "Epoch: ",
        "{}".format(epoch + 1),
        "loss = ",
        "{:.8f}".format(avg_loss),
        "train accuracy = ",
        "{:.4f}".format(avg_train_acc),
        "test accuracy = ",
        "{:.4f}".format(avg_test_acc),
    )

    # 에포크마다 체크포인트 저장
    checkpoint.save(file_prefix=checkpoint_prefix)

---

2. Fucntional Api

Sequential API의 한계

다음과 같은 모델 구조는 Sequential API로 구현할 수 없다.

• 다중 입력 모델
• 다중 출력 모델
• 공유 레이어를 사용하는 모델 (같은 레이어를 여러 번 호출)
• 비순차적 데이터 흐름을 갖는 모델 (예: Residual connections)

⇒ 이런 케이스를 다루기 위해 Functional API를 사용한다.

---

3. 신경망 모델 구축 (Functional API)

Sequential 대신 Functional API로 동일한 CNN을 구성한 예시이다.

def create_model():
    # input layer 선언
    inputs = keras.Input(shape=(28, 28, 1))

    # 이 레이어의 입력(inputs)을 명시해준다.
    conv1 = keras.layers.Conv2D(
        filters=32,
        kernel_size=3,
        padding="SAME",
        activation=tf.nn.relu,
    )(inputs)
    pool1 = keras.layers.MaxPool2D(padding="SAME")(conv1)

    conv2 = keras.layers.Conv2D(
        filters=64,
        kernel_size=3,
        padding="SAME",
        activation=tf.nn.relu,
    )(pool1)
    pool2 = keras.layers.MaxPool2D(padding="SAME")(conv2)

    conv3 = keras.layers.Conv2D(
        filters=128,
        kernel_size=3,
        padding="SAME",
        activation=tf.nn.relu,
    )(pool2)
    pool3 = keras.layers.MaxPool2D(padding="SAME")(conv3)

    pool3_flat = keras.layers.Flatten()(pool3)
    dense4 = keras.layers.Dense(units=256, activation=tf.nn.relu)(pool3_flat)
    drop4 = keras.layers.Dropout(rate=0.4)(dense4)
    logits = keras.layers.Dense(units=10)(drop4)

    # Functional API에서는 입력과 출력을 명시해서 Model을 만든다.
    return keras.Model(inputs=inputs, outputs=logits)


model = create_model()
model.summary()

모델 확인

---

Implementation of Residual Block (Functional API 예시)

ResNet 스타일의 Residual Block을 Functional API로 구현한 예시이다.

# 입력 feature map의 채널 수가 256이라고 가정
inputs = keras.Input(shape=(28, 28, 256))

conv1 = keras.layers.Conv2D(
    filters=64,
    kernel_size=1,
    padding="SAME",
    activation=keras.layers.ReLU(),
)(inputs)

conv2 = keras.layers.Conv2D(
    filters=64,
    kernel_size=3,
    padding="SAME",
    activation=keras.layers.ReLU(),
)(conv1)

conv3 = keras.layers.Conv2D(
    filters=256,
    kernel_size=1,
    padding="SAME",
)(conv2)

# 입력(inputs)과 conv3 출력을 더해 skip connection을 만든다.
add3 = keras.layers.add([conv3, inputs])
relu3 = keras.layers.ReLU()(add3)

residual_block = keras.Model(inputs=inputs, outputs=relu3)

---

3. Subclassing 사용

Model Subclassing

• tf.keras.Model을 상속하여 완전히 커스터마이즈 가능한 모델을 만든다.
• __init__ 메서드에서 레이어를 생성하고, 이를 클래스 인스턴스의 속성으로 설정한다.
• call 메서드에서 순전파(forward pass)를 정의한다.

---

3. 신경망 모델 구축 (Model Subclassing)

class MNISTModel(tf.keras.Model):
    def __init__(self):
        super(MNISTModel, self).__init__()
        self.conv1 = keras.layers.Conv2D(
            filters=32,
            kernel_size=3,
            padding="SAME",
            activation=tf.nn.relu,
        )
        self.pool1 = keras.layers.MaxPool2D(padding="SAME")

        self.conv2 = keras.layers.Conv2D(
            filters=64,
            kernel_size=3,
            padding="SAME",
            activation=tf.nn.relu,
        )
        self.pool2 = keras.layers.MaxPool2D(padding="SAME")

        self.conv3 = keras.layers.Conv2D(
            filters=128,
            kernel_size=3,
            padding="SAME",
            activation=tf.nn.relu,
        )
        self.pool3 = keras.layers.MaxPool2D(padding="SAME")

        self.pool3_flat = keras.layers.Flatten()
        self.dense4 = keras.layers.Dense(units=256, activation=tf.nn.relu)
        self.drop4 = keras.layers.Dropout(rate=0.4)
        self.dense5 = keras.layers.Dense(units=10)

    def call(self, inputs, training=False):
        net = self.conv1(inputs)
        net = self.pool1(net)
        net = self.conv2(net)
        net = self.pool2(net)
        net = self.conv3(net)
        net = self.pool3(net)
        net = self.pool3_flat(net)
        net = self.dense4(net)
        net = self.drop4(net)
        net = self.dense5(net)
        return net


model = MNISTModel()

---

전체 코드

# ============================================
# 0. Import Libraries
# ============================================

import os
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"  # 0: INFO, 1: WARNING, 2: ERROR, 3: FATAL
import tensorflow as tf
tf.get_logger().setLevel('ERROR')
from tensorflow import keras
from tensorflow.keras.utils import to_categorical
import numpy as np
import matplotlib.pyplot as plt

# ============================================
# 1. 하이퍼파라미터 설정
# ============================================
learning_rate = 0.001
training_epochs = 15
batch_size = 100

# 체크포인트를 저장할 디렉터리 생성
cur_dir = os.getcwd()
ckpt_dir_name = "checkpoints"
model_dir_name = "mnist_cnn_seq"

checkpoint_dir = os.path.join(cur_dir, ckpt_dir_name, model_dir_name)
os.makedirs(checkpoint_dir, exist_ok=True)

checkpoint_prefix = os.path.join(checkpoint_dir, model_dir_name)

# ============================================
# 2. 데이터 파이프라인 구성
# ============================================
mnist = keras.datasets.mnist

# 데이터 셋 분리
(train_images, train_labels), (test_images, test_labels) = mnist.load_data()

# 255로 나눠서 0~1로 scaling
train_images = train_images.astype(np.float32) / 255.0
test_images = test_images.astype(np.float32) / 255.0

# 3차원 데이터를 4차원 데이터로 변환
# (batch, width, height) => (batch, width, height, channel)
train_images = np.expand_dims(train_images, axis=-1)
test_images = np.expand_dims(test_images, axis=-1)

# one-hot encoding
train_labels = to_categorical(train_labels, 10)
test_labels = to_categorical(test_labels, 10)

train_dataset = (
    tf.data.Dataset.from_tensor_slices((train_images, train_labels))
    .shuffle(buffer_size=100000)  # training 할 때는 셔플이 필요함
    .batch(batch_size)
)

test_dataset = (
    tf.data.Dataset.from_tensor_slices((test_images, test_labels))
    .batch(batch_size)
)

# ============================================
# 3-1. 신경망 모델 구축 (Sequential CNN)
# ============================================
def create_model():
    model = keras.Sequential()  # Sequential API 사용 선언

    # 벽돌을 쌓듯 레이어를 쌓아올린다.
    model.add(
        keras.layers.Conv2D(
            filters=32,
            kernel_size=3,
            activation=tf.nn.relu,
            padding="same",
            input_shape=(28, 28, 1),
        )
    )
    model.add(keras.layers.MaxPool2D(padding="same"))  # 기본 2x2, stride 2

    model.add(
        keras.layers.Conv2D(
            filters=64,
            kernel_size=3,
            activation=tf.nn.relu,
            padding="same",
        )
    )
    model.add(keras.layers.MaxPool2D(padding="same"))

    model.add(
        keras.layers.Conv2D(
            filters=128,
            kernel_size=3,
            activation=tf.nn.relu,
            padding="same",
        )
    )
    model.add(keras.layers.MaxPool2D(padding="same"))

    # fc로 들어가기 전 벡터로 펴 주는 단계
    model.add(keras.layers.Flatten())
    model.add(keras.layers.Dense(256, activation=tf.nn.relu))
    model.add(keras.layers.Dropout(0.4))
    model.add(keras.layers.Dense(10))

    return model


model = create_model()
model.summary()

# ============================================
# 3-2. 신경망 모델 구축 (Fucntional Api)
# ============================================
def create_model_functional():
    # input layer 선언
    inputs = keras.Input(shape=(28, 28, 1))

    # 이 레이어의 입력(inputs)을 명시해준다.
    conv1 = keras.layers.Conv2D(
        filters=32,
        kernel_size=3,
        padding="same",
        activation=tf.nn.relu,
    )(inputs)
    pool1 = keras.layers.MaxPool2D(padding="same")(conv1)

    conv2 = keras.layers.Conv2D(
        filters=64,
        kernel_size=3,
        padding="same",
        activation=tf.nn.relu,
    )(pool1)
    pool2 = keras.layers.MaxPool2D(padding="same")(conv2)

    conv3 = keras.layers.Conv2D(
        filters=128,
        kernel_size=3,
        padding="same",
        activation=tf.nn.relu,
    )(pool2)
    pool3 = keras.layers.MaxPool2D(padding="same")(conv3)

    pool3_flat = keras.layers.Flatten()(pool3)
    dense4 = keras.layers.Dense(units=256, activation=tf.nn.relu)(pool3_flat)
    drop4 = keras.layers.Dropout(rate=0.4)(dense4)
    logits = keras.layers.Dense(units=10)(drop4)

    # Functional API에서는 입력과 출력을 명시해서 Model을 만든다.
    return keras.Model(inputs=inputs, outputs=logits)


functional_model = create_model_functional()
functional_model.summary()

# ============================================
# 3-3. 신경망 모델 구축 (Model Subclassing)
# ============================================
class MNISTModel(tf.keras.Model):
    def __init__(self):
        super(MNISTModel, self).__init__()
        self.conv1 = keras.layers.Conv2D(
            filters=32,
            kernel_size=3,
            padding="same",
            activation=tf.nn.relu,
        )
        self.pool1 = keras.layers.MaxPool2D(padding="same")

        self.conv2 = keras.layers.Conv2D(
            filters=64,
            kernel_size=3,
            padding="same",
            activation=tf.nn.relu,
        )
        self.pool2 = keras.layers.MaxPool2D(padding="same")

        self.conv3 = keras.layers.Conv2D(
            filters=128,
            kernel_size=3,
            padding="same",
            activation=tf.nn.relu,
        )
        self.pool3 = keras.layers.MaxPool2D(padding="same")

        self.pool3_flat = keras.layers.Flatten()
        self.dense4 = keras.layers.Dense(units=256, activation=tf.nn.relu)
        self.drop4 = keras.layers.Dropout(rate=0.4)
        self.dense5 = keras.layers.Dense(units=10)

    def call(self, inputs, training=False):
        net = self.conv1(inputs)
        net = self.pool1(net)
        net = self.conv2(net)
        net = self.pool2(net)
        net = self.conv3(net)
        net = self.pool3(net)
        net = self.pool3_flat(net)
        net = self.dense4(net)
        net = self.drop4(net)
        net = self.dense5(net)
        return net

model = MNISTModel()

# ============================================
# 4. 손실 함수 정의
# ============================================
def loss_fn(model, images, labels):
    # training=True : dropout 활성화
    logits = model(images, training=True)
    loss = tf.reduce_mean(
        tf.nn.softmax_cross_entropy_with_logits(
            logits=logits,
            labels=labels,
        )
    )
    return loss

# ============================================
# 5. 그래디언트 계산
# ============================================
def grad(model, images, labels):
    with tf.GradientTape() as tape:
        loss = loss_fn(model, images, labels)
    return tape.gradient(loss, model.variables)

# ============================================
# 6. 옵티마이저 선택
# ============================================
# optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)


# ============================================
# 7. 성능 지표 정의
# ============================================
def evaluate(model, images, labels):
    logits = model(images, training=False)
    correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(labels, 1))
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
    return accuracy

# ============================================
# 8. 체크포인트 저장 설정
# ============================================
checkpoint = tf.train.Checkpoint(cnn=model)

# ============================================
# 9. 신경망 모델 학습 및 검증
# ============================================
for epoch in range(training_epochs):
    avg_loss = 0.0
    avg_train_acc = 0.0
    avg_test_acc = 0.0
    train_step = 0
    test_step = 0

    # 학습 루프
    for images, labels in train_dataset:
        grads = grad(model, images, labels)
        optimizer.apply_gradients(zip(grads, model.variables))

        loss = loss_fn(model, images, labels)
        acc = evaluate(model, images, labels)

        avg_loss = avg_loss + loss
        avg_train_acc = avg_train_acc + acc
        train_step += 1

    avg_loss = avg_loss / float(train_step)
    avg_train_acc = avg_train_acc / float(train_step)

    # 평가 루프
    for images, labels in test_dataset:
        acc = evaluate(model, images, labels)
        avg_test_acc = avg_test_acc + acc
        test_step += 1

    avg_test_acc = avg_test_acc / float(test_step)

    print(
        "Epoch: ",
        "{}".format(epoch + 1),
        "loss = ",
        "{:.8f}".format(avg_loss),
        "train accuracy = ",
        "{:.4f}".format(avg_train_acc),
        "test accuracy = ",
        "{:.4f}".format(avg_test_acc),
    )

    # 에포크마다 체크포인트 저장
    checkpoint.save(file_prefix=checkpoint_prefix)

# ============================================
# 10. Residual Block (Functional API 예시)
# ============================================
# 입력 feature map의 채널 수가 256이라고 가정
inputs_res = keras.Input(shape=(28, 28, 256))

conv1 = keras.layers.Conv2D(
    filters=64,
    kernel_size=1,
    padding="same",
    activation=keras.layers.ReLU(),
)(inputs_res)

conv2 = keras.layers.Conv2D(
    filters=64,
    kernel_size=3,
    padding="same",
    activation=keras.layers.ReLU(),
)(conv1)

conv3 = keras.layers.Conv2D(
    filters=256,
    kernel_size=1,
    padding="same",
)(conv2)

# 입력(inputs_res)과 conv3 출력을 더해 skip connection을 만든다.
add3 = keras.layers.add([conv3, inputs_res])
relu3 = keras.layers.ReLU()(add3)

residual_block = keras.Model(inputs=inputs_res, outputs=relu3)

• 결과
  
  
  

---

출처: 모두를 위한 딥러닝 강좌 2
https://www.youtube.com/watch?v=7eldOrjQVi0&list=PLQ28Nx3M4Jrguyuwg4xe9d9t2XE639e5C