230522 Keras Sequential -> Functional type

import tensorflow as tf

import numpy as np

import matplotlib.pyplot as plt

from tensorflow.keras.layers import Input, Dense

from tensorflow.keras.models import Model

from keras.engine.training import optimizer

# mnist 데이터셋 불러오기

tf.keras.datasets.mnist.load_data()

# train, test 데이터 나누기

(x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()

# 데이터값 0~1 사이로 바꾸기

x_train = x_train / 255.

x_test = x_test / 255.

# 1. Sequential 모델

medel = tf.keras.Sequential([

    tf.keras.layers.Flatten(input_shape=x_train.shape[1:]),

    tf.keras.layers.Dense(256, activation="relu"),

    tf.keras.layers.Dense(16, activation="relu"),

    tf.keras.layers.Dense(10, activation="softmax"),

])

# 2. Functional 모델

inputs = Input(shape=x_train.shape[1:], name="inputlayer")

hidden1 = Flatten(input_shape=x_train.shape[1:])(inputs)

hidden2 = Dense(256, activation='relu')(hidden1)

hidden3 = Dense(16, activation='sigmoid')(hidden2)

output = Dense(10, activation="softmax", name="Outputlayer")(hidden3)

# 모델 저장

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

# 모델 요약 (일반)

model.summary()

# 모델 요약 (이미지)

tf.keras.utils.plot_model(model)

# 모델 compile(편집)

model.compile(

    optimizer="adam",

    loss="sparse_categorical_crossentropy", # sparse_ => 원핫인코딩해줌

    # loss="categorical_crossentropy",

    metrics=["accuracy"]

)

 

# callbacks.EarlyStopping => 더 이상 성능 나아지지 않는다 싶으면 훈련 끝냄

early_stopping_callback = tf.keras.callbacks.EarlyStopping(

    # monitor="val_accuracy", mode="max",

    # monitor="val_loss", mode="min"

    monitor="val_accuracy",

    mode = "auto",

    patience=10,

    restore_best_weights=True,

    verbose=1

)

# 모델 fit

history = model.fit(

    x_train,

    y_train,

    validation_split = .25,   # 검증용데이터 비율

    callbacks=[early_stopping_callback],    # 콜백

    batch_size=1024,        # 배치사이즈만큼씩 돌리겠다

    epochs=100,                # 전체 데이터를 에포크회 학습하겠다

    verbose=2

)

 

# 모델 평가

loss, acc = model.evaluate(x_test, y_test) # verbose=0 or 1 or 2

loss, acc

history.history

# history => 매 epoch 마다 저장되어있는 학습이력

# loss : 훈련 손실값

# acc : 훈련 정확도

# val_loss : 검증 손실값

# val_acc : 검증 정확도

# history 시각화

length = len(history.history["accuracy"])

plt.plot(range(length), history.history["accuracy"])

plt.plot(range(length), history.history["val_accuracy"])

plt.show()