18. Artificial Neural Network - Case Studies¶

# 基本環境設定
%matplotlib inline
import matplotlib.pyplot as plt
plt.style.use('seaborn-darkgrid')
plt.rcParams['figure.figsize'] = (10.0, 8.0)

import numpy as np
import pandas as pd
import seaborn as sns

# 工具函式： 畫出 k-fold 訓練歷程記錄的 metrics 曲線
def plot_kfold_metrics(histories):
    metrics_keys = histories[0].model.metrics_names
    fig, ax = plt.subplots(nrows=2, sharex=True, figsize=(12, 8))   
    ax[0].set_title('Training Metrics', fontsize=14)
    ax[0].set_ylabel(metrics_keys[0], fontsize=14)
    ax[1].set_ylabel(metrics_keys[1], fontsize=14)
    ax[1].set_xlabel("Epoch", fontsize=14)
    for k,h in enumerate(histories):
        ax[0].plot(h.history[metrics_keys[0]], label='fold '+str(k))
        ax[1].plot(h.history[metrics_keys[1]], label='fold '+str(k))
    ax[0].legend()
    ax[1].legend()

# 工具函式： 畫出 Kearas 訓練歷程記錄的 metrics 曲線
def plot_metrics(history):
    metrics_keys = history.model.metrics_names
    fig, ax = plt.subplots(nrows=2, sharex=True, figsize=(10, 6))
    ax[0].set_title('Training Metrics', fontsize=14)
    ax[0].set_ylabel(metrics_keys[0], fontsize=14)
    ax[1].set_ylabel(metrics_keys[1], fontsize=14)
    ax[1].set_xlabel("Epoch", fontsize=14)
    ax[0].plot(history.history[metrics_keys[0]], 'r')
    ax[1].plot(history.history[metrics_keys[1]], 'b')

from sklearn.metrics import confusion_matrix
# 工具函式： 畫出分類的混淆矩陣
def plot_confusion_matrix(y, y_pred):
    n_classes = np.unique(y)
    class_labels = ['Class'+str(int(n)) for n in reversed(n_classes)]
    confusion_table = pd.DataFrame(np.fliplr(np.rot90(confusion_matrix(y, y_pred))),
                                   index=class_labels,
                                   columns=class_labels)   
    plt.figure(figsize=(8, 6))
    ax = sns.heatmap(confusion_table, annot=True, fmt='d', cmap='RdYlBu', alpha=0.9, vmin=0)    
    ax.set_title('True label', fontsize=14)
    ax.set_ylabel('Predicted label', fontsize=14)

# 載入案例資料
dfbcw = pd.read_csv('./dataset/BreastCancerWisconsin/data.csv')

# TODO: 改成使用 scikit-learn 內建的
# sklearn.datasets.load_breast_cancer()

# 丟掉 "id" 與 "Unnamed: 32" 欄位
dfbcw.drop(columns=['id', 'Unnamed: 32'], inplace=True)

# 只有 diagnosis 欄位是類別資料，將良性與惡性的類別分別轉為 0 與 1
dfbcw['diagnosis'] = dfbcw['diagnosis'].map({'B':0, 'M':1})

# 分離特徵 X 與目標 Y
bcw_y = dfbcw['diagnosis']
bcw_X = dfbcw.drop(columns=['diagnosis'])

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Activation, BatchNormalization
from tensorflow.keras.regularizers import l1, l2
from tensorflow.keras.optimizers import Adam

def create_dnnlr_model():
    # 建立 Sequential 物件，指定網路結構配置
    model = Sequential([
        # 第一層隱藏層
        Dense(units=48, input_shape=(30,), kernel_initializer='he_normal', kernel_regularizer=l2(0.01)),
        BatchNormalization(),
        Activation('relu'),
        # 第二層隱藏層
        Dense(units=32, kernel_initializer='he_normal', kernel_regularizer=l2(0.01)),
        BatchNormalization(),
        Activation('relu'),
        # 第三層隱藏層
        Dense(units=24, kernel_initializer='he_normal', activation='relu', activity_regularizer=l1(0.01)),
        BatchNormalization(),
        # 輸出層
        Dense(units=1),
        Activation('sigmoid'),
    ])
    # 指定套用學習/訓練的方法
    model.compile(optimizer=Adam(lr=0.0005), loss='binary_crossentropy', metrics=['accuracy'])
    return model

from sklearn.model_selection import StratifiedKFold
from sklearn.preprocessing import StandardScaler

# 使用 3-fold 交叉驗證
skfold = StratifiedKFold(n_splits=3, shuffle=True, random_state=7)
histories = []
scores = []
for k, (train, test) in enumerate(skfold.split(bcw_X, bcw_y)):
    # 標準化
    stdscaler = StandardScaler()
    train_X = stdscaler.fit_transform(bcw_X.iloc[train,:])
    test_X = stdscaler.transform(bcw_X.iloc[test,:])
    train_y = bcw_y.iloc[train]
    test_y = bcw_y.iloc[test]
    # 建立 dnnlr 模型
    dnnlr = create_dnnlr_model()    
    # 指定訓練資料的配置，開始訓練
    history = dnnlr.fit(x=train_X, y=train_y, batch_size=None, epochs=200, verbose=1)
    histories.append(history)
    # 評估驗證集
    score = dnnlr.evaluate(x=test_X, y=test_y, batch_size=None, verbose=1)
    scores.append(score)

from sklearn.preprocessing import StandardScaler
# 標準化
stdscaler = StandardScaler()
X_train = stdscaler.fit_transform(bcw_X)
y_train = bcw_y
# 建立 annlr 模型
dnnlr = create_dnnlr_model()
# 指定套用學習/訓練的方法
dnnlr.compile(optimizer=Adam(lr=0.0005), loss='binary_crossentropy', metrics=['accuracy'])
# 指定訓練資料的配置，開始訓練
history = dnnlr.fit(x=X_train, y=y_train, batch_size=None, epochs=200, verbose=1)

from tensorflow.keras.models import save_model, load_model

# 輸出成檔案
save_model(dnnlr, 'breast_cancer_wisconsin_dnnlr.h5')

# 載入檔案
#bcw_dnnmodel = load_model('./breast_cancer_wisconsin_dnnlr.h5')

# 載入案例資料
dfiris = pd.read_csv('./dataset/Iris/data.csv')

# TODO: 改成使用 scikit-learn 內建的
# sklearn.datasets.load_iris()

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Activation, BatchNormalization
from tensorflow.keras.regularizers import l1, l2
from tensorflow.keras.optimizers import Adam

def create_dnnsc_model():
    # 建立 Sequential 物件，指定網路結構配置
    model = Sequential([
        # 第一層隱藏層
        Dense(units=32, activation='relu',input_shape=(4,)),
        # 第二層隱藏層
        Dense(units=16, activation='relu'),
        # 輸出層
        Dense(units=3, activation='softmax')
    ])
    # 指定套用學習/訓練的方法
    model.compile(optimizer=Adam(lr=0.005), loss='categorical_crossentropy', metrics=['accuracy'])
    return model

from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import cross_val_score
from tensorflow.keras.wrappers.scikit_learn import KerasClassifier

iris_y = dfiris['label'].map({'setosa':0, 'versicolor':1, 'virginica':2})
iris_Y_cat = pd.get_dummies(iris_y)
iris_X = dfiris.drop(columns=['label'])

# 使用 Scikit-Learn estimator 相容的 Keras API
estimator = KerasClassifier(build_fn=create_dnnsc_model,
                            epochs=100,
                            batch_size=None,
                            verbose=1)
# 使用 Scikit-Learn pipeline
dnnsc_pipe = Pipeline(steps=[
    ('normalize', StandardScaler()),
    ('dnnmodel', estimator)
])

scores = cross_val_score(dnnsc_pipe,
                         iris_X,
                         iris_Y_cat,
                         cv=5,
                         n_jobs=1,
                         verbose=3)

from sklearn.preprocessing import StandardScaler

# 建立 DNN 模型
dnnsc = create_dnnsc_model()
dnnsc.summary()

stdscaler = StandardScaler()
history = dnnsc.fit(x=stdscaler.fit_transform(iris_X), y=iris_Y_cat, batch_size=None, epochs=100, verbose=1)