[Toolkit] Keras - 鐵達尼號旅客資料及介紹

Source From Here (Ch11-12) 
Preface 
鐵達尼號的沉沒是歷史悲劇. 1912 年 4 月 15 日 鐵達尼在首航時撞上冰山沉沒, 乘客和船員共 2224 人, 造成 1502 人死亡. 這場悲劇震撼國際社會並未船舶制定了更好的安全規章. 鐵達尼號旅客資料完整保留下來. 這裡將利用 MLP (Multiple Layer Perception) 訓練模型來預測每一位乘客的存活率. 

下載鐵達尼號旅客資料集 
資料可以 這裡 下載. 這邊下載後分為以下兩個檔案並置放於目錄 datas 下: 

* titan_train.csv: 訓練資料
* titan_test.csv: 測試資料

STEP1. 使用 Pnadas dataframe 讀取資料 
這邊使用 pandas 的 API:read_csv 讀入資料: 
- ch11_1.py 

view plaincopy to clipboardprint?

1. import numpy as np  
2. import pandas as pd  
3.   
4. TRAIN_FILE_PATH='datas/titan_train.csv'  
5. TEST_FILE_PATH='datas/titan_test.csv'  
6.   
7. train_df = pd.read_csv(TRAIN_FILE_PATH)  
8. test_df = pd.read_csv(TEST_FILE_PATH)  

接著我們可以如下檢視資料: 

>>> from ch11_1 import *
>>> train_df.info()

RangeIndex: 891 entries, 0 to 890
Data columns (total 12 columns):
PassengerId 891 non-null int64
Survived 891 non-null int64 // 是否生存? 0=否; 1=是
Pclass 891 non-null int64 // 艙等. 1=頭等艙; 2=二等艙; 3=三等艙
Name 891 non-null object // 姓名
Sex 891 non-null object // 性別. female/male
Age 714 non-null float64 // 年齡
SibSp 891 non-null int64 // 手足或配偶也在船上人數
Parch 891 non-null int64 // 雙親或子女也在船上人數
Ticket 891 non-null object // 車票號碼
Fare 891 non-null float64 // 旅客費用
Cabin 204 non-null object // 艙位號碼
Embarked 889 non-null object // 登船港口. C=Cherbourg, Q=Queenstown, S=Southampton
dtypes: float64(2), int64(5), object(5)
memory usage: 83.6+ KB

STEP2. 進行資料前處理 
底下透過 DataFrame 上的 [] 運算子取出有興趣的欄位: 

view plaincopy to clipboardprint?

1. # Headers: PassengerId,Survived,Pclass,Name,Sex,Age,SibSp,Parch,Ticket,Fare,Cabin,Embarked  
2. # Filter out 'Ticket', 'PassengerId' and 'Cabin' columns  
3. cols = ['Survived','Pclass','Name','Sex','Age','SibSp','Parch','Fare','Embarked']  
4. train_df = train_df[cols]  
5.   
6. # Show top 2 records  
7. print("\t[Info] Show top 2 records:")  
8. pprint(train_df.as_matrix()[:2])  
9. print("")  

以上欄位還有以下問題, 必須進行前處理: 

* Name: 姓名欄位在預測階段會使用, 但是訓練時不需要, 要移除.
* Age: 有幾筆資料是 null 值, 這邊會將 null 改為平均值.
* Sex: 欄位是文字. 會將轉換成 0, 1 值表示.
* Embarked: 欄位是文字, 這邊會使用 Onehot Encoding 進行轉換.

轉換過程如下: 

view plaincopy to clipboardprint?

1. # Remove column 'Name'  
2. train_df.drop(['Name'], axis=1)  
3.   
4. # Show number of rows with null value  
5. print("\t[Info] Show number of rows with null value:")  
6. print(train_df.isnull().sum())  
7. print("")  
8.   
9. # Fill null with age mean value on 'Age' column  
10. print("\t[Info] Handle null value of Age column...")  
11. age_mean = train_df['Age'].mean()  
12. train_df['Age'] = train_df['Age'].fillna(age_mean)  
13.   
14. # Show number of rows with null value  
15. print("\t[Info] Show number of rows with null value:")  
16. print(train_df.isnull().sum())  
17. print("")  
18.   
19. print("\t[Info] Translate value of column Sex into (0,1)...")  
20. train_df['Sex'] = train_df['Sex'].map({'female':0, 'male':0}).astype(int)  
21.   
22. print("\t[Info] OneHot Encoding on column Embarked...")  
23. train_df = pd.get_dummies(data=train_df, columns=['Embarked'])  
24.   
25. # Show top 2 records  
26. print("\t[Info] Show top 2 records:")  
27. pprint(train_df.as_matrix()[:2])  
28. print("")  

STEP3. 將 dataframe 轉換為 array 

view plaincopy to clipboardprint?

1. ndarray = train_df.values  
2. print("\t[Info] Translate into ndarray(%s) with shape=%s" % (ndarray.__class__, str(ndarray.shape)))  
3. print("\t[Info] Show top 2 records:\n%s\n" % (ndarray[:2]))  
4.   
5. # Separate labels with features  
6. Label = ndarray[:,0]  
7. Features = ndarray[:,1:]  

STEP4. 將 ndarray 特徵欄位進行標準化 
我們將使用 sklearn 的 preprocessing 模組進行特徵值標準化: 

view plaincopy to clipboardprint?

1. # Normalized features  
2. print("\t[Info] Normalized features...")  
3. from sklearn import preprocessing  
4. minmax_scale = preprocessing.MinMaxScaler(feature_range=(0, 1))  
5. scaledFeatures = minmax_scale.fit_transform(Features)  
6. print("\t[Info] Show top 2 records:\n%s\n" % (scaledFeatures[:2]))  

STEP5. 切分資料為訓練資料與測試資料 
我們將以 8:2 比例切分訓練資料與測試資料: 

view plaincopy to clipboardprint?

1. # Splitting data into training/testing part  
2. print("\t[Info] Split data into training/testing part")  
3. msk = np.random.rand(len(scaledFeatures)) < 0.8  
4. trainFeatures = scaledFeatures[msk]  
5. trainLabels = Label[msk]  
6. testFeatures = scaledFeatures[~msk]  
7. testLabels = Label[~msk]  
8. print("\t[Info] Total %d training instances; %d testing instances!" % (trainFeatures.shape[0], testFeatures.shape[0]))  

STEP6. 包裝前處理流程為函式 preprocessData: 

view plaincopy to clipboardprint?

1. def preprocessData(data_df, sRatio=None):  
2.     r'''  
3.     Preprocess data frame  
4.   
5.     @param data_df(DataFrame):  
6.         Training DataFrame  
7.   
8.     @param sRatio(float):  
9.         if splitRation is not None:  
10.             (train_data, train_label, test_data, test_label)  
11.         else:  
12.             (train_data, train_label)  
13.     '''  
14.     # Remove column 'Name'  
15.     data_df = data_df.drop(['Name'], axis=1)  
16.   
17.     # Show number of rows with null value  
18.     print("\t[Info] Show number of rows with null value:")  
19.     print(data_df.isnull().sum())  
20.     print("")  
21.   
22.     # Fill null with age mean value on 'Age' column  
23.     print("\t[Info] Handle null value of Age column...")  
24.     age_mean = data_df['Age'].mean()  
25.     data_df['Age'] = data_df['Age'].fillna(age_mean)  
26.   
27.     # Show number of rows with null value  
28.     print("\t[Info] Show number of rows with null value:")  
29.     print(data_df.isnull().sum())  
30.     print("")  
31.   
32.     print("\t[Info] Translate value of column Sex into (0,1)...")  
33.     data_df['Sex'] = data_df['Sex'].map({'female':0, 'male':0}).astype(int)  
34.   
35.     print("\t[Info] OneHot Encoding on column Embarked...")  
36.     data_df = pd.get_dummies(data=data_df, columns=['Embarked'])  
37.   
38.     # Show top 2 records  
39.     print("\t[Info] Show top 2 records:")  
40.     pprint(data_df.as_matrix()[:2])  
41.     print("")  
42.   
43.     ndarray = data_df.values  
44.     print("\t[Info] Translate into ndarray(%s) with shape=%s" % (ndarray.__class__, str(ndarray.shape)))  
45.     print("\t[Info] Show top 2 records:\n%s\n" % (ndarray[:2]))  
46.   
47.     # Separate labels with features  
48.     Label = ndarray[:,0]  
49.     Features = ndarray[:,1:]  
50.   
51.     # Normalized features  
52.     print("\t[Info] Normalized features...")  
53.     from sklearn import preprocessing  
54.     minmax_scale = preprocessing.MinMaxScaler(feature_range=(0, 1))  
55.     scaledFeatures = minmax_scale.fit_transform(Features)  
56.     print("\t[Info] Show top 2 records:\n%s\n" % (scaledFeatures[:2]))  
57.   
58.     if sRatio:  
59.         # Splitting data into training/testing part  
60.         print("\t[Info] Split data into training/testing part")  
61.         msk = np.random.rand(len(scaledFeatures)) < sRatio  
62.         trainFeatures = scaledFeatures[msk]  
63.         trainLabels = Label[msk]  
64.         testFeatures = scaledFeatures[~msk]  
65.         testLabels = Label[~msk]  
66.         print("\t[Info] Total %d training instances; %d testing instances!" % (trainFeatures.shape[0], testFeatures.shape[0]))  
67.         return (trainFeatures, trainLabels, testFeatures, testLabels)  
68.     else:  
69.         return (scaledFeatures, Label)  

建立模型與訓練 
我們接下來程式碼將建立 MLP (Multiple Layer Perception) 包括: 輸入層 (9 個神經元); 隱藏層一 (40 個神經元); 隱藏層二 (30 個神經元); 輸出層 (1 個神經元) 

STEP1. 建立模型 

view plaincopy to clipboardprint?

1. # Building model  
2. print("\t[Info] Building MLP model")  
3. from keras.models import Sequential  
4. from keras.layers import Dense,Dropout  
5.   
6. model = Sequential()  
7. model.add(Dense(units=40, input_dim=9, kernel_initializer='uniform', activation='relu'))  
8. model.add(Dense(units=30, kernel_initializer='uniform', activation='relu'))  
9. model.add(Dense(units=1, kernel_initializer='uniform', activation='sigmoid'))  
10. print("\t[Info] Show model summary...")  
11. model.summary()  
12. print("")  

STEP2. 進行訓練 

view plaincopy to clipboardprint?

1. # Training  
2. print("\t[Info] Start training...")  
3. model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])  
4. train_history = model.fit(x=train_data, y=train_label, validation_split=0.1, epochs=50, batch_size=30, verbose=2)  

STEP3. 顯示訓練結果 

view plaincopy to clipboardprint?

1. # Show Training result  
2. from utils import *  
3. if isDisplayAvl():  
4.     show_train_history(train_history, 'acc', 'val_acc')  
5.     show_train_history(train_history, 'loss', 'val_loss')  

加入鐵達尼號電影 Jack 與 Rose 的資料 
在鐵達尼電影中男女主角 Jack 與 Rose 是虛擬人物, 我們希望透過剛剛訓練的模型來預測男女主角的生存機率. 以下是根據電影情節所構想的資料: 

* Jack 是 3 等艙; Rose 是頭等艙.
* Jack 是男性; Rose 是女性
* 票價 Jack 是 5; Rose 是 100
* 年齡 Jack 是 23; Rose 是 20

STEP1. 建立 Jack 與 Rose 的資料 
使用 pandas.Series 建立 Jack 與 Rose 資料如下: 

view plaincopy to clipboardprint?

1. Jack = pd.Series([0, 'Jack', 3, 'male', 23, 1, 0, 5.0, 'S'])  
2. Rose = pd.Series([1, 'Rose', 1, 'female', 28, 1, 0, 100.0, 'S'])  
3. JR_df = pd.DataFrame([list(Jack), list(Rose)], columns=['Survived','Name', 'Pclass','Sex','Age','SibSp','Parch','Fare','Embarked'])  
4.   
5. all_df = pd.concat([train_df, JR_df])  
6. print("\t[Info] Show last two records:\n%s\n" % (all_df[-2:]))  

STEP2. 進行預測與顯示結果 

view plaincopy to clipboardprint?

1. print("\t[Info] Making prediction...")  
2. features, labels = preprocessData(all_df)  
3. all_probability = model.predict(features)  
4. all_df.insert(len(all_df.columns), 'probability', all_probability)  
5. print("\t[Info] The prediction of last two records:\n%s\n" % (all_df[-2:]))  
6. print("")  

執行結果: 

完整代碼連結如下: 

* ch11_1.py:

Supplement 
* Apply one-hot encoding to a pandas DataFrame