【tensorflow2.0】處理結構化資料-titanic生存預測
阿新 • • 發佈：2020-04-05
1、準備資料
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import tensorflow as tf
from tensorflow.keras import models,layers
dftrain_raw = pd.read_csv('./data/titanic/train.csv')
dftest_raw = pd.read_csv('./data/titanic/test.csv')
dftrain_raw.head(10)
部分資料：
相關欄位說明：
- Survived:0代表死亡，1代表存活【y標籤】
- Pclass:乘客所持票類，有三種值(1,2,3) 【轉換成onehot編碼】
- Name:乘客姓名 【捨去】
- Sex:乘客性別 【轉換成bool特徵】
- Age:乘客年齡(有缺失) 【數值特徵，新增“年齡是否缺失”作為輔助特徵】
- SibSp:乘客兄弟姐妹/配偶的個數(整數值) 【數值特徵】
- Parch:乘客父母/孩子的個數(整數值)【數值特徵】
- Ticket:票號(字串)【捨去】
- Fare:乘客所持票的價格(浮點數，0-500不等) 【數值特徵】
- Cabin:乘客所在船艙(有缺失) 【新增“所在船艙是否缺失”作為輔助特徵】
- Embarked:乘客登船港口:S、C、Q(有缺失)【轉換成onehot編碼，四維度 S,C,Q,nan】
2、探索資料
（1）標籤分佈
%matplotlib inline
%config InlineBackend.figure_format = 'png'
ax = dftrain_raw['Survived'].value_counts().plot(kind = 'bar',
figsize = (12,8),fontsize=15,rot = 0)
ax.set_ylabel('Counts',fontsize = 15)
ax.set_xlabel('Survived',fontsize = 15)
plt.show()
（2） 年齡分佈
年齡分佈情況
%matplotlib inline
%config InlineBackend.figure_format = 'png'
ax = dftrain_raw['Age'].plot(kind = 'hist',bins = 20,color= 'purple',
figsize = (12,8),fontsize=15)
ax.set_ylabel('Frequency',fontsize = 15)
ax.set_xlabel('Age',fontsize = 15)
plt.show()（3） 年齡和標籤之間的相關性
%matplotlib inline
%config InlineBackend.figure_format = 'png'
ax = dftrain_raw.query('Survived == 0')['Age'].plot(kind = 'density',
figsize = (12,8),fontsize=15)
dftrain_raw.query('Survived == 1')['Age'].plot(kind = 'density',
figsize = (12,8),fontsize=15)
ax.legend(['Survived==0','Survived==1'],fontsize = 12)
ax.set_ylabel('Density',fontsize = 15)
ax.set_xlabel('Age',fontsize = 15)
plt.show()
3、資料預處理
（1）將Pclass轉換為one-hot編碼
dfresult=pd.DataFrame() #將船票型別轉換為one-hot編碼 dfPclass=pd.get_dummies(dftrain_raw["Pclass"]) #設定列名 dfPclass.columns =['Pclass_'+str(x) for x in dfPclass.columns] dfresult = pd.concat([dfresult,dfPclass],axis = 1) dfresult
（2） 將Sex轉換為One-hot編碼
#Sex dfSex = pd.get_dummies(dftrain_raw['Sex']) dfresult = pd.concat([dfresult,dfSex],axis = 1) dfresult
（3） 用0填充Age列缺失值，並重新定義一列Age_null用來標記缺失值的位置
#將缺失值用0填充
dfresult['Age'] = dftrain_raw['Age'].fillna(0)
#增加一列資料為Age_null，同時將不為0的資料用0，將為0的資料用1表示，也就是標記出現0的位置
dfresult['Age_null'] = pd.isna(dftrain_raw['Age']).astype('int32')
dfresult
（4） 直接拼接SibSp、Parch、Fare
dfresult['SibSp'] = dftrain_raw['SibSp'] dfresult['Parch'] = dftrain_raw['Parch'] dfresult['Fare'] = dftrain_raw['Fare'] dfresult
（5） 標記Cabin缺失的位置
#Carbin
dfresult['Cabin_null'] = pd.isna(dftrain_raw['Cabin']).astype('int32')
dfresult
（6）將Embarked轉換成one-hot編碼
#Embarked #需要注意的引數是dummy_na=True，將缺失值另外標記出來 dfEmbarked = pd.get_dummies(dftrain_raw['Embarked'],dummy_na=True) dfEmbarked.columns = ['Embarked_' + str(x) for x in dfEmbarked.columns] dfresult = pd.concat([dfresult,dfEmbarked],axis = 1) dfresult
最後，我們將上述操作封裝成一個函式：
def preprocessing(dfdata):
dfresult= pd.DataFrame()
#Pclass
dfPclass = pd.get_dummies(dfdata['Pclass'])
dfPclass.columns = ['Pclass_' +str(x) for x in dfPclass.columns ]
dfresult = pd.concat([dfresult,dfPclass],axis = 1)
#Sex
dfSex = pd.get_dummies(dfdata['Sex'])
dfresult = pd.concat([dfresult,dfSex],axis = 1)
#Age
dfresult['Age'] = dfdata['Age'].fillna(0)
dfresult['Age_null'] = pd.isna(dfdata['Age']).astype('int32')
#SibSp,Parch,Fare
dfresult['SibSp'] = dfdata['SibSp']
dfresult['Parch'] = dfdata['Parch']
dfresult['Fare'] = dfdata['Fare']
#Carbin
dfresult['Cabin_null'] = pd.isna(dfdata['Cabin']).astype('int32')
#Embarked
dfEmbarked = pd.get_dummies(dfdata['Embarked'],dummy_na=True)
dfEmbarked.columns = ['Embarked_' + str(x) for x in dfEmbarked.columns]
dfresult = pd.concat([dfresult,dfEmbarked],axis = 1)
return(dfresult)
然後進行資料預處理：
x_train = preprocessing(dftrain_raw)
y_train = dftrain_raw['Survived'].values
x_test = preprocessing(dftest_raw)
y_test = dftest_raw['Survived'].values
print("x_train.shape =", x_train.shape )
print("x_test.shape =", x_test.shape )
x_train.shape = (712, 15)
x_test.shape = (179, 15)
3、使用tensorflow定義模型
使用Keras介面有以下3種方式構建模型：使用Sequential按層順序構建模型，使用函式式API構建任意結構模型，繼承Model基類構建自定義模型。此處選擇使用最簡單的Sequential，按層順序模型。
tf.keras.backend.clear_session() model = models.Sequential() model.add(layers.Dense(20,activation = 'relu',input_shape=(15,))) model.add(layers.Dense(10,activation = 'relu' )) model.add(layers.Dense(1,activation = 'sigmoid' )) model.summary()
4、訓練模型
訓練模型通常有3種方法，內建fit方法，內建train_on_batch方法，以及自定義訓練迴圈。此處我們選擇最常用也最簡單的內建fit方法
# 二分類問題選擇二元交叉熵損失函式
model.compile(optimizer='adam',
loss='binary_crossentropy',
metrics=['AUC'])
history = model.fit(x_train,y_train,
batch_size= 64,
epochs= 30,
validation_split=0.2 #分割一部分訓練資料用於驗證
)
結果：
Epoch 1/30 WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/tensorflow/python/ops/resource_variable_ops.py:1817: calling BaseResourceVariable.__init__ (from tensorflow.python.ops.resource_variable_ops) with constraint is deprecated and will be removed in a future version. Instructions for updating: If using Keras pass *_constraint arguments to layers. 9/9 [==============================] - 0s 30ms/step - loss: 4.3524 - auc: 0.4888 - val_loss: 3.0274 - val_auc: 0.5492 Epoch 2/30 9/9 [==============================] - 0s 6ms/step - loss: 2.7962 - auc: 0.4710 - val_loss: 1.8653 - val_auc: 0.4599 Epoch 3/30 9/9 [==============================] - 0s 6ms/step - loss: 1.6765 - auc: 0.4040 - val_loss: 1.2673 - val_auc: 0.4067 Epoch 4/30 9/9 [==============================] - 0s 7ms/step - loss: 1.1195 - auc: 0.3799 - val_loss: 0.9501 - val_auc: 0.4006 Epoch 5/30 9/9 [==============================] - 0s 6ms/step - loss: 0.8156 - auc: 0.4874 - val_loss: 0.7090 - val_auc: 0.5514 Epoch 6/30 9/9 [==============================] - 0s 5ms/step - loss: 0.6355 - auc: 0.6611 - val_loss: 0.6550 - val_auc: 0.6502 Epoch 7/30 9/9 [==============================] - 0s 6ms/step - loss: 0.6308 - auc: 0.7169 - val_loss: 0.6502 - val_auc: 0.6546 Epoch 8/30 9/9 [==============================] - 0s 6ms/step - loss: 0.6088 - auc: 0.7156 - val_loss: 0.6463 - val_auc: 0.6610 Epoch 9/30 9/9 [==============================] - 0s 6ms/step - loss: 0.6066 - auc: 0.7163 - val_loss: 0.6372 - val_auc: 0.6644 Epoch 10/30 9/9 [==============================] - 0s 6ms/step - loss: 0.5964 - auc: 0.7253 - val_loss: 0.6283 - val_auc: 0.6646 Epoch 11/30 9/9 [==============================] - 0s 7ms/step - loss: 0.5876 - auc: 0.7326 - val_loss: 0.6253 - val_auc: 0.6717 Epoch 12/30 9/9 [==============================] - 0s 6ms/step - loss: 0.5827 - auc: 0.7409 - val_loss: 0.6195 - val_auc: 0.6708 Epoch 13/30 9/9 [==============================] - 0s 6ms/step - loss: 0.5769 - auc: 0.7489 - val_loss: 0.6170 - val_auc: 0.6762 Epoch 14/30 9/9 [==============================] - 0s 6ms/step - loss: 0.5719 - auc: 0.7555 - val_loss: 0.6156 - val_auc: 0.6803 Epoch 15/30 9/9 [==============================] - 0s 6ms/step - loss: 0.5662 - auc: 0.7629 - val_loss: 0.6119 - val_auc: 0.6826 Epoch 16/30 9/9 [==============================] - 0s 6ms/step - loss: 0.5627 - auc: 0.7694 - val_loss: 0.6107 - val_auc: 0.6892 Epoch 17/30 9/9 [==============================] - 0s 6ms/step - loss: 0.5586 - auc: 0.7753 - val_loss: 0.6084 - val_auc: 0.6927 Epoch 18/30 9/9 [==============================] - 0s 6ms/step - loss: 0.5539 - auc: 0.7837 - val_loss: 0.6051 - val_auc: 0.6983 Epoch 19/30 9/9 [==============================] - 0s 7ms/step - loss: 0.5479 - auc: 0.7930 - val_loss: 0.6011 - val_auc: 0.7056 Epoch 20/30 9/9 [==============================] - 0s 9ms/step - loss: 0.5451 - auc: 0.7986 - val_loss: 0.5996 - val_auc: 0.7128 Epoch 21/30 9/9 [==============================] - 0s 7ms/step - loss: 0.5406 - auc: 0.8047 - val_loss: 0.5962 - val_auc: 0.7192 Epoch 22/30 9/9 [==============================] - 0s 6ms/step - loss: 0.5357 - auc: 0.8123 - val_loss: 0.5948 - val_auc: 0.7212 Epoch 23/30 9/9 [==============================] - 0s 6ms/step - loss: 0.5295 - auc: 0.8181 - val_loss: 0.5928 - val_auc: 0.7267 Epoch 24/30 9/9 [==============================] - 0s 6ms/step - loss: 0.5275 - auc: 0.8223 - val_loss: 0.5910 - val_auc: 0.7296 Epoch 25/30 9/9 [==============================] - 0s 6ms/step - loss: 0.5263 - auc: 0.8227 - val_loss: 0.5884 - val_auc: 0.7325 Epoch 26/30 9/9 [==============================] - 0s 7ms/step - loss: 0.5199 - auc: 0.8313 - val_loss: 0.5860 - val_auc: 0.7356 Epoch 27/30 9/9 [==============================] - 0s 6ms/step - loss: 0.5145 - auc: 0.8356 - val_loss: 0.5835 - val_auc: 0.7386 Epoch 28/30 9/9 [==============================] - 0s 6ms/step - loss: 0.5138 - auc: 0.8383 - val_loss: 0.5829 - val_auc: 0.7402 Epoch 29/30 9/9 [==============================] - 0s 7ms/step - loss: 0.5092 - auc: 0.8405 - val_loss: 0.5806 - val_auc: 0.7416 Epoch 30/30 9/9 [==============================] - 0s 6ms/step - loss: 0.5082 - auc: 0.8394 - val_loss: 0.5792 - val_auc: 0.7424
5、評估模型
我們首先評估一下模型在訓練集和驗證集上的效果。
%matplotlib inline
%config InlineBackend.figure_format = 'svg'
import matplotlib.pyplot as plt
def plot_metric(history, metric):
train_metrics = history.history[metric]
val_metrics = history.history['val_'+metric]
epochs = range(1, len(train_metrics) + 1)
plt.plot(epochs, train_metrics, 'bo--')
plt.plot(epochs, val_metrics, 'ro-')
plt.title('Training and validation '+ metric)
plt.xlabel("Epochs")
plt.ylabel(metric)
plt.legend(["train_"+metric, 'val_'+metric])
plt.show()
plot_metric(history,"loss")
plot_metric(history,"auc")
然後看在在測試集上的效果：
model.evaluate(x = x_test,y = y_test)
結果：
6/6 [==============================] - 0s 2ms/step - loss: 0.5286 - auc: 0.7869 [0.5286471247673035, 0.786877453327179]
6、使用模型
（1）預測概率
model.predict(x_test[0:10])
結果：
array([[0.34822357],
[0.4793241 ],
[0.43986577],
[0.7916608 ],
[0.50268507],
[0.536609 ],
[0.29079646],
[0.6085641 ],
[0.34384924],
[0.17756936]], dtype=float32)
（2）預測類別
model.predict_classes(x_test[0:10])
結果：
WARNING:tensorflow:From <ipython-input-36-a161a0a6b51e>:1: Sequential.predict_classes (from tensorflow.python.keras.engine.sequential) is deprecated and will be removed after 2021-01-01.
Instructions for updating:
Please use instead:* `np.argmax(model.predict(x), axis=-1)`, if your model does multi-class classification (e.g. if it uses a `softmax` last-layer activation).* `(model.predict(x) > 0.5).astype("int32")`, if your model does binary classification (e.g. if it uses a `sigmoid` last-layer activation).
array([[0],
[0],
[0],
[1],
[1],
[1],
[0],
[1],
[0],
[0]], dtype=int32)
7、儲存模型
可以使用Keras方式儲存模型，也可以使用TensorFlow原生方式儲存。前者僅僅適合使用Python環境恢復模型，後者則可以跨平臺進行模型部署。推薦使用後一種方式進行儲存
1）使用keras方式儲存
# 儲存模型結構及權重
model.save('./data/keras_model.h5')
del model #刪除現有模型
（1）載入模型
# identical to the previous one
model = models.load_model('./data/keras_model.h5')
model.evaluate(x_test,y_test)
WARNING:tensorflow:Error in loading the saved optimizer state. As a result, your model is starting with a freshly initialized optimizer. 6/6 [==============================] - 0s 2ms/step - loss: 0.5286 - auc_1: 0.7869 [0.5286471247673035, 0.786877453327179]
（2）儲存模型結構和恢復模型結構
# 儲存模型結構 json_str = model.to_json() # 恢復模型結構 model_json = models.model_from_json(json_str)
（3）儲存模型權重
# 儲存模型權重
model.save_weights('./data/keras_model_weight.h5')
（4）恢復模型結構並載入權重
# 恢復模型結構
model_json = models.model_from_json(json_str)
model_json.compile(
optimizer='adam',
loss='binary_crossentropy',
metrics=['AUC']
)
# 載入權重
model_json.load_weights('./data/keras_model_weight.h5')
model_json.evaluate(x_test,y_test)
6/6 [==============================] - 0s 3ms/step - loss: 0.5217 - auc: 0.8123 [0.521678626537323, 0.8122605681419373]
2）tensorflow原生方式
# 儲存權重，該方式僅僅儲存權重張量
model.save_weights('./data/tf_model_weights.ckpt',save_format = "tf")
# 儲存模型結構與模型引數到檔案,該方式儲存的模型具有跨平臺性便於部署
model.save('./data/tf_model_savedmodel', save_format="tf")
print('export saved model.')
model_loaded = tf.keras.models.load_model('./data/tf_model_savedmodel')
model_loaded.evaluate(x_test,y_test)
INFO:tensorflow:Assets written to: ./data/tf_model_savedmodel/assets export saved model. 6/6 [==============================] - 0s 2ms/step - loss: 0.5286 - auc_1: 0.7869 [0.5286471247673035, 0.786877453327179]
參考：
開源電子書地址：https://lyhue1991.github.io/eat_tensorflow2_in_30_days/
GitHub 專案地址：https://github.com/lyhue1991/eat_tensorflow2_in_30_