Tensorflow 笔记 Ⅵ——Titanic Keras建模与应用

数据集

数据集记录了泰坦尼克号一部分人的信息，以及其存活率等，众所周知，泰坦尼克号是一场海难，也就造成了人员信息难以调查，所以数据集中具有一些缺失的数据，数据集可以在 Kaggle 下载，也可以点击此处下载

	pclass	survived	name	sex	age	sibsp	parch	ticket	fare	cabin	embarked	boat	body	home.dest
0	1	1	Allen, Miss. Elisabeth Walton	female	29.0000	0	0	24160	211.3375	B5	S	2	NaN	St Louis, MO
1	1	1	Allison, Master. Hudson Trevor	male	0.9167	1	2	113781	151.5500	C22 C26	S	11	NaN	Montreal, PQ / Chesterville, ON
2	1	0	Allison, Miss. Helen Loraine	female	2.0000	1	2	113781	151.5500	C22 C26	S	NaN	NaN	Montreal, PQ / Chesterville, ON
3	1	0	Allison, Mr. Hudson Joshua Creighton	male	30.0000	1	2	113781	151.5500	C22 C26	S	NaN	135.0	Montreal, PQ / Chesterville, ON
4	1	0	Allison, Mrs. Hudson J C (Bessie Waldo Daniels)	female	25.0000	1	2	113781	151.5500	C22 C26	S	NaN	NaN	Montreal, PQ / Chesterville, ON
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
1304	3	0	Zabour, Miss. Hileni	female	14.5000	1	0	2665	14.4542	NaN	C	NaN	328.0	NaN
1305	3	0	Zabour, Miss. Thamine	female	NaN	1	0	2665	14.4542	NaN	C	NaN	NaN	NaN
1306	3	0	Zakarian, Mr. Mapriededer	male	26.5000	0	0	2656	7.2250	NaN	C	NaN	304.0	NaN
1307	3	0	Zakarian, Mr. Ortin	male	27.0000	0	0	2670	7.2250	NaN	C	NaN	NaN	NaN
1308	3	0	Zimmerman, Mr. Leo	male	29.0000	0	0	315082	7.8750	NaN	S	NaN	NaN	NaN

1309 rows × 14 columns

泰坦尼克标签说明

字段	字段说明	数据说明
pclass	舱等级	1头等舱、2二等舱、3三等舱
survival	是否生存	0否、1是
name	姓名	None
sex	性别	Female女性、male男性
age	年龄	None
sibsp	siblings + parents	兄弟姐妹或父母是否同船
parch	parents + children	parents父母、children孩子
ticked	船票号码	None
fare	船票费用	None
cabin	舱位号码	None
embarked	登船港口	C=Cherbourg，Q=Queenstown，S=Southampton
home.dest	家、目的地	home、destination

基本原理

利用了多元线性回归问题的预处理方式，在最后一层输出层划分为分类问题，有关多元线性回归与逻辑回归问题的原理，屈尊移驾这里
和那里，由于此问题属于逻辑回归中的二分类问题，所以激活函数选用了 Sigmod 函数，在上诉的“那里”连接中解释了 Sigmod 函数

模型采用全连接方式，两个隐藏层，一个神经元输出作为预测的概率

Titanic TensorFLow 2.x Keras API 实现

前情函数介绍

在 data 文件夹下，新建一个 xls 文件，内容如下

	int	char	float
0	1	NaN	1.1
1	2	NaN	2.2
2	3	NaN	3.3
3	4	d	NaN
4	5	e	NaN
5	6	NaN	6.6
6	7	NaN	7.7

import pandas as pd


demo_data = pd.read_excel('./data/demo.xls')

demo_data

	int	char	float
0	1	NaN	1.1
1	2	NaN	2.2
2	3	NaN	3.3
3	4	d	NaN
4	5	e	NaN
5	6	NaN	6.6
6	7	NaN	7.7

Pandas 缺失值判断函数 isnull()、isnull().any()、isnull().sum()

isnull() 返回一个 bool 值的 dataframe
isnull().any() 判断特征列是否存在空值
isnull().sum() 获取特征列空值数目

demo_data.isnull()

	int	char	float
0	False	True	False
1	False	True	False
2	False	True	False
3	False	False	True
4	False	False	True
5	False	True	False
6	False	True	False

demo_data.isnull().any()

int      False
char      True
float     True
dtype: bool

demo_data.isnull().sum()

int      0
char     5
float    2
dtype: int64

Pandas 缺失值填充函数 fillna()

demo_data['char'] = demo_data['char'].fillna('A')
demo_data['float'] = demo_data['float'].fillna('5.5')

demo_data

	int	char	float
0	1	A	1.1
1	2	A	2.2
2	3	A	3.3
3	4	d	5.5
4	5	e	5.5
5	6	A	6.6
6	7	A	7.7

Pandas map() 映射函数

map 里面放入字典参数，将对应的键🗡替换成其值，注意 map 需要将所有值全部替换，否则会报错

try:
    demo_data['char'] = demo_data['char'].map({'A': 0, 'd': 1}).astype(int)
except ValueError:
    print('Error rising:Cannot convert non-finite values (NA or inf) to integer')
finally:
    demo_data['char'] = demo_data['char'].map({'A': 0, 'd': 1, 'e': 2}).astype(int)
    print('Mapping finished')

Error rising:Cannot convert non-finite values (NA or inf) to integer
Mapping finished

demo_data

	int	char	float
0	1	0	1.1
1	2	0	2.2
2	3	0	3.3
3	4	1	5.5
4	5	2	5.5
5	6	0	6.6
6	7	0	7.7

Pandas sample() 数据洗牌

sample 用于在原数据中提取数据，并进行洗牌操作，frac 代表提取的比例，为 1 表示 $100％$

shuffle_data_1 = demo_data.sample(frac = 1)

shuffle_data_1

	int	char	float
1	2	0	2.2
0	1	0	1.1
2	3	0	3.3
3	4	1	5.5
5	6	0	6.6
4	5	2	5.5
6	7	0	7.7

shuffle_data_2 = demo_data.sample(frac = 5 / 7)

shuffle_data_2

	int	char	float
6	7	0	7.7
0	1	0	1.1
2	3	0	3.3
5	6	0	6.6
1	2	0	2.2

Pandas drop() 删除数据

drop() 不改变原有数据，返回另一个 dataframe，使用 axis 可以指定行、列

demo_data = demo_data.drop(['char'], axis=1)

demo_data

	int	float
0	1	1.1
1	2	2.2
2	3	3.3
3	4	5.5
4	5	5.5
5	6	6.6
6	7	7.7

Pandas values 转换为 ndarray

经过上面处理，demo_data 中 data_frame 全部变成数字，使用 .values 将 data_frame 转换为 ndarray

nd_array = demo_data.values

print('nd_array:\n', nd_array,
      '\nnd_array type:', type(nd_array))

nd_array:
 [[1 1.1]
 [2 2.2]
 [3 3.3]
 [4 '5.5']
 [5 '5.5']
 [6 6.6]
 [7 7.7]] 
nd_array type: <class 'numpy.ndarray'>

正式开始

导入必要包

import numpy
import pandas as pd
import tensorflow as tf
import urllib.request
from sklearn import preprocessing
import matplotlib.pyplot as plt
import os
import datetime


tf.__version__

'2.0.0'

数据下载

data_url = 'http://biostat.mc.vanderbilt.edu/wiki/pub/Main/DataSets/titanic3.xls'

data_file = './data/titanic3.xls'

if not os.path.exists(data_file):
    operation = urllib.request.urlretrieve(data_url, data_file)
    print('downloading from %s' % data_url)
else:
    print('titanic3.xls is exists in the data directory')

titanic3.xls is exists in the data directory

读取数据

从数据摘要中发现 count 行的每一列数据不等，说明数据具有缺失项

dataframe = pd.read_excel(data_file)
dataframe.describe()

	pclass	survived	age	sibsp	parch	fare	body
count	1309.000000	1309.000000	1046.000000	1309.000000	1309.000000	1308.000000	121.000000
mean	2.294882	0.381971	29.881135	0.498854	0.385027	33.295479	160.809917
std	0.837836	0.486055	14.413500	1.041658	0.865560	51.758668	97.696922
min	1.000000	0.000000	0.166700	0.000000	0.000000	0.000000	1.000000
25%	2.000000	0.000000	21.000000	0.000000	0.000000	7.895800	72.000000
50%	3.000000	0.000000	28.000000	0.000000	0.000000	14.454200	155.000000
75%	3.000000	1.000000	39.000000	1.000000	0.000000	31.275000	256.000000
max	3.000000	1.000000	80.000000	8.000000	9.000000	512.329200	328.000000

dataframe

	pclass	survived	name	sex	age	sibsp	parch	ticket	fare	cabin	embarked	boat	body	home.dest
0	1	1	Allen, Miss. Elisabeth Walton	female	29.0000	0	0	24160	211.3375	B5	S	2	NaN	St Louis, MO
1	1	1	Allison, Master. Hudson Trevor	male	0.9167	1	2	113781	151.5500	C22 C26	S	11	NaN	Montreal, PQ / Chesterville, ON
2	1	0	Allison, Miss. Helen Loraine	female	2.0000	1	2	113781	151.5500	C22 C26	S	NaN	NaN	Montreal, PQ / Chesterville, ON
3	1	0	Allison, Mr. Hudson Joshua Creighton	male	30.0000	1	2	113781	151.5500	C22 C26	S	NaN	135.0	Montreal, PQ / Chesterville, ON
4	1	0	Allison, Mrs. Hudson J C (Bessie Waldo Daniels)	female	25.0000	1	2	113781	151.5500	C22 C26	S	NaN	NaN	Montreal, PQ / Chesterville, ON
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
1304	3	0	Zabour, Miss. Hileni	female	14.5000	1	0	2665	14.4542	NaN	C	NaN	328.0	NaN
1305	3	0	Zabour, Miss. Thamine	female	NaN	1	0	2665	14.4542	NaN	C	NaN	NaN	NaN
1306	3	0	Zakarian, Mr. Mapriededer	male	26.5000	0	0	2656	7.2250	NaN	C	NaN	304.0	NaN
1307	3	0	Zakarian, Mr. Ortin	male	27.0000	0	0	2670	7.2250	NaN	C	NaN	NaN	NaN
1308	3	0	Zimmerman, Mr. Leo	male	29.0000	0	0	315082	7.8750	NaN	S	NaN	NaN	NaN

1309 rows × 14 columns

数据预处理

去掉了 ticked、cabin，将 age、fare，空值用其列均值代替，sex 用0，1代替，embarked 用 S 代替，在将其所在字符值转为数字
注意：你应使用 .copy() 函数来防止严重警告⚠
selected_dataframe = dataframe[selected_cols].copy() ok
selected_dataframe = dataframe[selected_cols] not recommand

selected_cols = ['survived', 'name', 'pclass', 'sex', 'age', 'sibsp', 'parch', 'fare', 'embarked']
selected_dataframe = dataframe[selected_cols].copy()

空值填充

age_mean_value = selected_dataframe['age'].mean()
selected_dataframe['age'] = selected_dataframe['age'].fillna(age_mean_value)

fare_mean_value = selected_dataframe['fare'].mean()
selected_dataframe['fare'] = selected_dataframe['fare'].fillna(fare_mean_value)

selected_dataframe['embarked'] = selected_dataframe['embarked'].fillna('S')

selected_dataframe.describe()

	survived	pclass	age	sibsp	parch	fare
count	1309.000000	1309.000000	1309.000000	1309.000000	1309.000000	1309.000000
mean	0.381971	2.294882	29.881135	0.498854	0.385027	33.295479
std	0.486055	0.837836	12.883199	1.041658	0.865560	51.738879
min	0.000000	1.000000	0.166700	0.000000	0.000000	0.000000
25%	0.000000	2.000000	22.000000	0.000000	0.000000	7.895800
50%	0.000000	3.000000	29.881135	0.000000	0.000000	14.454200
75%	1.000000	3.000000	35.000000	1.000000	0.000000	31.275000
max	1.000000	3.000000	80.000000	8.000000	9.000000	512.329200

数字映射

selected_dataframe['sex'] = selected_dataframe['sex'].map({'female': 0, 'male': 1}).astype(int)
selected_dataframe['embarked'] = selected_dataframe['embarked'].map({'C': 0, 'Q': 1, 'S': 2}).astype(int)

创建训练数据集与标签

删除 name 列

selected_dataframe = selected_dataframe.drop(['name'], axis=1)

selected_dataframe[:3]

	survived	pclass	sex	age	sibsp	parch	fare	embarked
0	1	1	0	29.0000	0	0	211.3375	2
1	1	1	1	0.9167	1	2	151.5500	2
2	0	1	0	2.0000	1	2	151.5500	2

features 代表特征，第 1 到最后一列
label 代表标签，第 0 列

ndarray_data = selected_dataframe.values
features = ndarray_data[:, 1:]
label = ndarray_data[:, 0]

print('features:\n', features,
      '\nlabel:', label)

features:
 [[  1.       0.      29.     ...   0.     211.3375   2.    ]
 [  1.       1.       0.9167 ...   2.     151.55     2.    ]
 [  1.       0.       2.     ...   2.     151.55     2.    ]
 ...
 [  3.       1.      26.5    ...   0.       7.225    0.    ]
 [  3.       1.      27.     ...   0.       7.225    0.    ]
 [  3.       1.      29.     ...   0.       7.875    2.    ]] 
label: [1. 1. 0. ... 0. 0. 0.]

数据标准化

minmax_scale = preprocessing.MinMaxScaler(feature_range=(0, 1))
norm_features = minmax_scale.fit_transform(features)

print('norm_features:\n', norm_features,
      '\nlabel:', label)

norm_features:
 [[0.         0.         0.36116884 ... 0.         0.41250333 1.        ]
 [0.         1.         0.00939458 ... 0.22222222 0.2958059  1.        ]
 [0.         0.         0.0229641  ... 0.22222222 0.2958059  1.        ]
 ...
 [1.         1.         0.32985358 ... 0.         0.01410226 0.        ]
 [1.         1.         0.33611663 ... 0.         0.01410226 0.        ]
 [1.         1.         0.36116884 ... 0.         0.01537098 1.        ]] 
label: [1. 1. 0. ... 0. 0. 0.]

代码重构定义数据预处理函数

def prepare_data(df_data):
    df = df_data.drop(['name'], axis=1)
    age_mean = df['age'].mean()
    df['age'] = df['age'].fillna(age_mean)
    fare_mean = df['fare'].mean()
    df['fare'] = df['fare'].fillna(fare_mean)
    df['sex'] = df['sex'].map({'female':0, 'male':1}).astype(int)
    df['embarked'] = df['embarked'].fillna('S')
    df['embarked'] = df['embarked'].map({'C':0, 'Q':1, 'S':2}).astype(int)

    ndarray_data = df.values

    features = ndarray_data[:, 1:]
    label = ndarray_data[:, 0]

    minmax_scale = preprocessing.MinMaxScaler(feature_range=(0, 1))
    norm_features = minmax_scale.fit_transform(features)

    return norm_features, label

数据读取并洗牌

dataframe = pd.read_excel('./data/titanic3.xls')
selected_cols= ['survived', 'name', 'pclass', 'sex', 'age', 'sibsp', 'parch', 'fare', 'embarked']
selected_dataframe = dataframe[selected_cols].copy()
selected_dataframe = selected_dataframe.sample(frac=1)

x_data, y_data = prepare_data(selected_dataframe)

train_size = int(len(x_data) * 0.8)

x_train = x_data[:train_size]
y_train = y_data[:train_size]

x_test = x_data[train_size:]
y_test = y_data[train_size:]

model = tf.keras.models.Sequential([
    tf.keras.layers.Dense(units=64,
                          input_dim=7,
                          use_bias=True,
                          kernel_initializer='uniform',
                          bias_initializer='zeros',
                          activation='relu'),
    tf.keras.layers.Dropout(rate=0.3),
    tf.keras.layers.Dense(units=32, activation='sigmoid'),
    tf.keras.layers.Dropout(rate=0.3),
    tf.keras.layers.Dense(units=1, activation='sigmoid')
])

model.summary()

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense (Dense)                (None, 64)                512       
_________________________________________________________________
dropout (Dropout)            (None, 64)                0         
_________________________________________________________________
dense_1 (Dense)              (None, 32)                2080      
_________________________________________________________________
dropout_1 (Dropout)          (None, 32)                0         
_________________________________________________________________
dense_2 (Dense)              (None, 1)                 33        
=================================================================
Total params: 2,625
Trainable params: 2,625
Non-trainable params: 0
_________________________________________________________________

model.compile(optimizer=tf.keras.optimizers.Adam(0.003),
              loss='binary_crossentropy',
              metrics=['accuracy'])

train_history = model.fit(
                    x=x_train,
                    y=y_train,
                    validation_split=0.2,
                    epochs=100,
                    batch_size=40,
                    verbose=1)

Train on 837 samples, validate on 210 samples
Epoch 1/100
837/837 [==============================] - 2s 2ms/sample - loss: 0.6968 - accuracy: 0.5412 - val_loss: 0.5988 - val_accuracy: 0.6571
Epoch 2/100
837/837 [==============================] - 0s 147us/sample - loss: 0.6342 - accuracy: 0.6404 - val_loss: 0.5504 - val_accuracy: 0.7571
Epoch 3/100
837/837 [==============================] - 0s 153us/sample - loss: 0.5555 - accuracy: 0.7276 - val_loss: 0.4904 - val_accuracy: 0.8143
Epoch 4/100
837/837 [==============================] - 0s 148us/sample - loss: 0.5088 - accuracy: 0.7766 - val_loss: 0.4638 - val_accuracy: 0.8095
Epoch 5/100
837/837 [==============================] - 0s 137us/sample - loss: 0.4948 - accuracy: 0.7814 - val_loss: 0.4526 - val_accuracy: 0.8095
Epoch 6/100
837/837 [==============================] - 0s 124us/sample - loss: 0.4832 - accuracy: 0.7897 - val_loss: 0.4526 - val_accuracy: 0.8000
Epoch 7/100
837/837 [==============================] - 0s 140us/sample - loss: 0.4672 - accuracy: 0.7861 - val_loss: 0.4508 - val_accuracy: 0.8000
Epoch 8/100
837/837 [==============================] - 0s 142us/sample - loss: 0.4707 - accuracy: 0.7897 - val_loss: 0.4431 - val_accuracy: 0.8190
Epoch 9/100
837/837 [==============================] - 0s 117us/sample - loss: 0.4760 - accuracy: 0.8005 - val_loss: 0.4452 - val_accuracy: 0.8000
Epoch 10/100
837/837 [==============================] - 0s 121us/sample - loss: 0.4568 - accuracy: 0.8017 - val_loss: 0.4414 - val_accuracy: 0.7952
Epoch 11/100
837/837 [==============================] - 0s 132us/sample - loss: 0.4533 - accuracy: 0.8100 - val_loss: 0.4473 - val_accuracy: 0.7952
Epoch 12/100
837/837 [==============================] - 0s 122us/sample - loss: 0.4624 - accuracy: 0.7933 - val_loss: 0.4527 - val_accuracy: 0.8000
Epoch 13/100
837/837 [==============================] - 0s 119us/sample - loss: 0.4452 - accuracy: 0.8088 - val_loss: 0.4455 - val_accuracy: 0.8000
Epoch 14/100
837/837 [==============================] - 0s 122us/sample - loss: 0.4564 - accuracy: 0.7993 - val_loss: 0.4430 - val_accuracy: 0.7952
Epoch 15/100
837/837 [==============================] - 0s 161us/sample - loss: 0.4722 - accuracy: 0.8005 - val_loss: 0.4404 - val_accuracy: 0.8000
Epoch 16/100
837/837 [==============================] - 0s 141us/sample - loss: 0.4660 - accuracy: 0.8065 - val_loss: 0.4444 - val_accuracy: 0.8048
Epoch 17/100
837/837 [==============================] - 0s 124us/sample - loss: 0.4551 - accuracy: 0.8196 - val_loss: 0.4392 - val_accuracy: 0.8000
Epoch 18/100
837/837 [==============================] - 0s 118us/sample - loss: 0.4589 - accuracy: 0.8053 - val_loss: 0.4472 - val_accuracy: 0.8000
Epoch 19/100
837/837 [==============================] - 0s 116us/sample - loss: 0.4559 - accuracy: 0.8029 - val_loss: 0.4402 - val_accuracy: 0.8048
Epoch 20/100
837/837 [==============================] - 0s 122us/sample - loss: 0.4479 - accuracy: 0.8124 - val_loss: 0.4398 - val_accuracy: 0.7952
Epoch 21/100
837/837 [==============================] - 0s 122us/sample - loss: 0.4510 - accuracy: 0.8148 - val_loss: 0.4373 - val_accuracy: 0.8000
Epoch 22/100
837/837 [==============================] - 0s 126us/sample - loss: 0.4537 - accuracy: 0.8065 - val_loss: 0.4361 - val_accuracy: 0.8048
Epoch 23/100
837/837 [==============================] - 0s 119us/sample - loss: 0.4591 - accuracy: 0.8088 - val_loss: 0.4423 - val_accuracy: 0.7905
Epoch 24/100
837/837 [==============================] - 0s 119us/sample - loss: 0.4482 - accuracy: 0.8088 - val_loss: 0.4429 - val_accuracy: 0.7952
Epoch 25/100
837/837 [==============================] - 0s 121us/sample - loss: 0.4556 - accuracy: 0.8053 - val_loss: 0.4376 - val_accuracy: 0.8000
Epoch 26/100
837/837 [==============================] - 0s 128us/sample - loss: 0.4536 - accuracy: 0.8112 - val_loss: 0.4382 - val_accuracy: 0.8000
Epoch 27/100
837/837 [==============================] - 0s 119us/sample - loss: 0.4537 - accuracy: 0.7969 - val_loss: 0.4483 - val_accuracy: 0.8000
Epoch 28/100
837/837 [==============================] - 0s 120us/sample - loss: 0.4420 - accuracy: 0.8148 - val_loss: 0.4442 - val_accuracy: 0.7857
Epoch 29/100
837/837 [==============================] - 0s 125us/sample - loss: 0.4462 - accuracy: 0.8053 - val_loss: 0.4371 - val_accuracy: 0.7905
Epoch 30/100
837/837 [==============================] - 0s 125us/sample - loss: 0.4550 - accuracy: 0.8124 - val_loss: 0.4387 - val_accuracy: 0.7905
Epoch 31/100
837/837 [==============================] - 0s 139us/sample - loss: 0.4421 - accuracy: 0.8088 - val_loss: 0.4406 - val_accuracy: 0.7857
Epoch 32/100
837/837 [==============================] - 0s 122us/sample - loss: 0.4525 - accuracy: 0.8112 - val_loss: 0.4384 - val_accuracy: 0.7905
Epoch 33/100
837/837 [==============================] - 0s 126us/sample - loss: 0.4459 - accuracy: 0.8100 - val_loss: 0.4384 - val_accuracy: 0.7905
Epoch 34/100
837/837 [==============================] - 0s 133us/sample - loss: 0.4338 - accuracy: 0.8065 - val_loss: 0.4442 - val_accuracy: 0.7857
Epoch 35/100
837/837 [==============================] - 0s 143us/sample - loss: 0.4419 - accuracy: 0.8065 - val_loss: 0.4405 - val_accuracy: 0.7905
Epoch 36/100
837/837 [==============================] - 0s 137us/sample - loss: 0.4461 - accuracy: 0.8053 - val_loss: 0.4362 - val_accuracy: 0.7857
Epoch 37/100
837/837 [==============================] - 0s 118us/sample - loss: 0.4414 - accuracy: 0.8148 - val_loss: 0.4479 - val_accuracy: 0.7810
Epoch 38/100
837/837 [==============================] - 0s 120us/sample - loss: 0.4382 - accuracy: 0.8136 - val_loss: 0.4365 - val_accuracy: 0.7905
Epoch 39/100
837/837 [==============================] - 0s 125us/sample - loss: 0.4356 - accuracy: 0.8184 - val_loss: 0.4488 - val_accuracy: 0.7810
Epoch 40/100
837/837 [==============================] - 0s 143us/sample - loss: 0.4383 - accuracy: 0.8184 - val_loss: 0.4375 - val_accuracy: 0.7905
Epoch 41/100
837/837 [==============================] - 0s 150us/sample - loss: 0.4347 - accuracy: 0.8124 - val_loss: 0.4451 - val_accuracy: 0.7810
Epoch 42/100
837/837 [==============================] - 0s 143us/sample - loss: 0.4573 - accuracy: 0.8112 - val_loss: 0.4411 - val_accuracy: 0.7857
Epoch 43/100
837/837 [==============================] - 0s 119us/sample - loss: 0.4358 - accuracy: 0.8136 - val_loss: 0.4379 - val_accuracy: 0.7905
Epoch 44/100
837/837 [==============================] - 0s 164us/sample - loss: 0.4453 - accuracy: 0.8160 - val_loss: 0.4458 - val_accuracy: 0.7810
Epoch 45/100
837/837 [==============================] - 0s 126us/sample - loss: 0.4401 - accuracy: 0.8076 - val_loss: 0.4405 - val_accuracy: 0.7952
Epoch 46/100
837/837 [==============================] - 0s 142us/sample - loss: 0.4364 - accuracy: 0.8160 - val_loss: 0.4465 - val_accuracy: 0.7810
Epoch 47/100
837/837 [==============================] - 0s 129us/sample - loss: 0.4311 - accuracy: 0.8184 - val_loss: 0.4386 - val_accuracy: 0.8000
Epoch 48/100
837/837 [==============================] - 0s 117us/sample - loss: 0.4377 - accuracy: 0.8160 - val_loss: 0.4444 - val_accuracy: 0.7857
Epoch 49/100
837/837 [==============================] - 0s 133us/sample - loss: 0.4546 - accuracy: 0.7957 - val_loss: 0.4432 - val_accuracy: 0.7810
Epoch 50/100
837/837 [==============================] - 0s 145us/sample - loss: 0.4403 - accuracy: 0.8208 - val_loss: 0.4436 - val_accuracy: 0.7905
Epoch 51/100
837/837 [==============================] - 0s 144us/sample - loss: 0.4259 - accuracy: 0.8148 - val_loss: 0.4374 - val_accuracy: 0.7952
Epoch 52/100
837/837 [==============================] - 0s 155us/sample - loss: 0.4300 - accuracy: 0.8160 - val_loss: 0.4411 - val_accuracy: 0.7857
Epoch 53/100
837/837 [==============================] - 0s 160us/sample - loss: 0.4381 - accuracy: 0.8136 - val_loss: 0.4432 - val_accuracy: 0.7905
Epoch 54/100
837/837 [==============================] - 0s 136us/sample - loss: 0.4290 - accuracy: 0.8256 - val_loss: 0.4414 - val_accuracy: 0.7810
Epoch 55/100
837/837 [==============================] - 0s 148us/sample - loss: 0.4360 - accuracy: 0.8160 - val_loss: 0.4385 - val_accuracy: 0.7952
Epoch 56/100
837/837 [==============================] - 0s 114us/sample - loss: 0.4364 - accuracy: 0.8232 - val_loss: 0.4415 - val_accuracy: 0.7810
Epoch 57/100
837/837 [==============================] - 0s 127us/sample - loss: 0.4364 - accuracy: 0.8076 - val_loss: 0.4397 - val_accuracy: 0.7952
Epoch 58/100
837/837 [==============================] - 0s 122us/sample - loss: 0.4370 - accuracy: 0.8148 - val_loss: 0.4378 - val_accuracy: 0.7905
Epoch 59/100
837/837 [==============================] - 0s 139us/sample - loss: 0.4435 - accuracy: 0.8088 - val_loss: 0.4444 - val_accuracy: 0.7810
Epoch 60/100
837/837 [==============================] - 0s 122us/sample - loss: 0.4354 - accuracy: 0.8172 - val_loss: 0.4372 - val_accuracy: 0.7952
Epoch 61/100
837/837 [==============================] - 0s 137us/sample - loss: 0.4375 - accuracy: 0.8112 - val_loss: 0.4420 - val_accuracy: 0.7857
Epoch 62/100
837/837 [==============================] - 0s 135us/sample - loss: 0.4307 - accuracy: 0.8136 - val_loss: 0.4392 - val_accuracy: 0.7905
Epoch 63/100
837/837 [==============================] - 0s 158us/sample - loss: 0.4362 - accuracy: 0.8160 - val_loss: 0.4406 - val_accuracy: 0.7952
Epoch 64/100
837/837 [==============================] - 0s 124us/sample - loss: 0.4405 - accuracy: 0.8124 - val_loss: 0.4440 - val_accuracy: 0.7810
Epoch 65/100
837/837 [==============================] - 0s 119us/sample - loss: 0.4298 - accuracy: 0.8232 - val_loss: 0.4386 - val_accuracy: 0.8000
Epoch 66/100
837/837 [==============================] - 0s 108us/sample - loss: 0.4284 - accuracy: 0.8065 - val_loss: 0.4446 - val_accuracy: 0.7905
Epoch 67/100
837/837 [==============================] - 0s 129us/sample - loss: 0.4340 - accuracy: 0.8148 - val_loss: 0.4438 - val_accuracy: 0.7905
Epoch 68/100
837/837 [==============================] - 0s 114us/sample - loss: 0.4389 - accuracy: 0.8088 - val_loss: 0.4381 - val_accuracy: 0.8000
Epoch 69/100
837/837 [==============================] - 0s 124us/sample - loss: 0.4338 - accuracy: 0.8220 - val_loss: 0.4386 - val_accuracy: 0.8000
Epoch 70/100
837/837 [==============================] - 0s 115us/sample - loss: 0.4323 - accuracy: 0.8184 - val_loss: 0.4423 - val_accuracy: 0.7952
Epoch 71/100
837/837 [==============================] - 0s 122us/sample - loss: 0.4237 - accuracy: 0.8232 - val_loss: 0.4406 - val_accuracy: 0.7905
Epoch 72/100
837/837 [==============================] - 0s 137us/sample - loss: 0.4355 - accuracy: 0.8208 - val_loss: 0.4437 - val_accuracy: 0.7905
Epoch 73/100
837/837 [==============================] - 0s 116us/sample - loss: 0.4362 - accuracy: 0.8196 - val_loss: 0.4364 - val_accuracy: 0.7905
Epoch 74/100
837/837 [==============================] - 0s 128us/sample - loss: 0.4293 - accuracy: 0.8196 - val_loss: 0.4445 - val_accuracy: 0.7810
Epoch 75/100
837/837 [==============================] - 0s 124us/sample - loss: 0.4252 - accuracy: 0.8184 - val_loss: 0.4400 - val_accuracy: 0.7905
Epoch 76/100
837/837 [==============================] - 0s 119us/sample - loss: 0.4335 - accuracy: 0.8256 - val_loss: 0.4470 - val_accuracy: 0.7810
Epoch 77/100
837/837 [==============================] - 0s 119us/sample - loss: 0.4284 - accuracy: 0.8184 - val_loss: 0.4384 - val_accuracy: 0.8000
Epoch 78/100
837/837 [==============================] - 0s 147us/sample - loss: 0.4398 - accuracy: 0.8136 - val_loss: 0.4412 - val_accuracy: 0.7905
Epoch 79/100
837/837 [==============================] - 0s 129us/sample - loss: 0.4339 - accuracy: 0.8160 - val_loss: 0.4454 - val_accuracy: 0.7810
Epoch 80/100
837/837 [==============================] - 0s 127us/sample - loss: 0.4286 - accuracy: 0.8160 - val_loss: 0.4397 - val_accuracy: 0.7905
Epoch 81/100
837/837 [==============================] - 0s 120us/sample - loss: 0.4315 - accuracy: 0.8220 - val_loss: 0.4393 - val_accuracy: 0.7905
Epoch 82/100
837/837 [==============================] - 0s 138us/sample - loss: 0.4263 - accuracy: 0.8184 - val_loss: 0.4415 - val_accuracy: 0.7905
Epoch 83/100
837/837 [==============================] - 0s 136us/sample - loss: 0.4298 - accuracy: 0.8208 - val_loss: 0.4405 - val_accuracy: 0.8048
Epoch 84/100
837/837 [==============================] - 0s 129us/sample - loss: 0.4341 - accuracy: 0.8112 - val_loss: 0.4377 - val_accuracy: 0.7952
Epoch 85/100
837/837 [==============================] - 0s 127us/sample - loss: 0.4325 - accuracy: 0.8100 - val_loss: 0.4432 - val_accuracy: 0.8000
Epoch 86/100
837/837 [==============================] - 0s 132us/sample - loss: 0.4277 - accuracy: 0.8124 - val_loss: 0.4415 - val_accuracy: 0.7857
Epoch 87/100
837/837 [==============================] - 0s 113us/sample - loss: 0.4274 - accuracy: 0.8196 - val_loss: 0.4427 - val_accuracy: 0.7905
Epoch 88/100
837/837 [==============================] - 0s 112us/sample - loss: 0.4243 - accuracy: 0.8280 - val_loss: 0.4400 - val_accuracy: 0.7905
Epoch 89/100
837/837 [==============================] - 0s 118us/sample - loss: 0.4280 - accuracy: 0.8220 - val_loss: 0.4418 - val_accuracy: 0.7952
Epoch 90/100
837/837 [==============================] - 0s 112us/sample - loss: 0.4340 - accuracy: 0.8208 - val_loss: 0.4409 - val_accuracy: 0.8000
Epoch 91/100
837/837 [==============================] - 0s 121us/sample - loss: 0.4298 - accuracy: 0.8136 - val_loss: 0.4403 - val_accuracy: 0.8000
Epoch 92/100
837/837 [==============================] - 0s 122us/sample - loss: 0.4275 - accuracy: 0.8208 - val_loss: 0.4409 - val_accuracy: 0.7952
Epoch 93/100
837/837 [==============================] - 0s 136us/sample - loss: 0.4228 - accuracy: 0.8244 - val_loss: 0.4394 - val_accuracy: 0.8095
Epoch 94/100
837/837 [==============================] - 0s 119us/sample - loss: 0.4313 - accuracy: 0.8208 - val_loss: 0.4434 - val_accuracy: 0.8000
Epoch 95/100
837/837 [==============================] - 0s 130us/sample - loss: 0.4277 - accuracy: 0.8196 - val_loss: 0.4365 - val_accuracy: 0.8095
Epoch 96/100
837/837 [==============================] - 0s 118us/sample - loss: 0.4273 - accuracy: 0.8220 - val_loss: 0.4383 - val_accuracy: 0.8000
Epoch 97/100
837/837 [==============================] - 0s 113us/sample - loss: 0.4311 - accuracy: 0.8124 - val_loss: 0.4373 - val_accuracy: 0.8095
Epoch 98/100
837/837 [==============================] - 0s 123us/sample - loss: 0.4221 - accuracy: 0.8327 - val_loss: 0.4419 - val_accuracy: 0.7952
Epoch 99/100
837/837 [==============================] - 0s 124us/sample - loss: 0.4378 - accuracy: 0.8196 - val_loss: 0.4380 - val_accuracy: 0.8048
Epoch 100/100
837/837 [==============================] - 0s 122us/sample - loss: 0.4238 - accuracy: 0.8232 - val_loss: 0.4451 - val_accuracy: 0.7857

训练可视化

fig = plt.gcf()
fig.set_size_inches(10, 5)
ax1 = fig.add_subplot(111)
ax1.set_title('Train and Validation Picture')
ax1.set_ylabel('Loss value')
line1, = ax1.plot(train_history.history['loss'], color=(0.5, 0.5, 1.0), label='Loss train')
line2, = ax1.plot(train_history.history['val_loss'], color=(0.5, 1.0, 0.5), label='Loss valid')
ax2 = ax1.twinx()
ax2.set_ylabel('Accuracy value')
line3, = ax2.plot(train_history.history['accuracy'], color=(0.5, 0.5, 0.5), label='Accuracy train')
line4, = ax2.plot(train_history.history['val_accuracy'], color=(1, 0, 0), label='Accuracy valid')
plt.legend(handles=(line1, line2, line3, line4), loc='best')
plt.show()

测试模型

test_loss, test_acc = model.evaluate(x_test, y_test, verbose=2)
print('test_loss:', test_loss,
      '\ntest_acc:', test_acc,
      '\nmetrics_names:', model.metrics_names)

262/1 - 0s - loss: 0.3581 - accuracy: 0.7672
test_loss: 0.48995060536242624 
test_acc: 0.76717556 
metrics_names: ['loss', 'accuracy']

使用 Jack & Rose 测试

Jack_info = [0, 'Jack', 3, 'male', 23, 1, 0, 5.0000, 'S']
Rose_info = [1, 'Rose', 1, 'female', 20, 1, 0, 100.0000, 'S']

new_passenger_pd = pd.DataFrame([Jack_info, Rose_info], columns=selected_cols)
all_passenger_pd = selected_dataframe.append(new_passenger_pd)

pred = model.predict(prepare_data(all_passenger_pd)[0])

print('Rose survived probability:', pred[-1:][0][0],
      '\nJack survived probability:', pred[-2:][0][0])

Rose survived probability: 0.96711206 
Jack survived probability: 0.12514974

数据插入最后一列生存概率

all_passenger_pd.insert(len(all_passenger_pd.columns), 'surv_prob', pred)

all_passenger_pd

	survived	name	pclass	sex	age	sibsp	parch	fare	embarked	surv_prob
75	0	Colley, Mr. Edward Pomeroy	1	male	47.0	0	0	25.5875	S	0.221973
321	0	Wright, Mr. George	1	male	62.0	0	0	26.5500	S	0.194789
712	0	Celotti, Mr. Francesco	3	male	24.0	0	0	8.0500	S	0.130646
345	0	Berriman, Mr. William John	2	male	23.0	0	0	13.0000	S	0.211936
1298	0	Wittevrongel, Mr. Camille	3	male	36.0	0	0	9.5000	S	0.106074
...	...	...	...	...	...	...	...	...	...	...
10	0	Astor, Col. John Jacob	1	male	47.0	1	0	227.5250	C	0.213140
434	1	Hart, Miss. Eva Miriam	2	female	7.0	0	2	26.2500	S	0.881991
690	0	Brobeck, Mr. Karl Rudolf	3	male	22.0	0	0	7.7958	S	0.136104
0	0	Jack	3	male	23.0	1	0	5.0000	S	0.125150
1	1	Rose	1	female	20.0	1	0	100.0000	S	0.967112

1311 rows × 10 columns

form =pd.DataFrame(columns=[column for column in all_passenger_pd], data=all_passenger_pd)
form.to_excel('./data/result.xls', encoding='utf-8', index=None, header=True)

加入回调

def prepare_data(df_data):
    df = df_data.drop(['name'], axis=1)
    age_mean = df['age'].mean()
    df['age'] = df['age'].fillna(age_mean)
    fare_mean = df['fare'].mean()
    df['fare'] = df['fare'].fillna(fare_mean)
    df['sex'] = df['sex'].map({'female':0, 'male':1}).astype(int)
    df['embarked'] = df['embarked'].fillna('S')
    df['embarked'] = df['embarked'].map({'C':0, 'Q':1, 'S':2}).astype(int)

    ndarray_data = df.values

    features = ndarray_data[:, 1:]
    label = ndarray_data[:, 0]

    minmax_scale = preprocessing.MinMaxScaler(feature_range=(0, 1))
    norm_features = minmax_scale.fit_transform(features)

    return norm_features, label

dataframe = pd.read_excel('./data/titanic3.xls')
selected_cols= ['survived', 'name', 'pclass', 'sex', 'age', 'sibsp', 'parch', 'fare', 'embarked']
selected_dataframe = dataframe[selected_cols].copy()
selected_dataframe = selected_dataframe.sample(frac=1)

x_data, y_data = prepare_data(selected_dataframe)

train_size = int(len(x_data) * 0.8)

x_train = x_data[:train_size]
y_train = y_data[:train_size]

x_test = x_data[train_size:]
y_test = y_data[train_size:]

model = tf.keras.models.Sequential([
    tf.keras.layers.Dense(units=256,
                          input_dim=7,
                          use_bias=True,
                          kernel_initializer='uniform',
                          bias_initializer='zeros',
                          activation='relu'),
    tf.keras.layers.Dropout(rate=0.3),
    tf.keras.layers.Dense(units=128, activation='sigmoid'),
    tf.keras.layers.Dropout(rate=0.3),
    tf.keras.layers.Dense(units=64, activation='sigmoid'),
    tf.keras.layers.Dropout(rate=0.3),
    tf.keras.layers.Dense(units=32, activation='sigmoid'),
    tf.keras.layers.Dropout(rate=0.3),
    tf.keras.layers.Dense(units=1, activation='sigmoid')
])

model.summary()

Model: "sequential_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_3 (Dense)              (None, 256)               2048      
_________________________________________________________________
dropout_2 (Dropout)          (None, 256)               0         
_________________________________________________________________
dense_4 (Dense)              (None, 128)               32896     
_________________________________________________________________
dropout_3 (Dropout)          (None, 128)               0         
_________________________________________________________________
dense_5 (Dense)              (None, 64)                8256      
_________________________________________________________________
dropout_4 (Dropout)          (None, 64)                0         
_________________________________________________________________
dense_6 (Dense)              (None, 32)                2080      
_________________________________________________________________
dropout_5 (Dropout)          (None, 32)                0         
_________________________________________________________________
dense_7 (Dense)              (None, 1)                 33        
=================================================================
Total params: 45,313
Trainable params: 45,313
Non-trainable params: 0
_________________________________________________________________

model.compile(optimizer=tf.keras.optimizers.Adam(0.003),
              loss='binary_crossentropy',
              metrics=['accuracy'])

log_dir = os.path.join(
    'logs2.x',
    'train',
    'plugins',
    'profile',
    datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S'))

checkpoint_path = './checkpoint2.x/Titanic.{epoch:02d}.ckpt'
if not os.path.exists('./checkpoint2.x'):
    os.mkdir('./checkpoint2.x')

callbacks = [tf.keras.callbacks.TensorBoard(log_dir=log_dir,
                         histogram_freq=2),
       tf.keras.callbacks.ModelCheckpoint(filepath=checkpoint_path,
                         save_weights_only=True,
                         verbose=1,
                         period=5)]

WARNING:tensorflow:`period` argument is deprecated. Please use `save_freq` to specify the frequency in number of samples seen.

train_history = model.fit(x=x_train, y=y_train,
                          validation_split=0.2,
                          epochs=100,
                          batch_size=40,
                          callbacks=callbacks,
                          verbose=1)

部分训练如下
Train on 837 samples, validate on 210 samples
Epoch 80/100
760/837 [==========================>...] - ETA: 0s - loss: 0.4277 - accuracy: 0.8132
Epoch 00080: saving model to ./checkpoint2.x/Titanic.80.h5
837/837 [==============================] - 0s 302us/sample - loss: 0.4353 - accuracy: 0.8112 - val_loss: 0.4639 - val_accuracy: 0.7810
Epoch 81/100
837/837 [==============================] - 0s 270us/sample - loss: 0.4455 - accuracy: 0.8017 - val_loss: 0.4768 - val_accuracy: 0.7810
Epoch 82/100
837/837 [==============================] - 0s 211us/sample - loss: 0.4376 - accuracy: 0.7993 - val_loss: 0.4654 - val_accuracy: 0.7905
Epoch 83/100
837/837 [==============================] - 0s 278us/sample - loss: 0.4377 - accuracy: 0.8065 - val_loss: 0.4703 - val_accuracy: 0.7810
Epoch 84/100
837/837 [==============================] - 0s 232us/sample - loss: 0.4368 - accuracy: 0.8160 - val_loss: 0.4631 - val_accuracy: 0.7952
Epoch 85/100
360/837 [===========>..................] - ETA: 0s - loss: 0.4669 - accuracy: 0.8056
Epoch 00085: saving model to ./checkpoint2.x/Titanic.85.h5
837/837 [==============================] - 0s 292us/sample - loss: 0.4437 - accuracy: 0.8124 - val_loss: 0.4627 - val_accuracy: 0.7810
Epoch 86/100
837/837 [==============================] - 0s 197us/sample - loss: 0.4365 - accuracy: 0.8017 - val_loss: 0.4686 - val_accuracy: 0.7905
Epoch 87/100
837/837 [==============================] - 0s 288us/sample - loss: 0.4500 - accuracy: 0.8148 - val_loss: 0.4689 - val_accuracy: 0.7857
Epoch 88/100
837/837 [==============================] - 0s 208us/sample - loss: 0.4356 - accuracy: 0.8029 - val_loss: 0.4794 - val_accuracy: 0.7905
Epoch 89/100
837/837 [==============================] - 0s 239us/sample - loss: 0.4283 - accuracy: 0.8148 - val_loss: 0.4621 - val_accuracy: 0.7857
Epoch 90/100
440/837 [==============>...............] - ETA: 0s - loss: 0.4083 - accuracy: 0.8295
Epoch 00090: saving model to ./checkpoint2.x/Titanic.90.h5
837/837 [==============================] - 0s 258us/sample - loss: 0.4359 - accuracy: 0.8172 - val_loss: 0.4736 - val_accuracy: 0.7905
Epoch 91/100
837/837 [==============================] - 0s 299us/sample - loss: 0.4365 - accuracy: 0.8053 - val_loss: 0.4658 - val_accuracy: 0.7905
Epoch 92/100
837/837 [==============================] - 0s 319us/sample - loss: 0.4376 - accuracy: 0.8148 - val_loss: 0.4696 - val_accuracy: 0.7905
Epoch 93/100
837/837 [==============================] - 0s 355us/sample - loss: 0.4375 - accuracy: 0.8005 - val_loss: 0.4698 - val_accuracy: 0.7952
Epoch 94/100
837/837 [==============================] - 0s 205us/sample - loss: 0.4384 - accuracy: 0.8005 - val_loss: 0.4682 - val_accuracy: 0.7905
Epoch 95/100
440/837 [==============>...............] - ETA: 0s - loss: 0.4514 - accuracy: 0.7909
Epoch 00095: saving model to ./checkpoint2.x/Titanic.95.h5
837/837 [==============================] - 0s 344us/sample - loss: 0.4392 - accuracy: 0.8005 - val_loss: 0.4620 - val_accuracy: 0.7952
Epoch 96/100
837/837 [==============================] - 0s 219us/sample - loss: 0.4347 - accuracy: 0.8053 - val_loss: 0.4643 - val_accuracy: 0.7857
Epoch 97/100
837/837 [==============================] - 0s 309us/sample - loss: 0.4410 - accuracy: 0.8005 - val_loss: 0.4772 - val_accuracy: 0.7905
Epoch 98/100
837/837 [==============================] - 0s 230us/sample - loss: 0.4325 - accuracy: 0.8076 - val_loss: 0.4629 - val_accuracy: 0.7857
Epoch 99/100
837/837 [==============================] - 0s 267us/sample - loss: 0.4308 - accuracy: 0.8005 - val_loss: 0.4658 - val_accuracy: 0.7857
Epoch 100/100
360/837 [===========>..................] - ETA: 0s - loss: 0.4338 - accuracy: 0.8139
Epoch 00100: saving model to ./checkpoint2.x/Titanic.100.h5
837/837 [==============================] - 0s 265us/sample - loss: 0.4314 - accuracy: 0.8124 - val_loss: 0.4623 - val_accuracy: 0.7857

fig = plt.gcf()
fig.set_size_inches(10, 5)
ax1 = fig.add_subplot(111)
ax1.set_title('Train and Validation Picture')
ax1.set_ylabel('Loss value')
line1, = ax1.plot(train_history.history['loss'], color=(0.5, 0.5, 1.0), label='Loss train')
line2, = ax1.plot(train_history.history['val_loss'], color=(0.5, 1.0, 0.5), label='Loss valid')
ax2 = ax1.twinx()
ax2.set_ylabel('Accuracy value')
line3, = ax2.plot(train_history.history['accuracy'], color=(0.5, 0.5, 0.5), label='Accuracy train')
line4, = ax2.plot(train_history.history['val_accuracy'], color=(1, 0, 0), label='Accuracy valid')
plt.legend(handles=(line1, line2, line3, line4), loc='best')
plt.show()

Jack_info = [0, 'Jack', 3, 'male', 23, 1, 0, 5.0000, 'S']
Rose_info = [1, 'Rose', 1, 'female', 20, 1, 0, 100.0000, 'S']

new_passenger_pd = pd.DataFrame([Jack_info, Rose_info], columns=selected_cols)
all_passenger_pd = selected_dataframe.append(new_passenger_pd)

pred = model.predict(prepare_data(all_passenger_pd)[0])

print('Rose survived probability:', pred[-1:][0][0],
      '\nJack survived probability:', pred[-2:][0][0])

Rose survived probability: 0.9700622 
Jack survived probability: 0.12726058

加载模型

由于只保存了网络参数，没有保存网络结构，需要重新定义网络结构（当然，由于 jupyter 的缓存效应，你大可不必重新定义，对于独立的 py 文件则需要这么做）

model = tf.keras.models.Sequential([
    tf.keras.layers.Dense(units=256,
                          input_dim=7,
                          use_bias=True,
                          kernel_initializer='uniform',
                          bias_initializer='zeros',
                          activation='relu'),
    tf.keras.layers.Dropout(rate=0.3),
    tf.keras.layers.Dense(units=128, activation='sigmoid'),
    tf.keras.layers.Dropout(rate=0.3),
    tf.keras.layers.Dense(units=64, activation='sigmoid'),
    tf.keras.layers.Dropout(rate=0.3),
    tf.keras.layers.Dense(units=32, activation='sigmoid'),
    tf.keras.layers.Dropout(rate=0.3),
    tf.keras.layers.Dense(units=1, activation='sigmoid')
])

model.load_weights('./checkpoint2.x/Titanic.100.h5')

model.compile(optimizer=tf.keras.optimizers.Adam(0.003),
              loss='binary_crossentropy',
              metrics=['accuracy'])

loss, acc = model.evaluate(x_test, y_test, verbose=2)
print('Restore model accuracy:{:5.4f}%'.format(100 * acc))

262/1 - 0s - loss: 0.5042 - accuracy: 0.8511
Restore model accuracy:85.1145%

Titanic TensorFLow 1.x Keras API 实现

导入必要的包

import numpy
import pandas as pd
import tensorflow as tf
import urllib.request
from sklearn import preprocessing
import matplotlib.pyplot as plt
import os
import datetime


tf.__version__

'1.15.2'

定义预处理函数

def prepare_data(df_data):
    df = df_data.drop(['name'], axis=1)
    age_mean = df['age'].mean()
    df['age'] = df['age'].fillna(age_mean)
    fare_mean = df['fare'].mean()
    df['fare'] = df['fare'].fillna(fare_mean)
    df['sex'] = df['sex'].map({'female':0, 'male':1}).astype(int)
    df['embarked'] = df['embarked'].fillna('S')
    df['embarked'] = df['embarked'].map({'C':0, 'Q':1, 'S':2}).astype(int)

    ndarray_data = df.values

    features = ndarray_data[:, 1:]
    label = ndarray_data[:, 0]

    minmax_scale = preprocessing.MinMaxScaler(feature_range=(0, 1))
    norm_features = minmax_scale.fit_transform(features)

    return norm_features, label

读取数据，制作数据集

dataframe = pd.read_excel('./data/titanic3.xls')
selected_cols= ['survived', 'name', 'pclass', 'sex', 'age', 'sibsp', 'parch', 'fare', 'embarked']
selected_dataframe = dataframe[selected_cols].copy()
selected_dataframe = selected_dataframe.sample(frac=1)

x_data, y_data = prepare_data(selected_dataframe)

train_size = int(len(x_data) * 0.8)

x_train = x_data[:train_size]
y_train = y_data[:train_size]

x_test = x_data[train_size:]
y_test = y_data[train_size:]

搭建模型

model = tf.keras.models.Sequential([
    tf.keras.layers.Dense(units=64,
                          input_dim=7,
                          use_bias=True,
                          kernel_initializer='uniform',
                          bias_initializer='zeros',
                          activation='relu'),
    tf.keras.layers.Dropout(rate=0.3),
    tf.keras.layers.Dense(units=32, activation='sigmoid'),
    tf.keras.layers.Dropout(rate=0.3),
    tf.keras.layers.Dense(units=1, activation='sigmoid')
])

WARNING:tensorflow:From e:\anaconda3\envs\tensorflow1.x\lib\site-packages\tensorflow_core\python\keras\initializers.py:119: calling RandomUniform.__init__ (from tensorflow.python.ops.init_ops) with dtype is deprecated and will be removed in a future version.
Instructions for updating:
Call initializer instance with the dtype argument instead of passing it to the constructor
WARNING:tensorflow:From e:\anaconda3\envs\tensorflow1.x\lib\site-packages\tensorflow_core\python\ops\resource_variable_ops.py:1630: 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.

model.summary()

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense (Dense)                (None, 64)                512       
_________________________________________________________________
dropout (Dropout)            (None, 64)                0         
_________________________________________________________________
dense_1 (Dense)              (None, 32)                2080      
_________________________________________________________________
dropout_1 (Dropout)          (None, 32)                0         
_________________________________________________________________
dense_2 (Dense)              (None, 1)                 33        
=================================================================
Total params: 2,625
Trainable params: 2,625
Non-trainable params: 0
_________________________________________________________________

model.compile(optimizer=tf.keras.optimizers.Adam(0.003),
              loss='binary_crossentropy',
              metrics=['accuracy'])

WARNING:tensorflow:From e:\anaconda3\envs\tensorflow1.x\lib\site-packages\tensorflow_core\python\ops\nn_impl.py:183: where (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.
Instructions for updating:
Use tf.where in 2.0, which has the same broadcast rule as np.where

log_dir = os.path.join(
    'logs1.x',
    'train',
    'plugins',
    'profile',
    datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S'))

checkpoint_path = './checkpoint1.x/Titanic_{epoch:02d}-{val_loss:.2f}.ckpt'


callbacks = [tf.keras.callbacks.TensorBoard(log_dir=log_dir,
                         histogram_freq=2),
       tf.keras.callbacks.ModelCheckpoint(filepath=checkpoint_path,
                         save_weights_only=True,
                         verbose=1,
                         period=5)]

WARNING:tensorflow:`period` argument is deprecated. Please use `save_freq` to specify the frequency in number of samples seen.

开始训练

train_history = model.fit(x=x_train,
              y=y_train,
              validation_split=0.2,
              epochs=100,
              batch_size=40,
              callbacks=callbacks,
              verbose=2)

部分训练如下
Train on 837 samples, validate on 210 samples
837/837 - 0s - loss: 0.4398 - acc: 0.8124 - val_loss: 0.4671 - val_acc: 0.7857
Epoch 80/100

Epoch 00080: saving model to ./checkpoint1.x/Titanic_80-0.47.ckpt
837/837 - 0s - loss: 0.4360 - acc: 0.8076 - val_loss: 0.4673 - val_acc: 0.7857
Epoch 81/100
837/837 - 0s - loss: 0.4307 - acc: 0.8005 - val_loss: 0.4703 - val_acc: 0.7905
Epoch 82/100
837/837 - 0s - loss: 0.4401 - acc: 0.7981 - val_loss: 0.4666 - val_acc: 0.8000
Epoch 83/100
837/837 - 0s - loss: 0.4311 - acc: 0.8017 - val_loss: 0.4678 - val_acc: 0.7952
Epoch 84/100
837/837 - 0s - loss: 0.4296 - acc: 0.8172 - val_loss: 0.4673 - val_acc: 0.8000
Epoch 85/100

Epoch 00085: saving model to ./checkpoint1.x/Titanic_85-0.46.ckpt
837/837 - 0s - loss: 0.4384 - acc: 0.8029 - val_loss: 0.4634 - val_acc: 0.7857
Epoch 86/100
837/837 - 0s - loss: 0.4345 - acc: 0.8076 - val_loss: 0.4666 - val_acc: 0.7905
Epoch 87/100
837/837 - 0s - loss: 0.4307 - acc: 0.8053 - val_loss: 0.4650 - val_acc: 0.8000
Epoch 88/100
837/837 - 0s - loss: 0.4394 - acc: 0.8148 - val_loss: 0.4638 - val_acc: 0.8000
Epoch 89/100
837/837 - 0s - loss: 0.4355 - acc: 0.8053 - val_loss: 0.4648 - val_acc: 0.8000
Epoch 90/100

Epoch 00090: saving model to ./checkpoint1.x/Titanic_90-0.46.ckpt
837/837 - 0s - loss: 0.4326 - acc: 0.8100 - val_loss: 0.4623 - val_acc: 0.8000
Epoch 91/100
837/837 - 0s - loss: 0.4387 - acc: 0.8029 - val_loss: 0.4658 - val_acc: 0.7905
Epoch 92/100
837/837 - 0s - loss: 0.4285 - acc: 0.8065 - val_loss: 0.4613 - val_acc: 0.7905
Epoch 93/100
837/837 - 0s - loss: 0.4355 - acc: 0.8088 - val_loss: 0.4656 - val_acc: 0.7905
Epoch 94/100
837/837 - 0s - loss: 0.4318 - acc: 0.8136 - val_loss: 0.4629 - val_acc: 0.7952
Epoch 95/100

Epoch 00095: saving model to ./checkpoint1.x/Titanic_95-0.46.ckpt
837/837 - 0s - loss: 0.4386 - acc: 0.7981 - val_loss: 0.4639 - val_acc: 0.8000
Epoch 96/100
837/837 - 0s - loss: 0.4346 - acc: 0.8041 - val_loss: 0.4647 - val_acc: 0.7857
Epoch 97/100
837/837 - 0s - loss: 0.4256 - acc: 0.8160 - val_loss: 0.4608 - val_acc: 0.8048
Epoch 98/100
837/837 - 0s - loss: 0.4357 - acc: 0.8029 - val_loss: 0.4613 - val_acc: 0.8000
Epoch 99/100
837/837 - 0s - loss: 0.4265 - acc: 0.8041 - val_loss: 0.4614 - val_acc: 0.7952
Epoch 100/100

Epoch 00100: saving model to ./checkpoint1.x/Titanic_100-0.46.ckpt
837/837 - 0s - loss: 0.4243 - acc: 0.8148 - val_loss: 0.4611 - val_acc: 0.8000

训练可视化

fig = plt.gcf()
fig.set_size_inches(10, 5)
ax1 = fig.add_subplot(111)
ax1.set_title('Train and Validation Picture')
ax1.set_ylabel('Loss value')
line1, = ax1.plot(train_history.history['loss'], color=(0.5, 0.5, 1.0), label='Loss train')
line2, = ax1.plot(train_history.history['val_loss'], color=(0.5, 1.0, 0.5), label='Loss valid')
ax2 = ax1.twinx()
ax2.set_ylabel('Accuracy value')
line3, = ax2.plot(train_history.history['acc'], color=(0.5, 0.5, 0.5), label='Accuracy train')
line4, = ax2.plot(train_history.history['val_acc'], color=(1, 0, 0), label='Accuracy valid')
plt.legend(handles=(line1, line2, line3, line4), loc='best')
plt.show()

模型预测

Jack_info = [0, 'Jack', 3, 'male', 23, 1, 0, 5.0000, 'S']
Rose_info = [1, 'Rose', 1, 'female', 20, 1, 0, 100.0000, 'S']

new_passenger_pd = pd.DataFrame([Jack_info, Rose_info], columns=selected_cols)
all_passenger_pd = selected_dataframe.append(new_passenger_pd)

pred = model.predict(prepare_data(all_passenger_pd)[0])

print('Rose survived probability:', pred[-1:][0][0],
      '\nJack survived probability:', pred[-2:][0][0])

Rose survived probability: 0.9762004 
Jack survived probability: 0.10789904

加载模型进行预测

model = tf.keras.models.Sequential([
    tf.keras.layers.Dense(units=64,
                          input_dim=7,
                          use_bias=True,
                          kernel_initializer='uniform',
                          bias_initializer='zeros',
                          activation='relu'),
    tf.keras.layers.Dropout(rate=0.3),
    tf.keras.layers.Dense(units=32, activation='sigmoid'),
    tf.keras.layers.Dropout(rate=0.3),
    tf.keras.layers.Dense(units=1, activation='sigmoid')
])

model.compile(optimizer=tf.keras.optimizers.Adam(0.003),
              loss='binary_crossentropy',
              metrics=['accuracy'])

checkpoint_dir = os.path.dirname(checkpoint_path)
latest = tf.train.latest_checkpoint(checkpoint_dir)
model.load_weights(latest)

<tensorflow.python.training.tracking.util.CheckpointLoadStatus at 0x1f9aa2a0a20>

loss, acc = model.evaluate(x_test, y_test)

262/262 [==============================] - 0s 244us/sample - loss: 0.4393 - acc: 0.7977