tensorflow 2.0 学习 （十） 拟合与过拟合问题

 1 # encoding: utf-8
 2 
 3 import tensorflow as tf
 4 import numpy as np
 5 import seaborn as sns
 6 import os
 7 import matplotlib.pyplot as plt
 8 from sklearn.datasets import make_moons
 9 from sklearn.model_selection import train_test_split
10 from tensorflow.keras import layers, Sequential, optimizers, losses, metrics
11 from tensorflow.keras.layers import Dense
12 
13 N_SAMPLES = 1000  # 采样点数
14 N_Epochs = 300
15 TEST_SIZE = 0.3  # 测试数量比率
16 OUTPUT_DIR = r‘G:\2020\python‘
17 if not os.path.exists(OUTPUT_DIR):
18     os.mkdir(OUTPUT_DIR)
19 
20 # 产生一个简单的样本数据集，半环形图，类似的有make_circles,环形数据
21 X, y = make_moons(n_samples=N_SAMPLES, noise=0.25, random_state=100)  # (1000, 2),(1000, 1)
22 # 将矩阵随机划分训练集和测试集 (700,2),(300,2),(700,1),(300,1)
23 X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=TEST_SIZE, random_state=42)
24 print(X.shape, y.shape)
25 
26 
27 def make_plot(X, y, plot_name, file_name, XX=None, YY=None, preds=None):
28     plt.figure()
29     axes = plt.gca()
30     x_min = X[:, 0].min() - 1
31     x_max = X[:, 0].max() + 1
32     y_min = X[:, 1].min() - 1
33     y_max = X[:, 1].max() + 1
34     axes.set_xlim([x_min, x_max])
35     axes.set_ylim([y_min, y_max])
36     axes.set(xlabel="$x_l$", ylabel="$x_2$")
37 
38     # 根据网络输出绘制预测曲面
39     # markers = [‘o‘ if i == 1 else ‘s‘ for i in y.ravel()]
40     # plt.scatter(X[:, 0], X[:, 1], c=y.ravel(), s=20, cmap=plt.cm.Spectral, edgecolors=‘none‘, m=markers)
41     if XX is None and YY is None and preds is None:
42         yr = y.ravel()
43         for step in range(X[:, 0].size):
44             if yr[step] == 1:
45                 plt.scatter(X[step, 0], X[step, 1], c=‘b‘, s=20, cmap=plt.cm.Spectral, edgecolors=‘none‘, marker=‘o‘)
46             else:
47                 plt.scatter(X[step, 0], X[step, 1], c=‘r‘, s=20, cmap=plt.cm.Spectral, edgecolors=‘none‘, marker=‘s‘)
48         plt.savefig(OUTPUT_DIR+‘/‘+file_name)
49         # plt.show()
50     else:
51         plt.contour(XX, YY, preds, cmap=plt.cm.autumn, alpha=0.8)
52         plt.scatter(X[:, 0], X[:, 1], c=y, s=20, cmap=plt.cm.autumn, edgecolors=‘k‘)
53         plt.rcParams[‘font.sans-serif‘] = [‘SimHei‘]  # 解决plt.title乱码的问题
54         plt.rcParams[‘axes.unicode_minus‘] = False
55         plt.title(plot_name)
56         plt.savefig(OUTPUT_DIR+‘/‘+file_name)
57         # plt.show()
58 
59 
60 make_plot(X, y, None, "exam7_dataset.svg")
61 
62 # 创建网络 5种不同的网络
63 for n in range(5):
64     model = Sequential()  # 创建容器
65     model.add(Dense(8, input_dim=2, activation=‘relu‘))  # 第一层
66     for _ in range(n):
67         model.add(Dense(32, activation=‘relu‘))
68     model.add(Dense(1, activation=‘sigmoid‘))  # 创建末尾一层
69     model.compile(loss=‘binary_crossentropy‘, optimizer=‘adam‘, metrics=[‘accuracy‘])  # 模型的装配
70     history = model.fit(X_train, y_train, epochs=N_Epochs, verbose=1)
71     # 绘制不同层数的网络决策边界曲线
72     x_min = X[:, 0].min() - 1
73     x_max = X[:, 0].max() + 1
74     y_min = X[:, 1].min() - 1
75     y_max = X[:, 1].max() + 1
76     # XX(477, 600), YY(477, 600)
77     XX, YY = np.meshgrid(np.arange(x_min, x_max, 0.01), np.arange(y_min, y_max, 0.01))  # 创建网格
78     Z = model.predict_classes(np.c_[XX.ravel(), YY.ravel()])  # (286200, 1) [0 or 1]
79     preds = Z.reshape(XX.shape)
80     title = "网络层数({})".format(n)
81     file = "网络容量%f.png" % (2+n*1)
82     make_plot(X_train, y_train, title, file, XX, YY, preds)

 1 # 创建网络 5种不同数量的Dropout层的网络
 2 for n in range(5):
 3     model = Sequential()  # 创建容器
 4     model.add(Dense(8, input_dim=2, activation=‘relu‘))  # 第一层
 5     counter = 0
 6     for _ in range(5):  # 网络层数固定为5
 7         model.add(Dense(64, activation=‘relu‘))
 8         if counter < n:  # 添加n个Dropout层
 9             counter += 1
10             model.add(layers.Dropout(rate=0.5))
11     model.add(Dense(1, activation=‘sigmoid‘))  # 创建末尾一层
12     model.compile(loss=‘binary_crossentropy‘, optimizer=‘adam‘, metrics=[‘accuracy‘])  # 模型的装配
13     history = model.fit(X_train, y_train, epochs=N_Epochs, verbose=1)
14     # 绘制不同层数的网络决策边界曲线
15     x_min = X[:, 0].min() - 1
16     x_max = X[:, 0].max() + 1
17     y_min = X[:, 1].min() - 1
18     y_max = X[:, 1].max() + 1
19     # XX(477, 600), YY(477, 600)
20     XX, YY = np.meshgrid(np.arange(x_min, x_max, 0.01), np.arange(y_min, y_max, 0.01))  # 创建网格
21     Z = model.predict_classes(np.c_[XX.ravel(), YY.ravel()])  # (286200, 1) [0 or 1]
22     preds = Z.reshape(XX.shape)
23     title = "Dropout({})".format(n)
24     file = "Dropout%f.png" % (n)
25     make_plot(X_train, y_train, title, file, XX, YY, preds)