神经网络(NN)实现多分类-----Keras实现

IRIS数据集介绍

IRIS数据集（鸢尾花数据集），是一个经典的机器学习数据集，适合做为多分类问题的测试数据，它的下载地址为：http://archive.ics.uci.edu/ml/machine-learning-databases/iris/。
IRIS数据集是用来给鸢尾花作分类的数据集，一共150个样本，每一个样本包含了花萼长度（sepal length in cm）、花萼宽度（sepal width in cm）、花瓣长度（petal length in cm）、花瓣宽度（petal width in cm）四个特征，将鸢尾花分为三类，分别为Iris Setosa，Iris Versicolour，Iris Virginica，每一类都有50个样本。
IRIS数据集具体以下（只展现部分数据，顺序已打乱）：数组

读取数据集

import pandas as pd from sklearn.model_selection import train_test_split from sklearn.preprocessing import LabelBinarizer # 读取CSV数据集，并拆分为训练集和测试集 # 该函数的传入参数为CSV_FILE_PATH: csv文件路径
def load_data(CSV_FILE_PATH): IRIS = pd.read_csv(CSV_FILE_PATH) target_var = 'class'  # 目标变量
    # 数据集的特征
    features = list(IRIS.columns) features.remove(target_var) # 目标变量的类别
    Class = IRIS[target_var].unique() # 目标变量的类别字典
    Class_dict = dict(zip(Class, range(len(Class)))) # 增长一列target, 将目标变量进行编码
    IRIS['target'] = IRIS[target_var].apply(lambda x: Class_dict[x]) # 对目标变量进行0-1编码(One-hot Encoding)
    lb = LabelBinarizer() lb.fit(list(Class_dict.values())) transformed_labels = lb.transform(IRIS['target']) y_bin_labels = []  # 对多分类进行0-1编码的变量
    for i in range(transformed_labels.shape[1]): y_bin_labels.append('y' + str(i)) IRIS['y' + str(i)] = transformed_labels[:, i] # 将数据集分为训练集和测试集
    train_x, test_x, train_y, test_y = train_test_split(IRIS[features], IRIS[y_bin_labels], \ train_size=0.7, test_size=0.3, random_state=0) return train_x, test_x, train_y, test_y, Class_dict

搭建DNN

接下来，笔者将展现如何利用Keras来搭建一个简单的深度神经网络（DNN）来解决这个多分类问题。咱们要搭建的DNN的结构以下图所示：网络

DNN模型的结构示意图

咱们搭建的DNN由输入层、隐藏层、输出层和softmax函数组成，其中输入层由4个神经元组成，对应IRIS数据集中的4个特征，做为输入向量，隐藏层有两层，每层分别有5和6个神经元，以后就是输出层，由3个神经元组成，对应IRIS数据集的目标变量的类别个数，最后，就是一个softmax函数，用于解决多分类问题而建立。
对应以上的DNN结构，用Keras来搭建的话，其Python代码以下：app

import keras as K # 2. 定义模型
    init = K.initializers.glorot_uniform(seed=1) simple_adam = K.optimizers.Adam() model = K.models.Sequential() model.add(K.layers.Dense(units=5, input_dim=4, kernel_initializer=init, activation='relu')) model.add(K.layers.Dense(units=6, kernel_initializer=init, activation='relu')) model.add(K.layers.Dense(units=3, kernel_initializer=init, activation='softmax')) model.compile(loss='categorical_crossentropy', optimizer=simple_adam, metrics=['accuracy'])

在这个模型中，咱们选择的神经元激活函数为ReLU函数，损失函数为交叉熵（cross entropy），迭代的优化器（optimizer）选择Adam，最初各个层的链接权重（weights）和偏重（biases）是随机生成的。这样咱们就讲这个DNN的模型定义完毕了。dom

训练及预测

OK，定义完模型后，咱们须要对模型进行训练、评估及预测。对于模型训练，咱们每次训练的批数为1，共迭代100次，代码以下（接以上代码）：机器学习

# 3. 训练模型
    b_size = 1 max_epochs = 100
    print("Starting training ") h = model.fit(train_x, train_y, batch_size=b_size, epochs=max_epochs, shuffle=True, verbose=1) print("Training finished \n")

为了对模型有个评估，感知模型的表现，须要输出该DNN模型的损失函数的值以及在测试集上的准确率，其Python代码以下（接以上代码）：函数

# 4. 评估模型
    eval = model.evaluate(test_x, test_y, verbose=0) print("Evaluation on test data: loss = %0.6f accuracy = %0.2f%% \n" \ % (eval[0], eval[1] * 100) )

训练100次，输出的结果以下（中间部分的训练展现已忽略）：学习

Starting training Epoch 1/100

  1/105 [..............................] - ETA: 17s - loss: 0.3679 - acc: 1.0000
 42/105 [===========>..................] - ETA: 0s - loss: 1.8081 - acc: 0.3095 
 89/105 [========================>.....] - ETA: 0s - loss: 1.5068 - acc: 0.4270
105/105 [==============================] - 0s 3ms/step - loss: 1.4164 - acc: 0.4667 Epoch 2/100

  1/105 [..............................] - ETA: 0s - loss: 0.4766 - acc: 1.0000
 45/105 [===========>..................] - ETA: 0s - loss: 1.0813 - acc: 0.4889
 93/105 [=========================>....] - ETA: 0s - loss: 1.0335 - acc: 0.4839
105/105 [==============================] - 0s 1ms/step - loss: 1.0144 - acc: 0.4857 ...... Epoch 99/100

  1/105 [..............................] - ETA: 0s - loss: 0.0013 - acc: 1.0000
 43/105 [===========>..................] - ETA: 0s - loss: 0.0447 - acc: 0.9767
 84/105 [=======================>......] - ETA: 0s - loss: 0.0824 - acc: 0.9524
105/105 [==============================] - 0s 1ms/step - loss: 0.0711 - acc: 0.9619 Epoch 100/100

  1/105 [..............................] - ETA: 0s - loss: 2.3032 - acc: 0.0000e+00
 51/105 [=============>................] - ETA: 0s - loss: 0.1122 - acc: 0.9608    
 99/105 [===========================>..] - ETA: 0s - loss: 0.0755 - acc: 0.9798
105/105 [==============================] - 0s 1ms/step - loss: 0.0756 - acc: 0.9810 Training finished Evaluation on test data: loss = 0.094882 accuracy = 97.78%

能够看到，训练完100次后，在测试集上的准确率已达到97.78%，效果至关好。
最后是对新数据集进行预测，咱们假设一朵鸢尾花的4个特征为6.1,3.1,5.1,1.1，咱们想知道这个DNN模型会把它预测到哪一类，其Python代码以下：测试

import numpy as np # 5. 使用模型进行预测
    np.set_printoptions(precision=4) unknown = np.array([[6.1, 3.1, 5.1, 1.1]], dtype=np.float32) predicted = model.predict(unknown) print("Using model to predict species for features: ") print(unknown) print("\nPredicted softmax vector is: ") print(predicted) species_dict = {v:k for k,v in Class_dict.items()} print("\nPredicted species is: ") print(species_dict[np.argmax(predicted)])

输出的结果以下：优化

Using model to predict species for features: [[ 6.1  3.1  5.1  1.1]] Predicted softmax vector is: [[ 2.0687e-07   9.7901e-01   2.0993e-02]] Predicted species is: versicolor

若是咱们仔细地比对IRIS数据集，就会发现，这个预测结果使人至关满意，这个鸢尾花样本的预测结果，以人类的眼光来看，也应当是versicolor。编码

最后，附上该DNN模型的完整Python代码：

# iris_keras_dnn.py # Python 3.5.1, TensorFlow 1.6.0, Keras 2.1.5 # ======================================================== # 导入模块
import os import numpy as np import keras as K import tensorflow as tf import pandas as pd from sklearn.model_selection import train_test_split from sklearn.preprocessing import LabelBinarizer os.environ['TF_CPP_MIN_LOG_LEVEL']='2'

# 读取CSV数据集，并拆分为训练集和测试集 # 该函数的传入参数为CSV_FILE_PATH: csv文件路径
def load_data(CSV_FILE_PATH): IRIS = pd.read_csv(CSV_FILE_PATH) target_var = 'class'  # 目标变量
    # 数据集的特征
    features = list(IRIS.columns) features.remove(target_var) # 目标变量的类别
    Class = IRIS[target_var].unique() # 目标变量的类别字典
    Class_dict = dict(zip(Class, range(len(Class)))) # 增长一列target, 将目标变量进行编码
    IRIS['target'] = IRIS[target_var].apply(lambda x: Class_dict[x]) # 对目标变量进行0-1编码(One-hot Encoding)
    lb = LabelBinarizer() lb.fit(list(Class_dict.values())) transformed_labels = lb.transform(IRIS['target']) y_bin_labels = []  # 对多分类进行0-1编码的变量
    for i in range(transformed_labels.shape[1]): y_bin_labels.append('y' + str(i)) IRIS['y' + str(i)] = transformed_labels[:, i] # 将数据集分为训练集和测试集
    train_x, test_x, train_y, test_y = train_test_split(IRIS[features], IRIS[y_bin_labels], \ train_size=0.7, test_size=0.3, random_state=0) return train_x, test_x, train_y, test_y, Class_dict def main(): # 0. 开始
    print("\nIris dataset using Keras/TensorFlow ") np.random.seed(4) tf.set_random_seed(13) # 1. 读取CSV数据集
    print("Loading Iris data into memory") CSV_FILE_PATH = 'E://iris.csv' train_x, test_x, train_y, test_y, Class_dict = load_data(CSV_FILE_PATH) # 2. 定义模型
    init = K.initializers.glorot_uniform(seed=1) simple_adam = K.optimizers.Adam() model = K.models.Sequential() model.add(K.layers.Dense(units=5, input_dim=4, kernel_initializer=init, activation='relu')) model.add(K.layers.Dense(units=6, kernel_initializer=init, activation='relu')) model.add(K.layers.Dense(units=3, kernel_initializer=init, activation='softmax')) model.compile(loss='categorical_crossentropy', optimizer=simple_adam, metrics=['accuracy']) # 3. 训练模型
    b_size = 1 max_epochs = 100
    print("Starting training ") h = model.fit(train_x, train_y, batch_size=b_size, epochs=max_epochs, shuffle=True, verbose=1) print("Training finished \n") # 4. 评估模型
    eval = model.evaluate(test_x, test_y, verbose=0) print("Evaluation on test data: loss = %0.6f accuracy = %0.2f%% \n" \ % (eval[0], eval[1] * 100) ) # 5. 使用模型进行预测
    np.set_printoptions(precision=4) unknown = np.array([[6.1, 3.1, 5.1, 1.1]], dtype=np.float32) predicted = model.predict(unknown) print("Using model to predict species for features: ") print(unknown) print("\nPredicted softmax vector is: ") print(predicted) species_dict = {v:k for k,v in Class_dict.items()} print("\nPredicted species is: ") print(species_dict[np.argmax(predicted)]) main()

连接：https://www.jianshu.com/p/1d88a6ed707e 来源：简书