测试数据准确率：98%
采用一个隐藏层，30个节点（iris数据量并不大）

plot_train_BP_Netsiris.png

# encoding: utf-8

"""
@version: ??
@author: kaenlee  @contact: lichaolfm@163.com
@software: PyCharm Community Edition
@time: 2018/1/10 16:22
purpose: BP神经网络
"""

import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
from random import shuffle


class BP_Nets:
    def __init__(self, layers, optimizer=tf.train.AdamOptimizer()):
        assert len(layers) >= 3
        self.layers = layers
        self.X = tf.placeholder(dtype=tf.float32, shape=[None, layers[0]])
        self.Y_ = tf.placeholder(dtype=tf.float32, shape=[None, layers[-1]])
        count = 1
        tmp = self.X
        # 定义隐藏层
        for n_input, n_output in zip(layers[:-2], layers[1:-1]):
            tmp = self.__add_layer(tmp, n_input, n_output, name='Layer%d' % count)
        # 定义输出层
        self.Y = self.__add_layer(tmp, self.layers[-2], self.layers[-1], name='output_layer',
                                  acativefunction=tf.nn.softmax)
        # 交叉损失熵函数
        self.cross_entropy = tf.reduce_mean(- tf.reduce_sum(input_tensor=self.Y_ * tf.log(self.Y), axis=1),
                                            name='cross_entropy')
        # 优化器
        self.optimizer = optimizer.minimize(self.cross_entropy)
        # 初始化变量并启动session
        init = tf.global_variables_initializer()
        self.sess = tf.InteractiveSession()
        self.sess.run(init)
        # self.sess.close()

    def __init_weight(self, n_input, n_output):
        # 初始化权重
        total = n_input + n_output
        boundary_uniform = tf.sqrt(6.0 / total)
        weight = tf.random_uniform(shape=[n_input, n_output], minval=-boundary_uniform,
                                   maxval=boundary_uniform, dtype=tf.float32, name='weight')
        return weight

    def __add_layer(self, input, n_input, n_output, name, acativefunction=tf.nn.sigmoid):
        weight = self.__init_weight(n_input, n_output)
        weight = tf.Variable(initial_value=weight, name='W_' + name)
        bias = tf.Variable(tf.zeros(shape=(n_output), dtype=tf.float32), name='b_' + name)
        # print(input.shape, n_input, n_output)
        output = acativefunction(
            tf.add(tf.matmul(a=input, b=weight), bias), name='result' + name
        )
        return output

    def train(self, X_train, Y_train, n_iter, size_batch, name):
        # 迭代训练
        assert isinstance(X_train, np.ndarray)
        assert isinstance(Y_train, np.ndarray)
        size_sample = len(X_train)
        partition = int(size_sample / size_batch)
        cost_iter = []
        accuracy = []
        for i in range(n_iter):
            # 打乱数据的顺序
            indx = np.arange(size_sample)
            shuffle(indx)
            X_train = X_train[indx]
            Y_train = Y_train[indx]
            cost_batch = 0
            for j in range(partition):
                # 运行一个batch 更新权重
                X_batch = X_train[(j * size_batch):((j + 1) * size_batch)]
                Y_batch = Y_train[(j * size_batch):((j + 1) * size_batch)]
                # print(X_batch)
                opt, cost = self.sess.run(fetches=(self.optimizer, self.cross_entropy),
                                          feed_dict={self.X: X_batch, self.Y_: Y_batch})
                cost_batch += cost / partition
            cost_iter.append(cost_batch)
            Y = self.sess.run(fetches=self.Y, feed_dict={self.X: X_train})
            Y = np.array(Y)
            accuracy.append(self.accuracy(Y_train, Y))
        plt.figure()
        plt.plot(range(n_iter), cost_iter, color='green', label='cost')
        plt.plot(range(n_iter), accuracy, color='gray', label='accuracy')
        plt.title("model train result")
        plt.legend()
        plt.savefig('plot_train_BP_Nets%s.png' % name)

    def accuracy(self, Y_, Y):
        assert isinstance(Y_, np.ndarray)
        assert isinstance(Y, np.ndarray)
        boolean = tf.equal(x=tf.argmax(Y_, axis=1), y=tf.argmax(Y, axis=1))
        accu = tf.reduce_mean(tf.cast(x=boolean, dtype=tf.float32))
        return self.sess.run(accu)

    def evaluate(self, X_test, Y_test):
        assert isinstance(X_test, np.ndarray)
        assert isinstance(Y_test, np.ndarray)
        Y = self.sess.run(fetches=self.Y, feed_dict={self.X: X_test})
        Y = np.array(Y)
        return self.accuracy(Y_test, Y)


if __name__ == '__main__':
    # iris
    X = []
    Y = []
    with open("D:/Data/iris.txt", mode='r', encoding='utf-8') as f:
        f.readline()
        while 1:
            row = f.readline().strip().split(',')
            if len(row) <= 2:
                break
            X.append([float(i) for i in row[:-1]])
            Y.append(row[-1])
    X = np.array(X)
    # one-hot
    kind = np.unique(Y)
    size = len(Y)
    tmp = np.zeros(shape=(size, len(kind)))
    for indx in range(size):
        index_one, = np.where(kind == Y[indx])
        tmp[indx][index_one] = 1
    Y = np.array(tmp);
    del tmp
    # print(Y)
    # print(X)
    layers = [len(X[0]), 30, len(Y[0])]
    model = BP_Nets(layers)
    model.train(X_train=X, Y_train=Y, n_iter=300, size_batch=10, name='iris')
    print(model.evaluate(X_test=X, Y_test=Y))