TensorFlow 线性分类

原文链接: TensorFlow 线性分类

上一篇: Softmax函数 交叉熵 信息熵

下一篇: java lambda 表达式

构造直线 z = 2 * x - 3 * y + 4

x0*w0+x1*w1+b=0
x1=-x0* w0/w1-b/w1
斜率 k= -w0/w1 截距 -b/w1

随机生成数据，加入一定的偏差，用直线将二维平面分为两部分

使用线性模型拟合参数

损失函数计算

# 输出
output = tf.matmul(input_features, W) + b
# 交叉熵
cross_entropy = tf.nn.sigmoid_cross_entropy_with_logits(labels=input_labels, logits=output)

注意各个张量的shape一定满足运算要求

拟合结果

代码如下

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


# 模拟生成数据点, 返回np数组
def generate(sample_size):
    X = []
    Y = []
    for i in range(sample_size):
        x, y = random.random() * 10, random.random() * 10
        X.append([x, y])
        # 有一定的误差
        if 2 * x - 3 * y + 4 + (random.random() - 0.5) * 2 < 0:
            Y.append([0])
        else:
            Y.append([1])
    return np.array(X), np.array(Y)


X, Y = generate(1000)
# 显示散点图
colors = ['r' if i == 0 else 'b' for i in Y[:]]
plt.scatter(X[:, 0], X[:, 1], c=colors)
plt.show()

# 输入向量 NX2
input_features = tf.placeholder(tf.float32, [None, 2])
# 标签是NX1
input_labels = tf.placeholder(tf.float32, [None, 1])
# 权重2X1
W = tf.Variable(tf.random_normal([2, 1]), name="weight")
# 偏移是1X1
b = tf.Variable(tf.zeros([1]), name="bias")

# 输出
output = tf.matmul(input_features, W) + b
# 交叉熵
cross_entropy = tf.nn.sigmoid_cross_entropy_with_logits(labels=input_labels, logits=output)
# 损失函数
loss = tf.reduce_mean(cross_entropy)
# 平均错误率
err = tf.reduce_mean(tf.square(input_labels - output))
# 优化器尽量用这个--收敛快，会动态调节梯度
optimizer = tf.train.AdamOptimizer(0.04)
# 训练器
train = optimizer.minimize(loss)
# 训练总次数
epochs = 50
# 每次分多少批
batch = 100
#  每个批次训练样0本数目
batchSize = len(Y) // batch

# 启动session
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())

    for epoch in range(epochs):
        # 统计每批次的总误差
        sumerr = 0
        for i in range(batch):
            start = i * batchSize
            end = (i + 1) * batchSize
            x1 = X[start:end, :]
            y1 = Y[start:end]
            sess.run(
                train,
                feed_dict={input_features: x1, input_labels: y1}
            )
        wVal = sess.run(W)
        bVal = sess.run(b)
        # 输出斜率和截距
        print(wVal[0] / wVal[1], bVal / wVal[1])

    # 图形显示
    plt.scatter(X[:, 0], X[:, 1], c=colors)

    #    x0*w0+x1*w1+b=0
    #    x1=-x0* w0/w1-b/w1
    # 斜率 k= -w0/w1 截距 -b/w1
    x = np.linspace(-1, 10, 200)
    y = -x * (sess.run(W)[0] / sess.run(W)[1]) - sess.run(b) / sess.run(W)[1]
    plt.plot(x, y, label='Fitted line')
    plt.legend()
    plt.show()

斜率和截距基本上稳定在0.666和1.333附近

[-0.65097] [-1.4102162]
[-0.6506002] [-1.4103177]