操作步骤
导入所需的包。
import tensorflow as tf import numpy as np import matplotlib as mpl import matplotlib.pyplot as plt import sklearn.datasets as ds import sklearn.model_selection as ms
导入数据,并进行预处理。我们使用鸢尾花数据集所有样本,根据萼片长度和花瓣长度预测样本是不是山鸢尾(第一种)。注意,支持向量机只接受 1 和 -1 的标签。
iris = ds.load_iris() x_ = iris.data[:, [0, 2]] y_ = (iris.target == 0).astype(int) y_[y_ == 0] = -1 y_ = np.expand_dims(y_ , 1) x_train, x_test, y_train, y_test = \ ms.train_test_split(x_, y_, train_size=0.7, test_size=0.3)
定义超参数。
变量 | 含义 |
n_input |
样本特征数 |
n_epoch |
迭代数 |
lr |
学习率 |
lam |
L2 正则化项的系数 |
n_input = 2 n_epoch = 2000 lr = 0.05 lam = 0.05
搭建模型。
变量 | 含义 |
x |
输入 |
y |
真实标签 |
w |
权重 |
b |
偏置 |
z |
x 的线性变换 |
x = tf.placeholder(tf.float64, [None, n_input]) y = tf.placeholder(tf.float64, [None, 1]) w = tf.Variable(np.random.rand(n_input, 1)) b = tf.Variable(np.random.rand(1, 1)) z = x @ w + b
定义损失、优化操作、和准确率度量指标。分类问题有很多指标,这里只展示一种。
我们使用 Hinge 损失和 L2 损失的组合。Hinge 损失为:
mean(max(1−Z⊗Y,0))
在原始的模型中,约束是样本必须落在支持边界之外,也就是 y z > = 1 yz >= 1 yz>=1。我们将这个约束加到损失中,就得到了 Hinge 损失。它的意思是,对于满足约束的点,它的损失是零,对于不满足约束的点,它的损失是 1 − y z 1 - yz 1−yz。这样让样本尽可能到支持边界之外。
L2 损失用于最小化支持边界的几何距离,也就是 2 ∥ w ∥ \frac{2}{\|w\|} ∥w∥2。
变量 | 含义 |
hinge_loss |
Hinge 损失 |
l2_loss |
L2 损失 |
loss |
总损失 |
op |
优化操作 |
y_hat |
标签的预测值 |
acc |
准确率 |
hinge_loss = tf.reduce_mean(tf.maximum(1 - y * z, 0)) l2_loss = lam * tf.reduce_sum(w ** 2) loss = hinge_loss + l2_loss op = tf.train.AdamOptimizer(lr).minimize(loss) y_hat = tf.to_double(z > 0) - tf.to_double(z <= 0) acc = tf.reduce_mean(tf.to_double(tf.equal(y_hat, y)))
使用训练集训练模型。
losses = [] accs = [] with tf.Session() as sess: sess.run(tf.global_variables_initializer()) saver = tf.train.Saver(max_to_keep=1) for e in range(n_epoch): _, loss_ = sess.run([op, loss], feed_dict={x: x_train, y: y_train}) losses.append(loss_)
使用测试集计算准确率。
acc_ = sess.run(acc, feed_dict={x: x_test, y: y_test}) accs.append(acc_)
每一百步打印损失和度量值。
if e % 100 == 0: print(f'epoch: {e}, loss: {loss_}, acc: {acc_}') saver.save(sess,'logit/logit', global_step=e)
得到决策边界:
x_plt = x_[:, 0] y_plt = x_[:, 1] c_plt = y_.ravel() x_min = x_plt.min() - 1 x_max = x_plt.max() + 1 y_min = y_plt.min() - 1 y_max = y_plt.max() + 1 x_rng = np.arange(x_min, x_max, 0.05) y_rng = np.arange(y_min, y_max, 0.05) x_rng, y_rng = np.meshgrid(x_rng, y_rng) model_input = np.asarray([x_rng.ravel(), y_rng.ravel()]).T model_output = sess.run(y_hat, feed_dict={x: model_input}).astype(int) c_rng = model_output.reshape(x_rng.shape)
输出:
epoch: 0, loss: 4.511212919815273, acc: 0.2222222222222222 epoch: 100, loss: 0.0814942611949705, acc: 1.0 epoch: 200, loss: 0.07629443566925993, acc: 1.0 epoch: 300, loss: 0.07146107394130172, acc: 1.0 epoch: 400, loss: 0.06791927215796319, acc: 1.0 epoch: 500, loss: 0.06529065400047798, acc: 1.0 epoch: 600, loss: 0.06335060635876646, acc: 1.0 epoch: 700, loss: 0.061836271593737835, acc: 1.0 epoch: 800, loss: 0.06079800773555345, acc: 1.0 epoch: 900, loss: 0.06042716484730995, acc: 1.0 epoch: 1000, loss: 0.06091475237291386, acc: 1.0 epoch: 1100, loss: 0.06021069445352348, acc: 1.0 epoch: 1200, loss: 0.06019457351257251, acc: 1.0 epoch: 1300, loss: 0.06000348375369489, acc: 1.0 epoch: 1400, loss: 0.060206981088196394, acc: 1.0 epoch: 1500, loss: 0.060210741691625935, acc: 1.0 epoch: 1600, loss: 0.060570783158962985, acc: 1.0 epoch: 1700, loss: 0.06003457018203537, acc: 1.0 epoch: 1800, loss: 0.060203912161627175, acc: 1.0 epoch: 1900, loss: 0.06019910894894441, acc: 1.0
绘制整个数据集以及决策边界。
plt.figure() cmap = mpl.colors.ListedColormap(['r', 'b']) plt.scatter(x_plt, y_plt, c=c_plt, cmap=cmap) plt.contourf(x_rng, y_rng, c_rng, alpha=0.2, linewidth=5, cmap=cmap) plt.title('Data and Model') plt.xlabel('Petal Length (cm)') plt.ylabel('Sepal Length (cm)') plt.show()
绘制训练集上的损失。
plt.figure() plt.plot(losses) plt.title('Loss on Training Set') plt.xlabel('#epoch') plt.ylabel('Cross Entropy') plt.show()
绘制测试集上的准确率。
plt.figure() plt.plot(accs) plt.title('Accurary on Testing Set') plt.xlabel('#epoch') plt.ylabel('Accurary') plt.show()