使用TensorFlow实现SVM

较基础的SVM，后续会加上多分类以及高斯核，供大家参考。

Talk is cheap, show me the code

import tensorflow as tf
from sklearn.base import BaseEstimator, ClassifierMixin
import numpy as np

class TFSVM(BaseEstimator, ClassifierMixin):

 def __init__(self, 
  C = 1, kernel = 'linear', 
  learning_rate = 0.01, 
  training_epoch = 1000, 
  display_step = 50,
  batch_size = 50,
  random_state = 42):
  #参数列表
  self.svmC = C
  self.kernel = kernel
  self.learning_rate = learning_rate
  self.training_epoch = training_epoch
  self.display_step = display_step
  self.random_state = random_state
  self.batch_size = batch_size

 def reset_seed(self):
  #重置随机数
  tf.set_random_seed(self.random_state)
  np.random.seed(self.random_state)

 def random_batch(self, X, y):
  #调用随机子集，实现mini-batch gradient descent
  indices = np.random.randint(1, X.shape[0], self.batch_size)
  X_batch = X[indices]
  y_batch = y[indices]
  return X_batch, y_batch

 def _build_graph(self, X_train, y_train):
  #创建计算图
  self.reset_seed()

  n_instances, n_inputs = X_train.shape

  X = tf.placeholder(tf.float32, [None, n_inputs], name = 'X')
  y = tf.placeholder(tf.float32, [None, 1], name = 'y')

  with tf.name_scope('trainable_variables'):
   #决策边界的两个变量
   W = tf.Variable(tf.truncated_normal(shape = [n_inputs, 1], stddev = 0.1), name = 'weights')
   b = tf.Variable(tf.truncated_normal([1]), name = 'bias')

  with tf.name_scope('training'):
   #算法核心
   y_raw = tf.add(tf.matmul(X, W), b)
   l2_norm = tf.reduce_sum(tf.square(W))
   hinge_loss = tf.reduce_mean(tf.maximum(tf.zeros(self.batch_size, 1), tf.subtract(1., tf.multiply(y_raw, y))))
   svm_loss = tf.add(hinge_loss, tf.multiply(self.svmC, l2_norm))
   training_op = tf.train.AdamOptimizer(learning_rate = self.learning_rate).minimize(svm_loss)

  with tf.name_scope('eval'):
   #正确率和预测
   prediction_class = tf.sign(y_raw)
   correct_prediction = tf.equal(y, prediction_class)
   accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

  init = tf.global_variables_initializer()

  self._X = X; self._y = y
  self._loss = svm_loss; self._training_op = training_op
  self._accuracy = accuracy; self.init = init
  self._prediction_class = prediction_class
  self._W = W; self._b = b

 def _get_model_params(self):
  #获取模型的参数，以便存储
  with self._graph.as_default():
   gvars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
  return {gvar.op.name: value for gvar, value in zip(gvars, self._session.run(gvars))}

 def _restore_model_params(self, model_params):
  #保存模型的参数
  gvar_names = list(model_params.keys())
  assign_ops = {gvar_name: self._graph.get_operation_by_name(gvar_name + '/Assign') for gvar_name in gvar_names}
  init_values = {gvar_name: assign_op.inputs[1] for gvar_name, assign_op in assign_ops.items()}
  feed_dict = {init_values[gvar_name]: model_params[gvar_name] for gvar_name in gvar_names}
  self._session.run(assign_ops, feed_dict = feed_dict)

 def fit(self, X, y, X_val = None, y_val = None):
  #fit函数，注意要输入验证集
  n_batches = X.shape[0] // self.batch_size

  self._graph = tf.Graph()
  with self._graph.as_default():
   self._build_graph(X, y)

  best_loss = np.infty
  best_accuracy = 0
  best_params = None
  checks_without_progress = 0
  max_checks_without_progress = 20

  self._session = tf.Session(graph = self._graph)

  with self._session.as_default() as sess:
   self.init.run()

   for epoch in range(self.training_epoch):
    for batch_index in range(n_batches):
     X_batch, y_batch = self.random_batch(X, y)
     sess.run(self._training_op, feed_dict = {self._X:X_batch, self._y:y_batch})
    loss_val, accuracy_val = sess.run([self._loss, self._accuracy], feed_dict = {self._X: X_val, self._y: y_val})
    accuracy_train = self._accuracy.eval(feed_dict = {self._X: X_batch, self._y: y_batch})

    if loss_val < best_loss:
     best_loss = loss_val
     best_params = self._get_model_params()
     checks_without_progress = 0
    else:
     checks_without_progress += 1
     if checks_without_progress > max_checks_without_progress:
      break

    if accuracy_val > best_accuracy:
     best_accuracy = accuracy_val
     #best_params = self._get_model_params()

    if epoch % self.display_step == 0:
     print('Epoch: {}\tValidaiton loss: {:.6f}\tValidation Accuracy: {:.4f}\tTraining Accuracy: {:.4f}'
      .format(epoch, loss_val, accuracy_val, accuracy_train))
   print('Best Accuracy: {:.4f}\tBest Loss: {:.6f}'.format(best_accuracy, best_loss))
   if best_params:
    self._restore_model_params(best_params)
    self._intercept = best_params['trainable_variables/weights']
    self._bias = best_params['trainable_variables/bias']
   return self

 def predict(self, X):
  with self._session.as_default() as sess:
   return self._prediction_class.eval(feed_dict = {self._X: X})

 def _intercept(self):
  return self._intercept

 def _bias(self):
  return self._bias

实际运行效果如下（以Iris数据集为样本）：

画出决策边界来看看：