def SGD(self, training_data, epochs, mini_batch_size, eta, test_data=None):
  if test_data:
    n_test = len(test_data)#有多少个测试集
    n = len(training_data)
    for j in xrange(epochs):
      random.shuffle(training_data)
      mini_batches = [
        training_data[k:k+mini_batch_size] 
        for k in xrange(0,n,mini_batch_size)]
      for mini_batch in mini_batches:
        self.update_mini_batch(mini_batch, eta)
      if test_data:
        print "Epoch {0}: {1}/{2}".format(j, self.evaluate(test_data),n_test)
      else:
        print "Epoch {0} complete".format(j)  

def update_mini_batch(self, mini_batch,eta): 
  nabla_b = [np.zeros(b.shape) for b in self.biases]
  nabla_w = [np.zeros(w.shape) for w in self.weights]
  for x,y in mini_batch:
    delta_nabla_b, delta_nable_w = self.backprop(x,y)#目标函数对b和w的偏导数
    nabla_b = [nb+dnb for nb,dnb in zip(nabla_b,delta_nabla_b)]
    nabla_w = [nw+dnw for nw,dnw in zip(nabla_w,delta_nabla_w)]#累加b和w
  #最终更新权重为
  self.weights = [w-(eta/len(mini_batch))*nw for w, nw in zip(self.weights, nabla_w)]
  self.baises = [b-(eta/len(mini_batch))*nb for b, nb in zip(self.baises, nabla_b)]