def SGD(self, training_data, epochs, mini_batch_size, eta, test_data=None):
 if test_data:
  n_test = len(test_data)#有多少個測試集
  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)#目標函數(shù)對b和w的偏導(dǎo)數(shù)
  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
 #最終更新權(quán)重為
 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)]