如何正确实现动量和衰减 - SGD

问题描述

我正在尝试将动量和衰减应用于小批量 SGD： 更新权重的正确方法是什么，一旦设置衰减，我就会得到奇怪的结果..

  import numpy as np
  def _mini_batch(self,X,y,batch_size):
      # sack data for shuffle - mini batch
      rows = len(X)
      X_full = np.hstack(( np.ones((rows,1)),np.array(y).reshape(rows,-1) ))
      np.random.shuffle(X_full)
      # Performing minibatch
      num_batches = rows // batch_size  
      for rng in range(num_batches): 
        start_rng,end_rng = rng*batch_size,(rng+1)*batch_size
        yield X_full[start_rng:end_rng,:-1],X_full[start_rng:end_rng,-1]  # X_batch,y_batch
      if not rows % batch_size == 0: 
          yield X_full[end_rng:rows,X_full[end_rng:rows,-1] # X_batch,y_batch      
decay_rate = 0.2
alpha = 0.1 #learning rate
weights =  np.random.normal(size=X.shape[-1]) #np.zeros(X.shape[-1])   
rows = len(X)
for i in range(epochs): 
     # init mini-batch to update gradients for each batch
     for X_batch,y_batch in _mini_batch(X,batch_size):    
          
          train_predictions = np.dot(X,weights)  #y_hat
          errors =   np.subtract(train_predictions,y)
          
          self.weights = (1. - decay_rate) * weights - alpha * np.dot(X.T,errors) / rows 

  

解决方法

暂无找到可以解决该程序问题的有效方法，小编努力寻找整理中！

如果你已经找到好的解决方法，欢迎将解决方案带上本链接一起发送给小编。

小编邮箱:dio#foxmail.com (将#修改为@）