我在pytorch有一个深度学习模型。我训练了100个纪元的模型，随着纪元的增加，它需要更多的内存。你能帮我一下吗？据我所知，内存使用量应该是恒定的。我说错了吗？或者我在下面的代码中犯了一个错误？下面是内存使用情况的图表。还有我用来训练和验证迷你批处理和批处理的函数。- 谢谢你-谢谢
注：我的模型是一个条件变分自动编码器，损耗与kl损耗等有关。

def runs_for_validate(test_it, num_samples = 100):
    len_test_it = len(test_it)
    # set the evaluation mode
    model.eval()
    # add for validate 
    L_loss_b, L_loss, test_loss = 0, 0, 0
    # test loss for the data
    samples   = {}
    num_batch = 0
    
    with torch.no_grad():
        
        for local_batch, local_labels in test_it:
            
            length = local_batch.size(0)
            
            local_batch, local_labels = local_batch.to(device), local_labels.to(device)
            
            samples, mu_par, log_var_par, mu_r, log_var_r = model.test_for_validate(local_batch,num_samples)
            samples = np.transpose(np.array(samples),(1,0,2)) 
            
            truths  = local_labels
            
            num_batch+=1
            #break
            L_loss_b = log_lik(local_labels, mu_par, log_var_par)
            
            L_loss   += (torch.sum(L_loss_b))/length
            
        test_loss = L_loss/num_batch
        
    return samples,truths.cpu().numpy(), test_loss

def train_batch(model, optimizer, device, batch, labels):
    
    model.train()
    
    optimizer.zero_grad()
    
    length = float(batch.size(0))
    
    mu_x, log_var_x, mu_q, log_var_q, mu_r, log_var_r = model(batch,labels)

    kl_loss_b = KL(mu_r,log_var_r,mu_q,log_var_q)
    
    L_loss_b = log_lik(labels, mu_x, log_var_x)
    
    #print("Size of list3: " + str(sys.getsizeof(train_losses)) + "bytes")
    
    L_loss    = torch.sum(L_loss_b)
    
    kl_loss = torch.sum(kl_loss_b)
    
    loss = -(L_loss - kl_loss)/length
    
    loss.backward()
    
    # update the weights                                                                                                                              
    optimizer.step()
    
    # add for validation 
    return loss, kl_loss/length, L_loss/length

### when we want to check with validate data
def trainv(model, device, epochs, train_iterator, optimizer, validate_iterator):
    
    n_samples = 100

    train_losses, kl_losses, lik_losses, test_losses   = [], [],[], []
    
    for epoch in range(epochs): 
        
        ep_tr, ep_kl,ep_l, num_batch, iterator  = 0,0,0, 0, 0
        
        for local_batch, local_labels in train_iterator:
            
            local_batch, local_labels = local_batch.to(device), local_labels.to(device)
            
            train_loss, kl_loss, lik_loss = train_batch(model, optimizer, device, local_batch,local_labels)
            
            ep_tr += train_loss
            
            ep_kl += kl_loss
            
            ep_l += lik_loss
            
            num_batch+= 1
            
            iterator+=1
            
            del local_batch, local_labels
        
        train_losses.append(ep_tr/num_batch)
        
        #kl_losses.append(kl_loss/num_batch)
        
        #lik_losses.append(-lik_loss/num_batch)
        
        run_validate_flag = 0
        
        if run_validate_flag ==1:
            
            samples, truths, test_loss = runs_for_validate(validate_iterator, n_samples)
            
            test_losses.append(-test_loss)
            
        else:
            test_losses = f'run_validate_flag;{0}'
        
        
        if epoch % int(epochs/5) == 0:
            samples, truths, test_loss = runs_for_validate(validate_iterator, n_samples)
            print("Epoch: {}, Training loss: {}, val_loss : {} ".format(epoch,train_losses[epoch], -1*test_loss))
    #return train_losses, kl_losses, lik_losses, test_losses

    return train_losses, test_losses

字符串
x1c 0d1x的数据