脚本之家,脚本语言编程技术及教程分享平台!
分类导航

Python|VBS|Ruby|Lua|perl|VBA|Golang|PowerShell|Erlang|autoit|Dos|bat|

服务器之家 - 脚本之家 - Python - pytorch显存一直变大的解决方案

pytorch显存一直变大的解决方案

2021-10-07 00:22zhenggeaza Python

这篇文章主要介绍了pytorch显存一直变大的解决方案,具有很好的参考价值,希望对大家有所帮助。一起跟随小编过来看看吧

在代码中添加以下两行可以解决:

?
1
2
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True

补充:pytorch训练过程显存一直增加的问题

之前遇到了爆显存的问题,卡了很久,试了很多方法,总算解决了。

总结下自己试过的几种方法:

**1. 使用torch.cuda.empty_cache()

在每一个训练epoch后都添加这一行代码,可以让训练从较低显存的地方开始,但并不适用爆显存的问题,随着epoch的增加,最大显存占用仍然会提示out of memory 。

2.使用torch.backends.cudnn.enabled = True 和 torch.backends.cudnn.benchmark = True

原理不太清楚,用法和1一样。但是几乎没有效果,直接pass。

3.最重要的:查看自己的forward函数是否存在泄露。

常需要在forward函数里调用其他子函数,这时候要特别注意:

input尽量不要写在for循环里面!!!

子函数里如果有append()等函数,一定少用,能不用就不用!!!

子函数list一定少用,能不用就不用!!!

总之,子函数一般也不会太复杂,直接写出来,别各种for,嵌套,变量。!!!

补充:Pytorch显存不断增长问题的解决思路

这个问题,我先后遇到过两次,每次都异常艰辛的解决了。

在网上,关于这个问题,你可以找到各种看似不同的解决方案,但是都没能解决我的问题。所以只能自己摸索,在摸索的过程中,有了一个排查问题点的思路。

下面举个例子说一下我的思路。

大体思路

其实思路很简单,就是在代码的运行阶段输出显存占用量,观察在哪一块存在显存剧烈增加或者显存异常变化的情况。

但是在这个过程中要分级确认问题点,也即如果存在三个文件main.py、train.py、model.py。

在此种思路下,应该先在main.py中确定问题点,然后,从main.py中进入到train.py中,再次输出显存占用量,确定问题点在哪。

随后,再从train.py中的问题点,进入到model.py中,再次确认。

如果还有更深层次的调用,可以继续追溯下去。

具体例子

main.py

?
1
2
3
4
5
6
7
def train(model,epochs,data):
    for e in range(epochs):
        print("1:{}".format(torch.cuda.memory_allocated(0)))
        train_epoch(model,data)
        print("2:{}".format(torch.cuda.memory_allocated(0)))
        eval(model,data)
        print("3:{}".format(torch.cuda.memory_allocated(0)))

假设1与2之间显存增加极为剧烈,说明问题出在train_epoch中,进一步进入到train.py中。

train.py

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
def train_epoch(model,data):
    model.train()
    optim=torch.optimizer()
    for batch_data in data:
        print("1:{}".format(torch.cuda.memory_allocated(0)))
        output=model(batch_data)
        print("2:{}".format(torch.cuda.memory_allocated(0)))
        loss=loss(output,data.target)
        print("3:{}".format(torch.cuda.memory_allocated(0)))
        optim.zero_grad()
        print("4:{}".format(torch.cuda.memory_allocated(0)))
        loss.backward()
        print("5:{}".format(torch.cuda.memory_allocated(0)))
        utils.func(model)
        print("6:{}".format(torch.cuda.memory_allocated(0)))

如果在1,2之间,5,6之间同时出现显存增加异常的情况。此时需要使用控制变量法,例如我们先让5,6之间的代码失效,然后运行,观察是否仍然存在显存爆炸。如果没有,说明问题就出在5,6之间下一级的代码中。进入到下一级代码,进行调试:

utils.py

?
1
2
3
4
5
6
7
8
9
10
def func(model):
print("1:{}".format(torch.cuda.memory_allocated(0)))
a=f1(model)
print("2:{}".format(torch.cuda.memory_allocated(0)))
b=f2(a)
print("3:{}".format(torch.cuda.memory_allocated(0)))
c=f3(b)
print("4:{}".format(torch.cuda.memory_allocated(0)))
d=f4(c)
print("5:{}".format(torch.cuda.memory_allocated(0)))

此时我们再展示另一种调试思路,先注释第5行之后的代码,观察显存是否存在先训爆炸,如果没有,则注释掉第7行之后的,直至确定哪一行的代码出现导致了显存爆炸。假设第9行起作用后,代码出现显存爆炸,说明问题出在第九行,显存爆炸的问题锁定。

几种导致显存爆炸的情况

pytorch的hook机制可能导致,显存爆炸,hook函数取出某一层的输入输出跟权重后,不可进行存储,修改等操作,这会造成hook不能回收,进而导致取出的输入输出权重都可能不被pytorch回收,所以模型的负担越来也大,最终导致显存爆炸。

这种情况是我第二次遇到显存爆炸查出来的,非常让人匪夷所思。在如下代码中,p.sub_(torch.mm(k, torch.t(k)) / (alpha + torch.mm(r, k))),导致了显存爆炸,这个问题点就是通过上面的方法确定的。

这个P是一个矩阵,在使用p.sub_的方式更新P的时候,导致了显存爆炸。

将这行代码修改为p=p-(torch.mm(k, torch.t(k)) / (alpha + torch.mm(r, k))),显存爆炸的问题解决。

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
def pro_weight(p, x, w, alpha=1.0, cnn=True, stride=1):
              if cnn:
                  _, _, H, W = x.shape
                  F, _, HH, WW = w.shape
                  S = stride  # stride
                  Ho = int(1 + (H - HH) / S)
                  Wo = int(1 + (W - WW) / S)
                  for i in range(Ho):
                      for j in range(Wo):
                          # N*C*HH*WW, C*HH*WW = N*C*HH*WW, sum -> N*1
                          r = x[:, :, i * S: i * S + HH, j * S: j * S + WW].contiguous().view(1, -1)
                          # r = r[:, range(r.shape[1] - 1, -1, -1)]
                          k = torch.mm(p, torch.t(r))
                          p.sub_(torch.mm(k, torch.t(k)) / (alpha + torch.mm(r, k)))
                  w.grad.data = torch.mm(w.grad.data.view(F, -1), torch.t(p.data)).view_as(w)
              else:
                  r = x
                  k = torch.mm(p, torch.t(r))
                  p.sub_(torch.mm(k, torch.t(k)) / (alpha + torch.mm(r, k)))
                  w.grad.data = torch.mm(w.grad.data, torch.t(p.data))

以上为个人经验,希望能给大家一个参考,也希望大家多多支持服务器之家。如有错误或未考虑完全的地方,望不吝赐教。

原文链接:https://www.jianshu.com/p/8486076ddd10

延伸 · 阅读

精彩推荐