​inplace​​ 操作是 PyTorch 里面一个比较常见的错误,有的时候会比较好发现,例如下面的代码:

import torch
w = torch.rand(4, requires_grad=True)
w += 1
loss = w.sum()
loss.backward()

执行 ​​loss​​​ 对参数 ​​w​​​ 进行求导,会出现报错:​​RuntimeError: a leaf Variable that requires grad is being used in an in-place operation.​

导致这个报错的主要是第 3 行代码 ​​w += 1​​​,如果把这句改成 ​​w = w + 1​​​,再执行就不会报错了。这种写法导致的 ​​inplace operation​​ 是比较好发现的,但是有的时候同样类似的报错,会比较不好发现。例如下面的代码:

import torch
x = torch.zeros(4)
w = torch.rand(4, requires_grad=True)
x[0] = torch.rand(1) * w[0]
for i in range(3):
x[i+1] = torch.sin(x[i]) * w[i]
loss = x.sum()
loss.backward()

执行之后会出现报错:

>>> RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: 
[torch.FloatTensor []], which is output 0 of SelectBackward, is at version 4; expected version 3 instead.
Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).

根据提示我们可以使用 ​​with torch.autograd.set_detect_anomaly(True)​​ 来帮助我们定位具体的出错位置(这个方法会花费比较长的时间)。

with torch.autograd.set_detect_anomaly(True):
x = torch.zeros(4)
w = torch.rand(4, requires_grad=True)
x[0] = torch.rand(1) * w[0]
for i in range(3):
x[i+1] = torch.sin(x[i]) * w[i]
loss = x.sum()
loss.backward()

运行会增加这些报错:

>>> /Users/strongnine/anaconda3/lib/python3.8/site-packages/torch/autograd/__init__.py:130: 
UserWarning: Error detected in SinBackward. Traceback of forward call that caused the error:

可以看到出现了 ​​Error detected in SinBackward.​​​,这句描述,我们可以猜测大概是 ​​torch.sin()​​​ 这个函数出现了问题。实际上,这个报错的解决办法,就是将第 6 行代码 ​​x[i+1] = torch.sin(x[i]) * w[i]​​​ 改成 ​​x[i+1] = torch.sin(x[i].clone()) * w[i]​​,就行了。

import torch
x = torch.zeros(4)
w = torch.rand(4, requires_grad=True)
x[0] = torch.rand(1) * w[0]
for i in range(3):
x[i+1] = torch.sin(x[i].clone()) * w[i]
loss = x.sum()
loss.backward()

总结一下,遇到 ​​inplace operation​​ 的报错,一般可以通过:

  • ​x += 1​​​ 改成​​x = x + 1​​;
  • ​x[:, :, 0:3] = x[:, :, 0:3] + 1​​​ 改成​​x[:, :, 0:3] = x[:, :, 0:3].clone() + 1​​​;​1
  • ​x[i+1] = torch.sin(x[i]) * w[i]​​​ 改成​​x[i+1] = torch.sin(x[i].clone()) * w[i]​​;

如果自己检查不出是哪里出现了问题,可以使用 ​​with torch.autograd.set_detect_anomaly(True)​​ 来帮助我们定位具体的出错位置,但是要注意的是这个方法一般会运行比较长的时间。


  1. ​知乎:PyTorch debug 整理​​ ​​↩︎​