Encountered an error:
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation
This is generally a problem of non-guidance, that is, loss.backward() may exist:
1. In-place problem: if x= x+k is an inplace operation, the variable will not be guided
Other similar inplace operations
2. There is another possibility: the code of optimizer.step() is placed before loss.backward, resulting in the unsuccessful update of the gradient
In conclusion: The problem is all in loss.backward(), the gradient is not updated successfully. hope to help everyone
Error code:
This bug was solved by the help of fser, thank you very much!
fser的blog：The forest in the fog - Ruochen's Blog
My Zhihu homepage: German Viviane - Zhihu