迁移pytorch 4.0 问题 inplace 操作和tensor error

 

pytorch 4. 0   

KeyError: tensor

This is an obvious error. In the dict self.int_to_char, its keys are the type of integer rather than tensors. Therefore, you can change source code in line 169 as:

self.int_to_char[sequence[i].item()]

Should be fine.
说白了就是字典的key不是tensor而是值，所以在tensor后面加.item()

 

One of the variables needed for gradient computation has been modified by an inplace operation;

这是原地计算，去掉inplace操作，几个基本的要解决，一个是+=问题就是形如out+=x的改变成out = out+x，还有一个就是这个self.flatten_parameters()操作移动到init 中去，因为在GPU中计算时候回修改LSTM值，会使得backward不能进行。######

Move the self.flatten_parameters() call in def forward to def __init__. It has no effect on CPU, but on CUDA it will modify the LSTM’s parameters part way through the computation which would break backwards (leading to the exception).