Runtimeerror: Cudnn Rnn Backward Can Only Be Called in Training Mode
Trouble description ：
RuntimeError: cudnn RNN backward tin merely be chosen in training fashion
Problem assay ：
The reason is that information technology can't be right or wrong in back propagation train state .
My code added a before back propagation Alexnet,Alexnet Adding loss And so it will spread dorsum past itself , At this time, in the whole LSTM This non-linear phenomenon volition appear in the back propagation train Status problems .
solve ：
employ Baidu Search ：
I've tried all the methods on the first folio , Information technology'south all useless ...
Thorough solution ：
Add... Before you beginning preparation ：
torch.backends.cudnn.enabled=False Tin completely solve ！！！
to update ：
That method just now can merely forcibly change the state of the current model , The core trouble is that the lawmaking appears issues, Check carefully train What places trigger non train state . Information technology's common to trigger evaluation country .
A good way to debug is to enter each loop railroad train Record the current epoch And impress it out , If not train, It oft leads to repeated entry into the bicycle .
If there is a tensor stored in GPU, It oft causes the video retentiveness to explode first ！！
版权声明
本文为[A wind attorney fond to bicycles]所创，转载请带上原文链接，感谢
https://chowdera.com/2021/08/20210825022738918e.html
Source: https://chowdera.com/2021/08/20210830070053548n.html
0 Response to "Runtimeerror: Cudnn Rnn Backward Can Only Be Called in Training Mode"
Post a Comment