分类：loss.backward报错

2021
07-26

PyTorch训练LSTM时loss.backward()报错的解决方案

训练用PyTorch编写的LSTM或RNN时，在loss.backward()上报错：RuntimeError:Tryingtobackwardthroughthegraphasecondtime,butthebuffershavealreadybeenfreed.Specifyretain_graph=Truewhencallingbackwardthefirsttime.千万别改成loss.backward(retain_graph=True)，会导致显卡内存随着训练一直增加直到OOM：RuntimeError:CUDAoutofmemory.Triedtoallocate20.00MiB(GPU0;10.73GiBtotalcapa... 继续阅读 >