【Error】: Trying to backward through the graph a second time, but the buffers have already been

最新推荐文章于 2025-09-26 11:58:38 发布

codebrid

最新推荐文章于 2025-09-26 11:58:38 发布

阅读量1w

点赞数

CC 4.0 BY-SA版权

本文链接：https://blog.csdn.net/ccbrid/article/details/80050399

本文解决了PyTorch中因使用同一输入多次反向传播导致的运行时错误。通过在每次反向传播前重新计算特征，避免了释放缓冲区的问题。适用于需要对不同任务进行梯度累积的情况。

错误：

RuntimeError: Trying to backward through the graph a second time, but the buffers have already been freed. Specify retain_graph=True when calling backward the first time.

代码：

optimizer.zero_grad() # word
features = encoder(src)
loss = decoder(features, trg)
loss.backward()
grad_norm = torch.nn.utils.clip_grad_norm(decoder.parameters(), opt.grad_clip)############3
optimizer.step()

optimizer.zero_grad() # pos
# features = encoder(src)
loss = decoder.forward_pos(features, trg_pos)
loss.backward()
grad_norm = torch.nn.utils.clip_grad_norm(decoder.parameters(), opt.grad_clip)
optimizer.step()

代码更改为：

optimizer.zero_grad() # word
features = encoder(src)
loss = decoder(features, trg)
loss.backward()
grad_norm = torch.nn.utils.clip_grad_norm(decoder.parameters(), opt.grad_clip)
optimizer.step()

optimizer.zero_grad() # pos
features = encoder(src)
loss = decoder.forward_pos(features, trg_pos)
loss.backward()
grad_norm = torch.nn.utils.clip_grad_norm(decoder.parameters(), opt.grad_clip)
optimizer.step()

错误消失。

参考网址：https://blog.csdn.net/u010829672/article/details/79538853