秋月的私语 comments

Results 7 comments of


                                            秋月的私语

size mismatch for transformer_decoder.layers.0.0.fn.fn.to_q.weight: copying a param with shape torch.Size([512, 32]) from checkpoint, the shape in current model is torch.Size([64, 32]).

经过不断调试和测试，花了一天的时间才找到问题的所在。。。因为可能是作者故意留下的问题，所以这里就先不写出解决办法了，需要解决办法的可以私信我取得联系，9506#163.com

size mismatch for transformer_decoder.layers.0.0.fn.fn.to_q.weight: copying a param with shape torch.Size([512, 32]) from checkpoint, the shape in current model is torch.Size([64, 32]).

> 已经回复，请查收邮件

size mismatch for transformer_decoder.layers.0.0.fn.fn.to_q.weight: copying a param with shape torch.Size([512, 32]) from checkpoint, the shape in current model is torch.Size([64, 32]).

加你了可视化的问题我没注意测试可能我也帮不了你什么   秋月的私语 ***@***.***   ------------------ 原始邮件 ------------------ 发件人: ***@***.***>; 发送时间: 2022年9月12日(星期一) 中午11:16 收件人: ***@***.***>; 抄送: ***@***.***>; ***@***.***>; 主题: Re: [justchenhao/BIT_CD] size mismatch for transformer_decoder.layers.0.0.fn.fn.to_q.weight: copying a param with shape...

How could I upgrade the accuracy?

with the help of [stiwfmjX](https://github.com/stiwfmjX) ，i found that it is beacuse i don't cut the data into 256*256 so the result is not good, even if i tried to resize...

size mismatch for transformer_decoder.layers.0.0.fn.fn.to_q.weight: copying a param with shape torch.Size([512, 32]) from checkpoint, the shape in current model is torch.Size([64, 32]).

> 他把那个测试模型改了，测试模型换成和训练模型一样就行，改了半天，我以为是模型出错了呢，打印保存模型和现有模型才发现这个问题是的就是这个问题我当时也是搞了半天。。

How could I upgrade the accuracy?

> > with the help of [stiwfmjX](https://github.com/stiwfmjX) ，i found that it is beacuse i don't cut the data into 256_256 so the result is not good, even if i tried...

How could I upgrade the accuracy?

> > > > > > > 在 [stiwfmjX](https://github.com/stiwfmjX) 的帮助下，我发现这是因为我没有将数据剪切成 256_256，所以结果不好，即使我尝试将数据大小调整为 256_256，而它不起作用。我意识到在试用后调整数据大小而不是将数据剪切到 256_256 是不正确的，所以非常感谢。最重要的是，您应该在训练之前将数据（https://pan.baidu.com/s/1fzNiOE7elGRmIo2h6MIhZw 代码：L7IV）剪切成 256_256 以获得良好的结果。我的是“Historical_best_acc=0.9531（在纪元 458）”。 > > > > > > > > > 你能分享一下吗？ > >...