YOLOv3v4-ModelCompression-MultidatasetTraining-Multibackbone icon indicating copy to clipboard operation
YOLOv3v4-ModelCompression-MultidatasetTraining-Multibackbone copied to clipboard

使用知识蒸馏策略四进行训练出现feature mismatch

Open shaolingongfuhao opened this issue 5 years ago • 5 comments
trafficstars

怎么才能使feature_s和feature_t长度相同

shaolingongfuhao avatar Oct 26 '20 15:10 shaolingongfuhao

添加feature loss的整理方法都需要保证层数一致,一般使用原模型蒸馏通道剪植之后的模型。

SpursLipu avatar Oct 27 '20 03:10 SpursLipu

谢谢你的回复。我的老师模型是YOLOV3,学生模型经过通道和层剪枝的模型。 我是否能够将feature_t截取后半部分与feature_s相同层数的特征来和feature_s求loss,这种方法合适吗? 代码中你写的是哪些层会输出feature_out,有一部分没看懂。

------------------ 原始邮件 ------------------ 发件人: "SpursLipu"<[email protected]>; 发送时间: 2020年10月27日(星期二) 中午11:19 收件人: "SpursLipu/YOLOv3v4-ModelCompression-MultidatasetTraining-Multibackbone"<YOLOv3v4-ModelCompression-MultidatasetTraining-Multibackbone@noreply.github.com>; 抄送: "张晨阳"<[email protected]>; "Author"<[email protected]>; 主题: Re: [SpursLipu/YOLOv3v4-ModelCompression-MultidatasetTraining-Multibackbone] 使用知识蒸馏策略四进行训练出现feature mismatch (#79)

添加feature loss的整理方法都需要保证层数一致,一般使用原模型蒸馏层剪植之后的模型。

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

shaolingongfuhao avatar Oct 27 '20 03:10 shaolingongfuhao

理论上是可以的,你可以改改试试看

SpursLipu avatar Oct 27 '20 03:10 SpursLipu

我在compute_lost_KD4代码里改了一下,出现了这种错误,这是为什么?

------------------ 原始邮件 ------------------ 发件人: "SpursLipu"<[email protected]>; 发送时间: 2020年10月27日(星期二) 中午11:38 收件人: "SpursLipu/YOLOv3v4-ModelCompression-MultidatasetTraining-Multibackbone"<YOLOv3v4-ModelCompression-MultidatasetTraining-Multibackbone@noreply.github.com>; 抄送: "张晨阳"<[email protected]>; "Author"<[email protected]>; 主题: Re: [SpursLipu/YOLOv3v4-ModelCompression-MultidatasetTraining-Multibackbone] 使用知识蒸馏策略四进行训练出现feature mismatch (#79)

理论上是可以的,你可以改改试试看

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

shaolingongfuhao avatar Oct 27 '20 14:10 shaolingongfuhao

刚刚从tensorboard读数据时,还发现这个tensorboard记录数据好像会少记几个点,我微调和蒸馏都是50epoch,但是这些记录都是缺1-3个点,上图分别少了2个和3个,这是什么原因呢

------------------ 原始邮件 ------------------ 发件人: "SpursLipu"<[email protected]>; 发送时间: 2020年10月27日(星期二) 中午11:38 收件人: "SpursLipu/YOLOv3v4-ModelCompression-MultidatasetTraining-Multibackbone"<YOLOv3v4-ModelCompression-MultidatasetTraining-Multibackbone@noreply.github.com>; 抄送: "张晨阳"<[email protected]>; "Author"<[email protected]>; 主题: Re: [SpursLipu/YOLOv3v4-ModelCompression-MultidatasetTraining-Multibackbone] 使用知识蒸馏策略四进行训练出现feature mismatch (#79)

理论上是可以的,你可以改改试试看

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

shaolingongfuhao avatar Oct 27 '20 15:10 shaolingongfuhao