如何自定义MATLAB神经网络的误差函数
我想将此函数实现为训练神经网络的误差函数：
function err = MyErrorFunction(T,O)
d = T - O;
err = -d*( exp(-d) - 1 );
end
其中 T 是目标值，O 是输入的神经网络输出。
训练算法并不重要（显然 trainlm 的误差函数不可自定义，因此我可以使用 trainscg）。
我发现 本文建议使用template_performance.m 文件来定义新的性能函数。我说我只是复制这个文件并根据需要自定义它。
显然，从 MATLAB 2010 开始，template_performance.m 已被弃用。
那么，在训练神经网络时如何改变计算/评估误差/性能的方式呢？
I want to implement this function as the error function for training a neural network:
function err = MyErrorFunction(T,O)
d = T - O;
err = -d*( exp(-d) - 1 );
end
where T is target value and O is neural network output for an input.
The training algorithm doesn't matter (apparently error function for trainlm is not customizable, so I can go with the trainscg).
I've found this article that suggests using the template_performance.m file to define a new performance function. I says I just have copy this file and customize it as I want.
But as I have understood, `template_performance.m` is a template for **performance** function, not the **error** function: `template_performance.m` gets the error values and output a performance value, for instance it could sum up the square of the errors and output them (SSE).
Apparently, template_performance.m have been deprecated starting from MATLAB 2010.
So, how I can change the way that error/performance is calculated/evaluated when training a neural network?
如果你对这篇内容有疑问，欢迎到本站社区发帖提问 参与讨论，获取更多帮助，或者扫码二维码加入 Web 技术交流群。
绑定邮箱获取回复消息
由于您还没有绑定你的真实邮箱，如果其他用户或者作者回复了您的评论，将不能在第一时间通知您！
发布评论
评论（2）
我也遇到了类似的问题……定制性能函数的整个过程完全是一场灾难。很多东西都在贬值，而且关于我们应该做什么的文档为零。
我最终不得不破解我不打算使用的性能函数（SSE）的核心文件。在 matlab 目录
MATLAB\R2012b\toolbox\nnet\nnet\nnperformance下您可以找到它们。我根据t、t-1的变化修改了 apply.m（在 SSE+ 文件夹中）函数，并使用了一些方向权重。但后来我遇到了训练算法以与perform()不同的方式和格式向apply()发送参数的问题。我最终没有使用perform()并为此编写了自己的代码。天啊……一团糟。这是一个非常丑陋的黑客行为，我很想听听任何找到正确方法的人的意见。
I had a similar problem ... the whole thing about customizing a performance function is a total disaster. A lot of stuff is being depreciated and there is zero documentation on what we should do.
I ended up having to hack the core files of a performance function I wasn't planning on using (SSE). Under the matlab directory
MATLAB\R2012b\toolbox\nnet\nnet\nnperformanceyou can find them. I modified the apply.m (in the SSE+ folder) function with some directional weights based ont,t-1change. But then I ran into the problem of training algorithms sending parameters toapply()in a different way AND format thanperform(). I ended up not usingperform()and writing my own code for that. Jesus... Total mess.This was a very ugly hack and I'd love to hear from anyone who found the correct way to do this.
据我了解，性能函数既用于训练又用于测试/评估（除非某种训练算法被硬编码为特定函数）
From what I understand, the performance function is used both for training and testing/evaluation (unless a certain training algorithm is hard-coded to a specific function)