简体繁体 English

如何估计用Envg用Levenberg-Marquardt算法训练具有特定样本集的特定网络所需的RAM数量？

[英]How to estimate the amount of RAM it will take Encog to train a particular network with a particular sample set with Levenberg-Marquardt algorithm?

原文 2012-03-27 07:06:05 4 1 java/ algorithm/ neural-network/ encog

I am studying Encog (versions 3.0 and 3.1 for Java) neural network framework and would like to try Levenberg–Marquardt algorithm in my neural network training. 我正在研究Encog（Java的3.0和3.1版本）神经网络框架，并且想在我的神经网络培训中尝试使用Levenberg–Marquardt算法。 However my actual neural networks are pretty complex natural (featuring hundreds of input neurons) and not a single (I've been simplifying from tens of thousands to 136) network appears to be possible to be trained with Levenberg–Marquardt (saying there is not enough memory while I dedicate 2 GiBs to the JVM). 但是，我实际的神经网络是非常复杂的自然网络（拥有数百个输入神经元），并且没有一个（我已经从数万个简化为136个神经网络）似乎可以用Levenberg-Marquardt训练（说没有足够的内存，而我将2个GiB专用于JVM）。 But it performs very good when I try a simple synthetic task with a simple (just 3 input neurons) network. 但是，当我尝试使用简单的（仅3个输入神经元）网络进行简单的合成任务时，它的效果非常好。

The question is how can I estimate the amount of memory it will take Encog to train a particular network with a particular sample set with Levenberg-Marquardt algorithm? 问题是我该如何估计Encog使用Levenberg-Marquardt算法训练具有特定样本集的特定网络所需的内存量？ How can I guess what is the maximum complexity can be fit in 2 GiBs (the actual machine has 4, but the JVM seems unable to allocate more than 2)? 我怎么能猜到2个GiB可以适应的最大复杂度是多少（实际机器有4个，但是JVM似乎无法分配2个以上）？

1 个解决方案

I would try a small size, determine how much memory it uses and keep doubling the size until you see a pattern. 我将尝试使用较小的大小，确定其使用的内存量，并继续将大小增加一倍，直到看到模式为止。 ie is it linear or super linear. 即是线性还是超线性。

BTW: My TV has 4 GB of memory, perhaps its time to get more memory if you need it. 顺便说一句：我的电视有4 GB的内存，如果需要的话，也许是时候获取更多的内存了。 ;) ;）