How to catch middle-layer-output in cuda-convnet?

[Henri Shi]

Hi, all.

As my network (implemented in cuda-convnet) with an architecture ended by a cost.logreg layer, the default output is the softmax-classification error rate.

Now, I hope to get the output vector of one of the middle layers, in order to perform some specific task, such as face verification. 

So, how to "intercept" the output corresponding to a specific layer?