Question about Net surgery

[Eunjeong Ko]

Hello!

I have a question about net surgery.  

Our team wants to ....

1) load pre-trained weights from Caffe reference model, 

2) bring only weights of fully-connected layers among full CNN networks.

3) learn the modified networks (consists of only fully-connected layers) using other features (not image) or predict using other features.

(we want to replace the features defined by human. Its dimension can be smaller dimension than the input dimension of the fully-connected layer.)  

Our question is ...

1) Is it possible using Caffe framework?

2) We want to give features as input. But the dimension of our feature is not matched with input dimension of the fully-connected layer. In this case, how can we make them? 

3) If you have some similar experiences above tasks, could you give comments for our team? 

Thanks a lot in any advance and comments.

Thanks, 

Eunjeong Ko