Handling multiple inputs

[Arthur B.]

Hello all,

Thanks for this fantastic project!

I am trying to train a neural network on data having the following structure:
- a one dimensional signal of length 2048 (so width:2048, height:1, channels:1)
- a side information vector of length 32

I would like too separately place a few convolutional and max-pooling layers on top of the one dimensional signal,
and a few fully connected sigmoid layers on top of my side information vector

Then, I would like to concatenate the two outputs of these two networks, and build a few more fully
connected layers on top of that to make a prediction.

I'm not sure how to do this in the protobuf format specified by caffe....

So I have four questions,

1) I see in some examples that a DATA layer can have multiple "tops"... how does that work? Is it referring to some names in the input database?

2) My data is currently in a simple csv file. What is the easiest way for me to make it accessible to caffe? I'm happy to put it in hd5 format, but what structure should
my hd5 file have?

3) if I specify multiple "bottoms" for an INNER_PRODUCT layer, will the network simply concatenate them? 

4) is it possible, during training, and using only the command line interface, to regularly output a prediction on a validation test. Not actually validate it,
just output what the prediction would be?

Thanks in advance for your help,
Arthur