multiple regularizor on parameters of a single layer

[Yifan Wang]

I'd like to add an InnerProduct layer to the top of my net.

Y = XW

However, I need the following regularizor for this weights of this layer.

L2 norm/euclidean loss to a given target

||W-W*||_2

and L1 norm

|W|

I'm aware solver uses a global

weight_decay

to regularize the weights in network. There are also separate loss layers like euclidean_loss or smoothl1loss that could satisfy my need.

However, as far as I know, euclidean_loss can only take datablob as a input (?), how can I use the same concept for my weights?

On the other hand, how can I specify the regularization type for a single layer, or set use multiple regularization types?