CODEWITHSUNDEEP
pythondeep-learningcaffepycaffe
DankMasterDan
DankMasterDan

Reputation: 2133

How to implement Eltwise layer in pyCaffe

I am trying to implement the Eltwise layer in PyCaffe to sum 2 inputs.

My goal is to get this in prototext:

layer {
  name: "eltwise_sum"
  type: "Eltwise"
  bottom: "v1"
  bottom: "v2"
  top: "v1_v2_sum"
  eltwise_param { operation: SUM }
}

I couldn't find any documentation or even google examples on how to do this in PyCaffe. 

net.v1_v2_sum = caffe.layers.Eltwise( net.v1, 
                                      net.v2, 
                                      name='eltwise_sum', 
                                      param ={'operation': 'SUM'}
                                    )

However, I keep getting an error. I suspect this is due to not specifying the sum operation correctly, but I cant find any documentation on how to do this right?

Thanks,

Upvotes: 0

Views: 1960

Answers (2)

DankMasterDan
DankMasterDan

Reputation: 2133

Ok, so apparently there is an integer corresponding to each operation Eltwise performs, and 1 corresponds to sum, so the correct implementation is:

net.v1_v2_sum = caffe.layers.Eltwise( net.v1, 
                                      net.v2, 
                                      name='eltwise_sum', 
                                      operation = 1
                                    )

I figured this out by trial-and-error. See Parag S. Chandakkar's answer for more info on how to understand this from their git code.

Upvotes: 0

Autonomous
Autonomous

Reputation: 9075

The answer lies in caffe.proto. Each parameter that you include in a layer is listed in that file. For Eltwise layer, the parameters are operation, coeff, stable_prod_grad. The parameter operation takes 3 values 0, 1 or 2 whose mappings are defined by an enum EltwiseOp. As far as the parameter name is concerned, you will find it in LayerParameter. I couldn't find any documentation on this. It's just something you learn along the way. I hope this clears your doubts.

Upvotes: 1

Related Questions