Double layer params in train_val.prototxt (bvlc_reference_caffenet)

[Daniel Hamngren]

Hi,

I'm trying to get my head around how the prototxt files are structured. I noticed in the train_val.prototxt-files for bvlc_reference_caffenet there are for each layer two "param"-brackets. 

https://github.com/BVLC/caffe/blob/master/models/bvlc_reference_caffenet/train_val.prototxt#L61-L68

The learning rate multiplier and decay multiplier seems to be set twice. Why is that? There is surely something I'm missing.

Thank you.

/Daniel

[Abhishek Das]

Hi Daniel,

One is for the filters and the other for biases.

Cheers

[Daniel Hamngren]

Thank you!