How to add constraints L2-norm to the conv layers and L1-norm to the fc layers when training a model

[汤旭]

How to add constraints L2-norm to the conv layers and L1-norm to the fc layers when training a model using caffe?

I know that we can use regularization_type: 'L1'  (in solver.prototxt) to give global constraints to all the layers. And also we can adjust the value of weight_decay and decay_mult to rectify the performance.

  optional float weight_decay = 12; // The weight decay.

  // regularization types supported: L1 and L2

  // controlled by weight_decay

  optional string regularization_type = 29 [default = "L2"];

 // The multiplier on the global weight decay for this parameter.

  optional float decay_mult = 4 [default = 1.0];

But how can i add constraints L2-norm to the conv layers and L1-norm to the fc layers when training a model using caffe?

[汤旭]

Anybody know how to do ?

在 2015年10月3日星期六 UTC+8下午4:51:08，汤旭写道：