How to share the parameters of dropout layer in Siamese?

163 views
Skip to first unread message

Viveka Kulharia

unread,
Mar 6, 2016, 4:55:17 PM3/6/16
to Caffe Users
Hi, I am using the following to share dropout parameters in Siamese. But it gives the error: "Too many params specified for layer drop3_p ...". Is it the case that dropout layer parameters need not be changed? Thanks.

layer {
  name: "drop3_p"
  type: "Dropout"
  bottom: "fc3-conv_p"
  top: "fc3-conv_p"
  param { 
    name: "share_drop3"
  }
  dropout_param {
    dropout_ratio: 0.5
  }
}

Jan

unread,
Mar 7, 2016, 2:57:43 AM3/7/16
to Caffe Users
To the best of my knowledge, a dropout layer does not have any trainable parameters, so there is nothing to be shared. The only parameter of a dropout layer is the dropout_ratio, but that is fixed and not altered by training.

Jan
Reply all
Reply to author
Forward
0 new messages