CODEWITHSUNDEEP
caffevgg-nettransfer-learningfaster-rcnn
SomethingSomething
SomethingSomething

Reputation: 12266

Freezing weights in VGG19 network for transfer-learning in Caffe

The reason why I am asking this question is due to the fact that in VGG19 there are batch-normalization layers (unlike VGG16, for example).

I am trying to train a Faster-RCNN network in Caffe. I am doing it by:

I did not change anything regarding the lr_mult values of the convolutional layers. In the prototxt file, the convolutional layers (like conv1_1, etc. have non-zero lr_mult values, while the batch normalization layers' lr_mult values are set to 0 (layers named like conv1_1/bn).

Does the fact that the batch normalization layers are frozen means that the convolutional layers are frozen as well? Or should I set lr_mult to 0 also in the layers named convX_X?

Update: After running another training process while zeroing the lr_mult of all the convolutional layers, the training time reduced dramatically, which implies that the answer is that the lr_mult value needs to be set to 0 also in the convX_X layers.

Upvotes: 1

Views: 1120

Answers (1)

Dmytro Prylipko
Dmytro Prylipko

Reputation: 5084

To properly freeze convolutional layers with batchnorm in Caffe, you should:

  • For convolutional layers: Set lr_mult param to 0
  • For batchnorm layers: Set lr_mult params to 0, set use_global_stats to true:
layer {
  name: "bn1"
  type: "BatchNorm"
  bottom: "pool1"
  top: "bn1"
  batch_norm_param {
    use_global_stats: true
  }
  param {
    lr_mult: 0
  }
  param {
    lr_mult: 0
  }
  param {
    lr_mult: 0
  }
}

Does the fact that the batch normalization layers are frozen means that the convolutional layers are frozen as well?

Of course not. However, by using propagate_down param you can achieve this effect: How do I to prevent backward computation in specific layers in caffe.

Upvotes: 2

Related Questions