Digits3 GoogLeNet Batch Normalization?

[John Edwards]

I'm trying to get a custom GoogLeNet network put together under Digits3 that does batch normalization.  I found an example caffe model and tweaked it up to use the same input layer as the default GoogLeNet in Digits (see attached), but I'm getting the following error when I hit the visualize button:

ParseError: 70:3 : Message type "caffe.LayerParameter" has no field named "bn_param".

Based on other posts that I've found, it looks like BN is supported in cuda 7.5/cudNN 4.5.

Can anyone help me out on this or point me to a solution?

[Greg Heinrich]

Hi John,
I think the batch normalization parameter in Caffe's prototxt format is now named "batch_norm_param". This was changed from "bn_param" with this commit:
https://github.com/BVLC/caffe/commit/a52ee656a589313901560c87b65a570ee41c9fee#diff-e0bd8d028edf0f88febfebe3bf703904R353

The current definition of the prototxt format used in nv-caffe 0.14 may be found there:
https://github.com/NVIDIA/caffe/blob/caffe-0.14/src/caffe/proto/caffe.proto

Regards,
Greg.

[John Edwards]

Hi Greg,

Thanks for the reply.  After reviewing the latest prototxt format, I was able to update my network and train successfully.

[Lixin Su]

I am glad to find that bn_param was replaced with batch_norm_param.  What about the "frozen" field inside bn_param?  I have an old prototxt file that has the following code.

layer {
  name: "caffe.BN_896"
  type: "BN"
  bottom: "caffe.Eltwise_895"
  top: "caffe.BN_896"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  batch_norm_param { frozen: true
    slope_filler {
      value: 1
    }
    bias_filler {
      value: 0