Cannot explicitly "name" BatchNorm parameters for sharing (Siamese network)

[shaibagon]

TL;DR

It seems like commit a8ec123c00723df0d0ad897e1eea32a29201c81b broke something here: one can no longer pass param to "BatchNorm" layer (they are ignored).

I suppose auto-upgrade should not remove param completely from "BatchNorm" layers, but rather make sure lr_mult and decay_mult are zero.

Issue summary

I am trying to share BatchNorm internal parameters in a Siamese network: That is, having two BatchNorm layers sharing the same internal parameters.

I explicitly name the parameters using: param {'lr_mult':0 'name': 'bn_a'}

However, when building the net, the whole param section is missing, and caffe completely ignores the name of the parameters. No sharing is actually happening.

Steps to reproduce

Here is the relevant part of the prototxt:

layer {
  name: "bn_a"
  type: "BatchNorm"
  bottom: "conv1_a"
  top: "bn_a"
  param {
    name: "bn_m"
    lr_mult: 0
  }
  param {
    name: "bn_v"
    lr_mult: 0
  }
  param {
    name: "bn_b"
    lr_mult: 0
  }
  batch_norm_param {
    use_global_stats: false
  }
}
layer {
  name: "bn_b"
  type: "BatchNorm"
  bottom: "conv1_b"
  top: "bn_b"
  param {
    name: "bn_m"
    lr_mult: 0
  }
  param {
    name: "bn_v"
    lr_mult: 0
  }
  param {
    name: "bn_b"
    lr_mult: 0
  }
  batch_norm_param {
    use_global_stats: false
  }
}

The relevant part of the log:

layer {
  name: "bn_a"
  type: "BatchNorm"
  bottom: "conv1_a"
  top: "bn_a"
  batch_norm_param {
    use_global_stats: false
  }
}
layer {
  name: "bn_b"
  type: "BatchNorm"
  bottom: "conv1_b"
  top: "bn_b"
  batch_norm_param {
    use_global_stats: false
  }
}

Consequently, when building the layers the log reports:

I0111 22:28:11.185611 27369 layer_factory.hpp:77] Creating layer bn_b
I0111 22:28:11.185622 27369 net.cpp:100] Creating Layer bn_b
I0111 22:28:11.185631 27369 net.cpp:434] bn_b <- conv1_b
I0111 22:28:11.185642 27369 net.cpp:408] bn_b -> bn_b
I0111 22:28:11.185780 27369 net.cpp:150] Setting up bn_b
I0111 22:28:11.185803 27369 net.cpp:157] Top shape: 10 10 222 222 (4928400)
I0111 22:28:11.185813 27369 net.cpp:165] Memory required for data: 90896640

There is no
Sharing parameters 'bn_m' owned by layer 'bn1_a', param index 0
log entries as I would have expected.

It seems like BatchNorm ignores param entries completely!

Am I missing something?
How can I share the internal parameters of BatchNorm?

Your system configuration

Operating system: ubuntu 14.04
Compiler: gcc
CUDA version (if applicable):
CUDNN version (if applicable):
BLAS: mkl
Python or MATLAB version (for pycaffe and matcaffe respectively): python 2.7
Caffe: my caffe is updated with commit 365ac88

[D-X-Y]

I have the same question as #5120 ..

[shaibagon]

@D-X-Y I made an attempt at fixing this issue: PR#5184. Hope they'll approve.

[shelhamer]

Thanks for raising this issue. I was too focused on protecting the batch statistics from accidental gradients that I made sharing impossible across layers that could sensibly share statistics as in siamese nets. I'll double-check #5184 and come up with an auto-upgrade rule that handles both lr and name.

[shelhamer]

Fixed by #5184