How should "BatchNorm" layer be used in caffe?

Question

I am a little confused about how should I use/insert "BatchNorm" layer in my models.
I see several different approaches, for instance:

ResNets: "BatchNorm"+"Scale" (no parameter sharing)

"BatchNorm" layer is followed immediately with "Scale" layer:

layer {
    bottom: "res2a_branch1"
    top: "res2a_branch1"
    name: "bn2a_branch1"
    type: "BatchNorm"
    batch_norm_param {
        use_global_stats: true
    }
}

layer {
    bottom: "res2a_branch1"
    top: "res2a_branch1"
    name: "scale2a_branch1"
    type: "Scale"
    scale_param {
        bias_term: true
    }
}

cifar10 example: only "BatchNorm"

In the cifar10 example provided with caffe, "BatchNorm" is used without any "Scale" following it:

layer {
  name: "bn1"
  type: "BatchNorm"
  bottom: "pool1"
  top: "bn1"
  param {
    lr_mult: 0
  }
  param {
    lr_mult: 0
  }
  param {
    lr_mult: 0
  }
}

cifar10 Different batch_norm_param for TRAIN and TEST

batch_norm_param: use_global_scale is changed between TRAIN and TEST phase:

layer {
  name: "bn1"
  type: "BatchNorm"
  bottom: "pool1"
  top: "bn1"
  batch_norm_param {
    use_global_stats: false
  }
  param {
    lr_mult: 0
  }
  param {
    lr_mult: 0
  }
  param {
    lr_mult: 0
  }
  include {
    phase: TRAIN
  }
}
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
  }
  include {
    phase: TEST
  }
}

So what should it be?

How should one use"BatchNorm" layer in caffe?

Kingkingbird

Thanks for your information. I looked at some current prototxt. They do not use decay_mult in BN, just use lr_mult:0. Am I right? – Patricio

@Patricio decay_mult and lr_mult are meaningless for "BatchNorm" layer as its parameters are updated based on the input statistics, rather than the backprop gradients. AFAIK, recent versions of caffe automatically sets lr_mult to zero for this layer. – Kingkingbird

You means the default value can check at github.com/BVLC/caffe/blob/…? Because I want to check my current caffe is set to zero or not – Patricio

Answer 1

If you follow the original paper, the Batch normalization should be followed by Scale and Bias layers (the bias can be included via the Scale, although this makes the Bias parameters inaccessible). use_global_stats should also be changed from training (False) to testing/deployment (True) - which is the default behavior. Note that the first example you give is a prototxt for deployment, so it is correct for it to be set to True.

I'm not sure about the shared parameters.

I made a pull request to improve the documents on the batch normalization, but then closed it because I wanted to modify it. And then, I never got back to it.

Note that I think lr_mult: 0 for "BatchNorm" is no longer required (perhaps not allowed?), although I'm not finding the corresponding PR now.

Monanthous

(1) Why oh why didn't you got back to documenting "BatchNorm"?? (2) PR #4704 was meant to simplify lr_mult params in "BatchNorm" definition. IMHO, this only created a mess. – Kingkingbird

Thanks for the encouragement to get back to it :-). On the surface, I liked not specifying lr_mult (which I found confusing), but as you point out, it did cause a mess. – Monanthous

Just found your caffe.help webpage - awesome!! thanks! – Kingkingbird

Answer 2

After each BatchNorm, we have to add a Scale layer in Caffe. The reason is that the Caffe BatchNorm layer only subtracts the mean from the input data and divides by their variance, while does not include the γ and β parameters that respectively scale and shift the normalized distribution 1. Conversely, the Keras BatchNormalization layer includes and applies all of the parameters mentioned above. Using a Scale layer with the parameter “bias_term” set to True in Caffe, provides a safe trick to reproduce the exact behavior of the Keras version. https://www.deepvisionconsulting.com/from-keras-to-caffe/

Kylix