BatchNorm with CUDNN doesn't take effect, so there are no scale factors in BatchNorm layers?

[WenzhMicrosoft]

I found that CuDNNBatchNormLayer in your caffe branch ( https://github.com/Tongcheng/caffe ) is not registered as a Creator in layer_factory.cpp, But your BatchNorm Registered with REGISTER_LAYER_CLASS(BatchNorm);

An registered CuDNNBatchNormLayer seems like this:
BVLC/caffe@c9eda39#diff-6fe0622356ab61c001bcac36dd571e7d

So I guess the following setting wouldn't take affect, and there is no scale factors, and we need to add scale layers after each "BatchNorm" layer

    scale_filler {
      type: "constant"
      value: 1
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    engine: CUDNN

[Tongcheng]

Hello @WenzhMicrosoft , I am not sure what is the question we are having, but my understanding is that during layer initialization, it only reads the configuration from NeuralNetwork's proto based on the caffe.proto. Therefore I think layer_factory should not matter.

[WenzhMicrosoft]

Hi @Tongcheng, You specified "engine: CUDNN" in prototxt, I think the reason you specified this is because you want to use a CuDNNBatchNormLayer(Do you?), and it makes more sense if you use CuDNNBatchNormLayer because CuDNNBatchNormLayer actually is combination of BatchNormLayer and ScaleLayer, normally Caffe version of Neural Network need this ScaleLayer layer to implement the γ and β factor metioned in Batch normalization paper

This is an example: https://github.com/KaimingHe/deep-residual-networks/blob/master/prototxt/ResNet-50-deploy.prototxt

As what I said before. Caffe actually creates a BatchNormLayer without γ and β. Is this what you want?