The following is an example definition for training a BatchNorm layer with channel-wise scale and bias. Typically a BatchNorm layer is inserted between convolution and rectification layers. In this example, the convolution would output the blob layerx and the rectification would receive the layerx-bn blob.

layer { bottom: 'layerx' top: 'layerx-bn' name: 'layerx-bn' type: 'BatchNorm'
  batch_norm_param {
    use_global_stats: false  # calculate the mean and variance for each mini-batch
    moving_average_fraction: .999  # doesn't effect training 
  }
  param { lr_mult: 0 } 
  param { lr_mult: 0 } 
  param { lr_mult: 0 }}
# channel-wise scale and bias are separate
layer { bottom: 'layerx-bn' top: 'layerx-bn' name: 'layerx-bn-scale' type: 'Scale',
  scale_param { 
    bias_term: true
    axis: 1      # scale separately for each channel
    num_axes: 1  # ... but not spatially (default)
    filler { type: 'constant' value: 1 }           # initialize scaling to 1
    bias_filler { type: 'constant' value: 0.001 }  # initialize bias
}}

More information can be found in this thread.

The main change needed is to switch use_global_stats to true. This switches to using the moving average.

layer { bottom: 'layerx' top: 'layerx-bn' name: 'layerx-bn' type: 'BatchNorm'
  batch_norm_param {
    use_global_stats: true  # use pre-calculated average and variance
  }
  param { lr_mult: 0 } 
  param { lr_mult: 0 } 
  param { lr_mult: 0 }}
# channel-wise scale and bias are separate
layer { bottom: 'layerx-bn' top: 'layerx-bn' name: 'layerx-bn-scale' type: 'Scale',
  scale_param { 
    bias_term: true
    axis: 1      # scale separately for each channel
    num_axes: 1  # ... but not spatially (default)
}}