Contents

BNNSFilterCreateLayerNormalization(_:_:_:)

Returns a new normalization layer.

Declaration

func BNNSFilterCreateLayerNormalization(_ normType: BNNSFilterType, _ layer_params: UnsafePointer<BNNSLayerParametersNormalization>, _ filter_params: UnsafePointer<BNNSFilterParameters>?) -> BNNSFilter?

Parameters

  • normType:

    The normalization type.

  • layer_params:

    Layer parameters.

  • filter_params:

    The filter runtime parameters.

Discussion

Use a normalization loss layer to normalize tensor values, that is, calculate new elements to have a zero mean and a unit standard deviation.

For example, given a three-channel image that contains the following values:

let input: [Float] = [1, 2, 3, 4, 5, 6, 7, 8, 9,
                      10, 20, 30, 40, 50, 60, 70, 80, 90,
                      100, 200, 300, 400, 500, 600, 700, 800, 900]

The following code normalizes each channel and writes the result to the output array:

let count = input.count
var output = [Float](repeating: 0,
                     count: count)

let descriptor = BNNSNDArrayDescriptor(flags: BNNSNDArrayFlags(0),
                                       layout: BNNSDataLayoutImageCHW,
                                       size: (9, 1, 3, 0, 0, 0, 0, 0),
                                       stride: (0, 0, 0, 0, 0, 0, 0, 0),
                                       data: nil,
                                       data_type: BNNSDataType.float,
                                       table_data: nil,
                                       table_data_type: BNNSDataType.float,
                                       data_scale: 1,
                                       data_bias: 0)

var gamma = [Float](repeating: 1,
                    count: descriptor.size.2)

gamma.withUnsafeMutableBufferPointer { gammaPtr in
    
    let gammaDescriptor = BNNSNDArrayDescriptor(flags: BNNSNDArrayFlags(0),
                                                layout: BNNSDataLayoutVector,
                                                size: (descriptor.size.2, 0, 0, 0, 0, 0, 0, 0),
                                                stride: (0, 0, 0, 0, 0, 0, 0, 0),
                                                data: gammaPtr.baseAddress,
                                                data_type: BNNSDataType.float,
                                                table_data: nil,
                                                table_data_type: BNNSDataType.float,
                                                data_scale: 1,
                                                data_bias: 0)
    
    var layerParameters = BNNSLayerParametersNormalization(i_desc: descriptor,
                                                           o_desc: descriptor,
                                                           beta_desc: BNNSNDArrayDescriptor(),
                                                           gamma_desc: gammaDescriptor,
                                                           moving_mean_desc: BNNSNDArrayDescriptor(),
                                                           moving_variance_desc: BNNSNDArrayDescriptor(),
                                                           momentum: 1,
                                                           epsilon: 0,
                                                           activation: .identity,
                                                           num_groups: 1,
                                                           normalization_axis: 0)
    
    let filter = BNNSFilterCreateLayerNormalization(BNNSInstanceNorm, &layerParameters, nil)
    defer {
        BNNSFilterDestroy(filter)
    }
    
    BNNSNormalizationFilterApplyBatch(filter, 1,
                                      input, count,
                                      &output, count,
                                      false)
}

On return, the output contains the following values:

[-1.549, -1.162, -0.775, -0.387, 0.0, 0.387, 0.775, 1.162, 1.549,
 -1.549, -1.162, -0.775, -0.387, 0.0, 0.387, 0.775, 1.162, 1.549,
 -1.549, -1.162, -0.775, -0.387, 0.0, 0.387, 0.775, 1.162, 1.549]

See Also

Normalization layers