--- title: "BNNSFilterCreateLayerPooling(_:_:)" framework: accelerate role: symbol role_heading: Function path: "accelerate/bnnsfiltercreatelayerpooling(_:_:)" --- # BNNSFilterCreateLayerPooling(_:_:) Returns a new pooling layer. ## Declaration ```swift func BNNSFilterCreateLayerPooling(_ layer_params: UnsafePointer, _ filter_params: UnsafePointer?) -> BNNSFilter? ``` ## Parameters - `layer_params`: Layer parameters. - `filter_params`: The filter runtime parameters. ## Discussion Discussion Use a pooling layer to downsample an input, selecting, for example, the average or the maximum value in a specified kernel size. The following figure illustrates how a 2 x 2 maximum pooling kernel samples each 2 x 2 block of values of 4 x 4 source. The highlighted block contains the values [3, 4, 0, 0], so the maximum value passed to the destination element in output is 4. The following code shows how you apply pooling to the input in the above figure. Note that the input is a 2 x 2 x 1 BNNSDataLayoutImageCHW tensor. Definie x_padding and y_padding as 1 to add zero padding: let input: [Float] = [1, 2, 3, 4] var output = [Float](repeating: 0, count: 9) let inDescriptor = BNNSNDArrayDescriptor(flags: BNNSNDArrayFlags(0), layout: BNNSDataLayoutImageCHW, size: (2, 2, 1, 0, 0, 0, 0, 0), stride: (0, 0, 0, 0, 0, 0, 0, 0), data: nil, data_type: .float, table_data: nil, table_data_type: .float, data_scale: 0, data_bias: 0) let outDescriptor = BNNSNDArrayDescriptor(flags: BNNSNDArrayFlags(0), layout: BNNSDataLayoutImageCHW, size: (3, 3, 1, 0, 0, 0, 0, 0), stride: (0, 0, 0, 0, 0, 0, 0, 0), data: nil, data_type: .float, table_data: nil, table_data_type: .float, data_scale: 0, data_bias: 0) var parameters = BNNSLayerParametersPooling(i_desc: inDescriptor, o_desc: outDescriptor, bias: BNNSNDArrayDescriptor(), activation: .identity, pooling_function: .max , k_width: 2, k_height: 2, x_stride: 1, y_stride: 1, x_dilation_stride: 0, y_dilation_stride: 0, x_padding: 1, y_padding: 1, pad: (0, 0, 0, 0)) let filter = BNNSFilterCreateLayerPooling(¶meters, nil) defer { BNNSFilterDestroy(filter) } BNNSPoolingFilterApplyBatch(filter, 1, input, input.count, &output, output.count, nil, 0) On return, output contains the following values: [ 1.0, 2.0, 2.0, 3.0, 4.0, 4.0, 3.0, 4.0, 4.0 ] UnMax Pooling Use BNNSPoolingFunctionUnMax in conjunction with the indices generated by BNNSPoolingFunctionMax to partially recreate the original data of a maximum pooling operation. UnMax pooling is a partial inverse of maximum pooling that sets all non-maximal values to zero. For example, given the following input data: let input: [Float] = [1, 1, 1, 9, 1, 9, 9, 1, 1, 1, 1, 1, 1, 9, 1, 9] Use BNNSPoolingFilterApplyBatch(_:_:_:_:_:_:_:_:) to perform the maximum pooling and populate an indices array with positional information of maximum elements in each window: var output = [Float](repeating: 0, count: 3 * 3) let inDescriptor = BNNSNDArrayDescriptor(flags: BNNSNDArrayFlags(0), layout: BNNSDataLayoutImageCHW, size: (4, 4, 1, 0, 0, 0, 0, 0), stride: (0, 0, 0, 0, 0, 0, 0, 0), data: nil, data_type: .float, table_data: nil, table_data_type: .float, data_scale: 0, data_bias: 0) let outDescriptor = BNNSNDArrayDescriptor(flags: BNNSNDArrayFlags(0), layout: BNNSDataLayoutImageCHW, size: (3, 3, 1, 0, 0, 0, 0, 0), stride: (0, 0, 0, 0, 0, 0, 0, 0), data: nil, data_type: .float, table_data: nil, table_data_type: .float, data_scale: 0, data_bias: 0) var parameters = BNNSLayerParametersPooling(i_desc: inDescriptor, o_desc: outDescriptor, bias: BNNSNDArrayDescriptor(), activation: .identity, pooling_function: .max , k_width: 2, k_height: 2, x_stride: 1, y_stride: 1, x_dilation_stride: 0, y_dilation_stride: 0, x_padding: 0, y_padding: 0, pad: (0, 0, 0, 0)) let filter = BNNSFilterCreateLayerPooling(¶meters, nil) defer { BNNSFilterDestroy(filter) } var indices = [Int](repeating: 0, count: input.count) BNNSPoolingFilterApplyBatch(filter, 1, input, input.count, &output, output.count, &indices, indices.count) To perform the UnMax pooling, reuse the parameters structure, but swap the input and output descriptors: var recreatedInput = [Float](repeating: 0, count: input.count) parameters.i_desc = outDescriptor parameters.o_desc = inDescriptor parameters.pooling_function = BNNSPoolingFunctionUnMax let unMaxFilter = BNNSFilterCreateLayerPooling(¶meters, nil) defer { BNNSFilterDestroy(unMaxFilter) } BNNSPoolingFilterApplyBatch(unMaxFilter, 1, output, output.count, &recreatedInput, recreatedInput.count, &indices, indices.count) On return, recreatedInput is similar to input, but the non-maximal values are zero: [ 0.0, 0.0, 0.0, 9.0, 0.0, 9.0, 9.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 9.0, 0.0, 9.0 ] ## See Also ### Pooling layers - [BNNSPoolingLayerParameters](accelerate/bnnspoolinglayerparameters.md) - [BNNSFilterCreatePoolingLayer(_:_:_:_:)](accelerate/bnnsfiltercreatepoolinglayer(_:_:_:_:).md) - [BNNS.PoolingLayer](accelerate/bnns/poolinglayer.md) - [BNNSPoolingFunction](accelerate/bnnspoolingfunction.md) - [BNNSPoolingFunctionAverage](accelerate/bnnspoolingfunctionaverage.md) - [BNNSPoolingFunctionMax](accelerate/bnnspoolingfunctionmax.md) - [BNNSLayerParametersPooling](accelerate/bnnslayerparameterspooling.md) - [BNNSPoolingFilterApplyBatch(_:_:_:_:_:_:_:_:)](accelerate/bnnspoolingfilterapplybatch(_:_:_:_:_:_:_:_:).md) - [BNNSPoolingFilterApplyBackwardBatch(_:_:_:_:_:_:_:_:_:_:_:_:_:)](accelerate/bnnspoolingfilterapplybackwardbatch(_:_:_:_:_:_:_:_:_:_:_:_:_:).md) - [BNNSPoolingFilterApplyBatchEx(_:_:_:_:_:_:_:_:_:)](accelerate/bnnspoolingfilterapplybatchex(_:_:_:_:_:_:_:_:_:).md) - [BNNSPoolingFilterApplyBackwardBatchEx(_:_:_:_:_:_:_:_:_:_:_:_:_:_:)](accelerate/bnnspoolingfilterapplybackwardbatchex(_:_:_:_:_:_:_:_:_:_:_:_:_:_:).md)