/// <summary>
/// Setup the layer for both Engine.CUDNN and Engine.CAFFE modes.
/// </summary>
/// <param name="colBottom">Specifies the collection of bottom (input) Blobs.</param>
/// <param name="colTop">Specifies the collection of top (output) Blobs.</param>
public override void LayerSetUp(BlobCollection <T> colBottom, BlobCollection <T> colTop)
{
m_nSize = (int)m_param.lrn_param.local_size;
m_log.CHECK_EQ(m_nSize % 2, 1, "LRN only supports odd values for local_size.");
m_nPrePad = (m_nSize - 1) / 2;
m_dfAlpha = m_param.lrn_param.alpha;
m_dfBeta = m_param.lrn_param.beta;
m_dfK = m_param.lrn_param.k;
if (m_param.lrn_param.norm_region == LRNParameter.NormRegion.WITHIN_CHANNEL)
{
// Set up split_layer to use in the numerator and denominator.
m_colSplitTopVec = new BlobCollection <T>();
m_colSplitTopVec.Add(m_blobProductInput);
m_colSplitTopVec.Add(m_blobSquareInput);
LayerParameter split_param = new LayerParameter(LayerParameter.LayerType.SPLIT, "split");
m_splitLayer = new SplitLayer <T>(m_cuda, m_log, split_param);
m_splitLayer.Setup(colBottom, m_colSplitTopVec);
// Set up square_layer to square teh inputs.
m_colSquareBottomVec = new BlobCollection <T>();
m_colSquareTopVec = new BlobCollection <T>();
m_colSquareBottomVec.Add(m_blobSquareInput);
m_colSquareTopVec.Add(m_blobSquareOutput);
LayerParameter square_param = new LayerParameter(LayerParameter.LayerType.POWER, "square");
square_param.power_param.power = 2.0;
m_squareLayer = new PowerLayer <T>(m_cuda, m_log, square_param);
m_squareLayer.Setup(m_colSquareBottomVec, m_colSquareTopVec);
// Set up pool_layer to sum over square neighborhoods of the input.
m_colPoolTopVec = new BlobCollection <T>();
m_colPoolTopVec.Add(m_blobPoolOutput);
LayerParameter pool_param = new LayerParameter(LayerParameter.LayerType.POOLING, "pool");
pool_param.pooling_param.pool = PoolingParameter.PoolingMethod.AVE;
pool_param.pooling_param.pad.Add((uint)m_nPrePad);
pool_param.pooling_param.kernel_size.Add((uint)m_nSize);
m_poolLayer = new PoolingLayer <T>(m_cuda, m_log, pool_param);
m_poolLayer.Setup(m_colSquareTopVec, m_colPoolTopVec);
// Set up power_layer to compute (1 + alpha/N^2 s)^-beta, where s is
// the sum of the squared neighborhood (the output of pool_layer)
m_colPowerTopVec = new BlobCollection <T>();
m_colPowerTopVec.Add(m_blobPowerOutput);
LayerParameter power_param = new LayerParameter(LayerParameter.LayerType.POWER, "power");
power_param.power_param.power = -m_dfBeta;
power_param.power_param.scale = m_dfAlpha;
power_param.power_param.shift = 1.0;
m_powerLayer = new PowerLayer <T>(m_cuda, m_log, power_param);
m_powerLayer.Setup(m_colPoolTopVec, m_colPowerTopVec);
// Set up a product_layer to compute outputs by multiplying inputs by the
// inverse denominator computed by the power layer.
m_colProductBottomVec = new BlobCollection <T>();
m_colProductBottomVec.Add(m_blobProductInput);
m_colProductBottomVec.Add(m_blobPowerOutput);
LayerParameter product_param = new LayerParameter(LayerParameter.LayerType.ELTWISE, "product");
product_param.eltwise_param.operation = EltwiseParameter.EltwiseOp.PROD;
m_productLayer = new EltwiseLayer <T>(m_cuda, m_log, product_param);
m_productLayer.Setup(m_colProductBottomVec, colTop);
}
if (!m_param.lrn_param.useCudnn())
{
return;
}
m_hCuDnn = m_cuda.CreateCuDNN();
m_hNormDesc = m_cuda.CreateLRNDesc();
m_hBottomDesc = m_cuda.CreateTensorDesc();
m_hTopDesc = m_cuda.CreateTensorDesc();
}
/** @copydoc Layer::dispose */
protected override void dispose()
{
m_blobScale.Dispose();
m_blobSquareInput.Dispose();
m_blobSquareOutput.Dispose();
m_blobPoolOutput.Dispose();
m_blobPowerOutput.Dispose();
m_blobProductInput.Dispose();
if (m_splitLayer != null)
{
m_splitLayer.Dispose();
m_splitLayer = null;
}
if (m_squareLayer != null)
{
m_squareLayer.Dispose();
m_squareLayer = null;
}
if (m_poolLayer != null)
{
m_poolLayer.Dispose();
m_poolLayer = null;
}
if (m_powerLayer != null)
{
m_powerLayer.Dispose();
m_powerLayer = null;
}
if (m_powerLayer != null)
{
m_productLayer.Dispose();
m_powerLayer = null;
}
if (m_hNormDesc != 0)
{
m_cuda.FreeLRNDesc(m_hNormDesc);
m_hNormDesc = 0;
}
if (m_hBottomDesc != 0)
{
m_cuda.FreeTensorDesc(m_hBottomDesc);
m_hBottomDesc = 0;
}
if (m_hTopDesc != 0)
{
m_cuda.FreeTensorDesc(m_hTopDesc);
m_hTopDesc = 0;
}
if (m_hCuDnn != 0)
{
m_cuda.FreeCuDNN(m_hCuDnn);
m_hCuDnn = 0;
}
if (m_hTempData1 != 0)
{
m_cuda.FreeMemory(m_hTempData1);
m_hTempData1 = 0;
}
if (m_hTempData2 != 0)
{
m_cuda.FreeMemory(m_hTempData2);
m_hTempData2 = 0;
}
base.dispose();
}