/// <summary>
/// Reshape the bottom (input) and top (output) blobs.
/// </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 Reshape(BlobCollection <T> colBottom, BlobCollection <T> colTop)
{
m_log.CHECK_EQ(4, colBottom[0].num_axes, "Input must have 4 axes, corresponding to (num, channels, height, width)");
m_nNum = colBottom[0].num;
m_nChannels = colBottom[0].channels;
m_nHeight = colBottom[0].height;
m_nWidth = colBottom[0].width;
switch (m_param.lrn_param.norm_region)
{
case LRNParameter.NormRegion.ACROSS_CHANNELS:
colTop[0].Reshape(m_nNum, m_nChannels, m_nHeight, m_nWidth);
if (!m_param.lrn_param.useCudnn())
{
m_blobScale.Reshape(m_nNum, m_nChannels, m_nHeight, m_nWidth);
}
break;
case LRNParameter.NormRegion.WITHIN_CHANNEL:
m_splitLayer.Reshape(colBottom, m_colSplitTopVec);
m_squareLayer.Reshape(m_colSquareBottomVec, m_colSquareTopVec);
m_poolLayer.Reshape(m_colSquareTopVec, m_colPoolTopVec);
m_powerLayer.Reshape(m_colPoolTopVec, m_colPowerTopVec);
m_productLayer.Reshape(m_colProductBottomVec, colTop);
break;
}
if (!m_param.lrn_param.useCudnn())
{
return;
}
m_cuda.SetTensorDesc(m_hBottomDesc, m_nNum, m_nChannels, m_nHeight, m_nWidth);
m_cuda.SetTensorDesc(m_hTopDesc, m_nNum, m_nChannels, m_nHeight, m_nWidth);
m_cuda.SetLRNDesc(m_hNormDesc, (uint)m_nSize, m_dfAlpha, m_dfBeta, m_dfK);
if (m_param.lrn_param.norm_region == LRNParameter.NormRegion.WITHIN_CHANNEL)
{
int nTotalSize = m_nNum * m_nChannels * m_nHeight * m_nWidth;
if (nTotalSize > m_nTempDataSize)
{
if (m_hTempData1 != 0)
{
m_cuda.FreeMemory(m_hTempData1);
m_hTempData1 = 0;
}
if (m_hTempData2 != 0)
{
m_cuda.FreeMemory(m_hTempData2);
m_hTempData2 = 0;
}
m_hTempData1 = m_cuda.AllocMemory(nTotalSize);
m_hTempData2 = m_cuda.AllocMemory(nTotalSize);
m_nTempDataSize = nTotalSize;
}
}
}
/// <summary>
/// Reshape the bottom (input) and top (output) blobs.
/// </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 Reshape(BlobCollection <T> colBottom, BlobCollection <T> colTop)
{
m_log.CHECK_EQ(4, colBottom[0].num_axes, "Input must have 4 axes, corresponding to (num, channels, height, width)");
// Do nothing if bottom shape is unchanged since last Reshape.
if (m_nNum == colBottom[0].num &&
m_nChannels == colBottom[0].channels &&
m_nBottomH == colBottom[0].height &&
m_nBottomW == colBottom[0].width &&
m_bReshapedFirstTime)
{
return;
}
m_nNum = colBottom[0].num;
m_nChannels = colBottom[0].channels;
m_nBottomH = colBottom[0].height;
m_nBottomW = colBottom[0].width;
m_bReshapedFirstTime = true;
if (m_nPyramidHeight == 1)
{
LayerParameter pooling_param = getPoolingParam(0, m_nBottomH, m_nBottomW, m_param.spp_param);
if (m_rgPoolingLayers[0] != null)
{
m_rgPoolingLayers[0].Dispose();
}
m_rgPoolingLayers[0] = new PoolingLayer <T>(m_cuda, m_log, pooling_param);
m_rgPoolingLayers[0].Setup(colBottom, colTop);
m_rgPoolingLayers[0].Reshape(colBottom, colTop);
return;
}
m_split_layer.Reshape(colBottom, m_colBlobSplitTopVec);
for (int i = 0; i < m_nPyramidHeight; i++)
{
LayerParameter pooling_param = getPoolingParam(i, m_nBottomH, m_nBottomW, m_param.spp_param);
if (m_rgPoolingLayers[i] != null)
{
m_rgPoolingLayers[i].Dispose();
}
m_rgPoolingLayers[i] = new PoolingLayer <T>(m_cuda, m_log, pooling_param);
m_rgPoolingLayers[i].Setup(m_rgPoolingBottomVec[i], m_rgPoolingTopVecs[i]);
m_rgPoolingLayers[i].Reshape(m_rgPoolingBottomVec[i], m_rgPoolingTopVecs[i]);
m_rgFlattenLayers[i].Reshape(m_rgPoolingTopVecs[i], m_rgFlattenLayerTopVecs[i]);
}
m_concat_layer.Reshape(m_colBlobConcatBottomVec, colTop);
}
