GRNLayer.cs

using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using MyCaffe.basecode;
using MyCaffe.common;
using MyCaffe.param;
 
namespace MyCaffe.layers.beta
{
    public class GRNLayer<T> : Layer<T>
    {
        Blob<T> m_blobSumMultiplier;
        Blob<T> m_blobSquare;
        Blob<T> m_blobNorm;
        Blob<T> m_blobTempDot;
 
        public GRNLayer(CudaDnn<T> cuda, Log log, LayerParameter p)
            : base(cuda, log, p)
        {
            m_type = LayerParameter.LayerType.GRN;
            m_blobSumMultiplier = new common.Blob<T>(cuda, log, false);
            m_blobSumMultiplier.Name = m_param.name + " sum mult";
            m_blobSquare = new common.Blob<T>(cuda, log, false);
            m_blobSquare.Name = m_param.name + " sqr";
            m_blobNorm = new common.Blob<T>(cuda, log, false);
            m_blobNorm.Name = m_param.name + " norm";
            m_blobTempDot = new common.Blob<T>(cuda, log, false);
            m_blobTempDot.Name = m_param.name + " temp dot";
        }
 
        protected override void dispose()
        {
            base.dispose();
            dispose(ref m_blobSumMultiplier);
            dispose(ref m_blobSquare);
            dispose(ref m_blobNorm);
            dispose(ref m_blobTempDot);
        }
 
        protected override void setup_internal_blobs(BlobCollection<T> col)
        {
            if (col.Count > 0)
                return;
 
            col.Add(m_blobSquare);
            col.Add(m_blobNorm);
            col.Add(m_blobTempDot);
            col.Add(m_blobSumMultiplier);
        }
 
        public override int ExactNumBottomBlobs
        {
            get { return 1; }
        }
 
        public override int ExactNumTopBlobs
        {
            get { return 1; }
        }
 
        public override void LayerSetUp(BlobCollection<T> colBottom, BlobCollection<T> colTop)
        {
        }
 
        public override void Reshape(BlobCollection<T> colBottom, BlobCollection<T> colTop)
        {
            colTop[0].ReshapeLike(colBottom[0]);
 
            m_blobSumMultiplier.Reshape(1, colBottom[0].channels, 1, 1);
            m_blobSumMultiplier.SetData(1.0);
            m_blobSquare.ReshapeLike(colBottom[0]);
            m_blobNorm.Reshape(colBottom[0].num, 1, colBottom[0].height, colBottom[0].width);
            m_blobTempDot.Reshape(colBottom[0].num, 1, colBottom[0].height, colBottom[0].width);
        }
 
        protected override void forward(BlobCollection<T> colBottom, BlobCollection<T> colTop)
        {
            long hBottomData = colBottom[0].gpu_data;
            long hTopData = colTop[0].mutable_gpu_data;
            long hSquareData = m_blobSquare.mutable_gpu_data;
            long hNormData = m_blobNorm.mutable_gpu_data;
            int nCount = colBottom[0].count();
            int nNum = colBottom[0].num;
            int nChannels = colBottom[0].channels;
            int nSpatialDim = colBottom[0].height * colBottom[0].width;
 
            m_cuda.copy(nCount, hBottomData, hTopData);
            m_cuda.copy(nCount, hBottomData, hSquareData);
 
            // square
            m_cuda.powx(nCount, hSquareData, 2.0, hSquareData);
 
            // sum across channel
            m_cuda.channel_sum(nNum * nSpatialDim, nNum, nChannels, nSpatialDim, hSquareData, hNormData);
 
            // square root
            m_cuda.powx(nNum * nSpatialDim, hNormData, 0.5, hNormData);
 
            // divide
            m_cuda.channel_div(nNum * nSpatialDim, nNum, nChannels, nSpatialDim, hNormData, hTopData, 2);
        }
 
        protected override void backward(BlobCollection<T> colTop, List<bool> rgbPropagateDown, BlobCollection<T> colBottom)
        {
            long hTopDiff = colTop[0].gpu_diff;
            long hTopData = colTop[0].gpu_data;
            long hBottomDiff = colBottom[0].mutable_gpu_diff;
            long hBottomData = colBottom[0].gpu_data;
            long hNormData = m_blobNorm.gpu_data;
            long hTempDotData = m_blobTempDot.mutable_gpu_data;
            long hTempData = m_blobSquare.mutable_gpu_data;
            int nNum = colTop[0].num;
            int nChannels = colTop[0].channels;
            int nSpatialDim = colTop[0].height * colTop[0].width;
            int nCount = colTop[0].count();
 
            m_cuda.copy(nCount, hTopDiff, hBottomDiff);
            m_cuda.copy(nCount, hBottomData, hTempData);
 
            // b_diff = t_diff / norm - dot(t_diff, t_data) / (norm)^2 * bottom_data
            // temp_dot_data = dot(t_diff, t_data)
            m_cuda.channel_dot(nNum * nSpatialDim, nNum, nChannels, nSpatialDim, hTopDiff, hTopData, hTempDotData);
 
            // temp_dot_data /= (norm)^2
            m_cuda.div(nNum * nSpatialDim, hTempDotData, hNormData, hTempDotData);
            m_cuda.div(nNum * nSpatialDim, hTempDotData, hNormData, hTempDotData);
 
            // bottom_diff = top_diff, bottom_diff /= norm
            m_cuda.channel_div(nNum * nSpatialDim, nNum, nChannels, nSpatialDim, hNormData, hBottomDiff, 2);
 
            // temp_data = bottom_data, temp_data *= temp_dot_data
            m_cuda.channel_mul(nNum * nSpatialDim, nNum, nChannels, nSpatialDim, hTempDotData, hTempData, 2);
 
            // bottom_diff += -temp_data
            m_cuda.axpy(nCount, -1.0, hTempData, hBottomDiff);
        }
    }
}