GaussianFiller.cs

using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using MyCaffe.basecode;
using MyCaffe.param;
using MyCaffe.common;
 
namespace MyCaffe.fillers
{
    public class GaussianFiller<T> : Filler<T>
    {
        SyncedMemory<T> m_randVec = null;
 
        public GaussianFiller(CudaDnn<T> cuda, Log log, FillerParameter p)
            : base(cuda, log, p)
        {
            m_randVec = new SyncedMemory<T>(m_cuda, m_log);
        }
 
        public override void Fill(int nCount, long hMem, int nNumAxes = 1, int nNumOutputs = 1, int nNumChannels = 1, int nHeight = 1, int nWidth = 1)
        {
            m_log.CHECK(nCount > 0, "There is no data to fill!");
            m_cuda.rng_gaussian(nCount, m_param.mean, m_param.std, hMem);
 
            int nSparse = m_param.sparse;
            m_log.CHECK_GE(nSparse, -1, "The sparse value should be >= -1.");
 
            if (nSparse >= 0)
            {
                // Sparse initialization is implemented for 'weight' blobs; i.e. matrices.
                // These have num == channels == 1; width is number of inputs; height is
                // number of outputs.  The 'sparse' variable specifies the mean number
                // of non-zero input weights for a given output.
                m_log.CHECK_GE(nNumAxes, 1, "The blob must have at least one axis.");
 
                double dfNonZeroProbability = (double)nSparse / (double)nNumOutputs;
                T fNonZeroProbability = (T)Convert.ChangeType(dfNonZeroProbability, typeof(T));
 
                m_randVec.Allocate(nCount);
                m_cuda.rng_bernoulli(nCount, fNonZeroProbability, m_randVec.gpu_data);
                m_cuda.mul(nCount, hMem, m_randVec.gpu_data, hMem);
            }
        }
    }
}