private void softmax_fwd(Blob <float> blobBottom, Blob <float> blobClip, Blob <float> blobScale, Blob <float> blobTop, int nAxis)
{
int nCount = blobBottom.count();
int nOuterNum = blobBottom.count(0, nAxis);
int nInnerNum = blobBottom.count(nAxis + 1);
int nChannels = blobTop.shape(nAxis);
long hBottomData = blobBottom.gpu_data;
long hTopData = blobTop.mutable_gpu_data;
long hScaleData = blobScale.mutable_gpu_data;
CudaDnn <float> cuda = m_mycaffe.Cuda;
cuda.copy(nCount, hBottomData, hTopData);
// Apply clip.
if (blobClip != null)
{
cuda.channel_scale(nCount, blobTop.num, blobTop.channels, blobTop.count(2), blobTop.gpu_data, blobClip.gpu_data, blobTop.mutable_gpu_data);
}
// We need to subtract the max to avoid numerical issues, compute the exp
// and then normalize.
// compute max.
cuda.channel_max(nOuterNum * nInnerNum, nOuterNum, nChannels, nInnerNum, hTopData, hScaleData);
// subtract
cuda.channel_sub(nCount, nOuterNum, nChannels, nInnerNum, hScaleData, hTopData);
// exponentiate
cuda.exp(nCount, hTopData, hTopData);
// Apply clip to remove 1's.
if (blobClip != null)
{
cuda.channel_scale(nCount, blobTop.num, blobTop.channels, blobTop.count(2), blobTop.gpu_data, blobClip.gpu_data, blobTop.mutable_gpu_data);
}
// Sum after exp
cuda.channel_sum(nOuterNum * nInnerNum, nOuterNum, nChannels, nInnerNum, hTopData, hScaleData);
// divide
cuda.channel_div(nCount, nOuterNum, nChannels, nInnerNum, hScaleData, hTopData);
// Denan for divide by zero.
cuda.denan(nCount, blobTop.mutable_gpu_data, 0);
}
/// <summary>
/// 'softmax' operation runs the softmax on each item in the btm and places the results in the top.
/// </summary>
/// <param name="btm">Specifies the input data.</param>
/// <param name="top">Specifies the output data.</param>
/// <returns>The top blob is returned.</returns>
public Blob <T> softmax(Blob <T> btm, Blob <T> top)
{
string strMarker = marker;
top.ReshapeLike(btm);
int nOuterNum = btm.count(0, m_nAxis);
int nInnerNum = btm.count(m_nAxis + 1);
int nChannels = top.shape(m_nAxis);
int nCount = btm.count();
work.ReshapeLike(top);
m_cuda.copy(nCount, btm.gpu_data, top.mutable_gpu_data);
// We need to subtract the max to avoid numerical issues, compute the exp
// and then normalize.
// compute max.
m_cuda.channel_max(nOuterNum * nInnerNum, nOuterNum, nChannels, nInnerNum, top.gpu_data, work.mutable_gpu_data);
// subtract
m_cuda.channel_sub(nCount, nOuterNum, nChannels, nInnerNum, work.gpu_data, top.mutable_gpu_data);
// exponentiate
m_cuda.exp(nCount, top.gpu_data, top.mutable_gpu_data);
// Sum after exp
m_cuda.channel_sum(nOuterNum * nInnerNum, nOuterNum, nChannels, nInnerNum, top.gpu_data, work.mutable_gpu_data);
// divide
m_cuda.channel_div(nCount, nOuterNum, nChannels, nInnerNum, work.gpu_data, top.mutable_gpu_data);
if (m_bNeedsBackprop)
{
Action backward = () =>
{
work.ReshapeLike(top);
m_cuda.copy(nCount, top.gpu_diff, work.mutable_gpu_diff);
// Compute inner1d(top_diff, top_data) and subtract them from the bottom diff.
m_cuda.channel_dot(nOuterNum * nInnerNum, nOuterNum, nChannels, nInnerNum, top.gpu_diff, top.gpu_data, work.mutable_gpu_data);
m_cuda.channel_sub(nCount, nOuterNum, nChannels, nInnerNum, work.gpu_data, work.mutable_gpu_diff);
// elementwise multiplication
m_cuda.mul(nCount, work.gpu_diff, top.gpu_data, work.mutable_gpu_diff);
apply(work, btm);
if (m_bClipGradients)
{
clip_gradient(btm);
}
if (m_bCheckForNans)
{
check_nan(btm);
}
if (m_bAddDebug)
{
add_debug(strMarker + " - softmax", btm, top);
}
};
m_rgBackprop.Add(new Tuple <string, Action>(m_strMarker, backward));
}
return(top);
}