/// <summary>
/// Setup the layer.
/// </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)
{
base.LayerSetUp(colBottom, colTop);
GetIterationArgs args = getCurrentIteration();
if (args != null && args.CurrentPhase != Phase.TRAIN)
{
return;
}
m_log.CHECK(args != null, "WARNING: The OnGetIteration event is not connected!");
m_dfLowerBound = m_param.gradient_scale_param.lower_bound;
m_dfUpperBound = m_param.gradient_scale_param.upper_bound;
m_dfAlpha = m_param.gradient_scale_param.alpha;
m_dfMaxIter = m_param.gradient_scale_param.max_iter;
m_dfCoeff = 1.0; // default adaptation coefficient.
m_log.CHECK_LE(m_dfLowerBound, m_dfUpperBound, "The lower bound must be <= the upper bound.");
m_log.CHECK_GE(m_dfAlpha, 0, "The alpha value must be >= 0.0");
m_log.CHECK_GE(m_dfCoeff, 1, "The max_iter must be >= 1.0");
int nIteration = (args == null) ? 1 : args.Iteration;
double dfProgress = Math.Min(1.0, (double)nIteration / m_dfMaxIter);
double dfHeight = m_dfUpperBound - m_dfLowerBound;
m_dfCoeff = 2.0 * dfHeight / (1.0 + Math.Exp(-m_dfAlpha * dfProgress)) - dfHeight + m_dfLowerBound;
m_log.WriteLine("iter = " + nIteration.ToString() + " progress = " + dfProgress.ToString() + " coeff = " + m_dfCoeff.ToString());
m_swOutput.Start();
}
/// <summary>
/// Scales the error gradient w.r.t. the GRADIENTSCALER value inputs.
/// </summary>
/// <param name="colTop">top output blob vector, providing the error gradient
/// with respect to outputs
/// </param>
/// <param name="rgbPropagateDown">propagate_down see Layer::Backward.</param>
/// <param name="colBottom">bottom input blob vector
/// </param>
protected override void backward(BlobCollection <T> colTop, List <bool> rgbPropagateDown, BlobCollection <T> colBottom)
{
if (!rgbPropagateDown[0])
{
return;
}
GetIterationArgs args = getCurrentIteration();
if (args != null && args.CurrentPhase != Phase.TRAIN)
{
return;
}
int nCount = colTop[0].count();
long hTopDiff = colTop[0].gpu_diff;
long hBottomDiff = colBottom[0].mutable_gpu_diff;
int nIteration = (args == null) ? 1 : args.Iteration;
double dfProgress = Math.Min(1.0, (double)nIteration / m_dfMaxIter);
double dfHeight = m_dfUpperBound - m_dfLowerBound;
m_dfCoeff = 2.0 * dfHeight / (1.0 + Math.Exp(-m_dfAlpha * dfProgress)) - dfHeight + m_dfLowerBound;
if (m_swOutput.Elapsed.TotalMilliseconds > 1000)
{
m_log.WriteLine("iter = " + nIteration.ToString() + " progress = " + dfProgress.ToString() + " coeff = " + m_dfCoeff.ToString());
m_swOutput.Restart();
}
m_cuda.scale(nCount, -m_dfCoeff, hTopDiff, hBottomDiff);
}
/// <summary>
/// Fires the OnGetIteration event to query the current iteration.
/// </summary>
/// <returns>The GetIterationArgs is returned if the event is connected, otherwise <i>null</i> is returned.</returns>
protected GetIterationArgs getCurrentIteration()
{
GetIterationArgs args = null;
if (OnGetIteration != null)
{
args = new GetIterationArgs();
OnGetIteration(this, args);
}
return(args);
}