private TSNEState(TSNEState original, Cloner cloner) : base(original, cloner)
{
distance = cloner.Clone(original.distance);
random = cloner.Clone(original.random);
perplexity = original.perplexity;
exact = original.exact;
noDatapoints = original.noDatapoints;
finalMomentum = original.finalMomentum;
momSwitchIter = original.momSwitchIter;
stopLyingIter = original.stopLyingIter;
theta = original.theta;
eta = original.eta;
newDimensions = original.newDimensions;
if (original.valP != null)
{
valP = new double[original.valP.Length];
Array.Copy(original.valP, valP, valP.Length);
}
if (original.rowP != null)
{
rowP = new int[original.rowP.Length];
Array.Copy(original.rowP, rowP, rowP.Length);
}
if (original.colP != null)
{
colP = new int[original.colP.Length];
Array.Copy(original.colP, colP, colP.Length);
}
if (original.p != null)
{
p = new double[original.p.GetLength(0), original.p.GetLength(1)];
Array.Copy(original.p, p, p.Length);
}
newData = new double[original.newData.GetLength(0), original.newData.GetLength(1)];
Array.Copy(original.newData, newData, newData.Length);
iter = original.iter;
currentMomentum = original.currentMomentum;
gains = new double[original.gains.GetLength(0), original.gains.GetLength(1)];
Array.Copy(original.gains, gains, gains.Length);
uY = new double[original.uY.GetLength(0), original.uY.GetLength(1)];
Array.Copy(original.uY, uY, uY.Length);
dY = new double[original.dY.GetLength(0), original.dY.GetLength(1)];
Array.Copy(original.dY, dY, dY.Length);
}
public static double[,] Iterate(TSNEState state)
{
if (state.exact)
{
ComputeExactGradient(state.p, state.newData, state.noDatapoints, state.newDimensions, state.dY);
}
else
{
ComputeApproximateGradient(state.rowP, state.colP, state.valP, state.newData, state.noDatapoints, state.newDimensions, state.dY, state.theta);
}
// Update gains
for (var i = 0; i < state.noDatapoints; i++)
{
for (var j = 0; j < state.newDimensions; j++)
{
state.gains[i, j] = Math.Sign(state.dY[i, j]) != Math.Sign(state.uY[i, j])
? state.gains[i, j] + .2 // +0.2 nd *0.8 are used in two separate implementations of tSNE -> seems to be correct
: state.gains[i, j] * .8;
if (state.gains[i, j] < .01)
{
state.gains[i, j] = .01;
}
}
}
// Perform gradient update (with momentum and gains)
for (var i = 0; i < state.noDatapoints; i++)
{
for (var j = 0; j < state.newDimensions; j++)
{
state.uY[i, j] = state.currentMomentum * state.uY[i, j] - state.eta * state.gains[i, j] * state.dY[i, j];
}
}
for (var i = 0; i < state.noDatapoints; i++)
{
for (var j = 0; j < state.newDimensions; j++)
{
state.newData[i, j] = state.newData[i, j] + state.uY[i, j];
}
}
// Make solution zero-mean
ZeroMean(state.newData);
// Stop lying about the P-values after a while, and switch momentum
if (state.iter == state.stopLyingIter)
{
if (state.exact)
{
for (var i = 0; i < state.noDatapoints; i++)
{
for (var j = 0; j < state.noDatapoints; j++)
{
state.p[i, j] /= 12.0;
}
}
}
else
{
for (var i = 0; i < state.rowP[state.noDatapoints]; i++)
{
state.valP[i] /= 12.0;
}
}
}
if (state.iter == state.momSwitchIter)
{
state.currentMomentum = state.finalMomentum;
}
state.iter++;
return(state.newData);
}