public override void LearnUnits()
{
//I should probably make this a float array.
//I'm not sure what kind of memory restrictions we're looking at.
//This list of arrays is going to be reclaimed after the method finishes, as is the trainingSet
List <double[]> features = new List <double[]> ();
int featureLength = -1;
//Gets the relevant data out of the training set
foreach (PhantomUnit u in trainingSet)
{
Texture tex = u.tex;
if (tex is Texture2D)
{
Texture2D tex2D = (Texture2D)tex;
Color[] pixels = tex2D.GetPixels();
if (featureLength == -1)
{
featureLength = pixels.Length * 3;
unitWidth = tex2D.width;
unitHeight = tex2D.height;
}
//I don't actually check that the width and height are correct, just that width*height is correct. I'm going to assume it'll work out.
if (pixels.Length * 3 == featureLength)
{
//This is really IMPORTANT.
//The last spot in unit features is reserved for the unit identity (0 for Enemy, 1 for Ally)
//Don't iterate over unitFeatures or its full length.
double[] unitFeatures = new double[featureLength + 1];
//Assigns the pixel values for each of the colors in the pattern {r,g,b,r,g,b,...)
for (int i = 0; i < pixels.Length; i++)
{
Color c = pixels [i];
unitFeatures [3 * i] = c.r;
unitFeatures [3 * i + 1] = c.g;
unitFeatures [3 * i + 2] = c.b;
}
unitFeatures [featureLength] = u.identity;
features.Add(unitFeatures);
}
else
{
Debug.Log("Unit has incorrect dimensions to be analyzed: " + u.ToString());
}
}
else if (tex is RenderTexture)
{
RenderTexture texRend = (RenderTexture)tex;
Debug.Log("This is a RenderTexture");
//I don't have the code to get the pixels for this yet
}
else
{
Debug.Log("I'm not sure what Texture this is or what to do with it.");
}
} //for
weights = new double[featureLength];
b = 0;
for (int it = 0; it < iters; it++)
{
itersUsed = it + 1;
int misses = 0;
//Shuffles the feature list
NeuralNode.Shuffle(features);
foreach (double[] fs in features)
{
int trueIdentity = (int)Mathf.Round((float)fs [featureLength]);
int f = (int)Unit.ALLY_IDENTITY;
if (calculateZ(fs) < 0)
{
f = (int)Unit.ENEMY_IDENTITY;
}
misses += Mathf.Abs(f - trueIdentity);
for (int i = 0; i < featureLength; i++)
{
weights [i] = weights [i] + learningRate * fs[i] * (trueIdentity - f);
}
b = b + learningRate * (trueIdentity - f);
} //for each of the feature arrays
if (misses == 0)
{
break;
}
} //for iterations
actualWeights = new double[weights.Length];
System.Array.Copy(weights, actualWeights, weights.Length);
actualB = b;
SetNonzeroIndices();
//Notice that I am clearing out the training set here.
//Keeping the references is not necessary and will just take up space.
//Being able to get rid of them is sort of the point of making a model.
foreach (PhantomUnit u in trainingSet)
{
Destroy(u);
}
trainingSet.Clear();
isAILearned = true;
}
