void Start()
{
if (openSee == null)
{
openSee = GetComponent <OpenSee>();
}
if (openSee == null)
{
ResetInfo();
return;
}
ResetInfo();
expressions = new Dictionary <string, List <float[]> >();
model = new ThunderSVMModel();
rnd = new System.Random();
}
//private bool newModel = false;
static public void LoadSerialized(byte[] modelBytes, out Dictionary <string, List <float[]> > expressions, out SVMModel model, out string[] classLabels, out int[] indices, ref PointSelection pointSelection)
{
IFormatter formatter = new BinaryFormatter();
MemoryStream memoryStream = new MemoryStream(modelBytes);
memoryStream.Position = 0;
OpenSeeExpressionRepresentation oser;
using (GZipStream gzipStream = new GZipStream(memoryStream, CompressionMode.Decompress)) {
oser = formatter.Deserialize(gzipStream) as OpenSeeExpressionRepresentation;
}
expressions = oser.expressions;
if (oser.thunderSVM)
{
model = new ThunderSVMModel(oser.modelBytes);
}
else
{
model = new SVMModel(oser.modelBytes);
}
classLabels = oser.classLabels;
indices = oser.indices;
//pointSelection = oser.pointSelection;
if (indices == null)
{
indices = new int[1 + 1 + 3 + 4 + 3 + 3 * 66];
for (int i = 0; i < 1 + 1 + 3 + 4 + 3 + 3 * 66; i++)
{
indices[i] = i;
}
}
/*if (pointSelection == null)
* pointSelection = new PointSelection();
* if (!oser.newModel)
* pointSelection.features = true;*/
}
public bool TrainModel()
{
Debug.Log("---------------------");
train = false;
if (openSee == null)
{
ResetInfo();
return(false);
}
modelReady = false;
if (model is SVMModel)
{
model = new ThunderSVMModel();
}
SelectPoints();
List <string> keys = new List <string>();
List <string> accuracyWarnings = new List <string>();
int samples = 0;
foreach (string key in expressions.Keys)
{
List <float[]> list = new List <float[]>(expressions[key]);
list.RemoveAll(x => x.Length != colsFull);
if (list != null && list.Count == maxSamples)
{
Debug.Log("[Training info] Adding expression " + key + " to training data.");
keys.Add(key);
samples += list.Count;
}
else
{
if (list != null && list.Count > 20)
{
Debug.Log("[Training warning] Expression " + key + " has little data and might be inaccurate. (" + list.Count + ")");
accuracyWarnings.Add("Expression " + key + " has little data and might be inaccurate. (" + list.Count + ")");
samples += list.Count;
keys.Add(key);
}
else
{
Debug.Log("[Training warning] Skipping expression " + key + " due to lack of collected data. (" + list.Count + ")");
accuracyWarnings.Add("Skipping expression " + key + " due to lack of collected data. (" + list.Count + ")");
continue;
}
}
}
int classes = keys.Count;
if (classes < 2 || classes > 10)
{
Debug.Log("[Training error] The number of expressions that can be used for training is " + classes + ", which is either below 2 or higher than 10.");
accuracyWarnings.Add("[Training error] The number of expressions that can be used for training is " + classes + ", which is either below 2 or higher than 10.");
return(false);
}
keys.Sort();
classLabels = keys.ToArray();
float[] weights = null;
if (weightMap != null)
{
weights = new float[classLabels.Length];
for (int i = 0; i < classLabels.Length; i++)
{
if (weightMap.ContainsKey(classLabels[i]))
{
weights[i] = weightMap[classLabels[i]];
}
else
{
weights[i] = 1f;
}
Debug.Log("[Training info] Adding weight " + weights[i] + " for " + classLabels[i] + ".");
}
}
Debug.Log("[Training info] Preparing trainig data.");
int train_split = maxSamples * 3 / 4;
int test_split = maxSamples - train_split;
int rows_train = classes * train_split;
int rows_test = classes * test_split;
float[] X_train = new float[rows_train * cols];
float[] X_test = new float[rows_test * cols];
float[] y_train = new float[rows_train];
float[] y_test = new float[rows_test];
int i_train = 0;
int i_test = 0;
System.Random rnd = new System.Random();
for (int i = 0; i < classes; i++)
{
List <float[]> list = new List <float[]>(expressions[keys[i]]);
list.RemoveAll(x => x.Length != colsFull);
int local_train_split = list.Count * 3 / 4;
test_split = list.Count - local_train_split;
for (int j = list.Count; j > 1;)
{
j--;
int k = rnd.Next(j + 1);
float[] tmp = list[k];
list[k] = list[j];
list[j] = tmp;
}
int train_split_current = train_split;
if (classLabels[i] == "neutral")
{
train_split_current *= 1;
}
for (int j = 0; j < train_split_current; j++)
{
float factor = 1f;
float adder = 0f;
/*if (j > train_split || j > list.Count) {
* factor = UnityEngine.Random.Range(0.98f, 1.02f);
* adder = UnityEngine.Random.Range(-0.02f, 0.02f);
* }*/
for (int k = 0; k < cols; k++)
{
float v = list[j % local_train_split][indices[k]] * factor + adder;
X_train[i_train * cols + k] = v;
}
y_train[i_train] = i;
i_train++;
}
for (int j = local_train_split; j < local_train_split + test_split; j++)
{
for (int k = 0; k < cols; k++)
{
float v = list[j][indices[k]];
X_test[i_test * cols + k] = v;
}
y_test[i_test] = i;
i_test++;
}
}
Debug.Log("[Training info] Preparation complete.");
int probability = 0;
if (enableProbabilityTraining)
{
probability = 1;
}
model.TrainModel(X_train, y_train, weights, i_train, cols, probability, C);
confusionMatrix = model.ConfusionMatrix(X_test, y_test, i_test, out accuracy);
confusionMatrixString = SVMModel.FormatMatrix(confusionMatrix, classLabels);
for (int label = 0; label < classes; label++)
{
int total = 0;
for (int p = 0; p < classes; p++)
{
total += confusionMatrix[label, p];
}
float error = 100f * (1f - ((float)confusionMatrix[label, label] / (float)total));
if (error > warningThreshold)
{
accuracyWarnings.Add("[Training warning] The expression \"" + classLabels[label] + "\" is misclassified with a chance of " + error.ToString("0.00") + "%.");
Debug.Log("[Training warning] The expression \"" + classLabels[label] + "\" is misclassified with a chance of " + error.ToString("0.00") + "%.");
}
}
warnings = accuracyWarnings.ToArray();
modelReady = true;
Debug.Log("[Training info] Trained model.");
return(true);
}
static public byte[] ToSerialized(Dictionary <string, List <float[]> > expressions, SVMModel model, string[] classLabels, int[] indices, PointSelection pointSelection)
{
OpenSeeExpressionRepresentation oser = new OpenSeeExpressionRepresentation();
oser.expressions = expressions;
oser.modelBytes = model.SaveModel();
oser.classLabels = classLabels;
oser.indices = indices;
oser.pointSelection = pointSelection;
oser.thunderSVM = true;
//oser.newModel = true;
IFormatter formatter = new BinaryFormatter();
MemoryStream memoryStream = new MemoryStream();
using (GZipStream gzipStream = new GZipStream(memoryStream, CompressionMode.Compress)) {
formatter.Serialize(gzipStream, oser);
gzipStream.Flush();
}
return(memoryStream.ToArray());
}