copilot/OpenSeeFace
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

It is now possible to train expression models with probability support. This however is a lot slower! Also, the C parameter for SVM training was set incorrectly. Now training works much better.

Browse source
This commit is contained in:
Emiliana 2020-02-13 22:24:41 +01:00
parent 9abc3029e7
commit bb845098ca
4 changed files with 46 additions and 20 deletions
Download patch file Download diff file Expand all files Collapse all files

20
Source/SVMWrapper.cpp
View file

@ -15,6 +15,7 @@ extern "C" void __cdecl print_log(const char *str) {
struct SVMWrapper {
struct svm_model *model;
int cols;
int classes;
double *means;
double *sdevs;
struct svm_parameter param;
@ -86,7 +87,7 @@ void destroy_problem(struct svm_problem *problem) {
}
extern "C" {
__declspec(dllexport) void *__cdecl trainModel(float features[], float labels[], int rows, int cols) {
__declspec(dllexport) void *__cdecl trainModel(float features[], float labels[], int rows, int cols, int classes, int probability, float C) {
if (!(rows > 0 && cols > 0))
return NULL;
@ -94,6 +95,7 @@ extern "C" {
struct SVMWrapper *sw = new SVMWrapper();
sw->cols = cols;
sw->classes = classes;
sw->means = new double[cols];
sw->sdevs = new double[cols];
@ -104,11 +106,11 @@ extern "C" {
sw->param.coef0 = 0;
sw->param.nu = 0.5;
sw->param.cache_size = 100;
sw->param.C = 4;
sw->param.C = C;
sw->param.eps = 0.001;
sw->param.p = 0.1;
sw->param.shrinking = 1;
sw->param.probability = 0;
sw->param.probability = probability;
sw->param.nr_weight = 0;
sw->param.weight_label = NULL;
sw->param.weight = NULL;
@ -120,22 +122,23 @@ extern "C" {
return sw;
}
__declspec(dllexport) void __cdecl predict(void *ptr, float features[], float predictions[], int rows) {
__declspec(dllexport) void __cdecl predict(void *ptr, float features[], float predictions[], double probabilities[], int rows) {
struct SVMWrapper *sw = (struct SVMWrapper *)ptr;
int cols = sw->cols;
svmnode **x = make_svmnode_array(features, rows, cols, sw->means, sw->sdevs, 0);
for (int r = 0; r < rows; r++)
predictions[r] = (float)svm_predict(sw->model, x[r]);
predictions[r] = (float)svm_predict_probability(sw->model, x[r], &(probabilities[r * sw->classes]));
destroy_svmnode_array(x);
return;
}
__declspec(dllexport) void *__cdecl loadModel(char *filename, int cols, double means[], double sdevs[]) {
__declspec(dllexport) void *__cdecl loadModel(char *filename, int cols, int classes, double means[], double sdevs[]) {
struct SVMWrapper *sw = new SVMWrapper();
sw->cols = cols;
sw->model = svm_load_model(filename);
sw->classes = classes;
sw->means = new double[cols];
sw->sdevs = new double[cols];
for (int c = 0; c < cols; c++) {
@ -145,10 +148,11 @@ extern "C" {
return sw;
}
__declspec(dllexport) void *__cdecl loadModelString(char *modelString, int cols, double means[], double sdevs[]) {
__declspec(dllexport) void *__cdecl loadModelString(char *modelString, int cols, int classes, double means[], double sdevs[]) {
struct SVMWrapper *sw = new SVMWrapper();
sw->cols = cols;
sw->model = svm_load_model_string(modelString);
sw->classes = classes;
sw->means = new double[cols];
sw->sdevs = new double[cols];
for (int c = 0; c < cols; c++) {
@ -168,7 +172,7 @@ extern "C" {
}
}
__declspec(dllexport) char *__cdecl getModel(void *ptr) {
__declspec(dllexport) char *__cdecl saveModelString (void *ptr) {
struct SVMWrapper *sw = (struct SVMWrapper *)ptr;
return svm_save_model_string(sw->model);
}

18
Unity/OpenSeeExpression.cs
View file

@ -25,6 +25,8 @@ public class OpenSeeExpression : MonoBehaviour
public int overRecordingSkip = 5;
[Tooltip("This is the filename used for loading and saving expression data using the load and save flags.")]
public string filename = "";
[Tooltip("When enabled, training will be done in such a way that the model can output probabilities. This is a lot slower!")]
public bool enableProbabilityTraining = false;
[Header("Toggles")]
[Tooltip("When enabled, calibration data will be collected from the given OpenSee component.")]
public bool recording = false;
@ -54,6 +56,8 @@ public class OpenSeeExpression : MonoBehaviour
public float expressionTime = 0f;
[Tooltip("This is currently detected expression.")]
public string expression = "";
[Tooltip("These are the probabilities of all expressions. The labels in order can be retrieved through the GetClassLabels methods.")]
public double[] probabilities;
[Header("Training statistics")]
[Tooltip("This shows the accuracy result of the last training run.")]
public float accuracy = 0.0f;
@ -362,7 +366,10 @@ public class OpenSeeExpression : MonoBehaviour
i_test++;
}
}
model.TrainModel(X_train, y_train, rows_train, cols);
int probability = 0;
if (enableProbabilityTraining)
probability = 1;
model.TrainModel(X_train, y_train, rows_train, cols, probability);
confusionMatrix = model.ConfusionMatrix(X_test, y_test, i_test, out accuracy);
confusionMatrixString = SVMModel.FormatMatrix(confusionMatrix, classLabels);
for (int label = 0; label < classes; label++) {
@ -399,7 +406,7 @@ public class OpenSeeExpression : MonoBehaviour
float[] predictionData = new float[cols];
for (int i = 0; i < cols; i++)
predictionData[i] = faceData[indices[i]];
float[] prediction = model.Predict(predictionData, 1);
float[] prediction = model.Predict(predictionData, out probabilities, 1);
int predictedExpression = (int)Mathf.Round(prediction[0]);
if (predictedExpression == lastPrediction)
lastPredictionCount++;
@ -464,6 +471,13 @@ public class OpenSeeExpression : MonoBehaviour
return false;
}
}
public string[] GetClassLabels() {
List<string> labels = new List<string>();
foreach (string l in classLabels)
labels.Add(String.Copy(l));
return labels.ToArray();
}
public string[] GetExpressions() {
List<string> keys = new List<string>();

28
Unity/Plugins/SVMModel.cs
View file

@ -12,16 +12,16 @@ namespace OpenSee {
public class SVMModel {
#region DllImport
[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl)]
private static extern System.IntPtr trainModel(float[] features, float[] labels, int rows, int cols, float C);
private static extern System.IntPtr trainModel(float[] features, float[] labels, int rows, int cols, int classes, int probability, float C);
[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl)]
private static extern void predict(System.IntPtr model, float[] features, [Out] float[] predictions, int rows);
private static extern void predict(System.IntPtr model, float[] features, [Out] float[] predictions, [Out] double[] probabilities, int rows);
[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Ansi)]
private static extern System.IntPtr loadModel([MarshalAs(UnmanagedType.LPStr)]string filename, int cols, double[] means, double[] sdevs);
private static extern System.IntPtr loadModel([MarshalAs(UnmanagedType.LPStr)]string filename, int cols, int classes, double[] means, double[] sdevs);
[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Ansi)]
private static extern System.IntPtr loadModelString([MarshalAs(UnmanagedType.LPStr)]string modelString, int cols, double[] means, double[] sdevs);
private static extern System.IntPtr loadModelString([MarshalAs(UnmanagedType.LPStr)]string modelString, int cols, int classes, double[] means, double[] sdevs);
[DllImport("SVMModel", CallingConvention = CallingConvention.Cdecl)]
private static extern void getScales(System.IntPtr model, [Out] double[] means, [Out] double[] sdevs);
@ -63,7 +63,7 @@ public class SVMModel {
}
cols = smr.cols;
maxClasses = smr.maxClasses;
System.IntPtr model = loadModelString(smr.modelString, smr.cols, smr.means, smr.sdevs);
System.IntPtr model = loadModelString(smr.modelString, smr.cols, smr.maxClasses, smr.means, smr.sdevs);
return model;
}
@ -118,7 +118,7 @@ public class SVMModel {
}
// features is a rows*cols long float array with rows written one after another. labels is a rows long float array with the corresponding classes as integral numbers from 0 to num_classes-1, where all classes have to exist. Training on more than 10000 rows is disabled.
public bool TrainModel(float[] features, float[] labels, int rows, int cols) {
public bool TrainModel(float[] features, float[] labels, int rows, int cols, int probability) {
DestroyModel();
if (rows > 10000)
return false;
@ -143,7 +143,7 @@ public class SVMModel {
maxClasses = max + 1;
this.cols = cols;
model = trainModel(features, labels, rows, cols, 12f);
model = trainModel(features, labels, rows, cols, maxClasses, probability, 2f);
haveModel = true;
return true;
@ -163,7 +163,8 @@ public class SVMModel {
}
int[,] confusionMatrix = new int[maxClasses, maxClasses];
float[] predictions = new float[rows];
predict(model, features, predictions, rows);
double[] probabilities = new double[rows * maxClasses];
predict(model, features, predictions, probabilities, rows);
for (int i = 0; i < rows; i++) {
predictions[i] = (float)Math.Round(predictions[i]);
if (predictions[i] >= 0 && predictions[i] < maxClasses) {
@ -219,15 +220,22 @@ public class SVMModel {
}
// features is a rows*cols long float array with rows written one after another.
public float[] Predict(float[] features, int rows) {
public float[] Predict(float[] features, out double[] probabilities, int rows) {
probabilities = null;
if (!haveModel)
return null;
if (rows < 1 || features.Length != rows * cols)
return null;
float[] predictions = new float[rows];
predict(model, features, predictions, rows);
probabilities = new double[rows * maxClasses];
predict(model, features, predictions, probabilities, rows);
return predictions;
}
public float[] Predict(float[] features, int rows) {
double[] probabilities;
return Predict(features, out probabilities, rows);
}
public bool Ready() {
return haveModel;

BIN
Unity/Plugins/SVMModel.dll
View file

Binary file not shown.