public StdVeci(StdVeci other) : this(eogmaneoPINVOKE.new_StdVeci__SWIG_1(StdVeci.getCPtr(other)), true)
{
if (eogmaneoPINVOKE.SWIGPendingException.Pending)
{
throw eogmaneoPINVOKE.SWIGPendingException.Retrieve();
}
}
public StdVeciEnumerator(StdVeci collection)
{
collectionRef = collection;
currentIndex = -1;
currentObject = null;
currentSize = collectionRef.Count;
}
public void AddRange(StdVeci values)
{
eogmaneoPINVOKE.StdVeci_AddRange(swigCPtr, StdVeci.getCPtr(values));
if (eogmaneoPINVOKE.SWIGPendingException.Pending)
{
throw eogmaneoPINVOKE.SWIGPendingException.Retrieve();
}
}
public void create(StdVecPairi inputSizes, StdVeci inputChunkSizes, StdVecb predictInputs, StdVecLayerDesc layerDescs, uint seed)
{
eogmaneoPINVOKE.Hierarchy_create(swigCPtr, StdVecPairi.getCPtr(inputSizes), StdVeci.getCPtr(inputChunkSizes), StdVecb.getCPtr(predictInputs), StdVecLayerDesc.getCPtr(layerDescs), seed);
if (eogmaneoPINVOKE.SWIGPendingException.Pending)
{
throw eogmaneoPINVOKE.SWIGPendingException.Retrieve();
}
}
public void SetRange(int index, StdVeci values)
{
eogmaneoPINVOKE.StdVeci_SetRange(swigCPtr, index, StdVeci.getCPtr(values));
if (eogmaneoPINVOKE.SWIGPendingException.Pending)
{
throw eogmaneoPINVOKE.SWIGPendingException.Retrieve();
}
}
public StdVeci activate(StdVecf input, ComputeSystem system)
{
StdVeci ret = new StdVeci(eogmaneoPINVOKE.RandomEncoder_activate__SWIG_1(swigCPtr, StdVecf.getCPtr(input), ComputeSystem.getCPtr(system)), false);
if (eogmaneoPINVOKE.SWIGPendingException.Pending)
{
throw eogmaneoPINVOKE.SWIGPendingException.Retrieve();
}
return(ret);
}
public StdVecf reconstruct(StdVeci hiddenStates, ComputeSystem system)
{
StdVecf ret = new StdVecf(eogmaneoPINVOKE.RandomEncoder_reconstruct(swigCPtr, StdVeci.getCPtr(hiddenStates), ComputeSystem.getCPtr(system)), false);
if (eogmaneoPINVOKE.SWIGPendingException.Pending)
{
throw eogmaneoPINVOKE.SWIGPendingException.Retrieve();
}
return(ret);
}
public static StdVeci Repeat(int value, int count)
{
global::System.IntPtr cPtr = eogmaneoPINVOKE.StdVeci_Repeat(value, count);
StdVeci ret = (cPtr == global::System.IntPtr.Zero) ? null : new StdVeci(cPtr, true);
if (eogmaneoPINVOKE.SWIGPendingException.Pending)
{
throw eogmaneoPINVOKE.SWIGPendingException.Retrieve();
}
return(ret);
}
public StdVeci GetRange(int index, int count)
{
global::System.IntPtr cPtr = eogmaneoPINVOKE.StdVeci_GetRange(swigCPtr, index, count);
StdVeci ret = (cPtr == global::System.IntPtr.Zero) ? null : new StdVeci(cPtr, true);
if (eogmaneoPINVOKE.SWIGPendingException.Pending)
{
throw eogmaneoPINVOKE.SWIGPendingException.Retrieve();
}
return(ret);
}
public StdVeci getHiddenStates()
{
StdVeci ret = new StdVeci(eogmaneoPINVOKE.RandomEncoder_getHiddenStates(swigCPtr), false);
return(ret);
}
public StdVeci getHiddenStatesPrev()
{
StdVeci ret = new StdVeci(eogmaneoPINVOKE.Layer_getHiddenStatesPrev(swigCPtr), true);
return(ret);
}
public StdVeci getInputs(int v)
{
StdVeci ret = new StdVeci(eogmaneoPINVOKE.Layer_getInputs(swigCPtr, v), true);
return(ret);
}
internal static global::System.Runtime.InteropServices.HandleRef getCPtr(StdVeci obj)
{
return((obj == null) ? new global::System.Runtime.InteropServices.HandleRef(null, global::System.IntPtr.Zero) : obj.swigCPtr);
}
public StdVeci getPredictionsPrev(int v)
{
StdVeci ret = new StdVeci(eogmaneoPINVOKE.Layer_getPredictionsPrev(swigCPtr, v), true);
return(ret);
}
public StdVeci getPrediction(int i)
{
StdVeci ret = new StdVeci(eogmaneoPINVOKE.Hierarchy_getPrediction(swigCPtr, i), false);
return(ret);
}
public void LineSegmentDetector(StdVecf data, int width, int height, int chunkSize, StdVeci rotationSDR)
{
eogmaneoPINVOKE.OpenCVInterop_LineSegmentDetector__SWIG_1(swigCPtr, StdVecf.getCPtr(data), width, height, chunkSize, StdVeci.getCPtr(rotationSDR));
if (eogmaneoPINVOKE.SWIGPendingException.Pending)
{
throw eogmaneoPINVOKE.SWIGPendingException.Retrieve();
}
}
void Update()
{
if (applicationExiting)
{
return;
}
if (cameraTexture == null || predictionTexture == null || carController == null)
{
return;
}
// Remember currently active render texture
RenderTexture currentActiveRT = RenderTexture.active;
// Transfer the camera capture into the prediction texture (temporarily)
RenderTexture.active = cameraTexture;
predictionTexture.ReadPixels(new Rect(0, 0, _inputWidth, _inputHeight), 0, 0);
predictionTexture.Apply();
// Restore active render texture
RenderTexture.active = currentActiveRT;
// Edge Detection Convolution methods:
// - Canny - https://en.wikipedia.org/wiki/Canny_edge_detector
// Laplacian of the Gaussian (LoG) - https://en.wikipedia.org/wiki/Blob_detection#The_Laplacian_of_Gaussian
// - Sobel-Feldman and Sharr operators - https://en.wikipedia.org/wiki/Sobel_operator
// - Prewitt operator - https://en.wikipedia.org/wiki/Prewitt_operator
// Kirch operator - https://en.wikipedia.org/wiki/Kirsch_operator
bool useSobel = false;
bool useCanny = false && !useSobel;
bool useBlur = false && !useCanny; // Canny already includes Gaussian blurring
bool useThreholding = false;
bool useGaborFilter = false;
bool useLineSegmentDetector = true;
bool useFastFeatureDetector = !useLineSegmentDetector;
// Blur entire camera image?
if (useBlur)
{
Texture2D blurredTexture = ConvolutionFilter.Apply(predictionTexture, ConvolutionFilter.GaussianBlur);
predictionTexture.SetPixels(blurredTexture.GetPixels());
}
// Convert from RGB space to Y'UV (ignoring chrominance)
Color actualPixel = new Color();
Color yuvPixel = new Color();
for (int x = 0; x < _inputWidth; x++)
{
for (int y = 0; y < _inputHeight; y++)
{
actualPixel = predictionTexture.GetPixel(x, y);
// SDTV (BT.601) Y'UV conversion
yuvPixel.r = actualPixel.r * 0.299f + actualPixel.g * 0.587f + actualPixel.b * 0.114f; // Y' luma component
// Chrominance
// U = r * -0.14713 + g * -0.28886 + b * 0.436
yuvPixel.g = 0.0f;
// V = r * 0.615 + g * -0.51499 + b * -0.10001
yuvPixel.b = 0.0f;
predictionTexture.SetPixel(x, y, yuvPixel);
}
}
int pixelPos;
// Extract a portion of the camera image (half height)
int yOffset = 16; // Set to 0 for bottom half, _hiddenHeight for top half
int yHeight = _hiddenHeight;
for (int y = yOffset; y < yOffset + yHeight; y++)
{
for (int x = 0; x < _hiddenWidth; x++)
{
pixelPos = ((y - yOffset) * _hiddenWidth) + x;
_inputField[pixelPos] = predictionTexture.GetPixel(x, y).r;
}
}
if (useGaborFilter)
{
_openCV.GaborFilter(_inputField, 5, 4.0f, 0.0f, 10.0f, 0.5f, 0.0f);
Color tempPixel = new Color(0.0f, 0.0f, 0.0f);
for (int y = 0; y < yHeight; y++)
{
for (int x = 0; x < _hiddenWidth; x++)
{
pixelPos = (y * _hiddenWidth) + x;
tempPixel.r = _inputField[pixelPos];
tempPixel.g = tempPixel.r;
tempPixel.b = tempPixel.r;
predictionTexture.SetPixel(x, y + yHeight, tempPixel);
}
}
predictionTexture.Apply();
}
if (useThreholding)
{
//_openCV.Threshold(_inputField, 0.0f, 255.0f,
// eogmaneo.OpenCVInterop.CV_THRESH_TOZERO | eogmaneo.OpenCVInterop.CV_THRESH_OTSU);
_openCV.AdaptiveThreshold(_inputField, 255.0f,
eogmaneo.OpenCVInterop.CV_ADAPTIVE_THRESH_GAUSSIAN_C,
eogmaneo.OpenCVInterop.CV_THRESH_BINARY,
5, 2);
Color tempPixel = new Color(0.0f, 0.0f, 0.0f);
for (int y = 0; y < yHeight; y++)
{
for (int x = 0; x < _hiddenWidth; x++)
{
pixelPos = (y * _hiddenWidth) + x;
tempPixel.r = _inputField[pixelPos];
tempPixel.g = tempPixel.r;
tempPixel.b = tempPixel.r;
predictionTexture.SetPixel(x, y + yHeight, tempPixel);
}
}
predictionTexture.Apply();
}
if (useCanny)
{
_openCV.CannyEdgeDetection(_inputField, 50.0f, 50.0f * 3.0f);
Color tempPixel = new Color(0.0f, 0.0f, 0.0f);
for (int y = 0; y < yHeight; y++)
{
for (int x = 0; x < _hiddenWidth; x++)
{
pixelPos = (y * _hiddenWidth) + x;
tempPixel.r = _inputField[pixelPos];
tempPixel.g = tempPixel.r;
tempPixel.b = tempPixel.r;
predictionTexture.SetPixel(x, y + yHeight, tempPixel);
}
}
predictionTexture.Apply();
}
if (useSobel)
{
// Make sure that Sobel input and output uses a signed pixel data type,
// e.g. convert after to 8-bit unsigned
// sobelx64f = cv2.Sobel(img, cv2.CV_64F, 1, 0, ksize = 5)
// abs_sobel64f = np.absolute(sobelx64f)
// sobel_8u = np.uint8(abs_sobel64f)
Texture2D horzTexture = ConvolutionFilter.Apply(predictionTexture, ConvolutionFilter.Sobel3x3Horizontal);
Texture2D vertTexture = ConvolutionFilter.Apply(predictionTexture, ConvolutionFilter.Sobel3x3Vertical);
Texture2D convolvedTexture = new Texture2D(_inputWidth, _inputHeight, predictionTexture.format, false);
Color tempPixel = new Color(0.0f, 0.0f, 0.0f);
for (int y = yOffset; y < yOffset + yHeight; y++)
{
for (int x = 0; x < _hiddenWidth; x++)
{
Color horzPixel = horzTexture.GetPixel(x, y);
Color vertPixel = vertTexture.GetPixel(x, y);
tempPixel.r = Mathf.Sqrt((horzPixel.r * horzPixel.r) + (vertPixel.r * vertPixel.r));
tempPixel.g = tempPixel.r; // Mathf.Sqrt((horzPixel.g * horzPixel.g) + (vertPixel.g * vertPixel.g));
tempPixel.b = tempPixel.r; // Mathf.Sqrt((horzPixel.b * horzPixel.b) + (vertPixel.b * vertPixel.b));
convolvedTexture.SetPixel(x, (y - yOffset) + _hiddenHeight, tempPixel);
pixelPos = ((y - yOffset) * _hiddenWidth) + x;
_inputField[pixelPos] = (int)(tempPixel.r * 255.0f);
}
}
predictionTexture.SetPixels(convolvedTexture.GetPixels());
predictionTexture.Apply();
}
if (useLineSegmentDetector)
{
// Pass filtered image into the Line Segment Detector (optionally drawing found lines),
// and construct the rotation SDR for passing into the hierarchy
bool drawLines = true;
_openCV.LineSegmentDetector(_inputField, _hiddenWidth, _hiddenHeight, 6, _rotationSDR, drawLines);
if (drawLines)
{
// With drawLines enabled, the _inputField gets overriden with black background
// pixels and detected white lines drawn ontop.
// Transfer back into the predictionTexture for display (top half, bottom will show SDRs)
Color tempPixel = new Color(0.0f, 0.0f, 0.0f);
for (int y = yOffset; y < yOffset + yHeight; y++)
{
for (int x = 0; x < _hiddenWidth; x++)
{
pixelPos = ((y - yOffset) * _hiddenWidth) + x;
tempPixel.r = _inputField[pixelPos];
tempPixel.g = tempPixel.r;
tempPixel.b = tempPixel.r;
predictionTexture.SetPixel(x, (y - yOffset) + _hiddenHeight, tempPixel);
}
}
predictionTexture.Apply();
}
}
if (useFastFeatureDetector)
{
// Pass filtered image into the FAST Feature Detector (optionally drawing points found),
// and construct the feature SDR for passing into the hierarchy
bool drawPoints = true;
_openCV.FastFeatureDetector(_inputField, _hiddenWidth, _hiddenHeight, 6, _rotationSDR, drawPoints, 0, 1, true);
if (drawPoints)
{
// With drawPoints enabled, the _inputField gets overriden with black background
// pixels and detected white points drawn ontop.
// Transfer back into the predictionTexture for display (top half, bottom will show SDRs)
Color tempPixel = new Color(0.0f, 0.0f, 0.0f);
for (int y = yOffset; y < yOffset + yHeight; y++)
{
for (int x = 0; x < _hiddenWidth; x++)
{
pixelPos = ((y - yOffset) * _hiddenWidth) + x;
tempPixel.r = _inputField[pixelPos];
tempPixel.g = tempPixel.r;
tempPixel.b = tempPixel.r;
predictionTexture.SetPixel(x, (y - yOffset) + _hiddenHeight, tempPixel);
}
}
predictionTexture.Apply();
}
}
Color predictedPixel = new Color();
// Plot pre-encoder SDR output just underneath the input filtered image
int onState = 0;
for (int y = 16; y < 32; y++)
{
for (int x = 0; x < _inputWidth; x++)
{
if (x < _rotationSDR.Count)
{
predictedPixel.r = _rotationSDR[x];
if (y == 16)
{
onState += (int)predictedPixel.r;
}
}
else
{
predictedPixel.r = 0.0f;
}
predictedPixel.g = predictedPixel.r;
predictedPixel.b = predictedPixel.r;
predictionTexture.SetPixel(x, y, predictedPixel);
}
}
// Plot predicted SDR output at the bottom
int ccState = 0;
for (int y = 0; y < 16; y++)
{
for (int x = 0; x < _inputWidth; x++)
{
if (x < _rotationSDR.Count)
{
predictedPixel.r = _predictedSDR[x];
if (y == 0)
{
ccState += _rotationSDR[x] & _predictedSDR[x];
}
}
else
{
predictedPixel.r = 0.0f;
}
predictedPixel.g = predictedPixel.r;
predictedPixel.b = predictedPixel.r;
predictionTexture.SetPixel(x, y, predictedPixel);
}
}
predictionTexture.Apply();
_onStates.Add(onState);
_ccStates.Add(ccState);
// Trim lists?
if (_onStates.Count > _maxNumStates)
{
_onStates.RemoveAt(0);
_ccStates.RemoveAt(0);
}
NCC = 0.0f;
for (int i = 0; i < _onStates.Count; i++)
{
if (_ccStates[i] == 0 && _onStates[i] == 0)
{
NCC += 1.0f;
}
else if (_onStates[i] == 0)
{
NCC += 1.0f;
}
else
{
NCC += (float)_ccStates[i] / (float)_onStates[i];
}
}
NCC /= (float)_onStates.Count;
// Encode scalar values from the car controller
Steer = carController.CurrentSteerAngle / carController.m_MaximumSteerAngle;
Accel = carController.AccelInput;
Brake = carController.BrakeInput;
//for (int i = 0; i < 6 * 6; i++)
// _inputValues[i] = 0;
//int index = (int)((Steer * 0.5f + 0.5f) * (6 * 6 - 1) + 0.5f);
//_inputValues[index] = 1;
_inputValues[0] = (int)((Steer * 0.5f + 0.5f) * (6.0f * 6.0f - 1.0f) + 0.5f);
// Setup the hierarchy input vector
Std2DVeci input = new Std2DVeci();
input.Add(_rotationSDR);
input.Add(_inputValues);
// Step the hierarchy
_hierarchy.step(input, _system, Training);
StdVeci predictions = _hierarchy.getPrediction(0);
for (int i = 0; i < _predictedSDR.Count; i++)
{
_predictedSDR[i] = predictions[i];
}
// Wait for physics to settle
if (_time < 1.0f)
{
_time += Time.deltaTime;
// Apply hand brake
carSteer = 0.0f;
carAccel = 0.0f;
carBrake = -1.0f;
HandBrake = 1.0f;
}
else
{
// Release hand brake
HandBrake = 0.0f;
Accel = -1.0f;
Brake = Accel;
// Update the car controller
StdVeci steeringPredictions = _hierarchy.getPrediction(1);
//int maxIndex = 0;
//for (int i = 1; i < 6 * 6; i++)
// if (steeringPredictions[i] > steeringPredictions[maxIndex])
// maxIndex = i;
//PredictedSteer = (float)(maxIndex) / (float)(6 * 6 - 1) * 2.0f - 1.0f;
PredictedSteer = (steeringPredictions[0] / (6.0f * 6.0f - 1.0f)) * 2.0f - 1.0f;
PredictedAccel = Accel;
PredictedBrake = Brake;
carSteer = PredictedSteer;
carAccel = PredictedAccel;
carBrake = PredictedBrake;
// Search along the spline for the closest point to the current car position
float bestT = 0.0f, minDistance = 100000.0f;
Vector3 carPosition = carController.gameObject.transform.localPosition;
// When not training use the track spline
BezierSpline spline = trackSpline;
if (Training)
{
spline = splineList[SplineIndex];
}
float totalDistance = 0.0f;
for (float t = 0.0f; t <= 1.0f; t += 0.001f)
{
Vector3 position = spline.GetPoint(t);
Vector3 positionPrev = spline.GetPoint(t - 0.001f);
float distance = Vector3.Distance(position, carPosition);
totalDistance += Vector3.Distance(position, positionPrev);
if (distance <= minDistance)
{
minDistance = distance;
bestT = t;
}
}
// Assume +-2 units is maximum distance the car is allowed to be from the center spline
NCC = Mathf.Max(0.0f, NCC - (1.0f - ((2.0f - Vector3.Distance(carPosition, spline.GetPoint(bestT))) / 2.0f)));
//NCC = ((2.0f - Vector3.Distance(carPosition, spline.GetPoint(bestT))) / 2.0f);
// Reset car position and direction?
if (Input.GetKeyUp(KeyCode.R) || carController.Collided)
{
if (ForcePredictionMode == false)
{
Training = true;
}
carController.ResetCollided();
// Spline 0 is usually set as the spline used to create the track
SplineIndex = 0;
Vector3 position = spline.GetPoint(bestT);
position.y = carController.gameObject.transform.localPosition.y;
carController.gameObject.transform.localPosition = position;
Vector3 splineDirection = spline.GetDirection(bestT).normalized;
carController.gameObject.transform.forward = -splineDirection;
}
// Toggle training on iff too divergent?
if (Training == false && ForcePredictionMode == false && NCC < 0.25f)
{
Training = true;
}
// Toggle training off iff quite confident?
if (Training == true && NCC > 0.85f && LapCount >= initialTrainingLaps)
{
Training = false;
}
if (carController.CurrentSpeed < 2.0f)
{
Training = true;
}
if (Training)
{
_trainingCount++;
}
else
{
_predictingCount++;
}
if (Training && spline != null)
{
Vector3 carDirection = -carController.gameObject.transform.forward.normalized;
Vector3 targetPosition = spline.GetPoint(bestT + (SteerAhead / totalDistance));
//Vector3 splineDirection = spline.GetDirection(bestT).normalized;
Vector3 targetDirection = (targetPosition - carPosition).normalized;
float angle = (1.0f - Vector3.Dot(carDirection, targetDirection));// * Mathf.Rad2Deg;
Vector3 right = Vector3.Cross(carDirection, Vector3.up);
float angle2 = Vector3.Dot(right, targetDirection);
float newCarSteer = Mathf.Exp(256.0f * angle) - 1.0f;
if (Mathf.Abs(minDistance) > 0.01f)//newCarSteer > Mathf.PI / 64.0f)
{
newCarSteer += angle2 * Mathf.Abs(minDistance);
}
if (angle2 > 0.0f)
{
newCarSteer = -newCarSteer;
}
if (newCarSteer > 1.0f)
{
newCarSteer = 1.0f;
}
else
if (newCarSteer < -1.0f)
{
newCarSteer = -1.0f;
}
float steerBlend = 0.5f;
carSteer = (steerBlend * newCarSteer) + ((1.0f - steerBlend) * carSteer);
if (enableDebugLines)
{
debugLinePositions[0] = carController.gameObject.transform.localPosition;
debugLinePositions[1] = debugLinePositions[0] + carDirection * 10.0f;
debugLinePositions[2] = carController.gameObject.transform.localPosition;
debugLinePositions[3] = debugLinePositions[2] + targetDirection * 10.0f;
debugLine.SetPositions(debugLinePositions);
}
}
float totalCount = _trainingCount + _predictingCount;
if (totalCount == 0.0f)
{
TrainingPercent = 1.0f;
PredictionPercent = 0.0f;
}
else
{
TrainingPercent = (float)_trainingCount / totalCount;
PredictionPercent = (float)_predictingCount / totalCount;
}
if (bestT < prevBestT)
{
LapCount++;
_trainingCount = 0;
_predictingCount = 0;
if ((LapCount % lapsPerSpline) == 0)
{
SplineIndex++;
if (SplineIndex >= splineList.Length)
{
SplineIndex = 0;
}
}
}
prevBestT = bestT;
}
if (connectToNeoVis && _neoVis != null)
{
_neoVis.update(0.01f);
}
if (userControl)
{
// Control overides
// pass the input to the car!
float h = CrossPlatformInputManager.GetAxis("Horizontal");
float v = CrossPlatformInputManager.GetAxis("Vertical");
#if !MOBILE_INPUT
float handbrake = CrossPlatformInputManager.GetAxis("Jump");
#endif
carSteer = h;
carAccel = v;
carBrake = v;
HandBrake = handbrake;
}
// Toggle training mode?
if (Input.GetKeyUp(KeyCode.T))
{
Training = !Training;
ForcePredictionMode = false;
}
else
// Force prediction mode?
if (Input.GetKeyUp(KeyCode.F))
{
Training = false;
ForcePredictionMode = true;
}
// Save out the current state of the hierarchy?
if (Input.GetKeyUp(KeyCode.O) && hierarchyFileName.Length > 0)
{
_hierarchy.save(hierarchyFileName);
print("Saved OgmaNeo hierarchy to " + hierarchyFileName);
}
}
// Use this for initialization
void Start()
{
_time = 0.0f;
LapCount = 0;
if (carController == null)
{
return;
}
print("Initializing EOgmaNeo");
_system = new eogmaneo.ComputeSystem(4, 1111);
_inputWidth = cameraTexture.width;
_inputHeight = cameraTexture.height;
print("Capture image size: " + _inputWidth + "x" + _inputHeight);
_hiddenWidth = _inputWidth;
_hiddenHeight = _inputHeight / 2;
// Y' input field
_inputField = new eogmaneo.StdVecf(_hiddenWidth * _hiddenHeight);
for (int i = 0; i < _hiddenWidth * _hiddenHeight; i++)
{
_inputField.Add(0.0f);
}
// Scalar values input field
_inputValues = new eogmaneo.StdVeci(1);
_inputValues.Add(0);
print("Constructing hierarchy");
// Hierarchy layer descriptions
const int layerSize = 36;
const int numLayers = 3;
StdVecLayerDesc lds = new StdVecLayerDesc(numLayers);
for (int l = 0; l < numLayers; l++)
{
lds.Add(new LayerDesc());
lds[l]._width = layerSize;
lds[l]._height = layerSize;
lds[l]._chunkSize = 6;
lds[l]._forwardRadius = 9;
lds[l]._backwardRadius = 9;
lds[l]._ticksPerUpdate = 2;
lds[l]._temporalHorizon = 2;
lds[l]._alpha = 0.065f;
lds[l]._beta = 0.1f;
// Disable reinforcement learning
lds[l]._delta = 0.0f;
}
// Encoder output sizes, as input sizes to hierarchy
StdVecPairi inputSizes = new eogmaneo.StdVecPairi();
inputSizes.Add(new StdPairi(_hiddenWidth, _hiddenHeight));
inputSizes.Add(new StdPairi(6, 6));
StdVeci inputChunkSizes = new StdVeci();
inputChunkSizes.Add(6);
inputChunkSizes.Add(6);
// Whether to predict an input
StdVecb predictInputs = new StdVecb();
predictInputs.Add(true);
predictInputs.Add(true);
print("Generating hierarchy");
// Generate the hierarchy
_hierarchy = new eogmaneo.Hierarchy();
_hierarchy.create(inputSizes, inputChunkSizes, predictInputs, lds, 41);
if (reloadHierarchy && hierarchyFileName.Length > 0)
{
_hierarchy.load(hierarchyFileName);
print("Reloaded OgmaNeo hierarchy from " + hierarchyFileName);
}
if (connectToNeoVis)
{
_neoVis = new eogmaneo.VisAdapter();
_neoVis.create(_hierarchy, neoVisPort);
}
_openCV = new eogmaneo.OpenCVInterop();
int numSDRbits = (_hiddenWidth / inputChunkSizes[0]) * (_hiddenHeight / inputChunkSizes[1]);
_rotationSDR = new StdVeci(numSDRbits);
_predictedSDR = new StdVeci(numSDRbits);
for (int i = 0; i < numSDRbits; i++)
{
_rotationSDR.Add(0);
_predictedSDR.Add(0);
}
// For calculating the normalized cross-corrolation
_ccStates = new List <int>();
_onStates = new List <int>();
NCC = 0.0f;
_trainingCount = 0;
_predictingCount = 0;
TrainingPercent = 1.0f;
PredictionPercent = 0.0f;
Training = false;
ForcePredictionMode = false;
if (splineList != null && splineList[SplineIndex] != null)
{
Vector3 position = splineList[SplineIndex].GetPoint(0.0f);
Vector3 splineDirection = splineList[SplineIndex].GetDirection(0.0f);
carController.gameObject.transform.localPosition = position;
carController.gameObject.transform.LookAt(position - splineDirection);
Training = true;
}
if (enableDebugLines)
{
debugLine = gameObject.AddComponent <LineRenderer>();
debugLine.startWidth = 0.1f;
debugLine.endWidth = 0.1f;
debugLinePositions = new Vector3[4];
debugLine.numPositions = debugLinePositions.Length;
}
PredictedSteer = 0.0f;
PredictedAccel = 0.0f;
PredictedBrake = -1.0f;
HandBrake = 1.0f;
carSteer = PredictedSteer;
carAccel = PredictedAccel;
carBrake = PredictedBrake;
}
public StdVeci getFeedBackPrev(int f)
{
StdVeci ret = new StdVeci(eogmaneoPINVOKE.Layer_getFeedBackPrev(swigCPtr, f), true);
return(ret);
}
public void FastFeatureDetector(StdVecf data, int width, int height, int chunkSize, StdVeci featuresSDR, bool drawKeypoints, int threshold)
{
eogmaneoPINVOKE.OpenCVInterop_FastFeatureDetector__SWIG_2(swigCPtr, StdVecf.getCPtr(data), width, height, chunkSize, StdVeci.getCPtr(featuresSDR), drawKeypoints, threshold);
if (eogmaneoPINVOKE.SWIGPendingException.Pending)
{
throw eogmaneoPINVOKE.SWIGPendingException.Retrieve();
}
}
public StdVeci getHiddenStates(int order)
{
StdVeci ret = new StdVeci(eogmaneoPINVOKE.CornerEncoder_getHiddenStates(swigCPtr, order), false);
return(ret);
}