/// <summary> /// Query actions' probabilities based on curren states. The first dimension of the array must be batch dimension. Note that it is the normalized log probability /// </summary> public virtual float[,] EvaluateProbability(float[,] vectorObservation, float[,] actions, List <float[, , , ]> visualObservation, List <float[, ]> actionsMask = null) { Debug.Assert(mode == Mode.PPO, "This method is for PPO mode only"); Debug.Assert(TrainingEnabled == true, "The model needs to initalized with Training enabled to use EvaluateProbability()"); List <Array> inputLists = new List <Array>(); if (HasVectorObservation) { Debug.Assert(vectorObservation != null, "Must Have vector observation inputs!"); inputLists.Add(vectorObservation); } if (HasVisualObservation) { Debug.Assert(visualObservation != null, "Must Have visual observation inputs!"); inputLists.AddRange(visualObservation); } var actionProbs = new float[actions.GetLength(0), ActionSpace == SpaceType.continuous ? actions.GetLength(1) : 1]; if (ActionSpace == SpaceType.continuous) { inputLists.Add(actions); var result = ActionProbabilityFunction.Call(inputLists); actionProbs = ((float[, ])result[0].eval()); } else if (ActionSpace == SpaceType.discrete) { List <float[, ]> masks = actionsMask; int batchSize = vectorObservation.GetLength(0); int branchSize = ActionSizes.Length; //create all 1 mask if the input mask is null. if (masks == null) { masks = CreateDummyMasks(ActionSizes, batchSize); } inputLists.AddRange(masks); var result = ActionFunction.Call(inputLists); //get the log probabilities actionProbs = new float[batchSize, branchSize]; for (int b = 0; b < branchSize; ++b) { var tempProbs = ((float[, ])result[b + 1].eval()); int actSize = ActionSizes[b]; for (int i = 0; i < batchSize; ++i) { actionProbs[i, b] = tempProbs[i, Mathf.RoundToInt(actions[i, b])]; } } } return(actionProbs); }
/// <summary> /// Query actions' probabilities based on curren states. The first dimension of the array must be batch dimension /// </summary> public virtual float[,] EvaluateProbability(float[,] vectorObservation, float[,] actions, List <float[, , , ]> visualObservation) { Debug.Assert(mode == Mode.PPO, "This method is for PPO mode only"); Debug.Assert(TrainingEnabled == true, "The model needs to initalized with Training enabled to use EvaluateProbability()"); List <Array> inputLists = new List <Array>(); if (HasVectorObservation) { Debug.Assert(vectorObservation != null, "Must Have vector observation inputs!"); inputLists.Add(vectorObservation); } if (HasVisualObservation) { Debug.Assert(visualObservation != null, "Must Have visual observation inputs!"); inputLists.AddRange(visualObservation); } var actionProbs = new float[actions.GetLength(0), ActionSpace == SpaceType.continuous ? actions.GetLength(1) : 1]; if (ActionSpace == SpaceType.continuous) { inputLists.Add(actions); var result = ActionProbabilityFunction.Call(inputLists); actionProbs = ((float[, ])result[0].eval()); } else if (ActionSpace == SpaceType.discrete) { var result = ActionFunction.Call(inputLists); var outputAction = ((float[, ])result[0].eval()); for (int j = 0; j < outputAction.GetLength(0); ++j) { actionProbs[j, 0] = outputAction.GetRow(j)[Mathf.RoundToInt(actions[j, 0])]; } } return(actionProbs); }