/// <summary>
/// THis is implemented for ISupervisedLearingModel so that this model can also be used for TrainerMimic
/// </summary>
/// <param name="vectorObservation"></param>
/// <param name="visualObservation"></param>
/// <returns>(mean, var) var will be null for discrete</returns>
public ValueTuple <float[, ], float[, ]> EvaluateAction(float[,] vectorObservation, List <float[, , , ]> visualObservation, List <float[, ]> actionsMask)
{
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);
}
if (ActionSpace == SpaceType.discrete)
{
int batchSize = vectorObservation != null?vectorObservation.GetLength(0) : visualObservation[0].GetLength(0);
int branchSize = ActionSizes.Length;
List <float[, ]> masks = actionsMask;
//create all 1 mask if the input mask is null.
if (masks == null)
{
masks = CreateDummyMasks(ActionSizes, batchSize);
}
inputLists.AddRange(masks);
}
var result = PolicyFunction.Call(inputLists);
float[,] actions = ((float[, ])result[0].eval());
float[,] outputVar = null;
if (SLHasVar)
{
object varTemp = result[1].eval();
if (varTemp is float[, ])
{
outputVar = (float[, ])result[1].eval();
}
else
{
outputVar = null;
}
}
//normlaized the input observations in every calll of eval action
if (useInputNormalization && HasVectorObservation)
{
UpdateNormalizerFunction.Call(new List <Array>()
{
vectorObservation
});
}
return(ValueTuple.Create(actions, outputVar));
}
/// <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(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 = PolicyFunction.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 based on curren states. The first dimension of the array must be batch dimension
/// </summary>
/// <param name="vectorObservation">current vector states. Can be batch input</param>
/// <param name="actionProbs">output actions' probabilities. note that it is the normalized log probability</param>
/// <param name="actoinsMask">action mask for discrete action. </param>
/// <returns></returns>
public virtual float[,] EvaluateAction(float[,] vectorObservation, out float[,] actionProbs, List <float[, , , ]> visualObservation, List <float[, ]> actionsMask = null)
{
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);
}
float[,] actions = null;
actionProbs = null;
if (ActionSpace == SpaceType.continuous)
{
var result = PolicyFunction.Call(inputLists);
actions = ((float[, ])result[0].eval());
actionProbs = ((float[, ])result[1].eval());
}
else if (ActionSpace == SpaceType.discrete)
{
int batchSize = vectorObservation != null?vectorObservation.GetLength(0) : visualObservation[0].GetLength(0);
int branchSize = ActionSizes.Length;
List <float[, ]> masks = actionsMask;
//create all 1 mask if the input mask is null.
if (masks == null)
{
masks = CreateDummyMasks(ActionSizes, batchSize);
}
inputLists.AddRange(masks);
var result = PolicyFunction.Call(inputLists);
actions = ((float[, ])result[0].eval());
//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])];
}
}
}
//normlaized the input observations in every calll of eval action
if (useInputNormalization && HasVectorObservation)
{
UpdateNormalizerFunction.Call(new List <Array>()
{
vectorObservation
});
}
return(actions);
}
