public virtual float[] TrainBatch(float[,] vectorObservations, List <float[, , , ]> visualObservations, float[,] actions, float[,] actionProbs, float[] targetValues, float[] oldValues, float[] advantages)
{
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 TrainBatch()");
List <Array> inputs = new List <Array>();
if (vectorObservations != null)
{
inputs.Add(vectorObservations);
}
if (visualObservations != null)
{
inputs.AddRange(visualObservations);
}
if (ActionSpace == SpaceType.continuous)
{
inputs.Add(actions);
}
else if (ActionSpace == SpaceType.discrete)
{
int[,] actionsInt = actions.Convert(t => Mathf.RoundToInt(t));
inputs.Add(actionsInt);
}
inputs.Add(actionProbs);
inputs.Add(targetValues);
inputs.Add(oldValues);
inputs.Add(advantages);
inputs.Add(new float[] { ClipEpsilon });
inputs.Add(new float[] { ValueLossWeight });
inputs.Add(new float[] { EntropyLossWeight });
var loss = UpdatePPOFunction.Call(inputs);
var result = new float[] { (float)loss[0].eval(), (float)loss[1].eval(), (float)loss[2].eval(), (float)loss[3].eval() };
//float[,] outActionProbs = (float[,])loss[4].eval();
return(result);
//Debug.LogWarning("test save graph");
//((UnityTFBackend)K).ExportGraphDef("SavedGraph/PPOTest.pb");
//return new float[] { 0, 0, 0 }; //test for memeory allocation
}
public virtual float[] TrainBatch(float[,] vectorObservations, List <float[, , , ]> visualObservations, float[,] actions, float[,] actionProbs, float[] targetValues, float[] oldValues, float[] advantages, 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 TrainBatch()");
List <Array> inputs = new List <Array>();
inputs.Add(actionProbs);
inputs.Add(targetValues);
inputs.Add(oldValues);
inputs.Add(advantages);
inputs.Add(new float[] { ClipEpsilon });
inputs.Add(new float[] { ClipValueLoss });
inputs.Add(new float[] { ValueLossWeight });
inputs.Add(new float[] { EntropyLossWeight });
if (vectorObservations != null)
{
inputs.Add(vectorObservations);
}
if (visualObservations != null)
{
inputs.AddRange(visualObservations);
}
if (ActionSpace == SpaceType.continuous)
{
inputs.Add(actions);
}
else if (ActionSpace == SpaceType.discrete)
{
inputs.AddRange(actionsMask);
int[,] actionsInt = actions.Convert(t => Mathf.RoundToInt(t));
inputs.Add(actionsInt);
}
var loss = UpdatePPOFunction.Call(inputs);
var result = new float[] { (float)loss[0].eval(), (float)loss[1].eval(), (float)loss[2].eval(), (float)loss[3].eval() };
return(result);
}
