public static void ResizeTensor(TensorProxy tensor, int batch, ITensorAllocator allocator)
{
if (tensor.shape[0] == batch &&
tensor.data != null && tensor.data.batch == batch)
{
return;
}
tensor.data?.Dispose();
tensor.shape[0] = batch;
if (tensor.shape.Length == 4)
{
tensor.data = allocator.Alloc(
new TensorShape(
batch,
(int)tensor.shape[1],
(int)tensor.shape[2],
(int)tensor.shape[3]));
}
else
{
tensor.data = allocator.Alloc(
new TensorShape(
batch,
(int)tensor.shape[tensor.shape.Length - 1]));
}
}
public void Apply(TensorProxy tensorProxy, IEnumerable <AgentIdActionPair> actions)
{
//var tensorDataProbabilities = tensorProxy.Data as float[,];
var idActionPairList = actions as List <AgentIdActionPair> ?? actions.ToList();
var batchSize = idActionPairList.Count;
var actionValues = new float[batchSize, m_ActionSize.Length];
var startActionIndices = Utilities.CumSum(m_ActionSize);
for (var actionIndex = 0; actionIndex < m_ActionSize.Length; actionIndex++)
{
var nBranchAction = m_ActionSize[actionIndex];
var actionProbs = new TensorProxy()
{
valueType = TensorProxy.TensorType.FloatingPoint,
shape = new long[] { batchSize, nBranchAction },
data = m_Allocator.Alloc(new TensorShape(batchSize, nBranchAction))
};
for (var batchIndex = 0; batchIndex < batchSize; batchIndex++)
{
for (var branchActionIndex = 0;
branchActionIndex < nBranchAction;
branchActionIndex++)
{
actionProbs.data[batchIndex, branchActionIndex] =
tensorProxy.data[batchIndex, startActionIndices[actionIndex] + branchActionIndex];
}
}
var outputTensor = new TensorProxy()
{
valueType = TensorProxy.TensorType.FloatingPoint,
shape = new long[] { batchSize, 1 },
data = m_Allocator.Alloc(new TensorShape(batchSize, 1))
};
Eval(actionProbs, outputTensor, m_Multinomial);
for (var ii = 0; ii < batchSize; ii++)
{
actionValues[ii, actionIndex] = outputTensor.data[ii, 0];
}
actionProbs.data.Dispose();
outputTensor.data.Dispose();
}
var agentIndex = 0;
foreach (var idActionPair in idActionPairList)
{
var actionVal = new float[m_ActionSize.Length];
for (var j = 0; j < m_ActionSize.Length; j++)
{
actionVal[j] = actionValues[agentIndex, j];
}
idActionPair.action.Invoke(new AgentAction {
vectorActions = actionVal
});
agentIndex++;
}
}
public void Apply(TensorProxy tensorProxy, Dictionary <Agent, AgentInfo> agentInfo)
{
//var tensorDataProbabilities = tensorProxy.Data as float[,];
var batchSize = agentInfo.Keys.Count;
var actions = new float[batchSize, m_ActionSize.Length];
var startActionIndices = Utilities.CumSum(m_ActionSize);
for (var actionIndex = 0; actionIndex < m_ActionSize.Length; actionIndex++)
{
var nBranchAction = m_ActionSize[actionIndex];
var actionProbs = new TensorProxy()
{
valueType = TensorProxy.TensorType.FloatingPoint,
shape = new long[] { batchSize, nBranchAction },
data = m_Allocator.Alloc(new TensorShape(batchSize, nBranchAction))
};
for (var batchIndex = 0; batchIndex < batchSize; batchIndex++)
{
for (var branchActionIndex = 0;
branchActionIndex < nBranchAction;
branchActionIndex++)
{
actionProbs.data[batchIndex, branchActionIndex] =
tensorProxy.data[batchIndex, startActionIndices[actionIndex] + branchActionIndex];
}
}
var outputTensor = new TensorProxy()
{
valueType = TensorProxy.TensorType.FloatingPoint,
shape = new long[] { batchSize, 1 },
data = m_Allocator.Alloc(new TensorShape(batchSize, 1))
};
Eval(actionProbs, outputTensor, m_Multinomial);
for (var ii = 0; ii < batchSize; ii++)
{
actions[ii, actionIndex] = outputTensor.data[ii, 0];
}
actionProbs.data.Dispose();
outputTensor.data.Dispose();
}
var agentIndex = 0;
foreach (var agent in agentInfo.Keys)
{
var action = new float[m_ActionSize.Length];
for (var j = 0; j < m_ActionSize.Length; j++)
{
action[j] = actions[agentIndex, j];
}
agent.UpdateVectorAction(action);
agentIndex++;
}
}
public void Generate(TensorProxy tensorProxy, int batchSize, IEnumerable <AgentInfoSensorsPair> infos)
{
tensorProxy.shape = new long[0];
tensorProxy.data?.Dispose();
tensorProxy.data = m_Allocator.Alloc(new TensorShape(1, 1));
tensorProxy.data[0] = 1;
}
public void Generate(TensorProxy tensorProxy, int batchSize, Dictionary <Agent, AgentInfo> agentInfo)
{
tensorProxy.shape = new long[0];
tensorProxy.data?.Dispose();
tensorProxy.data = m_Allocator.Alloc(new TensorShape(1, 1));
tensorProxy.data[0] = 1;
}
// @TODO: choose approach to handle case when tensors after Flatten/Reshape are written into OR taken ownership of
// 1) owns data, copy on PrepareCacheForAccess() and PinForWrite()
// 2) always copy data in Flatten()/Reshape(), remove from Tensor interface
// 2) always copy data in Flatten()/Reshape(), implement ICloneable for GPU ITensorData
/// <summary>
/// Create a flattened copy of the current Tensor ie of shape [B,1,1,H*W*CH]
/// </summary>
public Tensor Flatten()
{
var newShape = shape.Flatten();
Tensor copy;
if (m_TensorAllocator != null)
{
copy = m_TensorAllocator.Alloc(newShape, m_TensorOnDevice);
}
else
{
copy = new Tensor(newShape, m_TensorOnDevice);
}
copy.name = $"flatten of {name}";
copy.m_Cache = m_Cache;
copy.m_CacheIsDirty = m_CacheIsDirty;
return(copy);
}
// @TODO: choose approach to handle case when tensors after Flatten/Reshape are written into OR taken ownership of
// 1) owns data, copy on PrepareCacheForAccess() and PinForWrite()
// 2) always copy data in Flatten()/Reshape(), remove from Tensor interface
// 2) always copy data in Flatten()/Reshape(), implement ICloneable for GPU ITensorData
private Tensor ShallowCopy(TensorShape newShape, string newName)
{
Tensor copy;
if (m_TensorAllocator != null)
{
copy = m_TensorAllocator.Alloc(newShape, m_TensorOnDevice);
}
else
{
copy = new Tensor(newShape, m_TensorOnDevice);
}
copy.name = newName;
copy.m_Cache = m_Cache;
copy.m_CacheIsDirty = m_CacheIsDirty;
return(copy);
}
public void Generate(TensorProxy tensorProxy, int batchSize, IEnumerable <Agent> agents)
{
tensorProxy.data?.Dispose();
tensorProxy.data = m_Allocator.Alloc(new TensorShape(1, 1));
tensorProxy.data[0] = batchSize;
}
public void Generate(TensorProxy tensorProxy, int batchSize, IList <AgentInfoSensorsPair> infos)
{
tensorProxy.data?.Dispose();
tensorProxy.data = m_Allocator.Alloc(new TensorShape(1, 1));
tensorProxy.data[0] = batchSize;
}
public void Apply(TensorProxy tensorProxy, IEnumerable <int> actionIds, Dictionary <int, float[]> lastActions)
{
//var tensorDataProbabilities = tensorProxy.Data as float[,];
var idActionPairList = actionIds as List <int> ?? actionIds.ToList();
var batchSize = idActionPairList.Count;
var actionValues = new float[batchSize, m_ActionSize.Length];
var startActionIndices = Utilities.CumSum(m_ActionSize);
for (var actionIndex = 0; actionIndex < m_ActionSize.Length; actionIndex++)
{
var nBranchAction = m_ActionSize[actionIndex];
var actionProbs = new TensorProxy()
{
valueType = TensorProxy.TensorType.FloatingPoint,
shape = new long[] { batchSize, nBranchAction },
data = m_Allocator.Alloc(new TensorShape(batchSize, nBranchAction))
};
for (var batchIndex = 0; batchIndex < batchSize; batchIndex++)
{
for (var branchActionIndex = 0;
branchActionIndex < nBranchAction;
branchActionIndex++)
{
actionProbs.data[batchIndex, branchActionIndex] =
tensorProxy.data[batchIndex, startActionIndices[actionIndex] + branchActionIndex];
}
}
var outputTensor = new TensorProxy()
{
valueType = TensorProxy.TensorType.FloatingPoint,
shape = new long[] { batchSize, 1 },
data = m_Allocator.Alloc(new TensorShape(batchSize, 1))
};
Eval(actionProbs, outputTensor, m_Multinomial);
for (var ii = 0; ii < batchSize; ii++)
{
actionValues[ii, actionIndex] = outputTensor.data[ii, 0];
}
actionProbs.data.Dispose();
outputTensor.data.Dispose();
}
var agentIndex = 0;
foreach (int agentId in actionIds)
{
if (lastActions.ContainsKey(agentId))
{
var actionVal = lastActions[agentId];
if (actionVal == null)
{
actionVal = new float[m_ActionSize.Length];
lastActions[agentId] = actionVal;
}
for (var j = 0; j < m_ActionSize.Length; j++)
{
actionVal[j] = actionValues[agentIndex, j];
}
}
agentIndex++;
}
}
public void Generate(TensorProxy tensorProxy, int batchSize, Dictionary <Agent, AgentInfo> agentInfo)
{
tensorProxy.Data = _allocator.Alloc(new TensorShape(1, 1));
tensorProxy.Data[0] = batchSize;
}
