public override void Forward(Executor executor)
{
var ctx = executor.Context;
var x = executor.GetTensor(Input);
var y = executor.GetTensor(Output, x.Shape);
if (ctx.Type == ContextType.Gpu && x.Layout.IsInnerChangeMostFullyPacked)
{
var dnn = ctx.ToGpuContext().Dnn;
var n = (int)x.Shape[0];
var classes = (int)x.Shape[1];
using (var xDesc = executor.TensorDescRepo.Acquire())
using (var yDesc = executor.TensorDescRepo.Acquire())
{
xDesc.Value.SetND(Dnn.DataTypeOf(typeof(T)), new[] { n, classes, 1, 1 }, new[] { classes, 1, 1, 1 });
yDesc.Value.SetND(Dnn.DataTypeOf(typeof(T)), new[] { n, classes, 1, 1 }, new[] { classes, 1, 1, 1 });
var xPtr = x.Buffer.Ptr;
var yPtr = y.Buffer.Ptr;
var alpha = ScalarOps.Conv <T>(1.0);
var beta = ScalarOps.Conv <T>(0.0);
const SoftmaxAlgorithm algorithm = SoftmaxAlgorithm.ACCURATE;
const SoftmaxMode mode = SoftmaxMode.INSTANCE;
dnn.SoftmaxForward(algorithm, mode, alpha, xDesc.Value, xPtr, beta, yDesc.Value, yPtr);
}
return;
}
throw new NotImplementedException();
}
public override void Forward(Executor executor)
{
var data = executor.GetTensor(Data);
var output = executor.GetTensor(Output, Shape.Create(data.Shape[0], Output.Shape[1], Output.Shape[2], Output.Shape[3]));
if (executor.Context.Type == ContextType.Gpu)
{
var dnn = executor.Context.ToGpuContext().Dnn;
using (var dataDescRcpt = executor.TensorDescRepo.Acquire())
using (var outputDescRcpt = executor.TensorDescRepo.Acquire())
{
var dataDesc = dataDescRcpt.Value;
var outputDesc = outputDescRcpt.Value;
var dataType = Dnn.DataTypeOf <T>();
dataDesc.Set4D(dataType, TensorFormat.CUDNN_TENSOR_NCHW, (int)data.Shape[0], (int)data.Shape[1], (int)data.Shape[2], (int)data.Shape[3]);
outputDesc.Set4D(dataType, TensorFormat.CUDNN_TENSOR_NCHW, (int)data.Shape[0], (int)Output.Shape[1], (int)Output.Shape[2], (int)Output.Shape[3]);
dnn.PoolingForward(Descriptor, ScalarOps.Conv <T>(1.0), dataDesc, data.Buffer.Ptr, ScalarOps.Conv <T>(0.0), outputDesc, output.Buffer.Ptr);
return;
}
}
throw new NotImplementedException();
}
public void Backward(Executor executor)
{
var context = executor.Context.ToGpuContext();
var dnn = context.Dnn;
var rnnDesc = RnnDesc;
var filterDesc = WDesc;
Util.EnsureTrue(IsTraining);
dnn.RNNBackwardData(
rnnDesc,
1,
YDesc,
Output.Buffer.Ptr,
YDesc,
DOutput.Buffer.Ptr,
StateDesc,
DHY.Buffer.Ptr,
StateDesc,
DCY.Buffer.Ptr,
filterDesc,
executor.GetTensor(W).Buffer.Ptr,
StateDesc,
HX.Buffer.Ptr,
StateDesc,
CX.Buffer.Ptr,
XDesc,
DInput.Buffer.Ptr,
StateDesc,
DHX.Buffer.Ptr,
StateDesc,
DCX.Buffer.Ptr,
Workspace.Buffer.Ptr,
(IntPtr)Workspace.Shape.Length,
ReserveSpace.Buffer.Ptr,
(IntPtr)ReserveSpace.Shape.Length);
if (executor.IncreaseGradientAggregationCounter(W) == 0)
{
executor.AssignGradient(W, ScalarOps.Conv <T>(0.0).AsScalar(), replace: true);
}
dnn.RNNBackwardWeights(
rnnDesc,
1,
XDesc,
Input.Buffer.Ptr,
StateDesc,
HX.Buffer.Ptr,
YDesc,
Output.Buffer.Ptr,
Workspace.Buffer.Ptr,
(IntPtr)Workspace.Shape.Length,
WDesc,
executor.GetGradient(W).Buffer.Ptr,
ReserveSpace.Buffer.Ptr,
(IntPtr)ReserveSpace.Shape.Length);
}
public override void InitBias <T>(Context ctx, int layerId, int linLayerId, Tensor <T> tensor)
{
// cuDNN LSTM layout is: IFAO, bias has 2x4, we ignore the second set of bias
// so only the first forget bias needs to be set, which is linLayerId = 1
if (linLayerId == 1)
{
ctx.Assign(tensor, ScalarOps.Conv <T>(ForgetBiasInit));
}
else
{
ctx.Assign(tensor, ScalarOps.Conv <T>(0.0));
}
}
public override void Forward(Executor executor)
{
var z = executor.GetTensor(Input);
var y = executor.GetTensor(Label);
Util.EnsureTrue(z.Shape.Rank == 2);
Util.EnsureTrue(Dnn.IsAvailable, "TODO: make non-cuDnn implementation.");
var n = (int)z.Shape[0];
var classes = (int)z.Shape[1];
using (var xDesc = executor.TensorDescRepo.Acquire())
using (var yDesc = executor.TensorDescRepo.Acquire())
{
var dnn = executor.Context.ToGpuContext().Dnn;
xDesc.Value.SetND(Dnn.DataTypeOf(typeof(T)), new[] { n, classes, 1, 1 }, new[] { classes, 1, 1, 1 });
yDesc.Value.SetND(Dnn.DataTypeOf(typeof(T)), new[] { n, classes, 1, 1 }, new[] { classes, 1, 1, 1 });
var xPtr = executor.GetTensor(Input).Buffer.Ptr;
var yPtr = executor.GetTensor(LogPred, Shape.Create(n, classes)).Buffer.Ptr;
var alpha = ScalarOps.Conv <T>(1.0);
var beta = ScalarOps.Conv <T>(0.0);
const SoftmaxAlgorithm algorithm = SoftmaxAlgorithm.LOG;
const SoftmaxMode mode = SoftmaxMode.INSTANCE;
dnn.SoftmaxForward(algorithm, mode, alpha, xDesc.Value, xPtr, beta, yDesc.Value, yPtr);
}
// TODO: make it expression
var logPred = executor.GetTensor(LogPred);
var temp = executor.GetTensor(Temp, Shape.Create(n));
var ctx = executor.Context;
if (ctx.Type == ContextType.Gpu && logPred.Layout.IsInnerChangeMostFullyPacked)
{
var stream = ctx.ToGpuContext().Stream;
var tempPtr = temp.Buffer.Ptr;
var logPredPtr = logPred.Buffer.Ptr;
var idxPtr = y.Buffer.Ptr;
DeviceFor.For(stream, 0, n, i =>
{
var idx = idxPtr[i];
tempPtr[i] = logPredPtr[i * classes + idx];
});
executor.AssignTensor(Loss, -ReduceSum(temp));
return;
}
throw new NotImplementedException();
}
public override void Forward(Executor executor)
{
var data = executor.GetTensor(Data);
var weight = executor.GetTensor(Weight);
var bias = executor.GetTensor(Bias);
var output = executor.GetTensor(Output, Shape.Create(data.Shape[0], Output.Shape[1], Output.Shape[2], Output.Shape[3]));
if (executor.Context.Type == ContextType.Gpu)
{
var convDesc = ConvolutionDesc;
var dnn = executor.Context.ToGpuContext().Dnn;
using (var dataDescRcpt = executor.TensorDescRepo.Acquire())
using (var weightDescRcpt = executor.FilterDescRepo.Acquire())
using (var biasDescRcpt = executor.TensorDescRepo.Acquire())
using (var outputDescRcpt = executor.TensorDescRepo.Acquire())
{
var dataDesc = dataDescRcpt.Value;
var weightDesc = weightDescRcpt.Value;
var biasDesc = biasDescRcpt.Value;
var outputDesc = outputDescRcpt.Value;
var dataType = Dnn.DataTypeOf <T>();
dataDesc.Set4D(dataType, TensorFormat.CUDNN_TENSOR_NCHW, (int)data.Shape[0], (int)Data.Shape[1], (int)Data.Shape[2], (int)Data.Shape[3]);
weightDesc.Set4D(dataType, TensorFormat.CUDNN_TENSOR_NCHW, (int)weight.Shape[0], (int)weight.Shape[1], (int)weight.Shape[2], (int)weight.Shape[3]);
biasDesc.Set4D(dataType, TensorFormat.CUDNN_TENSOR_NCHW, 1, (int)output.Shape[1], 1, 1);
outputDesc.Set4D(dataType, TensorFormat.CUDNN_TENSOR_NCHW, (int)output.Shape[0], (int)output.Shape[1], (int)output.Shape[2], (int)output.Shape[3]);
ConvolutionFwdAlgo algo;
IntPtr workspaceSize;
dnn.GetConvolutionForwardAlgorithm(dataDesc, weightDesc, convDesc, outputDesc,
ConvolutionFwdPreference.PREFER_FASTEST, IntPtr.Zero, out algo);
dnn.GetConvolutionForwardWorkspaceSize(dataDesc, weightDesc, convDesc, outputDesc, algo, out workspaceSize);
var workspace = workspaceSize.ToInt64() > 0L
? executor.GetTensor(Workspace1, Shape.Create(workspaceSize.ToInt64()))
: null;
//Console.WriteLine($"==> {algo} {workspaceSize}");
// step 1, convolute
dnn.ConvolutionForward(ScalarOps.Conv <T>(1.0), dataDesc, data.Buffer.Ptr, weightDesc, weight.Buffer.Ptr,
convDesc, algo, workspace?.Buffer.Ptr ?? new deviceptr <byte>(), workspaceSize, ScalarOps.Conv <T>(0.0), outputDesc, output.Buffer.Ptr);
// step 2, add bias
dnn.AddTensor(ScalarOps.Conv <T>(1.0), biasDesc, bias.Buffer.Ptr, ScalarOps.Conv <T>(1.0), outputDesc, output.Buffer.Ptr);
return;
}
}
throw new NotImplementedException();
}
public static Tensor <T> GetGradient(Executor executor, Variable <T> var, Shape shape, bool zero = false)
{
var ctx = executor.Context;
var data = executor.GetData(var);
Util.EnsureTrue(data.GradientAggregationCounter == 0);
data.GradientAggregationCounter++;
var gradient = executor.GetGradient(var, shape);
if (zero)
{
ctx.Assign(gradient, Fill(shape, ScalarOps.Conv <T>(0.0)));
}
return(gradient);
}
public Convolution2D(Variable <T> data, int kernelH, int kernelW, int numFilter)
{
Util.EnsureTrue(data.Shape.Rank == 4);
Util.EnsureTrue(data.Shape[1] > 0);
Util.EnsureTrue(data.Shape[2] > 0);
Util.EnsureTrue(data.Shape[3] > 0);
var numInputFilter = data.Shape[1];
var numOutputFilter = numFilter;
var height = data.Shape[2];
var width = data.Shape[3];
// fixed padding and stride now
ConvolutionDesc = new ConvolutionDescriptor();
ConvolutionDesc.Set2D(0, 0, 1, 1, 1, 1, ConvolutionMode.CROSS_CORRELATION);
using (var dataDesc = new TensorDescriptor())
using (var weightDesc = new FilterDescriptor())
{
var dataType = Dnn.DataTypeOf <T>();
var tempN = 100; // for temp mini batch size
dataDesc.Set4D(dataType, TensorFormat.CUDNN_TENSOR_NCHW, tempN, (int)numInputFilter, (int)height, (int)width);
weightDesc.Set4D(dataType, TensorFormat.CUDNN_TENSOR_NCHW, numOutputFilter, (int)numInputFilter, kernelH, kernelW);
// get output dimension
int n, c, h, w;
ConvolutionDesc.Get2DForwardOutputDim(dataDesc, weightDesc, out n, out c, out h, out w);
//Console.WriteLine($"{c},{h},{w}");
// Create variables
var scale = Sqrt(3.0.AsScalar <T>() / ((double)(numInputFilter * kernelH * kernelW)).AsScalar <T>());
Data = data;
Weight = Parameter(scale * (2.0.AsScalar <T>() * RandomUniform <T>(Shape.Create(numOutputFilter, numInputFilter, kernelH, kernelW), 0UL, 0UL) - 1.0.AsScalar <T>()));
Bias = Parameter(Fill(Shape.Create(c), ScalarOps.Conv <T>(0.1)));
Output = Variable <T>(PartialShape.Create(-1, c, h, w));
Workspace1 = AuxVariable <byte>();
Workspace2 = AuxVariable <byte>();
AddInput(Data);
AddInput(Weight);
AddInput(Bias);
AddOutput(Output);
AddAuxVar(Workspace1);
AddAuxVar(Workspace2);
}
}
public FullyConnected(Variable <T> data, long numHidden)
{
Util.EnsureTrue(data.HasShape);
Util.EnsureEqual(2, data.Shape.Rank, "Input must be matrix.");
Util.EnsureTrue(data.Shape[1] > 0L);
Data = data;
var numInput = data.Shape[1];
var scale = Sqrt(12.0.AsScalar <T>() / ((double)(numInput + numHidden)).AsScalar <T>());
Weights = Parameter(scale * (RandomUniform <T>(Shape.Create(numInput, numHidden), 0UL, 0UL) - 0.5.AsScalar <T>()));
Bias = Parameter(Fill(Shape.Create(numHidden), ScalarOps.Conv <T>(0.0)));
Output = Variable <T>(PartialShape.Create(data.Shape[0], numHidden));
AddInput(Data);
AddInput(Weights);
AddInput(Bias);
AddOutput(Output);
}
public override void Backward(Executor executor)
{
var ctx = executor.Context;
var indices = executor.GetTensor(Indices);
var gradout = executor.GetGradient(Output);
// for performance fix.
if (ctx.Type == ContextType.Gpu && gradout.Layout.IsInnerChangeMostFullyPacked && indices.Layout.IsInnerChangeMostFullyPacked)
{
var embedDim = EmbedDim;
var batchSize = (int)indices.Shape.Length;
var threadSize = 256;
// first set all to 0
executor.AssignGradient(Weights, Fill(executor.GetTensor(Weights).Shape, ScalarOps.Conv <T>(0.0)));
var dW = executor.GetGradient(Weights);
// then use a 1 block kernel to update it, cause usually the batch size is not huge, but the embedsize is huge!
var stream = ctx.ToGpuContext().Stream;
var iPtr = indices.Buffer.Ptr;
// the following kernel is for 1 block, so there is no need for synchornization,
// there could be further optimized.
if (typeof(T) == typeof(float))
{
var dOPtr = gradout.Buffer.Ptr.Reinterpret <float>();
var dWPtr = dW.Buffer.Ptr.Reinterpret <float>();
var lp = new LaunchParam(1, threadSize);
//Console.WriteLine($"{indices.Shape} {gradout.Shape} {dW.Shape}");
stream.Launch(() =>
{
for (var i = 0; i < batchSize; ++i)
{
var row = iPtr[i];
for (var k = threadIdx.x; k < embedDim; k += blockDim.x)
{
dWPtr[row * embedDim + k] += dOPtr[i * embedDim + k];
}
}
}, lp);
return;
}
throw new NotImplementedException();
}
else
{
executor.AssignGradient(Weights, TakeGrad(indices, gradout, EmbedSize));
}
}
public override void Backward(Executor executor)
{
var data = executor.GetTensor(Data);
var weight = executor.GetTensor(Weight);
var dOutput = executor.GetGradient(Output);
var dWeight = executor.GetGradient(Weight, Shape.Create(Weight.Shape.AsArray));
var dBias = executor.GetGradient(Bias, Shape.Create(Bias.Shape.AsArray));
var dData = executor.GetGradient(Data, Shape.Create(data.Shape.AsArray));
if (executor.Context.Type == ContextType.Gpu)
{
var convDesc = ConvolutionDesc;
var dnn = executor.Context.ToGpuContext().Dnn;
using (var dataDescRcpt = executor.TensorDescRepo.Acquire())
using (var weightDescRcpt = executor.FilterDescRepo.Acquire())
using (var dDataDescRcpt = executor.TensorDescRepo.Acquire())
using (var dOutputDescRcpt = executor.TensorDescRepo.Acquire())
using (var dBiasDescRcpt = executor.TensorDescRepo.Acquire())
using (var dWeightDescRcpt = executor.FilterDescRepo.Acquire())
{
var dataDesc = dataDescRcpt.Value;
var weightDesc = weightDescRcpt.Value;
var dDataDesc = dDataDescRcpt.Value;
var dOutputDesc = dOutputDescRcpt.Value;
var dBiasDesc = dBiasDescRcpt.Value;
var dWeightDesc = dWeightDescRcpt.Value;
var dataType = Dnn.DataTypeOf <T>();
dataDesc.Set4D(dataType, TensorFormat.CUDNN_TENSOR_NCHW, (int)data.Shape[0], (int)Data.Shape[1], (int)Data.Shape[2], (int)Data.Shape[3]);
dDataDesc.Set4D(dataType, TensorFormat.CUDNN_TENSOR_NCHW, (int)data.Shape[0], (int)Data.Shape[1], (int)Data.Shape[2], (int)Data.Shape[3]);
dOutputDesc.Set4D(dataType, TensorFormat.CUDNN_TENSOR_NCHW, (int)dOutput.Shape[0], (int)dOutput.Shape[1], (int)dOutput.Shape[2], (int)dOutput.Shape[3]);
dBiasDesc.Set4D(dataType, TensorFormat.CUDNN_TENSOR_NCHW, 1, (int)dOutput.Shape[1], 1, 1);
dWeightDesc.Set4D(dataType, TensorFormat.CUDNN_TENSOR_NCHW, (int)weight.Shape[0], (int)weight.Shape[1], (int)weight.Shape[2], (int)weight.Shape[3]);
weightDesc.Set4D(dataType, TensorFormat.CUDNN_TENSOR_NCHW, (int)weight.Shape[0], (int)weight.Shape[1], (int)weight.Shape[2], (int)weight.Shape[3]);
ConvolutionBwdFilterAlgo filterAlgo;
IntPtr filterWorkspaceSize;
dnn.GetConvolutionBackwardFilterAlgorithm(dataDesc, dOutputDesc, convDesc, dWeightDesc,
ConvolutionBwdFilterPreference.PREFER_FASTEST, IntPtr.Zero, out filterAlgo);
dnn.GetConvolutionBackwardFilterWorkspaceSize(dataDesc, dOutputDesc, convDesc, dWeightDesc, filterAlgo, out filterWorkspaceSize);
var filterWorkspace = filterWorkspaceSize.ToInt64() > 0L
? executor.GetTensor(Workspace1, Shape.Create(filterWorkspaceSize.ToInt64()))
: null;
//Console.WriteLine($"==> {filterAlgo} {filterWorkspaceSize}");
ConvolutionBwdDataAlgo dataAlgo;
IntPtr dataWorkspaceSize;
dnn.GetConvolutionBackwardDataAlgorithm(weightDesc, dOutputDesc, convDesc, dDataDesc,
ConvolutionBwdDataPreference.PREFER_FASTEST, IntPtr.Zero, out dataAlgo);
dnn.GetConvolutionBackwardDataWorkspaceSize(dWeightDesc, dOutputDesc, convDesc, dDataDesc, dataAlgo, out dataWorkspaceSize);
var dataWorkspace = dataWorkspaceSize.ToInt64() > 0L
? executor.GetTensor(Workspace2, Shape.Create(dataWorkspaceSize.ToInt64()))
: null;
//Console.WriteLine($"==> {dataAlgo} {dataWorkspaceSize}");
// filter
dnn.ConvolutionBackwardFilter(ScalarOps.Conv <T>(1.0), dataDesc, data.Buffer.Ptr, dOutputDesc,
dOutput.Buffer.Ptr, convDesc, filterAlgo, filterWorkspace?.Buffer.Ptr ?? new deviceptr <byte>(), filterWorkspaceSize,
ScalarOps.Conv <T>(0.0), dWeightDesc, dWeight.Buffer.Ptr);
// data
dnn.ConvolutionBackwardData(ScalarOps.Conv <T>(1.0), weightDesc, weight.Buffer.Ptr, dOutputDesc,
dOutput.Buffer.Ptr, convDesc, dataAlgo, dataWorkspace?.Buffer.Ptr ?? new deviceptr <byte>(), dataWorkspaceSize,
ScalarOps.Conv <T>(0.0), dDataDesc, dData.Buffer.Ptr);
// bias
dnn.ConvolutionBackwardBias(ScalarOps.Conv <T>(1.0), dOutputDesc, dOutput.Buffer.Ptr, ScalarOps.Conv <T>(0.0), dBiasDesc, dBias.Buffer.Ptr);
return;
}
}
throw new NotImplementedException();
}
public override void Backward(Executor executor)
{
Util.EnsureTrue(IsTraining);
var context = executor.Context.ToGpuContext();
var dnn = context.Dnn;
if (executor.IncreaseGradientAggregationCounter(X) != 0)
{
throw new InvalidOperationException();
}
if (executor.IncreaseGradientAggregationCounter(HX) != 0)
{
throw new InvalidOperationException();
}
if (executor.IncreaseGradientAggregationCounter(CX) != 0)
{
throw new InvalidOperationException();
}
dnn.RNNBackwardData(
executor.RnnDescDict[RnnDesc],
SeqLength,
YDesc,
executor.GetTensor(Y).Buffer.Ptr,
YDesc,
executor.GetGradient(Y).Buffer.Ptr,
StateDesc,
new deviceptr <T>(), // executor.GetGradient(HY).Buffer.Ptr,
StateDesc,
new deviceptr <T>(), // executor.GetGradient(CY).Buffer.Ptr,
executor.FilterDescDict[WDesc],
executor.GetTensor(W).Buffer.Ptr,
StateDesc,
executor.GetTensor(HX).Buffer.Ptr,
StateDesc,
executor.GetTensor(CX).Buffer.Ptr,
XDesc,
executor.GetGradient(X).Buffer.Ptr,
StateDesc,
executor.GetGradient(HX).Buffer.Ptr,
StateDesc,
executor.GetGradient(CX).Buffer.Ptr,
executor.GetTensor(Workspace).Buffer.Ptr,
(IntPtr)executor.GetTensor(Workspace).Shape.Length,
executor.GetTensor(ReserveSpace).Buffer.Ptr,
(IntPtr)executor.GetTensor(ReserveSpace).Shape.Length);
if (executor.IncreaseGradientAggregationCounter(W) == 0)
{
executor.AssignGradient(W, ScalarOps.Conv <T>(0.0).AsScalar(), replace: true);
}
dnn.RNNBackwardWeights(
executor.RnnDescDict[RnnDesc],
SeqLength,
XDesc,
executor.GetTensor(X).Buffer.Ptr,
StateDesc,
executor.GetTensor(HX).Buffer.Ptr,
YDesc,
executor.GetTensor(Y).Buffer.Ptr,
executor.GetTensor(Workspace).Buffer.Ptr,
(IntPtr)executor.GetTensor(Workspace).Shape.Length,
executor.FilterDescDict[WDesc],
executor.GetGradient(W).Buffer.Ptr,
executor.GetTensor(ReserveSpace).Buffer.Ptr,
(IntPtr)executor.GetTensor(ReserveSpace).Shape.Length);
}
public override void Initialize(Executor executor)
{
var context = executor.Context.ToGpuContext();
var dnn = context.Dnn;
// dropout
var dropoutDesc = executor.DropoutDescDict[DropoutDesc];
IntPtr dropoutStatesSize;
dnn.DropoutGetStatesSize(out dropoutStatesSize);
var dropoutStates = executor.GetTensor(DropoutStates, Shape.Create(dropoutStatesSize.ToInt64()));
dropoutDesc.Set(dnn, (float)Dropout, dropoutStates.Buffer.Ptr, dropoutStatesSize, DropoutSeed);
// rnn descriptor
var rnnDesc = executor.RnnDescDict[RnnDesc];
var mode = Type.Mode;
rnnDesc.Set(HiddenSize, NumLayers, dropoutDesc, RNNInputMode.LINEAR_INPUT, DirectionMode.UNIDIRECTIONAL, mode, Dnn.DataTypeOf <T>());
// weight
var wDesc = executor.FilterDescDict[WDesc];
IntPtr weightsSize;
dnn.GetRNNParamsSize(rnnDesc, XDesc[0], out weightsSize, Dnn.DataTypeOf <T>());
Util.EnsureTrue(weightsSize.ToInt64() % Gpu.SizeOf <T>() == 0);
var shapeW = Shape.Create(weightsSize.ToInt64() / Alea.Gpu.SizeOf <T>());
wDesc.SetND(Dnn.DataTypeOf <T>(), TensorFormat.CUDNN_TENSOR_NCHW, new [] { (int)shapeW[0], 1, 1 });
// workspace and reserved space
IntPtr workSize;
dnn.GetRNNWorkspaceSize(rnnDesc, SeqLength, XDesc, out workSize);
executor.GetTensor(Workspace, Shape.Create(workSize.ToInt64()));
if (IsTraining)
{
IntPtr reserveSize;
dnn.GetRNNTrainingReserveSize(rnnDesc, SeqLength, XDesc, out reserveSize);
executor.GetTensor(ReserveSpace, Shape.Create(reserveSize.ToInt64()));
}
// since we are using cuDNN, we'd better make sure these varaibles are allocated
executor.GetTensor(W, shapeW);
if (IsTraining)
{
executor.GetGradient(W, shapeW);
}
executor.GetTensor(Y, Shape.Create(Y.Shape.AsArray));
executor.GetTensor(HX, Shape.Create(HX.Shape.AsArray));
executor.GetTensor(CX, Shape.Create(CX.Shape.AsArray));
executor.GetTensor(HY, Shape.Create(HY.Shape.AsArray));
executor.GetTensor(CY, Shape.Create(CY.Shape.AsArray));
if (IsTraining)
{
executor.GetGradient(X, Shape.Create(X.Shape.AsArray));
executor.GetGradient(Y, Shape.Create(Y.Shape.AsArray));
executor.GetGradient(HX, Shape.Create(HX.Shape.AsArray));
executor.GetGradient(CX, Shape.Create(CX.Shape.AsArray));
}
// init weights
var numLinearLayers = Type.NumLinLayers;
using (var filterDesc = new FilterDescriptor())
{
var w = executor.GetTensor(W);
var filterDimA = new int[3];
for (var layer = 0; layer < NumLayers; ++layer)
{
for (var linLayerId = 0; linLayerId < numLinearLayers; ++linLayerId)
{
int nbDims;
DataType dataType;
TensorFormat format;
deviceptr <T> linLayerMat;
dnn.GetRNNLinLayerMatrixParams(rnnDesc, layer, XDesc[0], wDesc, w.Buffer.Ptr, linLayerId,
filterDesc, out linLayerMat);
filterDesc.GetND(out dataType, out format, out nbDims, filterDimA);
var length = filterDimA.Aggregate(ScalarOps.Mul);
var linLayerMatBuffer = new Buffer <T>(context.Device, w.Memory, new Layout(Shape.Create(length)), linLayerMat);
var linLayerMatTensor = new Tensor <T>(linLayerMatBuffer);
context.Assign(linLayerMatTensor, RandomNormal <T>(Shape.Create(length)) / (Math.Sqrt(HiddenSize + InputSize).AsScalar <T>()));
deviceptr <T> linLayerBias;
dnn.GetRNNLinLayerBiasParams(rnnDesc, layer, XDesc[0], wDesc, w.Buffer.Ptr, linLayerId, filterDesc, out linLayerBias);
filterDesc.GetND(out dataType, out format, out nbDims, filterDimA);
length = filterDimA.Aggregate(ScalarOps.Mul);
var linLayerBiasBuffer = new Buffer <T>(context.Device, w.Memory, new Layout(Shape.Create(length)), linLayerBias);
var linLayerBiasTensor = new Tensor <T>(linLayerBiasBuffer);
Type.InitBias(context, layer, linLayerId, linLayerBiasTensor);
}
}
}
base.Initialize(executor);
const double value = 0.0;
executor.AssignTensor(HX, Fill(Shape.Create(HX.Shape.AsArray), ScalarOps.Conv <T>(value)));
executor.AssignTensor(CX, Fill(Shape.Create(CX.Shape.AsArray), ScalarOps.Conv <T>(value)));
}
public void ZeroInitialStates(Executor executor)
{
executor.AssignTensor(Rnn.HX, Fill(Shape.Create(Rnn.HX.Shape.AsArray), ScalarOps.Conv <T>(0.0)));
executor.AssignTensor(Rnn.CX, Fill(Shape.Create(Rnn.CX.Shape.AsArray), ScalarOps.Conv <T>(0.0)));
}
public void ForwardBasic(Executor executor)
{
var ctx = executor.Context;
var w = executor.GetTensor(W);
var xphpb = w.Shape[0];
var x = executor.GetTensor(X);
var b = x.Shape[1];
var n = x.Shape[0];
var d = HiddenSize;
var y = executor.GetTensor(Y, Shape.Create(n, b, d));
var inputSize = InputSize;
var one = 1.0.AsScalar <T>();
// inital states
var cx = executor.GetTensor(CX);
var hx = executor.GetTensor(HX);
Util.EnsureTrue(cx.Shape.SequenceEqual(Shape.Create(b, d)));
Util.EnsureTrue(hx.Shape.SequenceEqual(Shape.Create(b, d)));
// we assign output states to inital states, and later we update it
var cy = executor.GetTensor(CY, Shape.Create(b, d));
var hy = executor.GetTensor(HY, Shape.Create(b, d));
ctx.Assign(cy, cx);
ctx.Assign(hy, hx);
var prevc = cy.Reshape(1, b, d);
var prevh = hy.Reshape(1, b, d);
var hin = executor.GetTensor(Hin, Shape.Create(n, b, xphpb));
var ifoa1 = executor.GetTensor(IFOA1, Shape.Create(n, b, d * 4));
var ifoa2 = executor.GetTensor(IFOA2, Shape.Create(n, b, d * 4));
var c = executor.GetTensor(C, Shape.Create(n, b, d));
for (var t = 0; t < n; ++t)
{
// stack input
ctx.Assign(hin.Slice(t, -1, 0), Fill(Shape.Create(1, b, 1), ScalarOps.Conv <T>(1.0))); // bias
ctx.Assign(hin.Slice(t, -1, Range(1, inputSize + 1)), x.Slice(t));
ctx.Assign(hin.Slice(t, -1, Range(inputSize + 1, -1)), prevh);
// dot
ctx.Assign(ifoa1.Slice(t), Dot(hin.Slice(t).Reshape(b, xphpb), w));
// values for applying element-wise transformation
// they are of shape (1, b, d)
var ct = c.Slice(t);
var ht = y.Slice(t);
var it = ifoa2.Slice(t, -1, Range(0, d));
var ft = ifoa2.Slice(t, -1, Range(d, 2 * d));
var ot = ifoa2.Slice(t, -1, Range(2 * d, 3 * d));
var at = ifoa2.Slice(t, -1, Range(3 * d, 4 * d));
// non-linearities
// first 3 matrices are IFO, we apply sigmoid
var ifot = ifoa2.Slice(t, -1, Range(0, 3 * d));
var _ifot = ifoa1.Slice(t, -1, Range(0, 3 * d));
ctx.Assign(ifot, one / (one + Exp(-_ifot)));
// last one is for activation gate, we apply tanh
var _at = ifoa1.Slice(t, -1, Range(3 * d, 4 * d));
ctx.Assign(at, Tanh(_at));
// c_t = i_t * a_t + f_t * c_t-1
ctx.Assign(ct, it * at + ft * prevc);
// h_t = o_t * tanh(c_t)
ctx.Assign(ht, ot * Tanh(ct));
// update states
ctx.Assign(prevh, y.Slice(t));
ctx.Assign(prevc, c.Slice(t));
}
}
public override void Initialize(Executor executor)
{
base.Initialize(executor);
// set bias to zero
var ctx = executor.Context;
var w = executor.GetTensor(W);
// first set 4 bias to 0.0
ctx.Assign(w.Slice(0), 0.0.AsScalar <T>());
// set forget bias is needed, layout: IFOA, so forget index is 1
if (ForgetBiasInit != 0.0)
{
ctx.Assign(w.Slice(0, Range(HiddenSize, 2 * HiddenSize)), Fill(Shape.Create(1, HiddenSize), ScalarOps.Conv <T>(ForgetBiasInit)));
}
}
public void ZeroTerminalGradient(Executor executor)
{
executor.AssignGradient(HY, Fill(Shape.Create(HY.Shape.AsArray), ScalarOps.Conv <T>(0.0)), replace: true);
executor.AssignGradient(CY, Fill(Shape.Create(CY.Shape.AsArray), ScalarOps.Conv <T>(0.0)), replace: true);
}