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 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();
}
