public static DataSet BuildSet(IAllocator allocator, DigitImage[] images)
{
var inputs = new Tensor(allocator, DType.Float32, images.Length, MnistParser.ImageSize, MnistParser.ImageSize);
var outputs = new Tensor(allocator, DType.Float32, images.Length, MnistParser.LabelCount);
var cpuAllocator = new TensorSharp.Cpu.CpuAllocator();
for (int i = 0; i < images.Length; ++i)
{
var target = inputs.TVar().Select(0, i);
TVar.FromArray(images[i].pixels, cpuAllocator)
.AsType(DType.Float32)
.ToDevice(allocator)
.Evaluate(target);
target.Div(255)
.Evaluate(target);
}
Ops.FillOneHot(outputs, MnistParser.LabelCount, images.Select(x => (int)x.label).ToArray());
var targetValues = Tensor.FromArray(allocator, images.Select(x => (float)x.label).ToArray());
return(new DataSet()
{
inputs = inputs, targets = outputs, targetValues = targetValues
});
}
public static TVar Softmax(this TVar x)
{
long[] shape = x.Evaluate().Shape;
List <float> data = new List <float>();
for (long i = 0; i < shape[0]; i++)
{
var s_x = x.Select(0, i);
var exp = s_x.Exp();
var sum = exp.SumAll();
var s_t = (exp / sum.ToScalar()).View(1, shape[1]).Evaluate();
data.AddRange(s_t.ToArray().Cast <float>());
}
return(TVar.FromArray(data.ToArray(), Global.Device).View(shape));
}
