public override Tensor log_prob(Tensor value)
{
value = value.@long().unsqueeze(-1);
var valLogPmf = torch.broadcast_tensors(value, logits);
value = valLogPmf[0][TensorIndex.Ellipsis, TensorIndex.Slice(null, 1)];
return(valLogPmf[1].gather(-1, value).squeeze(-1));
}
static (Tensor, Tensor) GetBatch(Tensor source, int index, int bptt)
{
var len = Math.Min(bptt, source.shape[0] - 1 - index);
var data = source[TensorIndex.Slice(index, index + len)];
var target = source[TensorIndex.Slice(index + 1, index + 1 + len)].reshape(-1);
return(data, target);
}
private static List <Tensor> LoadImages(IList <string> images, int batchSize, int channels, int height, int width)
{
List <Tensor> tensors = new List <Tensor>();
var imgSize = channels * height * width;
bool shuffle = false;
Random rnd = new Random();
var indices = !shuffle?
Enumerable.Range(0, images.Count).ToArray() :
Enumerable.Range(0, images.Count).OrderBy(c => rnd.Next()).ToArray();
// Go through the data and create tensors
for (var i = 0; i < images.Count;)
{
var take = Math.Min(batchSize, Math.Max(0, images.Count - i));
if (take < 1)
{
break;
}
var dataTensor = torch.zeros(new long[] { take, imgSize }, ScalarType.Byte);
// Take
for (var j = 0; j < take; j++)
{
var idx = indices[i++];
var lblStart = idx * (1 + imgSize);
var imgStart = lblStart + 1;
using (var stream = new SKManagedStream(File.OpenRead(images[idx])))
using (var bitmap = SKBitmap.Decode(stream)) {
using (var inputTensor = torch.tensor(GetBytesWithoutAlpha(bitmap))) {
Tensor finalized = inputTensor;
var nz = inputTensor.count_nonzero().DataItem <long>();
if (bitmap.Width != width || bitmap.Height != height)
{
var t = inputTensor.reshape(1, channels, bitmap.Height, bitmap.Width);
finalized = torchvision.transforms.functional.resize(t, height, width).reshape(imgSize);
}
dataTensor.index_put_(finalized, TensorIndex.Single(j));
}
}
}
tensors.Add(dataTensor.reshape(take, channels, height, width));
dataTensor.Dispose();
}
return(tensors);
}
public PositionalEncoding(long dmodel, double dropout, int maxLen = 5000) : base("PositionalEncoding")
{
this.dropout = Dropout(dropout);
var pe = torch.zeros(new long[] { maxLen, dmodel });
var position = torch.arange(0, maxLen, 1).unsqueeze(1);
var divTerm = (torch.arange(0, dmodel, 2) * (-Math.Log(10000.0) / dmodel)).exp();
pe[TensorIndex.Ellipsis, TensorIndex.Slice(0, null, 2)] = (position * divTerm).sin();
pe[TensorIndex.Ellipsis, TensorIndex.Slice(1, null, 2)] = (position * divTerm).cos();
this.pe = pe.unsqueeze(0).transpose(0, 1);
RegisterComponents();
}
public static Tensor crop(this Tensor image, int top, int left, int height, int width)
{
var dims = image.Dimensions;
var hoffset = dims - 2;
long h = image.shape[hoffset], w = image.shape[hoffset + 1];
var right = left + width;
var bottom = top + height;
if (left < 0 || top < 0 || right > w || bottom > h)
{
var slice = image.index(TensorIndex.Ellipsis, TensorIndex.Slice(Math.Max(top, 0), bottom), TensorIndex.Slice(Math.Max(left, 0), right));
var padding_ltrb = new long[] { Math.Max(-left, 0), Math.Max(-top, 0), Math.Max(right - w, 0), Math.Max(bottom - h, 0) };
return(TorchSharp.torch.nn.functional.pad(slice, padding_ltrb));
}
return(image.index(TensorIndex.Ellipsis, TensorIndex.Slice(top, bottom), TensorIndex.Slice(left, right)));
}
public override Tensor forward(Tensor t)
{
using var x = t + pe[TensorIndex.Slice(null, t.shape[0]), TensorIndex.Slice()];
return(dropout.forward(x));
}
