public TorchTensor forward(TorchTensor t, TorchTensor mask)
{
var src = pos_encoder.forward(encoder.forward(t) * MathF.Sqrt(ninputs));
var enc = transformer_encoder.forward(src, mask);
return(decoder.forward(enc));
}
private torch.Tensor ForwardEmbedding(torch.Tensor tokens, torch.Tensor segmentLabels, torch.Tensor positions)
{
using var disposeScope = torch.NewDisposeScope();
var x = TokenEmbedding.forward(tokens);
if (EmbedScale != null)
{
x.mul_(EmbedScale);
}
if (PositionalEmbedding != null)
{
var positionalEmbedding = PositionalEmbedding.forward(tokens,
new Dictionary <string, object> {
{ PositionalEmbedding.PositionKey, positions }
});
x.add_(positionalEmbedding);
}
if (SegmentEmbedding != null && segmentLabels.IsNotNull())
{
var segmentEmbedding = SegmentEmbedding.forward(segmentLabels);
x.add_(segmentEmbedding);
}
if (EmbeddingLayerNorm != null)
{
x = EmbeddingLayerNorm.forward(x);
}
x = EmbedTransfer.forward(x, (int)x.size()[x.size().Length - 1]);
x = DropoutLayer.forward(x);
return(x.MoveToOuterDisposeScope());
}
public override torch.Tensor forward(torch.Tensor input, Dictionary <string, object> param = null)
{
using var disposeScope = torch.NewDisposeScope();
ParseArguments(param, out var incrementalState, out var positions);
if (positions.IsNull())
{
positions = incrementalState
? torch.tensor(PadPositionIndex + input.size(1))
: MakePositions(input, PadTokenIndex);
}
var embedding = Embedding.forward(positions);
return(embedding.MoveToOuterDisposeScope());
}