in javacpp-pytorch tensor slice batch broadcast opration how to do like python pytorch style ?

[mullerhai]

HI ,
in python code , tensor slice batch broadcast opration is very easy and convient ,like this

//  encoding[::, 0 :: 2] = torch.sin(position * div_term)
//  encoding[::, 1 :: 2] = torch.cos(position * div_term)

but now in javacpp-pytorch ,I can not do that to set batch value ,so how to do like this

thanks

[mullerhai]

if I write this code like this

    for (int ti = 0; ti < pTimeSize; ti++) {
        for (int ba = 0; ba < pBatchSize; ba++) {
            for (int ch = 0; ch < pChannelSize; ch++) {
                for (int ro = 0; ro < pRowSize; ro++) {
                    for (int co = 0; co < pColumnSize; co++) {
                        int   i = 0, pos = 0;
                        float div_term = 0, pe = 0;
                        i   = co / 2;
                        pos = ch;
                        div_term = pow(10000, 2 * i / (double)pColumnSize);
                        if (co % 2 == 0)
                            pe = sin(pos / div_term);
                        else
                            pe = cos(pos / div_term);
                        (*pPositionalEncoding)[Index5D(peShape, ti, ba, ch, ro, co)] = pe;
                    }
                }
            }
        }
    }

I think not suit pytorch

[mullerhai]

or like this

  private def initializeEncodings(): Unit = {
    for (pos <- 0 until config.maxSeqLen; i <- 0 until config.embeddingDim) {
      val angle = pos / math.pow(10000, 2 * (i / 2) / config.embeddingDim)
      encodings(pos)(i) = if (i % 2 == 0) sin(angle) else cos(angle)
    }
  }

[HGuillemet]

See https://pytorch.org/cppdocs/notes/tensor_indexing.html.
One again, javacpp pytorch is a binding to C++ api, not to Python API.

[mullerhai]

I see , thanks

See https://pytorch.org/cppdocs/notes/tensor_indexing.html. One again, javacpp pytorch is a binding to C++ api, not to Pytorch API.

in cpp libtorch , I see them write code like this

              auto pe = torch::zeros({1, this->max_len, this->d_model});
                    auto position = torch::arange(0, max_len).unsqueeze(1);
                    auto div_term = torch::exp(torch::arange(0, d_model, 2) * -std::log(10000.0) / d_model);
                    pe.slice(2, 0, pe.size(2), 2) = torch::sin(position * div_term);
                    pe.slice(2, 1, pe.size(2), 2) = torch::cos(position * div_term);

[HGuillemet]

In Java you don't have operator overloading, so, compared to C++, you must replace the operators like *, = ... on tensor objects to explicit method calls.
Also you don't have curly brace initializers. You must use explicit new.

[mullerhai]

在 Java 中，没有运算符重载，因此与 C++ 相比，您必须将张量对象上的运算符（如*、=...）替换为显式方法调用。 此外，您没有花括号初始化程序。您必须使用显式new。

ok, I know ,thanks ,I will try another way do that

[mullerhai]

last year ,I found I had write some code to do tensor index set value
python origin code

        batch_size, num_fields, embed_dim = x.shape
        a_prev = torch.ones((batch_size, num_fields + 1, embed_dim), dtype=torch.float).to(x.device)
        for t in range(self.order):
            a = torch.zeros((batch_size, num_fields + 1, embed_dim), dtype=torch.float).to(x.device)
            a[:, t+1:, :] += x[:, t:, :] * a_prev[:, t:-1, :]
            a = torch.cumsum(a, dim=1)
            a_prev = a

use javacpp-pytorch scala code

   val shape = xs.shape()
    val batch_size = shape(0)
    val num_fields = shape(1)
    val embed_dim = shape(2)
    var a_prev = torch.ones(Array(batch_size, num_fields + 1, embed_dim), new ScalarTypeOptional(ScalarType.Float), new LayoutOptional, new DeviceOptional(xs.device()), new BoolOptional())
    var a = torch.zeros(Array(batch_size, num_fields + 1, embed_dim), new ScalarTypeOptional(ScalarType.Float), new LayoutOptional, new DeviceOptional(xs.device()), new BoolOptional())
    for (t <- 0 until this.order by 1) {
      val sliceNone = new Slice()
      val sliceA = new Slice(new SymIntOptional(new SymInt(t+1)), new SymIntOptional(new SymInt(0)), new SymIntOptional(new SymInt(1)))
      val sliceA_prev = new Slice(new SymIntOptional(new SymInt(t)), new SymIntOptional(new SymInt(-1)), new SymIntOptional(new SymInt(1)))
      val sliceX = new Slice(new SymIntOptional(new SymInt(t)), new SymIntOptional(new SymInt(0)), new SymIntOptional(new SymInt(1)))
      val indexSliceNone = new TensorIndex(sliceNone)
      val indexSliceA = new TensorIndex(sliceA)
      val indexSliceA_prev = new TensorIndex(sliceA_prev)
      val indexSliceX = new TensorIndex(sliceX)

      val indexVectorA = new TensorIndexVector(indexSliceNone, indexSliceA, indexSliceNone)
      val indexArrayRefA = new TensorIndexArrayRef(indexVectorA)

      val indexVectorX = new TensorIndexVector(indexSliceNone, indexSliceX, indexSliceNone)
      val indexArrayRefX = new TensorIndexArrayRef(indexVectorX)

      val indexVectorA_prev = new TensorIndexVector(indexSliceNone, indexSliceA_prev, indexSliceNone)
      val indexArrayRefA_prev = new TensorIndexArrayRef(indexVectorA_prev)

      a.index(indexArrayRefA).add_(torch.mul(x.index(indexArrayRefX), a_prev.index(indexArrayRefA_prev)))
      a = torch.cumsum(a, 1)
      a_prev = a

use javacpp-pytorch for tensor index slice batch broadcast opration is so hard and more line than python pytorch, do we has easy way ? thanks