Feature Request: Add LogCumSumExp Operation to TorchSharp

[m-usman98]

The LogCumSumExp operation available in PyTorch (torch.logcumsumexp) is currently missing in TorchSharp. This function provides a numerically stable way to compute the cumulative log-sum-exp along a specified dimension, which is essential for log-domain computations in probabilistic models and sequence processing. I found a reference implementation by agadetsk.

Python Code:

def logcumsumexp(x, dim):
    # slow implementation, but ok for now
    if (dim != -1) or (dim != x.ndimension() - 1):
        x = x.transpose(dim, -1)

    out = []
    for i in range(1, x.size(-1) + 1):
        out.append(torch.logsumexp(x[..., :i], dim=-1, keepdim=True))
    out = torch.cat(out, dim=-1)

    if (dim != -1) or (dim != x.ndimension() - 1):
        out = out.transpose(-1, dim)
    return out

I've reimplemented this functionality in C#, achieving equivalent behavior.

C# Code:

 public static Tensor LogCumSumExp(Tensor x, long dim)
     {
         int ndim = (int)x.ndim;
         int lastDim = ndim - 1;
         bool needTranspose = (dim != -1) && (dim != lastDim);

         if (needTranspose)
         {
             x = x.transpose((int)dim, lastDim);
         }

         int size = (int)x.size(lastDim);
         List<Tensor> outputs = new List<Tensor>();

         for (int i = 1; i <= size; i++)
         {
             Tensor slice = x.slice(lastDim, 0, i, 1);
             Tensor lse = torch.logsumexp(slice, dim: lastDim, keepdim: true);
             outputs.Add(lse);
         }

         Tensor[] outputArray = outputs.ToArray();
         Tensor result = torch.cat(outputArray, dim: lastDim);

         if (needTranspose)
         {
             result = result.transpose(lastDim, (int)dim);
         }

         return result;}

I would like this implementation to be considered for inclusion in TorchSharp.