Overview
Hello and thank you for all the work that goes into TorchSharp — it has been a very useful tool for us!
I recently needed to compute an N-D histogram from a Tensor object. In PyTorch, this is provided by torch.histogramdd (introduced in 1.11 and still present in the latest 2.7): https://pytorch.org/docs/stable/generated/torch.histogramdd.html
In TorchSharp, the method signature was added 2 years ago but remains as only a stub which throws a NotImplementedException:
|
static (Tensor hist_values, Tensor[] edges) histogramdd( |
Why is it useful?
Building N-D histograms from tensor data is quite powerful. For instance, this is useful for calculating occupancy grids over multiple dimensions, computing information loss, color image histograms, etc. Currently, to do this with TorchSharp requires work-arounds using other libraries or manual implementation. It could also benefit from running on the same hardware and minimizing unnecessary CPU-GPU transfers.
Feasibility
The native function already ships with LibTorch (ATen/ops/histogramdd.h), so perhaps only the C++/CLI binding and C# wrapper is needed? If possible, it would be great to have this added into an upcoming release.
If it's any help, I'm happy to take a first pass at a PR for this. Please let me know if there are design constraints or coding conventions I should follow before I start.
Overview
Hello and thank you for all the work that goes into TorchSharp — it has been a very useful tool for us!
I recently needed to compute an N-D histogram from a
Tensorobject. In PyTorch, this is provided bytorch.histogramdd(introduced in 1.11 and still present in the latest 2.7): https://pytorch.org/docs/stable/generated/torch.histogramdd.htmlIn TorchSharp, the method signature was added 2 years ago but remains as only a stub which throws a
NotImplementedException:TorchSharp/src/TorchSharp/Tensor/torch.OtherOperations.cs
Line 468 in 75eab9f
Why is it useful?
Building N-D histograms from tensor data is quite powerful. For instance, this is useful for calculating occupancy grids over multiple dimensions, computing information loss, color image histograms, etc. Currently, to do this with TorchSharp requires work-arounds using other libraries or manual implementation. It could also benefit from running on the same hardware and minimizing unnecessary CPU-GPU transfers.
Feasibility
The native function already ships with LibTorch (ATen/ops/histogramdd.h), so perhaps only the C++/CLI binding and C# wrapper is needed? If possible, it would be great to have this added into an upcoming release.
If it's any help, I'm happy to take a first pass at a PR for this. Please let me know if there are design constraints or coding conventions I should follow before I start.