a) Feature Request: Function sample_dirichlet, b) State of probabilistic model support?

[andife]

I am very interested in converting DeepLearning models, that contain the PowerSpherical function (https://github.com/andife/power_spherical) to ONNX.

Currently this fails because of the Dirichlet function (https://github.com/pytorch/pytorch/issues/116336). After my research, I came across https://github.com/onnx/onnxmltools/issues/549, among others and wondered whether it would be useful to have gamma, dirichlet, beta available in general? For this reason, the question arises as to what the current state of probabilistic model support looks like?

Dirichlet is available in pytorch: https://pytorch.org/docs/stable/distributions.html#dirichlet tensorflow: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/ops/distributions/dirichlet_multinomial.py

Would it be a suitable direction, e.g. to create a sample Dirichlet method as an Onnx function based on RandomUniform (https://onnx.ai/onnx/operators/onnx__RandomUniform.html#l-onnx-doc-randomuniform)?