Köster Manuel

There are tons of ways to do this. I've attached an example on how you could approach it.

The important factor is the "Piece Attribute" on the Copy to Points node which basically looks up the attribute on the points and tries to find the corresponding Objects with the same attribute value.

I recently had a similar issue.

The problem is because of the uncommon version of HDF5 Houdini ships with, which is HDF5-1.8.11 [www.sidefx.com]. There are no h5py versions available on package management systems like pip or conda which are built with HDF5-1.8.11.

Additionally this prevents working with TensorFlow within Houdini, since TF also requires h5py.

I've recently logged a RFE for this: Shipping Houdini with a HDF5 version that has a matching & available h5py version [www.sidefx.com]

Feel free to add another RFE and mention RFE#109905 from above if this affects you as well.

#RFE109905
As we now have stable Python 3 builds of Houdini, machine learning within Houdini becomes more and more attractive. The most popular Python machine learning frameworks are Tensorflow and Pytorch.

Both Tensorflow and Pytorch (with torch-geometric) depend on h5py (a HDF5 Python wrapper).
Houdini ships with HDF5-1.8.11. There are no builds of h5py available via any package manager (pip, conda..) that are built upon HDF5-1.8.11, which makes the usage of Tensorflow/Pytorch within Houdini hard since there are no Houdini compatible versions of h5py available. The only remaining option is to build h5py from source yourself, which is very hard (especially on Windows).

My proposal is to ship Houdini with a HDF5 version that is used by some h5py distributions. One of these versions would be HDF5-1.8.18. This would make the implementation of ML processes within Houdini a lot easier/possible.

I do not know what implications this brings for you as a lot of other things that you guys ship are probably built with this exact HDF5 version.

Any feedback would be appreciated on how we could better integrate ML processes within Houdini.

Take a look at the SOP Solvers like Vellum or RBD. They have a "dive target" where the user can dive into the HDA and pop down their own nodes.

Would that workflow be a good option for you?