About Spaghetti

[dainel40911]

Sorry for bothering you. Do you have training script for the part of spaghetti, or you only use the pre-trained model. Cause I want to apply this model on multiple class

[AlexandreBinninger]

Dear Dainel,

Actually, I got asked a few weeks ago about it, and after contacting SPAGHETTI's author, I got the following answer:

spaghetti_v2.zip is the code of Spaghetti that includes the pre-processing and training pipeline.

For training on shapenet, first, you need to parse the data:

• Download the shapenet data (ShapeNetCore.v2) https://shapenet.org/
• install Watertight Manifold
• Change the paths in constants.py to the location of the ShapeNetCore.v2 and the MANIFOLD build path.
• run make_data.py. You can specify the category name to parse. It will run the watertight script and sample points for training afterward.
• Run train.py for training
• if you want to only fine-tune part of an existing model, you can filter out the parameters you want to keep in the trainer:line 177

I hope this helps.

Best Regards,
Alexandre Binninger.

[dainel40911]

Thank you so much for the comprehensive explanation. However, my linux crash when running train.py on certain step of epoch 0 and I'm running on RTX 4090. Do you have any ideas causing this issue? Thank you so much.

[AlexandreBinninger]

You're welcome.
No, I didn't have to train the models myself. According to the paper, the models were trained using RTX A6000 GP.

[dainel40911]

Could I ask how your multiclass model work? It seems like that spaghetti only have single class for each category.

[AlexandreBinninger]

I got the trained model from my collaborator. The interface that he gave me was a bit different so it does not work as a plug-and-play with the current single-class model in the repo. It works by just training the model on several classes at the same time. There is no way to indicate the class to spaghetti or condition the generation to a specific class.