PolyGen Question: Train-Test Split

Hello,

I'm working with 3D shapes with Transformers and currently looking closer at the PolyGen paper.

In [Colab](https://colab.research.google.com/github/deepmind/deepmind-research/blob/master/polygen/sample-pretrained.ipynb) a pre-trained model is given.
I would like to get some quantitative metrics on the generated quality of the meshes, e.g. [MMD/COV scores](https://arxiv.org/abs/1812.02822), but for this task I need to know exactly which samples of ShapeNet were used for training and which for testing. The idea is to sample a generated set G and compare it to the test set A.

The information given in the paper is sadly not enought.
> "We train all our models on the ShapeNet Core V2 dataset (Chang et al., 2015), which we subdivide into 92.5% training, 2.5% validation and 5% testing splits." - PolyGen, 3.1. Training Details

@charlienash
Is the train-test split following any rules, which would allow to reconstruct the test set or are they just randomly sampled?

Sincerely Gregor