Is there a multi-output or multi-label example for graph classification?

[Boehm92]

Hi,
i'm working on a GNN to classify 3D geometric shapes in a 3D cube. The cube and its geometric shape is mesh based.

I'm allready capable to classify a cube with one geometric shape and it works realy well. For that i'm using the dataset from this paper https://github.com/zibozzb/FeatureNet and this PyG examle https://colab.research.google.com/drive/1I8a0DfQ3fI7Njc62__mVXUlcAleUclnb?usp=sharing

My graphs look something like that: Data(x=[14, 3], edge_index=[2, 144], y=[24])

Now, i want to classify a cube with two or more shapes. I tried following steps:

• reconstructed the target to an array with 24 element: y = [0, 0, 1, 0, ..., 0, 1]
• switched BCE to BCEWithLogitsLoss.

But i'm not shure how to continue with the training method or if this is the right way anyway.

Is there some example for multi-output or multi-label classification?

Kind Regards
Stefan

[rusty1s]

The PPI example denotes a multi-label classification problem (although it is not mesh-based). It uses torch.nn.BCEWithLogitsLoss() for loss computation.