Recommended way of serializing HeteroData instance

[genesis-jamin]

Are there any recommended ways of serializing a HeteroData instance? The use case here is that we have a featurization pipeline that produces HeteroData, and we want to store it on Google Cloud Storage for use in a training pipeline. We're currently just saving things as bytestrings but I'm wondering if there are any official ways of doing it.

[rusty1s]

If you want to be PyG version agnostic, it is recommended to save data instances as dictionary, e.g.:

torch.save(data.to_dict(), path)
HeteroData.from_dict(torch.load(path))

I think using pickling is the easiest and best approach to do it, but there certainly exists other ways, e.g., via databases, text, csv, etc.