[Question] Is it possible to "freeze" embeddings of a certain entity type? #270
Description
Thank you for making this library! I'm a big fan.
When training or finetuning Embedding tables it can be common to freeze all or part of it: https://pytorch.org/docs/stable/generated/torch.nn.Embedding.html#torch.nn.Embedding
I'm wondering if that is possible to do in PyTorch-BigGraph out of the box? E.G. If I have 3 entity types (Users, Tweets, Communities) but the Tweet nodes come with some pre-trained embeddings, and then I want to "freeze" or not update the Tweet node embeddings. As an aside, it seems like warm starting some set of nodes with a pre-trained embedding is not first-class supported but is mockable by manually generating a first checkpoint file.
Reading through some code and documentation suggests "no, you can't freeze some entity embedding table". I would appreciate if somebody can confirm this, and also give me some pointers on what I'd have to fiddle with to fork and add this functionality (happy to submit as a well documented PR as well, if there is interest from maintainers).