Add to existing embeddings and retrain

[ssharpe42]

Is there an easy way or framework that is used to

1. Load existing embeddings
2. Add to node vocabulary
3. Retrain with data that contains new nodes

[lw]

I don't think there's anything out-of-the-box for that. Though you should be able to build it yourself. The init_path argument in the config could be a good place to start: it's used to provide a checkpoint of a previous run which will be used to "warmstart" the new run. That old checkpoint must have the exact same entities as the new run though. However, you could try to manually alter a previous checkpoint to artificially insert some entities that weren't there, and give them a random embedding.

Even better would be to do this as part of your importing/exporting scripts. The scripts we provide don't do it AFAIK but you could modify them or write your own ones so that the importer looks up, for each entity, whether the previous exporter has stored an embedding for that entity, and if so it includes it into a "fake" checkpoint that can be passed to init_path.

Hope this helps.