Refit a trained model with new triplets

[aastha19]

I have a dataset with new triplets and want to fit this data on an already trained model.
The new dataset is in alliance with the data the model was trained on.
Can this be achieved? If so, how?

[lukostaz]

Unfortunately this is not possible at the moment. I agree it would be a useful feature to open source. Keep an eye on future releases 😉

[sumitpai]

You can try this using constant initializer, to initialize your new model with old embeddings. you also need to ensure that model dictionary(model.ent_to_idx, model.rel_to_idx) stays the same with the new data.

However this is not same as continuing training, since the optimizer states wont be the same.

[AbhiPawar5]

Unfortunately this is not possible at the moment. I agree it would be a useful feature to open source. Keep an eye on future releases 😉

@lukostaz does this mean we have to retrain the model from start if we want to add new triplets? What do you suggest for adding a KG in production that needs to be updated frequently? Thanks.