search

UKPLab / sentence-transformers

State-of-the-Art Text Embeddings
https://www.sbert.net
Apache License 2.0
15.25k stars 2.47k forks source link

Is there a way to invoke tensorboard for cross encoder model training? #2430

Open mucun1988 opened 9 months ago

mucun1988 commented 9 months ago

I am wondering whether there is a way to invoke tensorboard for visualization when training cross encoder model? E.g., https://github.com/UKPLab/sentence-transformers/blob/master/examples/training/cross-encoder/training_nli.py#L73

tomaarsen commented 9 months ago

Hello!

I'm afraid that at this time it is not easily possible to get the loss in Tensorboard or W&B during training. The only option currently is to use the callback option and create a function that logs the scores to Tensorboard. However, this is likely more difficult than imagined.

I'm actively working on implementing integrations with third-parties like Tensorboard to make this much, much easier in the future.

l4b4r4b4b4 commented 5 months ago

same question here. Did not see it, so guess one way it to get some info out is the evaluation callback function.

tomaarsen commented 5 months ago

Indeed, that's currently the only way. I'm planning on improving the training over the CrossEncoder much like I've done for the SentenceTransformer. By then, the tensorboard integration will be automatic.

l4b4r4b4b4 commented 5 months ago

sounds great. Can you give a time estimate on when you think youll release it?

Will you include Hyper-param optimisation like with sentence transformer?