Warning when loading the model

#2
by Pavelrst - opened

First of all - Great model, cool library! Thanks!

code:

from span_marker import SpanMarkerModel
model = SpanMarkerModel.from_pretrained("tomaarsen/span-marker-bert-base-uncased-acronyms")
entities = model.predict("compression algorithms like principal component analysis (pca) can reduce noise and complexity.")
print(entities)

warning:

Some weights of SpanMarkerModel were not initialized from the model checkpoint at tomaarsen/span-marker-bert-base-uncased-acronyms and are newly initialized: ['encoder.embeddings.position_ids']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.

The output look OK, the spans are correct.
How to fix the warning?

Hmm, that is odd! I can reproduce this locally, and also with the cased variant of this model (tomaarsen/span-marker-bert-base-acronyms), but not with any of my other models, even if they use the same encoder (tomaarsen/span-marker-bert-base-fewnerd-fine-super). I'll try to dig into this more.

It's interesting that the performance seems unaffected: I ran the model against the validation set and it's still 91.96 F1, despite getting the warning.

For the time being - I wouldn't worry too much about the warning. There's probably ways to silence it, but it shouldn't affect the model performance.

I've trained and uploaded a new version - this one doesn't have the warning. It performs equivalently, and comes with a fancy new model card!
I'll close this for now :)

tomaarsen changed discussion status to closed

Sign up or log in to comment