Closed hannah348 closed 1 month ago
Yes that's exactly what's happening. 100% agree it's suboptimal (although not really hindering the perfs). I'm going to update the code to export everything and not just the adapters, not as lightweight but at least reproducible and will push a full checkpoint. It's actually a lot more tricky than I thought with the PEFT library but I'll make it work and post here !
Hey - so everything should be deterministic now ! Would be awesome if you guys can confirm using this new model: https://huggingface.co/vidore/colpali-v1.1
and the code in branch: https://github.com/illuin-tech/colpali/tree/hard-negs (optional but should get you better performance and fixes a padding issue)
The base model version is fixed !
I am trying to run some evaluation but with the example scripts
scripts/infer/run_inference_with_python.py
I get different scores every time I run it.My current hypothesis is that the weights of the
custom_text_proj
are randomly initialized and loading the adapter only adds a delta to the weights in form of the LoRA. Hence, the projection would be different every time I load the model. Could that be the case or is something else going on? How do I load or initialize the weights of the projection?