--model and --rescale-with-baseline can not be used together

[L-Zhe]

I use --model to load local model and I use --lang en --rescale-with-baseline to use rescale-with-baseline. The output shows :(hug_trans=3.0.2)-rescaled P: R: F: , but I found the numerical is same with the model without rescale-with-baseline. The command I used is as follow:
bert-score -r XXX -c XXX --model XXX --num_layers 17 --lang en --rescale-with-baseline

[Tiiiger]

hi @L-Zhe because we couldn't precompute the baselines for your local model, you cannot do rescale-with-baseline. This is an intended feature.

[L-Zhe]

hi @L-Zhe because we couldn't precompute the baselines for your local model, you cannot do rescale-with-baseline. This is an intended feature.

Because of network limination, downloading the model directly is very slow, so I have to use the local model downloaded from BaiduPan. How could I do to use rescale with baseline?

[Tiiiger]

hi @L-Zhe, I see the issue here. One workaround would be to put your downloaded model into the huggingface cache folder. You might open an issue to the transformers repo to see how to do this. Here's a. relevant repo huggingface/transformers#2323 .

A quick but less principled hack is to change the model type within this code branch. You would need to clone this repo, modify the code, and install locally from source.

[felixgwu]

Hi @L-Zhe, we just added a --baseline_path argument to specify a custom path of the baseline file in our master branch.
For example, you can just set it to this file if you use roberta-large for English sentences.
We hope this helps.