[QUESTION] Load the model from path about comet

[avaz]

From my experience what the load_from_checkpoint downloads extra is the models--xlm-roberta-large. You can download it yourself and save alongside the wmt20-comet-da model and set the environment variable TRANSFORMERS_CACHE to point to where the models--xlm-roberta-large is. This will avoid the download.
I'm using comet version 1.1.3.

[ricardorei]

Yep exactly, The downloads you are observing are the hugging face XLM-R weights. You can also cache those weights to avoid download them multiple times.

[ricardorei]

hmmm thats strange because the load_from_checkpoint function does not perform any download. Can you provide more information? for example a snipped of the script?

[ymoslem]

Thanks, Ricardo!

Here is the whole script that I use:

# Calculate COMET
df = pd.DataFrame({"src":source_sentences, "mt":translations, "ref":references})
data = df.to_dict('records')
model_path = "/content/drive/MyDrive/models/wmt20-comet-da/checkpoints/model.ckpt"
model = load_from_checkpoint(model_path)
seg_scores, sys_score = model.predict(data, batch_size=128, gpus=1)
print("COMET:", round(sys_score*100, 2))

Here is a screenshot of what I see.

I think it does not download wmt20-comet-da, but it downloads that 2.24GB every time, and I am not sure what it is. Thanks!

[ymoslem]

Thanks, @avaz and @ricardorei! This worked.

First, download XLM-RoBERTa.

from transformers import AutoModelForMaskedLM 
model = AutoModelForMaskedLM.from_pretrained("xlm-roberta-large", cache_dir="/content/drive/MyDrive/models/")

Then, set the cache folder for Transformers:

import os
os.environ['TRANSFORMERS_CACHE'] = "/content/drive/MyDrive/models/"

OR

export TRANSFORMERS_CACHE="/content/drive/MyDrive/models/"

Finally, install COMET

pip3 install unbabel-comet -q