[QUESTION] Training my own metric using Comet regression model about comet

[ricardorei]

Yes, typically when you use comet-train you are just loading XLM-R weights (or any other encoder) and finetuning on top.

The only exception is if you use the model flag load_weights_from_checkpoint. This flag can be used to load an existing COMET checkpoint.

[ricardorei]

For our models from WMT21 shared task we actually did that.. we train a checkpoint on DA's and then we fine-tune them in MQM z-scores using that same flag: Are References Really Needed?
Unbabel-IST 2021 Submission for the Metrics Shared Task

[ricardorei]

Btw the advantage of load_weights_from_checkpoint over resume_from_checkpoint is that the architecture doesn't need to be exactly the same! we only load the matching (key, values) between the new model and the checkpoint. While in resume_from_checkpoint the model architecture needs to be exactly the same. Does this make sense?

Let's imagine you want to reuse the weights learned with a Ranking Model... with load_weights_from_checkpoint you will be able to do that and then fine-tune a Regression Model.

[xu1998hz]

Thank you so much for your detailed explanations!