MEDS/FEMR/CLMBR-T-Base Stack #224
-
|
Hello, all: When I was working with CLMBR-T-Base this spring, I got all tangled up in versioning incompatibilities between the model itself, FEMR, and MEDS. I see you'vve all been working hard on MEDS (very cool work!) Before I pick this up and start work again, can you please confirm what version of each need to be together for me to test CLMBR-T-Base and tune it over my OMOP dataset? Your input would be much appreciated |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
|
Last I checked, the versions specified on the huggingface README ( Note that FEMR 0.2.3 requires MEDS v1 (https://github.com/som-shahlab/femr/blob/0.2.3/pyproject.toml#L18). Note that I would generally recommend training from scratch if at all possible (https://github.com/som-shahlab/femr/blob/0.2.3/tutorials/6_Train%20MOTOR.ipynb). That's the best way to avoid vocabulary issues (where CLMBR-T-Base's vocab doesn't match the vocab of your dataset). |
Beta Was this translation helpful? Give feedback.
-
|
Thanks, Ethan, much appreciated. |
Beta Was this translation helpful? Give feedback.
Last I checked, the versions specified on the huggingface README (
pip install torch==2.1.1 femr==0.2.3 datasets==2.15.0 xformers transformers==4.35.2) are still correct.Note that FEMR 0.2.3 requires MEDS v1 (https://github.com/som-shahlab/femr/blob/0.2.3/pyproject.toml#L18).
Note that I would generally recommend training from scratch if at all possible (https://github.com/som-shahlab/femr/blob/0.2.3/tutorials/6_Train%20MOTOR.ipynb). That's the best way to avoid vocabulary issues (where CLMBR-T-Base's vocab doesn't match the vocab of your dataset).