search

OpenBioML / protein-lm-scaling

Other
54 stars 15 forks source link

fix: training #66

Closed othertea closed 5 months ago

othertea commented 5 months ago

After the recently merged PRs, training with usual configs no longer worked. (I'm not sure how others have been running training scripts; I did not see any new training configs.) This PR makes minimal changes so that we can once again train with the two provided configs: protein_lm/configs/train/toy_localcsv.yaml, protein_lm/configs/train/toy_hf.yaml.

talkhanz commented 4 months ago

Hey @othertea.

I'm running into issues when do_curriculum_learning is true. The fields removed from protein_lm/configs/train/toy_localcsv.yaml and protein_lm/configs/train/toy_hfyaml as per this issue are required for curriculum learning to take place.

Can you let me know if you can reproduce an error when do_curriculum_learning is true in the config file?

othertea commented 4 months ago

Hi @talkhanz , Please feel free to create a new yaml file that does curriculum learning with the desired specifications! As their names indicate, the yamls in the repo are "toy" yamls that are used for sanity checking, not yamls intended to be used for the full training workflows. In particular, they are not guaranteed to work when you modify a single value, such as setting do_curriculum_learning to true. This PR updated the yamls because the previous versions of the yamls did not run.

talkhanz commented 4 months ago

makes sense. Thankyou @othertea!