Skip to content

Ume model, per-modality loss, UmeTokenizerTransform#52

Merged
ncfrey merged 24 commits intomainfrom
lobster-ume-modality-loss
Mar 25, 2025
Merged

Ume model, per-modality loss, UmeTokenizerTransform#52
ncfrey merged 24 commits intomainfrom
lobster-ume-modality-loss

Conversation

@karinazad
Copy link
Collaborator

@karinazad karinazad commented Mar 24, 2025

  • expands the Ume model with modality information and switches training configs from FlexBERT to Ume
  • Ume logs loss and perplexity per modality
  • implements UmeTokenizerTransform that adds modality information to batches
              {
                    "input_ids": tensor(...)
                    "attention_mask": tensor(...)
                    "modality": [Modality.AMINO_ACID, Modality.SMILES, ...]
                }
  • other small housekeeping: remove unused kwargs arguments from FlexBERT, remove unused imports

@karinazad karinazad changed the title [draft] Ume monitor per-modality loss Ume monitor per-modality loss Mar 24, 2025
@karinazad karinazad changed the title Ume monitor per-modality loss [draft] Ume monitor per-modality loss Mar 24, 2025
@karinazad karinazad changed the title [draft] Ume monitor per-modality loss Ume monitor per-modality loss Mar 25, 2025
@karinazad karinazad changed the title Ume monitor per-modality loss Ume model, per-modality loss, UmeTokenizerTransform Mar 25, 2025
----------
modality : str | Modality
The modality of the tokenizer
Can be either one of {"SMILES", "amino_acid", "nucleotide", "3d_coordinates"}
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

reference the constants.ModalityType where these are defined

@ncfrey ncfrey merged commit 183c6c6 into main Mar 25, 2025
5 checks passed
@ncfrey ncfrey deleted the lobster-ume-modality-loss branch March 25, 2025 14:33
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants