Skip to content

Fix/multi pathogen training#89

Merged
jeffreyHoelzel merged 12 commits into
mainfrom
fix/multi-pathogen-training
Jun 23, 2026
Merged

Fix/multi pathogen training#89
jeffreyHoelzel merged 12 commits into
mainfrom
fix/multi-pathogen-training

Conversation

@jeffreyHoelzel

Copy link
Copy Markdown
Collaborator

Summary

  • added embedding verification before back propagation and optimizer step
  • implemented global valid loss fix so DDP training back-propagates to a globally normalized loss
  • validation no longer uses overlapping windowed residues to reduce duplication chances
  • implemented automatic positive class weighting during training
  • added threshold policy selection arguments
  • added majority vs mean probability aggregation in final classification
  • added label stratification method attempting to balance label support with family groups so each rank gets an equivalent number of positive and negative cases
  • deprecated FFNN-only training in favor of generalized entry-point
  • added convolutional head option to enforce peptide-level consistency
  • implemented CLI flags for various way to append sequence length or not
  • implemented a fix for uneven data loading in DDP using dummy tensors to fill gaps
  • removed stale CLI arguments
  • simple bug fixes and CLI updates

Test

  • run all unit tests
  • run all integration tests
  • verify correctness of outputs on Monsoon HPC

…reducing chances of duplicated residues being counted as one
…sing train loader instead of labels, removed stale CLI args
…an-prob aggregation to prediction/evaluation CLIs
… label support with family grouping so each fold gets a better idea of positive vs negative classes
…ts, added new convolutional head to enforce peptide-level consistency
…of each residue embedding; none, raw (i.e., 1476), or inverse (i.e., 1 / 1476)
…ors in uneven data distribution across ranks, implemented small doc changes
@jeffreyHoelzel jeffreyHoelzel merged commit f147b12 into main Jun 23, 2026
1 check passed
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.

1 participant