What's Changed (Initial release)
- infrastructure and shmoof models by @matsen in #1
- Hyperparam opt, more models, more flexible training by @matsen in #2
- Fix branch length optimization using correct loss; WiggleAct by @matsen in #5
- Adding per-base inference by @matsen in #9
- Branch length optimization by @matsen in #11
- masking from child sequences; DNSM model bugfix; Yun branch lengths by @matsen in #14
- Add CI Tests by @willdumm in #21
- Standardization fixes; less radical LR reset by @matsen in #20
- Delete experiment.py by @matsen in #23
- syncing with epam update; better tensorboard; refactoring by @matsen in #26
- parallel branch length optimization by @matsen in #28
- Ability to parallelize between GPUs by @matsen in #30
- Simplified handling of DNSM datasets by @matsen in #32
- Bring over epam code to avoid circular dependency by @matsen in #34
- More flexible training; device bugfix; more hyperparams in yml; renaming to weight_decay by @matsen in #36
- Bring tests/test_sequences.py from epam by @matsen in #38
- Bringing over tests/test_molevol.py by @matsen in #40
- Try "warm up" phase by @matsen in #41
- Add ability to get attention maps by @matsen in #43
- Further development of attention maps; no weight decay for 1D parameters by @matsen in #45
- Minor fixes; pyproject.toml file by @matsen in #47
- Fix issues having to do with standardization by @matsen in #49
- don't record loss before training, which fixes incorrect logging by @matsen in #53
- Add
codon_prob.py
with a model to adjust codon probs by hit class by @willdumm in #50 - Format Docstrings by @willdumm in #59
- Refactor neutral_aa_mut_probs to return per-AA information by @matsen in #62
- First approximation to a DASM: 20 output dimensions but same loss by @matsen in #64
- Adding a per-AA loss to the DASM by @matsen in #66
- Better DASM handling of ambiguous amino acids by @matsen in #68
- Integrate the multihit model into the DNSM framework by @willdumm in #71
- Docstrings; multihit device fix by @matsen in #73
- Split forward functions in DXSM models by @willdumm in #74
- Renaming to CSP where appropriate, and other related things by @matsen in #75
- Replacing normalize_sub_probs with a check; fixing consistency problems by @matsen in #77
- Add loss weights keyword argument to DASMBurrito constructor by @willdumm in #78
- Release cleanup; weights-only crepe loading by @matsen in #80
- Release of thrifty models; pretrained module; demo notebook by @matsen in #81
- moving shared code to dxsm.py; add_shm_model_outputs_to_pcp_df by @matsen in #83
- Publish to PyPI by @willdumm in #82
New Contributors
Full Changelog: https://github.com/matsengrp/netam/commits/v0.2.0