v0.2.1

What's Changed

• allow strong preferences from pick_device using a string by @matsen in #85
• Support for dnsm-experiments-1 PR#35 by @willdumm in #86
• dnsm-experiments-1 #39 supporting PR by @willdumm in #87
• DXSM ambiguous sequences by @willdumm in #88
• New Dataset superclass by @willdumm in #89
• Chunked model evaluation by @willdumm in #91

Full Changelog: v0.2.0...v0.2.1

v0.2.0

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

• @matsen made their first contribution in #1
• @willdumm made their first contribution in #21

Full Changelog: https://github.com/matsengrp/netam/commits/v0.2.0