v1.3.2

What's Changed

• Hotfix SGMCMC metrics by @svwingerden in #101

Full Changelog: v1.3.1...v1.3.2

v1.3.1

FIX SweepConfig missing init method

v1.3.0

What's Changed

• Fix seeding across chains & in init_loss (NB seeding not backwards reproducible) by @svwingerden in #98
• added an SGMCMC sampler, with prior and preconditioning. This subsumes the earlier SGLD and SGHNT implementations, and allows f.e. RMSProp as a sampler
• deprecation warnings for SGLD and SGNHT (calling SGMCMC with the same arguments should work)
• a small refactor of sampler requirements (backwards compatible)
• cleaned up the requirements
• Fixed small bugs with LLCs on TPU with BF16 and TPUv2/v3s
• more tests, including snapshot tests so we know if we break something

Full Changelog: v1.2.0...v1.3.0

v1.2.0

What's Changed

• Port over changes from monorepo by @svwingerden in #97
• Early September Improvements + Fixes by @wz-ml in #95

Full Changelog: v1.1.0...v1.2.0

v1.1.0

What's Changed

• Add Basic Support for TPUs, refactor temp -> nbeta by @jqhoogland in #81
• Will/Non-TPU dependent Integration with Aether by @wz-ml in #92
• Quick fixes (use nbeta by default, though allow passing in temperature as a fallback as well) by @wz-ml in #93

Full Changelog: v1.0.0...v1.1.0

v1.0.0

What's Changed

• All estimate functions now use evaluate instead of criterion, which can be constructed using the helper function make_evaluate(criterion). A custom make_evaluate or evaluate_function allows preprocessing the data, making the package compatible with transformers. Example evaluate functions evaluate_mse and evaluate_ce can be imported from utils.py.
• Update estimate_learning_coeff functions by @georgeyw in #77
• Add support for non-regression loss functions by @jqhoogland in #80
• RLLC v1 by @svwingerden in #74
• Add gradient accumulation for LLC estimation by @georgeyw in #76
• Fixing two issues in MalaAcceptanceRate callback by @rohan-hitchcock in #87
• Epsilon-Beta Sweeping & Plotting by @wz-ml in #86

New Contributors

• @rohan-hitchcock made their first contribution in #87
• @wz-ml made their first contribution in #86

Full Changelog: v0.2.2...v1.0.0

v0.2.2

Removed an unnecessary warning, and fixed requirements to work with Python 3.8

Full Changelog: v0.2.0...v0.2.2

v0.2.0

Merge pull request #69 from timaeus-research/stan/fixes

Refactor num_samples / temperature, rename elasticity to localization, add warnings, init_loss before sampling, OnlineLLC fix