Trak Versions Save

Bug fixes and enhancements:

• fix excessive CPU memory load from logging
• remove unnecessary dependencies between featurizing & scoring methods
• fix abstraction violation in projectors
• bring back an example of iterative (as opposed to functional) gradient computer
• add a ridge regularization option for the computation of the XTX inverse term in the TRAK estimator.

• community extensions in trak/contrib (see CONTRIBUTING.md)
• updates to docs, tests, and README

0.3.0 by @kristian-georgiev and @AlaaKhaddaj in https://github.com/MadryLab/trak/pull/50

• Added support for large models and datasets (ChunkedCudaProjector, (much) faster scoring by removing I/O bottlenecks)
• Allow taking gradients with respect to a selected set of parameter groups (e.g., only wrt last layer)
• black codestyle
• bug fixes

What's Changed

• 0.2.2 by @kristian-georgiev in https://github.com/MadryLab/trak/pull/49

• more tests

• better formatting

• minor bug fixes:

  • controllable random seed for projector

  • fix bug with dtype and device of gradients

  • fix bug with init_projector when device is CPU

Co-authored-by: Sung Min Park spark@mslurm Co-authored-by: Alaa Khaddaj [email protected]

Full Changelog: https://github.com/MadryLab/trak/compare/v0.2.1...v0.2.2

What's Changed

• bug fixes

• updated docs

Full Changelog: https://github.com/MadryLab/trak/compare/v0.2.0...v0.2.1

Some (mild) backward incompatibilities introduced. In particular, exp_name is now a required argument when scoring.

What's Changed

• handle pre-emption for featurizing

• support scoring & featurizing data shards in parallel

• reduce memory footprint by ~1.5x

• migrate to torch.func

• bump torch dep requirement to 2.0.0 because of torch.func

• python >=3.8 for pytorch 2.0

• project and store in float16 by default

• tie experiment name to scoring targets; simplify saver; add logging

• save scores as mmap

• normalization factor for numerical stability

• clean up quickstart

• no-op projector

• pass in an instance of a class for tasks, rather than init inside of gradientcomputer

• bug fixes

New Contributors

@AlaaKhaddaj made their first contribution in #38

Full Changelog: https://github.com/MadryLab/trak/compare/v0.1.3...v0.2.0

What's Changed

• 0.1.3 by @kristian-georgiev in https://github.com/MadryLab/trak/pull/32

  • allow skipping model IDs in finalize scores

  • allow subclassing of saver and score_computer directly from traker args

  • default to BasicProjector if CudaProjector projeciton step errors out

  • add another type of error that sometime occurs when fast_jl has issues

  • update quickstart notebook

  • Add link to colab with pre-computed trak scores to readme

  • add dropbox links to quickstart nb

  • update training code in quickstart tutorial

  • bump version

Full Changelog: https://github.com/MadryLab/trak/compare/v0.1.2...v0.1.3

What's Changed

• Fix a bit typo README by @guspan-tanadi in https://github.com/MadryLab/trak/pull/18
• Fix bug with custom model output and add test by @jvendrow in https://github.com/MadryLab/trak/pull/25
• 0.1.2 by @kristian-georgiev in https://github.com/MadryLab/trak/pull/28
  • allow capping the CudaProjector batch size; more helpful error msg for too large batch size
  • Fix mismatch of model parameters during scoring
  • update quickstart notebook
  • Require torch 2.0.0 for fast_jl

New Contributors

• @guspan-tanadi made their first contribution in https://github.com/MadryLab/trak/pull/18
• @jvendrow made their first contribution in https://github.com/MadryLab/trak/pull/25

Full Changelog: https://github.com/MadryLab/trak/compare/v0.1.1...v0.1.2

trak v0.1.1

What's Changed

• 0.1.1 by @kristian-georgiev in https://github.com/MadryLab/trak/pull/17

Full Changelog: https://github.com/MadryLab/trak/compare/v0.1.0...v0.1.1

trak v0.1.0