Prepare 1.1.0 release
#23
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…#9) * [ADD] Inverse linear operator based on CG * [DOC] Specify requirements of linop to be inverted * [FIX] Correct return arguments of CG * [ADD] Add CGInverseLinearOperator to API * [DOC] Unfinished example for natural gradient with TODOs * [FIX] Use correct tol * [DOC] Finish content of NGD example * [CI] Ignore darglint for CG hyperparameters * [DOC] Pass over natural gradient example, rename Co-authored-by: Felix Dangel <[email protected]>
* [REF] Extract method to report mismatches in array comparisons * [REF] Extract functorch methods to compute the Hessian, GGN & gradient * [FIX] Overlong lines * [DOC] Add docstrings, ignore if redundant Co-authored-by: Felix Dangel <[email protected]>
* [FIX] Print correct array element * [DOC] Mention more accurate inversion by setting CG parameters * [DOC] Improve documentation * [FMT] Shorten lines * [REF] Improve readability of first description line * [FIX] rst Syntax * [FIX] Reference to eigsh Co-authored-by: Felix Dangel <[email protected]>
* [FIX] Print correct array element * [DOC] Mention more accurate inversion by setting CG parameters * [DOC] Improve documentation * [FMT] Shorten lines * [REF] Improve readability of first description line * [FIX] rst Syntax * [FIX] Reference to eigsh Co-authored-by: Felix Dangel <[email protected]>
Implements tests for - Matrix-vector/matrix products with the GGN/Hessian - Matrix-vector/matrix products with the inverse damped GGN and refactors existing tests. * [FIX] Print correct array element * [DOC] Mention more accurate inversion by setting CG parameters * [DOC] Improve documentation * [FMT] Shorten lines * [REF] Improve readability of first description line * [FIX] rst Syntax * [FIX] Reference to eigsh * [TEST] Add parametrized tests for GGN/Hessian matvec/mat * [REF] Use fixture In order for a fixture to be visible, it has to live in a `conftest.py` file (see https://stackoverflow.com/a/36457475). * [REQ] Add functorch to test requirements * [REF] Clean up and move tests for `check_deterministic` * [REF] Improve BN initialization code * [TEST] Matrix-vector/matrix products with the inverse damped GGN Co-authored-by: Felix Dangel <[email protected]>
Includes an example that reproduces Fig. 15 from the paper. Updates API. Smaller improvements in the documentation. * [REF] Rename LowRank → OuterProductLinearOperator * [ADD] Spectral analysis methods from papyan2020traces * [DOC] Add numpy link to intersphinx * [FIX] Add logo in .png format, specify as default thumb * [DOC] Add spectral density methods to API * [DOC] Add example (reproducing Figure 15 of papyan2020traces) * [FIX] Typo in API definition * [TEST] Integration test for spectral density estimation * [REQ] Add matplotlib to test requirements Co-authored-by: Felix Dangel <[email protected]>
…ns (#15) * [REF] Extract one repeat of spectral approximation * [ADD] Class for density approximation with cached Lanczos * [DOC] Add docstrings * [DOC] Make keyword arguments more explicit * [FIX] Missing argument in docstring * [REF] Extract log spectrum construction from Lanczos iteration * [ADD] Class for log-spectrum approximation with cached Lanczos iterations * [DOC] Fix indentation Co-authored-by: Felix Dangel <[email protected]>
* [DOC] Add setup for visual tour * [ADD] Finish content of visual tour, add matplotlib to intersphinx * [REF] Extract transforms * [FIX] flake8
…17) * [DOC] Add mathematical definitions of GGN and Hessian * [FIX] Set dtype of random vector in deterministic check * [ADD] Linear operator for empirical Fisher * [DEL] Remove autograd double-check in functorch EF implementation * [DOC] Add empirical Fisher to visual tour * [DEL] Remove TODO after failed attempt with `is_grad_batched` * [FIX] Remove unused import * [DOC] Add missing docstrings
Requires time-consuming tests due to verifying the Monte-Carlo approximation. Introduces `test` and `test-light`, the former of which is run when pushing to `master` and `development`. `test` executes all tests, `test-light` skips the expensive Monte-Carlo tests (realized with the `pytest-optional-tests` package). * [ADD] Regression test cases * [ADD] MC approximation of the Fisher * [FIX] Docstring, extract number of classes * [REF] Use larger rtol * [CI] Mark MC tests as expensive, only run on `development` and `master` * [REF] Use smaller rtol, improve convergence check * [REF] Integrate MSE normalization over classes into standard deviation * [REF] Increase atol, decrease number of vectors for matmat test * [DOC] Update features in README * [REF] Only create _generator if it does not exist yet * [REF] Extract batch size, refactor if-else statement
* [CI] Update code to latest flake8 (6.0.0) and black (23.1.0) * [CI] Make flake8 ignore B905 * [FMT] Use TAB instead of two spaces
* [DEL] Remove functorch dependency as it's part of `torch>=1.13` * [DOC] Update changelog * [DOC] Update future supported features
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
See the changelog for details.