f-dangel / curvlinops

89%

main: 92%

LAST BUILD ON BRANCH ggn-linop

branch: main

CHANGE BRANCH

x

Sync Branches

• main
• development
• papyan20traces-density
• lanczos-cache
• partial-boundary-specification
• visual-tour
• empirical-fisher
• doc-eigenvalues
• test-regression
• fisher-mc
• example-fisher-mc
• ci-update
• update-changelog
• submatrices
• submatrix-example
• neumann-inverse
• neumann-inverse-example
• adjoint
• jacobian
• jacobian-transposed
• adjoint-jacobians
• use-built-in-functorch
• fisher-averaging
• hutchinson
• hutchpp
• diagonal
• kfac
• kfac-ce-loss
• kfac-arbitrary-params-order
• kfac-ef
• only-bias-or-weight
• kfac-type-2
• joint-weight-and-bias
• pydocstyle-coverage
• kfac-with-inplace-activations
• fix-typo
• high-dim-outputs
• expand-reduce
• kfac-conv2d
• activation-hessian
• specify-num-data
• kfac-inv
• update-black
• matvec-to-matmat
• hessian-blocks
• fix-device
• bug-kfac-device-change
• ref-internal-mapping
• frobenius-norm
• source-button
• faster-ef
• faster-fisher-mc
• parallelize-grad-output
• parallelize-sample-grad-output
• no-cpu
• kfac-torch-only
• kfac-damping
• kfac-trace
• update-ci
• foof
• bcewithlogitsloss
• retry-kfac-inv
• fix-device-errors
• remove-kfac-side-effect
• lsmr-inverse
• fix-scale
• fix-kfac-scale
• fix-mcfisher-scale
• kfac-state-dict
• adjoint-submatrix
• remove-loss-average
• fix-enum
• improve-frobenius
• deprecate-pkg-resources
• matmat-list-shapes
• ekfac
• deprecate-python38
• update-github-actions
• pytorch-linop
• curvature-linop
• hessian-linop
• fix-mc-tests
• bug-dtype-check-deterministic
• improve-visual-tour
• fix-markdown-syntax
• power-iteration-comparison
• issue-138-improve-links
• improve-spectral-density
• kfac-linop
• remove-to_device
• kfac-ema
• ggn-linop

Build # 11670932829

Build Type

Pull #146

github

Committed by f-dangel

Commit Message

[FIX] Convert `bfloat16` to `float32` before calling `.numpy`

Pull Request Pull Request #146: [ADD] Change GGN to purely PyTorch

Run Details

19 of 20 new or added lines in 3 files covered. (95.0%)

56 existing lines in 2 files now uncovered.

1515 of 1710 relevant lines covered (88.6%)

0.89 hits per line