f-dangel / curvlinops

89%

main: 92%

LAST BUILD ON BRANCH parallelize-grad-output

branch: parallelize-grad-output

CHANGE BRANCH

x

Reset

Sync Branches

• parallelize-grad-output
• 3.0.1
• activation-hessian
• activation-hessian-linop
• add-ekfac-benchmark
• add-kfoc-linear-operator
• adjoint
• adjoint-jacobians
• adjoint-submatrix
• alt-ekfac
• always-full-tests
• arbitrary-inputs
• bcewithlogitsloss
• benchmark-compile-everywhere
• benchmark-with-backends
• blockdiagonal-linear-operator
• bug-dtype-check-deterministic
• bug-kfac-device-change
• callable-model-func-curv-operators
• callable-model-func-ef
• callable-model-func-ggn
• callable-model-func-ggn-diagonal
• callable-model-func-hessian
• callable-model-func-jacobian
• callable-model-func-kfac-fx
• ci-update
• codex/replace-darglint-with-ruff-for-linting
• composition-rules
• copilot/add-fishertype-kfactype-enums
• curvature-linop
• deprecate-pkg-resources
• deprecate-python38
• development
• diagonal
• diagonal-linear-operator
• dict-params
• dict-params-test-fixture
• doc-compiler-unfriendly-operators
• doc-eigenvalues
• docs/ekfac-influence-example
• docstring-linting-with-ruff
• double-gramian-contraction
• dtype-device-properties
• ef-linop
• eigh-linear-operator
• ekfac
• ekfac-fx-backend
• ekfac-fx-layer-io
• empirical-fisher
• empirical-gradient-computer
• empirical-risk-mixin
• eva
• example-benchmark
• example-fisher-mc
• example-trace
• execute-tests-in-float64
• expand-reduce
• extract-base-kfac-computer
• extract-gradient-and-loss
• fail-first
• faster-ef
• faster-fisher-mc
• feat/astra-preconditioner-support
• feat/kfac-adaptive-damping
• fisher-averaging
• fisher-linop
• fisher-mc
• fix-compiled-memory-measurement
• fix-coveralls-badge
• fix-device
• fix-device-errors
• fix-ekfac-performance
• fix-enum
• fix-fx-requires-grad
• fix-inverse-kfac
• fix-kfac-scale
• fix-markdown-syntax
• fix-mc-tests
• fix-mcfisher-scale
• fix-python-badge
• fix-scale
• fix-typo
• flake8-to-ruff
• foof
• frobenius-norm
• functional-estimators
• functional-ggnmp
• functorch-efmp
• functorch-fmp
• functorch-ggnmp
• functorch-hmp
• functorch-jacobians
• further-fix-ekfac-performance
• fx-kfac-reshape-wrapped-bias
• fx-requires-grad-cleanup
• ggn-diagonal
• ggn-diagonal-pr
• ggn-linop
• ggn-mc
• hessian-blocks
• hessian-linop
• high-dim-outputs
• hutchinson
• hutchpp
• implement-rmatmul
• improve-benchmark-results
• improve-frobenius
• improve-spectral-density
• improve-tests
• improve-visual-tour
• io-collector
• io-collector-linear-reshapes
• io-collector-pr1
• io-collector-pr2
• issue-138-improve-links
• jacobian
• jacobian-linop
• jacobian-transposed
• joint-weight-and-bias
• kfac
• kfac-arbitrary-params-order
• kfac-as-blockdiag-kronecker
• kfac-canonicalization
• kfac-ce-loss
• kfac-computer
• kfac-conv2d
• kfac-damping
• kfac-ef
• kfac-ema
• kfac-fx-backend-v2
• kfac-inv
• kfac-linop
• kfac-properties-to-methods
• kfac-state-dict
• kfac-torch-only
• kfac-trace
• kfac-type-2
• kfac-with-inplace-activations
• kfoc-fx-layer-io
• kfoc-mc-support
• kronecker-linear-operator
• kronecker-product
• lanczos-cache
• layer-io-abstraction
• lifebook-rescue/ggn-log-spectrum-example
• load-state-dict-model-func
• logo
• loss-hessian-any-loss
• loss-hessian-core-routines
• lsmr-inverse
• main
• matmat-list-shapes
• matvec-to-matmat
• merge-adjoint-and-is-vec-tests
• merge-linting-formatting
• migrate-params-to-dict
• modernize-type-annotations
• move-util-linops
• neumann-inverse
• neumann-inverse-example
• no-cpu
• only-bias-or-weight
• overwrite-model-warning
• papyan20traces-density
• parallelize-sample-grad-output
• parallelize-tests
• parameter-groups
• parameter-usage
• partial-boundary-specification
• power-iteration-comparison
• pr1_deprecate_py
• pr2
• pr3
• prepare-3-0-1
• pydocstyle-coverage
• pytorch-cg
• pytorch-inverse
• pytorch-linop
• pytorch-spectral-density
• reduce-scipy
• reduce-test-parametrize
• ref-internal-mapping
• refactor-data-prediction-loss-gradient
• remove-activation-hessian
• remove-block-hessian
• remove-caching
• remove-ekfac-activation-caching
• remove-kfac-inverse
• remove-kfac-side-effect
• remove-loss-average
• remove-to_device
• remover-iterator-conversion
• rename-model-func-to-model-module
• retry-kfac-inv
• rla-pytorch
• ruff-latest-conventions
• runame-patch-1
• shorten-benchmark-labels
• simplify-grad-output-fn
• simplify-make-functional-call
• source-button
• specify-num-data
• submatrices
• submatrix-example
• submatrix-pytorch
• support-bce-non-binary
• support-division
• support-sequences-in-grad-output-sampler
• support-sequences-in-hessian-sqrt
• test-consecutive-matmats
• test-regression
• torch-compile-ef
• torch-compile-ggn
• torch-compile-hessian
• torch-compile-kfac-fx
• torch-compile-kfac-fx-precompute
• torch-compile-test
• update-benchmark-a6000
• update-black
• update-changelog
• update-ci
• update-github-actions
• use-built-in-functorch
• use-model-func-in-make-batch
• use-params-context-manager
• use-pyproject-toml
• use-tuples-in-functionals
• v3-release-notes
• visual-tour
• weight-tying
• weighted-ce-kfac
• xdiag
• xtrace

Build # 8249825057

Build Type

Pull #88

github

Committed by f-dangel

Commit Message

[REF] Parallelize `grad_output` computation in empirical Fisher

Pull Request Pull Request #88: [REF] Parallelize `grad_output` computation in empirical Fisher

Coverage Stats

967 of 1085 relevant lines covered (89.12%)

0.89 hits per line