0.3.0 - Loss & Callback Refactors

Config Updates

Due to large refactors of the loss functions and callbacks, it is highly advised you reset your configs.
anemoi-training config generate --override

In particular, the training and diagnostics configurations have greatly changed.

Added

• training_loss
• validation_metrics
• variable_loss_scaling

Removed

• loss_scaling

Changed

• callbacks

What's Changed

• Feature/improve loss functions by @HCookie in #70
• Refactor Callbacks by @HCookie in #60
• Bind max steps and lr iterations by @Rilwan-Adewoyin in #67
• feature: sub hour timesteping by @JesperDramsch in #63
• added config for bounding by @gabrieloks in #10
• fix: metric ranges in the validation space not the normalized space by @sahahner in #116
• fix: enable learningrate monitor automatically by @sahahner in #119
• Mlflow benchmark profiler update by @anaprietonem in #38
• Rename frequency to batch_frequency in RolloutEval by @HCookie in #118
• Add expansion of params to logger by @HCookie in #91
• Fix missing checkpoint callbacks by @HCookie in #125
• Change how mlflow measures CPU memory usage by @cathalobrien in #94
• Graph config for stretched grid graph by @havardhhaugen in #133
• Graph config for limited area models (LAMs) by @JPXKQX in #134
• Feature/update graph callbacks by @JPXKQX in #135
• fix: Rename loss_scaling to variable_loss_scaling by @HCookie in #138
• fix: Allow updates to scalars & other improvements by @HCookie in #137

New Contributors

• @Rilwan-Adewoyin made their first contribution in #67
• @gabrieloks made their first contribution in #10
• @cathalobrien made their first contribution in #94
• @havardhhaugen made their first contribution in #133

Full Changelog: 0.2.2...0.3.0

0.2.2 - Maintenance: pin python <3.13

What's Changed

This release pins python <3.13 due to a missing dependency distribution. Support for python 3.13 is expected in the near future.

Changed

• Lock python version <3.13 #107

Commit Log

• Fix version import by @HCookie in #104
• Add contributors file. by @mchantry in #106
• Lock python version to <3.13 by @HCookie in #107

New Contributors

• @mchantry made their first contribution in #106

Full Changelog: https://github.com/ecmwf/anemoi-training/blob/develop/CHANGELOG.md

0.2.1 - Bugfix: resuming mlflow runs

What's Changed

Added

• Mlflow-sync to include new tag for server to server syncing #83
• Mlflow-sync to include functionality to resume and fork server2server runs #83
• Rollout training for Limited Area Models. #79
• Feature: New Boolean1DMask class. Enables rollout training for limited area models. #79

Fixed

• Mlflow-sync to handle creation of new experiments in the remote server #83
• Fix for multi-gpu when using mlflow due to refactoring of _get_mlflow_run_params function #99
• ci: fix pyshtools install error #100

Changed

• Update copyright notice

Commit Log

• Add output mask by @JPXKQX in #79
• Fix/mlflow sync tag by @anaprietonem in #83
• Ci/fix pyshtools installation error by @theissenhelen in #100
• Update copyright notice by @b8raoult in #101
• Fix/mlflow multi gpu by @anaprietonem in #99

Full Changelog: https://github.com/ecmwf/anemoi-training/blob/develop/CHANGELOG.md#021---bugfix-resuming-mlflow-runs---2024-10-24

0.2.0 - Feature release

What's Changed

This release brings some changes to the default config. Use the anemoi-training config CLI to re-generate your default config.

Added

• Add anemoi-transform link to documentation
• Codeowners file (#56)
• Changelog merge strategy (#56)

Miscellaneous

• Introduction of remapper to anemoi-models leads to changes in the data indices. Some preprocessors cannot be applied in-place anymore.

Functionality

• Enable the callback for plotting a histogram for variables containing NaNs
• Enforce same binning for histograms comparing true data to predicted data
• Fix: Inference checkpoints are now saved according the frequency settings defined in the config #37
• Feature: Add configurable models #50
• Feature: Authentication support for mlflow sync - #51
• Feature: Support training for datasets with missing time steps #48
• Feature: AnemoiMlflowClient, an mlflow client with authentication support #86
• Long Rollout Plots

Fixed

• Fix TypeError raised when trying to JSON serialise datetime.timedelta object - #43
• Bugfixes for CI (#56)
• Fix mlflow subcommand on python 3.9 #62
• Show correct subcommand in MLFlow - Addresses #39 in #61
• Fix interactive multi-GPU training #82
• Allow 500 characters in mlflow logging #88

Changed

• Updated configuration examples in documentation and corrected links - #46
• Remove credential prompt from mlflow login, replace with seed refresh token via web - #78
• Update CODEOWNERS

Commit Log

• fix: histogram only on non-nan values by @sahahner in #15
• fix: saving frequency bug for inference checkpoints by @anaprietonem in #37
• [pre-commit.ci] pre-commit autoupdate by @pre-commit-ci in #36
• feature: make models instantiateable by @JesperDramsch in #50
• Feat: adding in ability to configure precip like plots by @da-ewanp in #49
• Changes to data indices in anemoi models by @sahahner in #17
• fix: remapped variables in tests need to be dictionaries by @sahahner in #52
• Fix - datetime.timedelta to string conversion for JSON by @gareth-j in #43
• 40 support dataset with missing timesteps by @JPXKQX in #48
• Expanded intro docs and examples by @gareth-j in #46
• Chore/multiple fixes ci precommit by @theissenhelen in #56
• 23 plot a validation sample after long rollout by @sahahner in #26
• [fix] Capture Anemoi Training subcommands in MLFlow by @JesperDramsch in #61
• fix version pinning by @MeraX in #66
• Fix mlflow command on python 3.9 by @gmertes in #62
• Authentication support for mlflow sync by @gmertes in #51
• New mlflow authentication API by @gmertes in #78
• add link to transform by @b8raoult in #81
• Allow for longer truncation when mlflow > 1.28 by @HCookie in #88
• Update CODEOWNERS by @b8raoult in #90
• Fix interactive multi-GPU training by @gmertes in #82
• Add AnemoiMlflowClient with auth support by @gmertes in #86
• [pre-commit.ci] pre-commit autoupdate by @pre-commit-ci in #74
• Make pin_memory of the Dataloader configurable by @MeraX in #64

New Contributors

• @anaprietonem made their first contribution in #37
• @da-ewanp made their first contribution in #49
• @MeraX made their first contribution in #66
• @b8raoult made their first contribution in #81
• @HCookie made their first contribution in #88

Full Changelog: https://github.com/ecmwf/anemoi-training/blob/develop/CHANGELOG.md#020---feature-release---2024-10-16

0.1.0 - Anemoi training - First release

What's Changed

Added

Subcommands

• Subcommand for training anemoi-training train
• Subcommand for config generation of configs
• Subcommand for mlflow: login and sync
• Subcommand for checkpoint handling

Functionality

• Searchpaths for Hydra configs, to enable configs in CWD, ANEMOI_CONFIG_PATH env, and .config/anemoi/training in addition to package defaults
• MlFlow token authentication
• Configurable pressure level scaling

Continuous Integration / Deployment

• Downstream CI to test all dependencies with changes
• Changelog Status check
• Readthedocs PR builder
• Changelog Release Updater Workflow

Miscellaneous

• Extended ruff Ruleset
• Added Docsig pre-commit hook
• __future__ annotations for typehints
• Added Typehints where missing
• Added Changelog
• Correct errors in callback plots
• fix error in the default config
• example slurm config

Changed

Move to Anemoi Ecosystem

• Fixed PyPI packaging
• Use of Anemoi models
• Use of Anemoi graphs
• Adjusted tests to work with new Anemoi ecosystem
• Adjusted configs to reasonable common defaults

Functionality

• Changed hardware-specific keys from configs to ??? to trigger "missing"
• __len__ of NativeGridDataset
• Configurable dropout in attention layer

Docs

• First draft on Read the Docs
• Fixed docstrings

Miscellaneous

• Moved callbacks into folder to facilitate future refactor
• Adjusted PyPI release infrastructure to common ECMWF workflow
• Bumped versions in Pre-commit hooks
• Fix crash when logging hyperparameters with missing values in the config
• Fixed "null" tracker metadata when tracking is disabled, now returns an empty dict
• Pinned numpy<2 until we can test all migration
• ci: path ignore of docs for downstream ci
• ci: make python QA reusable
• ci: permissions on changelog updater

Removed

• Dependency on mlflow-export-import
• Specific user configs
• len function of NativeGridDataset as it lead to bugs

Release Work

• @gmertes made their first contribution in #2
• @theissenhelen made their first contribution in #11
• @sahahner made their first contribution in #24
• @JesperDramsch made their first contribution in #28
• @mc4117 and @gareth-j made their first contribution in #25

Original Contributions in AIFS by

• @ssmmnn11
• @mishooax
• @mchantry
• @JPXKQX
• @anaprietonem
• @gabrieloks
• @Rilwan-Adewoyin
• @stietsche
• @jakob-schloer
• The ECMWF AIFS team.

Full Changelog: https://github.com/ecmwf/anemoi-training/blob/develop/CHANGELOG.md#0.1.0