- Updated minimum Python version to 3.9
- Fixed GitHub Actions workflow (quoted Python version, added setup-python step)
- Added support for MLFlow with
mlflow>=3.10andazureml-mlflow>=1.60as dependencies - Separated AzureML and MLflow initialization paths in
parse_user_args - Fixed process hanging after
--offlinecompletion when no TensorBoard server is needed - Made gradient norm (
gNorm) optional in training log regex to support older Marian log formats - Updated unit tests to match current parser output
- Updated protobuf to >=5.29.5, tensorboard to >=2.15 and tensorboardX to >=2.6
- Made azureml-core and mlflow optional dependencies (install with
pip install marian-tensorboard[azureml]or[mlflow]) - Updated minimum Python version to 3.8
- Suppressed deprecation warnings from TensorBoard's vendored dependencies
- Fixed a bug when parsing logs without batch or total labels (GitHub issue #12)
- Added
--stepoption for choosing which stat to use for TensorBoard step - Parsing additional statistics: WPS, gradient norm, label counts
- Updated logic for
--toolin cases where AzureML is not detected
- Added
--tooloption for choosing logging into TensorBoard or AzureML or both - Removed
--azuremloption - Setting
--port 0will skip starting a local TensorBoard server - Fixed a bug with loading version number when calling the script directly
- Trying to set working directory on Azure ML to
/tb_logs - Creating working directory if it does not exist
- Added
--update-freqoption
- Added parsing learning rates
- Added parsing log files with logical epochs
- Exit with error code if log file does not exist
- Set
--azuremlautomatically when running on Azure ML - Created package: marian-tensorboard
- Added integration for TensorBoard and Azure ML Metrics
- Started the project at MTMA 2022