Marian Tensorboard

TensorBoard integration for Marian NMT. marian-tensorboard generates charts for TensorBoard or Azure ML Metrics from Marian's training logs.

It started as a project at MTMA 2022 and conceptually at MTM 2019.

Installation

Using PyPI:

pip install marian-tensorboard

With optional extras for MLflow or Azure ML support:

pip install marian-tensorboard[mlflow]    # requires Python >= 3.10
pip install marian-tensorboard[azureml]

Locally:

git clone https://github.com/marian-nmt/marian-tensorboard
cd marian-tensorboard
virtualenv -p python3 venv
source ./venv/bin/activate
python3 setup.py install

Both will add new marian-tensorboard command.

Usage

Local machine

marian-tensorboard -f examples/train.encs.*.log

Open a web browser at https://localhost:6006. The script will update the TensorBoard charts every --update-freq seconds unless --offline is used.

Azure ML

marian-tensorboard -f path/to/train.log [-t tb azureml]

Then on Azure Machine Learning VM go to the Metrics tab or start a TensorBoard server under the Endpoints tab.

Note that logging into Azure ML Metrics is automatically enabled if Azure ML Run ID is detected. Specify -t azureml to disable TensorBoard logging. If Azure ML is enabled, the script will not start an own TensorBoard server instance.

Contributors

• Amr Hendy
• Kevin Duh
• Roman Grundkiewicz
• Marcin Junczys-Dowmunt

See CHANGELOG.md.

License

See LICENSE.md.