Experiment Manager#

NeMo’s Experiment Manager leverages PyTorch Lightning for model checkpointing, TensorBoard Logging, Weights and Biases, and MLFlow logging. The Experiment Manager is included by default in all NeMo example scripts.

To use the experiment manager simply call exp_manager and pass in the PyTorch Lightning Trainer.

exp_manager(trainer, cfg.get("exp_manager", None))

And is configurable via YAML with Hydra.

    exp_dir: /path/to/my/experiments
    name: my_experiment_name
    create_tensorboard_logger: True
    create_checkpoint_callback: True

Optionally, launch TensorBoard to view the training results in ./nemo_experiments (by default).

tensorboard --bind_all --logdir nemo_experiments

If create_checkpoint_callback is set to True, then NeMo automatically creates checkpoints during training using PyTorch Lightning’s ModelCheckpoint. We can configure the ModelCheckpoint via YAML or CLI.

    # configure the PyTorch Lightning ModelCheckpoint using checkpoint_call_back_params
    # any ModelCheckpoint argument can be set here

    # save the best checkpoints based on this metric

    # choose how many total checkpoints to save

We can auto-resume training as well by configuring the exp_manager. Being able to auto-resume is important when doing long training runs that are premptible or may be shut down before the training procedure has completed. To auto-resume training, set the following via YAML or CLI:

    # resume training if checkpoints already exist
    resume_if_exists: True

    # to start training with no existing checkpoints
    resume_ignore_no_checkpoint: True

    # by default experiments will be versioned by datetime
    # we can set our own version with
    exp_manager.version: my_experiment_version