Skip to content

Run

Stages

Lighter uses PyTorch Lightning's Trainer to manage deep learning experiments. The available stages are:

  • fit: Train your model on training data.
  • validate: Evaluate model performance on validation data
  • test: Evaluate final model performance on test data
  • predict: Generate predictions on new data
  • lr_find: Find optimal learning rate
  • scale_batch_size: Find largest batch size that fits in GPU memory

For documentation on each, please refer to PyTorch Lightning: fit, validate, test, predict, lr_find, scale_batch_size.

Running a Stage

The basic command to run a stage of an experiment is:

lighter <stage> config.yaml

Where:

  • <stage> is one of the stages mentioned above (e.g., fit, validate, test, predict, lr_find, scale_batch_size).
  • config.yaml is the configuration file that defines your experiment. You can also define multiple config files separated by commas, which will be merged (e.g., config1.yaml,config2.yaml).

For example, to train your model, you would use:

lighter fit config.yaml

Passing arguments to a stage

To pass arguments to a stage, use the args section in in your config. For example, to set the ckpt_path argument of the fit stage/method in your config:

args:
    fit:
        ckpt_path: "path/to/checkpoint.ckpt"

or pass/override it from the command line:

lighter fit experiment.yaml --args#fit#ckpt_path="path/to/checkpoint.ckpt"

Recap and Next Steps

You now know how to run different stages of your experiment using Lighter. Next, explore the Project Module to learn how to organize your project and reference custom modules in your configuration.