Command Line Interface

lightning-template provides several cli scripts for you to use. Generally, you can run all scripts under lightning_tempalet/tools by python -m lightning_template.tools.<path.to.scripts>, for example, use python -m lightning_template.tools.cli to run tools/cli.py. For convenience, we also provide some aliases for these scripts. All of them are as follows:

alias	scripts
cli	lightning_template/tools/cli.py
lr_finder	lightning_template/tools/models/lr_finder.py
batch_size_finder	lightning_template/tools/models/batch_size_finder.py
model_statistics	lightning_template/tools/models/model_statistics.py

We recommend that you use the alias method to launch scripts in the terminal and use the module method to debug scripts in vscode etc.

For commands and options, you can get all available options and commands with a help argument, for example, cli --help.

CLI

Train

You can start an experiment with a command using the alias as follows, in which, gpu_ids is a comma-separated id list or just one int.

CUDA_VISIBLE_DEVICES=<gpu_ids> cli fit --config configs/runs/path/to/config

Samely, in the module method.

CUDA_VISIBLE_DEVICES=<gpu_ids> python -m lightning_template.tools.cli fit --config configs/runs/path/to/config

Validation Test and Predict

Just replace fit by validation, test, or predict as follows:

CUDA_VISIBLE_DEVICES=<gpu_ids> cli {validation, test, predict} --config configs/runs/path/to/config

Tune

auto lr finder

Automatically finding the best learning rate for models currently only supports the first optimizer. You can get more information from the doc.

You can use lr_finder alias or lightning_template.tools.lr_finder module to use this feature with this project.

CUDA_VISIBLE_DEVICES=<gpu_id> lr_finder --config configs/runs/path/to/config

CUDA_VISIBLE_DEVICES=<gpu_id> python -m lightning_template.tools.models.lr_finder --config configs/runs/path/to/config

auto scale batch size

Auto finds the largest batch size or largest power of two as batch size. You can get more information from the doc.

Same as auto lr finder, you can use batch_size_finder alias or lightning_template.tools.batch_size_finder module to use this feature with this project.

Note that, the batch_size_finder script only supports running on only one device, which means even if you list many gpu_ids in CUDA_VISIBLE_DEVICES env var, it will only select the first one to run.

CUDA_VISIBLE_DEVICES=<gpu_id> batch_size_finder --config configs/runs/path/to/config

CUDA_VISIBLE_DEVICES=<gpu_id> python -m lightning_template.tools.models.batch_size_finder --config configs/runs/path/to/config

Model statistics

You can use model_statistics alias or lightning_template.tools.model_statistics module to calculate the params and MACs used by your model.

CUDA_VISIBLE_DEVICES=<gpu_id> model_statistics --config configs/runs/path/to/config

CUDA_VISIBLE_DEVICES=<gpu_id> python -m lightning_template.tools.models.model_statistics --config configs/runs/path/to/config