Marianne Stecklina Say goodbye to configuration hassles jsonargparse 

● Deep Learning Engineer at omni:us ● colleague of Mauricio Villegas, creator of jsonargparse ● early user of jsonargparse How I came to give this talk That’s me! That’s Mauricio! 

Introduction 1 

jsonargparse is an open-source library that helps to configure your code by creating a CLI. *CLI = command line interface 

Why CLIs are great compared to… no interface ● run program differently without modifying code ● easier to use for others graphical user interface (GUI) ● easier to create ● perfect for batch tasks ● perfect for scheduled tasks ● useful in terminal-only environments 

Survey What tools are you currently using? Take the poll in #b05-b06 on Discord 

jsonargparse basics 2 

https://github.com/stecklin/pycon22-jsonargparse ● Compare python 0_argparse.py --help and python 1_cli.py --help ● Run python 1_cli.py PyCON --year 2022 ● Print config using python 1_cli.py --print_config > config.yaml ● Edit config.yaml ● Run python 1_cli.py --config config.yaml ● Run python 1_cli.py --config config.yaml --year 2024 ● Run python 1_cli.py --year 2024 --config config.yaml ● Run PYCON_CONFERENCE_NAME=PyCON PYCON_YEAR=2024 python 2_cli_with_env_variables.py Live Demo 

Key features ● CLIs with minimal code overhead ● possible to parse ○ command line arguments ○ config files ○ environment variables 

Hierarchical configs 3 

Use case: Train image classifier def train(dataset: Cifar10Dataset, model: ConvNet): ... frog 

https://github.com/stecklin/pycon22-jsonargparse ● Run python 3_image_classification.py --help ● Run python 3_image_classification.py --dataset.help Cifar10Dataset ● Run python 3_image_classification.py --config configs/3_config.yaml --dataset.augment maybe ● Run python 3_image_classification.py --config configs/3_config.yaml --dataset.data_dir /home/ghosts/ Live Demo 

Key features ● config (and help) is hierarchical ● config structure always in sync with code ● type checks 

Subclasses 4 

Different datasets & models ConvNet kernel_size: int num_layers: int VisionTransformer num_layers: int num_heads: int hidden_dim: int mlp_dim: int {abstract} Model num_classes: int Cifar10Dataset ImageNetDataset {abstract} Dataset data_dir: path_dr batch_size: int augment: bool 

https://github.com/stecklin/pycon22-jsonargparse ● Run python 4_subclasses.py --help ● Run python 4_subclasses.py --dataset ImageNetDataset --model ConvNet --print_config ● Run python 4_subclasses.py --config configs/4_config.yaml ● Run python 4_subclasses.py --dataset configs/4_cifar10.yaml --model configs/4_conv_net.yaml ● Run python 4_subclasses.py --config configs/4_config_with_subconfigs.yaml ● Run python 5_model_composition.py --config configs/5_config.yaml Live Demo 

Key features ● switch between subclasses ● compose config into multiple files 

Adoption 5 

PyTorch Lightning 

Links ● jsonargparse: https://github.com/omni-us/jsonargparse ● Example code: https://github.com/stecklin/pycon22-jsonargparse ● Slides: https://speakerdeck.com/stecklin ● Twitter: @MStecklina Thanks for listening!