Quickstart

Quickstart#

Overview#

This guide walks through the minimum steps to get started with Neural-LAM:

  1. Set up example data

  2. Create a graph

  3. Train a model

  4. Evaluate the model

Step 1: Download Example Data#

When running tests for the first time, example data is automatically downloaded from S3. You can trigger this download by running a minimal test.

# Run a minimal test to trigger the data download
pytest tests/test_training.py -vv -s -k "test_training"

Step 2: Create a Graph#

Before training, you must construct a graph mesh for your data.

python -m neural_lam.create_graph_with_wmg --config_path <path-to-config> --name <graph-name>

This script builds the mesh graph required by the GNN models.

Step 3: Train a Model#

Now you can train a model using the graph and configuration.

python -m neural_lam.train_model --config_path <path-to-config> --model graph_lam --graph <graph-name>

Neural-LAM supports several models like graph_lam, hi_lam, and hi_lam_parallel.

Step 4: Evaluate#

After training, you can evaluate the model on the test set by loading the saved checkpoint.

python -m neural_lam.train_model --eval test --config_path <path-to-config> --model graph_lam --graph <graph-name> --load <checkpoint-path>

Configuration#

Neural-LAM uses a YAML configuration system powered by dataclass-wizard. The configuration selects the datastore and sets training options such as state feature weighting and output clamping; model hyperparameters are train_model CLI flags. For complete details, see the API reference for neural_lam.config.NeuralLAMConfig.