Train models

Input

To train a model, users need to provide:

• a demographic model in DEMES format
• a GAISHI configuration file in YAML format

Feature-vector models

Feature-vector training is used for:

• logistic_regression
• extra_trees_classifier

Run training with:

gaishi train \
    --demes examples/demog/ArchIE_3D19.yaml \
    --config examples/configs/lr.config.yaml \
    --output examples/results/models/example.lr.onnx

To train an extra-trees classifier, use the same command with model.name set to extra_trees_classifier in the configuration file.

UNet++ models

UNet++ training uses genotype matrices stored in HDF5 format.

Run training with:

gaishi train \
    --demes examples/demog/ArchIE_3D19.yaml \
    --config examples/configs/unet.config.yaml \
    --output examples/results/models/example.unet.safetensors

Only simulate training data

Users can run simulation without training a model:

gaishi train \
    --demes examples/demog/ArchIE_3D19.yaml \
    --config examples/configs/lr.config.yaml \
    --only-simulation

Output

For feature-vector models, gaishi train creates:

• a training table: <output_dir>/<output_prefix>.tsv
• a trained model file in the ONNX format specified by --output

For UNet++ models, gaishi train creates:

• a training HDF5 file: <output_dir>/<output_prefix>.h5
• a trained model checkpoint in the Safetensors format specified by --output

Settings

Argument	Description
--demes	Path to the DEMES demographic model file.
--config	Path to the GAISHI configuration YAML file.
--output	Path to the trained model output file. Required unless --only-simulation is used.
--only-simulation	Run simulation only and skip model training.