Fits the configured model and exports results to model.output_dir. Model files and exported splits depend on the selected implementation.
mosa train --config <your-config>.yaml
| Flag | Type | Required | Description |
|---|---|---|---|
-c, --config |
path | yes | Path to the YAML configuration file. |
-r, --resume |
CKPT |
no | Resume training from an existing Lightning checkpoint (.ckpt). |
-d, --debug |
flag | no | Add effective settings, preprocessing details, epoch metrics, timings and tracebacks. |
Description¶
The command loads data, creates a holdout stratified by model_type when evaluation.test_size > 0, builds the registered model, and calls fit() followed by save_outputs().
For the VAE, setting evaluation.test_size: 0 disables validation, early stopping and validation-monitored checkpoints. The run still produces its final output parquets, but no monitored checkpoints/last.ckpt. Lightning may write an unmonitored checkpoint in its log directory. With a validation split, model.checkpoint_top_k: 0 disables the monitored checkpoints too.
MOFA reports backend iterations. It writes mofa_model.hdf5; validation samples do not drive fitting or early stopping, and --resume is unsupported and warns before fitting from scratch. Non-Gaussian views may not produce reconstruction parquets.
Outputs¶
See Saved outputs for Parquet conventions and model-specific directory layouts.
Limitations¶
- Training does not resume automatically if interrupted. A new run can overwrite exported results while leaving older logs and checkpoints in place. Use a fresh output directory for an independent run.
- A nonzero stratified holdout requires at least two samples per source and enough training and validation samples for all sources. Invalid splits raise an error.