validate

Checks a configuration against the input data structure without training.

mosa validate --config <your-config>.yaml

Outputs¶

Config OK
  data:    <data>.h5mu
  views:   ['gexp', 'meth'] (discrete: [])
  model:   MOSAConfig
  output:  <output-dir>
  seed:    42
  eval:    test_size=0.1, n_folds=5, strategy=stratified, shuffle=True
  arch:    fusion=concat, latent=8
  train:   epochs=2, batch_size=8

Checks run cheapest first: YAML parsing and key names, then path existence, then a structural summary of the data, then model-specific requirements.

Shared failures exit with status 1: a missing path or view, a missing mask layer, or no model_type column. Model-specific checks add further failures, such as a VAE target_batch absent from the data when conditional reconstruction is enabled.

Model-specific warnings still exit with status 0. For the MOSA VAE, these include no tissue column while tissue conditioning is on, no mutation_* columns while mutation conditioning is on, or adversarial correction configured with fewer than two model_type categories.

Progress and warnings are written to stderr; the configuration summary stays on stdout. For example, a missing optional mutation input produces [WARNING] No 'mutation_*' columns in data; mutation conditioning will be disabled.

Limitations¶

Validation reads structural metadata without loading matrices and cannot check numerical values. Use inspect for value ranges and observed-entry counts.