Language

unilab.algos.torch.hora.distill_config¶

HORA distillation config and teacher-owner resolution helpers.

Functions

`apply_teacher_defaults`(cfg, *[, root_dir])	Merge teacher-owner defaults under the user distillation config.
`get_teacher_owner_spec`(cfg)	Resolve the teacher algo family and task owner from distillation config.
`load_teacher_owner_config`(algo_family, task, *)	Load a HORA teacher owner config and its direct owner defaults.
`resolve_teacher_checkpoint_path`(cfg, *[, ...])	Resolve the selected HORA teacher checkpoint through owner metadata.
`resolved_distill_runtime_cfg`(cfg)	Return checkpoint runtime fields needed to rebuild the student model.
`teacher_default_cfg`(cfg, *[, root_dir])	Build HORA student defaults from the selected teacher owner YAML.
`teacher_run_metadata`(cfg, *, ...[, root_dir])	Build explicit teacher provenance metadata for distillation outputs.

unilab.algos.torch.hora.distill_config.load_teacher_owner_config(algo_family, task, *, root_dir=None)[source]¶

Load a HORA teacher owner config and its direct owner defaults.

Parameters:

algo_family (str)
task (str)
root_dir (str | Path | None)

Return type:

DictConfig

unilab.algos.torch.hora.distill_config.get_teacher_owner_spec(cfg)[source]¶

Resolve the teacher algo family and task owner from distillation config.

Parameters:: cfg (DictConfig)
Return type:: tuple[str | None, str | None]

unilab.algos.torch.hora.distill_config.teacher_default_cfg(cfg, *, root_dir=None)[source]¶

Build HORA student defaults from the selected teacher owner YAML.

Parameters:

cfg (DictConfig)
root_dir (str | Path | None)

Return type:

DictConfig

unilab.algos.torch.hora.distill_config.apply_teacher_defaults(cfg, *, root_dir=None)[source]¶

Merge teacher-owner defaults under the user distillation config.

Parameters:

cfg (DictConfig)
root_dir (str | Path | None)

Return type:

DictConfig

unilab.algos.torch.hora.distill_config.resolved_distill_runtime_cfg(cfg)[source]¶

Return checkpoint runtime fields needed to rebuild the student model.

Stage-2 checkpoints intentionally do not persist owner runtime settings such as env, reward, or domain randomization. Replay should use the currently composed owner config for those fields.

Parameters:: cfg (DictConfig)
Return type:: DictConfig

unilab.algos.torch.hora.distill_config.teacher_run_metadata(cfg, *, teacher_algo_family, teacher_checkpoint, root_dir=None)[source]¶

Build explicit teacher provenance metadata for distillation outputs.

Parameters:

cfg (DictConfig)
teacher_algo_family (str)
teacher_checkpoint (Path)
root_dir (str | Path | None)

Return type:

dict[str, Any]

unilab.algos.torch.hora.distill_config.resolve_teacher_checkpoint_path(cfg, *, root_dir=None)[source]¶

Resolve the selected HORA teacher checkpoint through owner metadata.

Parameters:

cfg (DictConfig)
root_dir (str | Path | None)

Return type:

tuple[Path | None, Path | None]