Pipeline API

Public API for the training pipeline driver.

class pstrain.api.pipeline.PipelineContext(project_dir, experiment='default', config_name='default', feat=<factory>, train=<factory>, split=<factory>, runner=<factory>, sharding=<factory>, description='', resolved_config=None)

Bases: object

Per-run configuration for the training pipeline.

Parameters:

• project_dir (Path)

• experiment (str)

• config_name (str)

• feat (FeatParams)

• train (TrainParams)

• split (SplitParams)

• runner (RunnerParams)

• sharding (ShardingParams)

• description (str)

• resolved_config (ResolvedConfig | None)

__init__(project_dir, experiment='default', config_name='default', feat=<factory>, train=<factory>, split=<factory>, runner=<factory>, sharding=<factory>, description='', resolved_config=None)

Parameters:

• project_dir (Path)

• experiment (str)

• config_name (str)

• feat (FeatParams)

• train (TrainParams)

• split (SplitParams)

• runner (RunnerParams)

• sharding (ShardingParams)

• description (str)

• resolved_config (ResolvedConfig | None)

Return type:

None

all_fileids()

Train + test fileids (post-split). Use audio_fileids() if you need the full corpus before split has run.

Return type:

list[str]

property all_transcription: Path

Master (pre-split) transcription, input to the split task.

property architecture_dir: Path

property audio_dir: Path

audio_fileids(extension='.wav')

All audio fileids, recursively derived from audio/**/*<ext>.

Returns sorted POSIX-style paths relative to the audio directory, without the extension. This is the canonical set of files that feature extraction operates on and is independent of whether the train/test split has run yet.

Parameters:

extension (str)

Return type:

list[str]

config_name: str = 'default'

description: str = ''

property dist_dir: Path

property etc_dir: Path

experiment: str = 'default'

property experiment_dir: Path

feat: FeatParams

property features_dir: Path

property filler_dict: Path | None

Optional filler dictionary; None if not present.

property filler_dict_path: Path

Stable optional dependency path, including when the file is absent.

fingerprint_payload(stage)

Return the explicitly declared resolved values and identity keys.

Parameters:

stage (str)

Return type:

dict[str, Any]

classmethod from_config(project_dir, *, experiment='default', config_name='default', cli_overrides=None)

Build a context by reading project/etc/configs.yaml.

Parameters:

• project_dir (Path | str)

• experiment (str)

• config_name (str)

• cli_overrides (dict[str, Any] | None)

Return type:

Self

property lm_dir: Path

model_dir(target)

Directory for an acoustic model output, e.g. cd-8g.

Parameters:

target (str)

Return type:

Path

model_files(target)

Standard set of files that constitute a trained model directory.

Parameters:

target (str)

Return type:

list[Path]

property models_dir: Path

project_dir: Path

provenance_document(stage)

Serializable provenance, including its effective-config fingerprint.

Parameters:

stage (str)

Return type:

dict[str, Any]

provenance_path(stage)

Content-addressed path for a stage’s effective configuration.

Parameters:

stage (str)

Return type:

Path

provenance_payload(stage)

Canonical effective configuration governing a pipeline stage.

Parameters:

stage (str)

Return type:

dict[str, Any]

read_fileids(split)

Read a fileid list (e.g. ‘train’, ‘test’, ‘dev’).

Returns an empty list if the file doesn’t exist (planning before corpus setup is allowed; tasks that need fileids will fail at run time with a clearer message).

Parameters:

split (str)

Return type:

list[str]

property reports_dir: Path

resolved_config: ResolvedConfig | None = None

runner: RunnerParams

sharding: ShardingParams

property shared_dir: Path

split: SplitParams

train: TrainParams

property trees_dir: Path

validate_training_features()

Reject unsupported native BW front ends before building dependencies.

Return type:

None

exception pstrain.api.pipeline.UnknownTargetError

Bases: KeyError

Raised when a build target is not registered with the pipeline.

pstrain.api.pipeline.build_pipeline(ctx)

Build a pipeline with all pstrain training tasks registered for ctx.

Feature-extraction tasks are expanded to one task per audio file found recursively under the audio directory. An empty or missing audio directory fails immediately during pipeline construction.

Parameters:

ctx (PipelineContext)

Return type:

Pipeline

pstrain.api.pipeline.create_pipeline_context(project_dir, *, experiment='default', config_name='default', cli_overrides=None)

Build a context by reading project/etc/configs.yaml.

PipelineContext is re-exported unchanged, so the public name and the library name are the same class. This factory gives the command line a public-API call frame without altering that class.

Parameters:

• project_dir (Path | str)

• experiment (str)

• config_name (str)

• cli_overrides (dict[str, Any] | None)

Return type:

PipelineContext

pstrain.api.pipeline.run_pipeline(pipeline, target, *, dry_run=False, force=False, jobs=None, verbose=False)

Run a training graph through a concrete public-API call frame.

Parameters:

• pipeline (Pipeline)

• target (str | Path)

• dry_run (bool)

• force (bool)

• jobs (int | None)

• verbose (bool)

Return type:

int

pstrain.api.pipeline.TARGETS

Registered pipeline targets.

pstrain.api.pipeline.DEFAULT_TARGET

Default target selected for a pipeline build.

pstrain.api.pipeline.DEFAULT_CONFIGS

Names of the bundled pipeline configurations.