intelligence.acuirt.workflow

ConversionWorkflow Objects

class ConversionWorkflow()

Workflow for converting and evaluating models with AcuiRT. This class manages the conversion of a PyTorch model to an optimized format using AcuiRT, evaluates converted models, and profiles the converted model's performance.

__init__

def __init__(model: nn.Module,
             evaluator: AcuiRTEvalProtocol,
             eval_dataset: Iterable[Any],
             eval_dataset_post_process: Optional[Callable] = None,
             conversion_dataset: Optional[Any] = None,
             model_post_process: Optional[Callable] = None,
             model_post_process_rt: Optional[Callable] = None,
             eval_non_converted_model: bool = False,
             settings_torch_profiler: Optional[dict] = None,
             profile_max_depth: int = -1,
             exporters: Optional[List[ExporterProtocol]] = None,
             *,
             logger: Optional[AcuiRTDefaultLogger] = None)

Initialize the conversion workflow.

Arguments:

• model nn.Module - The PyTorch model to convert.
• evaluator AcuiRTEvalProtocol - Evaluator for model performance.
• eval_dataset Iterable[Any] - Dataset for evaluating the model.
• eval_dataset_post_process Optional[Callable] - Post-processing function for evaluation dataset.
• conversion_dataset Optional[Any] - Dataset for model conversion. Defaults to eval_dataset if None.
• model_post_process Optional[Callable] - Post-processing function for the original model outputs.
• model_post_process_rt Optional[Callable] - Post-processing function for the converted model outputs.
• eval_non_converted_model bool - Whether to evaluate the non-converted model. Defaults to False.
• settings_torch_profiler Optional[dict] - Settings for torch profiler during profiling. It is strongly recommended to set 'activities' and 'schedule' keys.
• profile_max_depth int - Maximum depth for profiling nn.Module inference. Defaults to -1 (no limit).
• exporters Optional[List[ExporterProtocol]] - List of exporters to export the workflow report.
• logger Optional[AcuiRTDefaultLogger] - Logger for logging information during conversion.

run

def run(config: Union[dict, AcuiRTBaseConversionConfig,
                      AcuiRTBaseConversionReport],
        export_path: str,
        skip_conversion: bool = False)

Run the conversion workflow.

Arguments:

• config Union[dict, AcuiRTBaseConversionConfig, AcuiRTBaseConversionReport] - Configuration for model conversion.
• export_path str - Path to export the converted model and report. skip_conversion (bool): Whether to skip the conversion step. Defaults to False.

Returns:

Tuple[nn.Module, AcuiRTWorkFlowReport]: The converted model and workflow report