Skip to content

PyTorch

model_navigator.api.torch

Torch optimize API.

optimize(model, dataloader, sample_count=DEFAULT_SAMPLE_COUNT, batching=True, input_names=None, output_names=None, target_formats=None, target_device=DeviceKind.CUDA, runners=None, profiler_config=None, workspace=None, verbose=False, debug=False, verify_func=None, custom_configs=None)

Entrypoint for Torch optimize.

Perform export, conversion, correctness testing, profiling and model verification.

Parameters:

Name Type Description Default
model torch.nn.Module

PyTorch model object

required
dataloader SizedDataLoader

Sized iterable with data that will be feed to the model

required
sample_count Optional[int]

Limits how many samples will be used from dataloader

DEFAULT_SAMPLE_COUNT
batching Optional[bool]

Enable or disable batching on first (index 0) dimension of the model

True
input_names Optional[Tuple[str, ...]]

Model input names

None
output_names Optional[Tuple[str, ...]]

Model output names

None
target_formats Optional[Union[Union[str, Format], Tuple[Union[str, Format], ...]]]

Target model formats for optimize process

None
target_device Optional[DeviceKind]

Target device for optimize process, default is CUDA

DeviceKind.CUDA
runners Optional[Union[Union[str, Type[NavigatorRunner]], Tuple[Union[str, Type[NavigatorRunner]], ...]]]

Use only runners provided as parameter

None
profiler_config Optional[ProfilerConfig]

Profiling config

None
workspace Optional[Path]

Workspace where packages will be extracted

None
verbose Optional[bool]

Enable verbose logging

False
debug Optional[bool]

Enable debug logging from commands

False
verify_func Optional[VerifyFunction]

Function for additional model verification

None
custom_configs Optional[Sequence[CustomConfig]]

Sequence of CustomConfigs used to control produced artifacts

None

Returns:

Type Description
Package

Package descriptor representing created package.

Source code in model_navigator/api/torch.py
def optimize(
    model: torch.nn.Module,
    dataloader: SizedDataLoader,
    sample_count: Optional[int] = DEFAULT_SAMPLE_COUNT,
    batching: Optional[bool] = True,
    input_names: Optional[Tuple[str, ...]] = None,
    output_names: Optional[Tuple[str, ...]] = None,
    target_formats: Optional[Union[Union[str, Format], Tuple[Union[str, Format], ...]]] = None,
    target_device: Optional[DeviceKind] = DeviceKind.CUDA,
    runners: Optional[Union[Union[str, Type[NavigatorRunner]], Tuple[Union[str, Type[NavigatorRunner]], ...]]] = None,
    profiler_config: Optional[ProfilerConfig] = None,
    workspace: Optional[Path] = None,
    verbose: Optional[bool] = False,
    debug: Optional[bool] = False,
    verify_func: Optional[VerifyFunction] = None,
    custom_configs: Optional[Sequence[CustomConfig]] = None,
) -> Package:
    """Entrypoint for Torch optimize.

    Perform export, conversion, correctness testing, profiling and model verification.

    Args:
        model: PyTorch model object
        dataloader: Sized iterable with data that will be feed to the model
        sample_count: Limits how many samples will be used from dataloader
        batching: Enable or disable batching on first (index 0) dimension of the model
        input_names: Model input names
        output_names: Model output names
        target_formats: Target model formats for optimize process
        target_device: Target device for optimize process, default is CUDA
        runners: Use only runners provided as parameter
        profiler_config: Profiling config
        workspace: Workspace where packages will be extracted
        verbose: Enable verbose logging
        debug: Enable debug logging from commands
        verify_func: Function for additional model verification
        custom_configs: Sequence of CustomConfigs used to control produced artifacts

    Returns:
        Package descriptor representing created package.
    """
    if workspace is None:
        workspace = get_default_workspace()
    if target_formats is None:
        target_formats = DEFAULT_TORCH_TARGET_FORMATS
        if batching:
            target_formats, custom_configs = update_allowed_batching_parameters(
                target_formats=target_formats,
                custom_configs=custom_configs,
            )
        LOGGER.info(f"Using default target formats: {[tf.name for tf in target_formats]}")

    sample = next(iter(dataloader))
    if isinstance(sample, Mapping):
        forward_kw_names = tuple(sample.keys())
    else:
        forward_kw_names = None

    if runners is None:
        runners = default_runners(device_kind=target_device)

    target_formats = enums.parse(target_formats, Format)
    runner_names = enums.parse(runners, lambda runner: runner if isinstance(runner, str) else runner.name())

    if profiler_config is None:
        profiler_config = ProfilerConfig()

    if Format.TORCH not in target_formats:
        target_formats = (Format.TORCH,) + target_formats

    config = CommonConfig(
        framework=Framework.TORCH,
        model=model,
        dataloader=dataloader,
        target_formats=target_formats,
        workspace=workspace,
        sample_count=sample_count,
        _input_names=input_names,
        _output_names=output_names,
        target_device=target_device,
        forward_kw_names=forward_kw_names,
        batch_dim=0 if batching else None,
        runner_names=runner_names,
        profiler_config=profiler_config,
        verbose=verbose,
        debug=debug,
        verify_func=verify_func,
        custom_configs=map_custom_configs(custom_configs=custom_configs),
    )

    models_config = ModelConfigBuilder.generate_model_config(
        framework=Framework.TORCH,
        target_formats=target_formats,
        custom_configs=custom_configs,
    )

    builders = [
        preprocessing_builder,
        torch_export_builder,
        find_device_max_batch_size_builder,
        torch_conversion_builder,
        correctness_builder,
    ]
    if profiler_config.run_profiling:
        builders.append(profiling_builder)
    builders.append(verify_builder)

    package = PipelineManager.run(
        pipeline_builders=builders,
        config=config,
        models_config=models_config,
    )

    return package