AutoModelPipeline

inference_models.AutoModelPipeline

Functions

from_pretrained classmethod

from_pretrained(pipline_id, models_parameters=None, weights_provider='roboflow', api_key=None, max_package_loading_attempts=None, verbose=False, model_download_file_lock_acquire_timeout=FILE_LOCK_ACQUIRE_TIMEOUT, allow_untrusted_packages=False, trt_engine_host_code_allowed=True, allow_local_code_packages=True, verify_hash_while_download=True, download_files_without_hash=False, use_auto_resolution_cache=True, auto_resolution_cache=None, allow_direct_local_storage_loading=True, model_access_manager=None, weights_provider_extra_query_params=None, weights_provider_extra_headers=None, **kwargs)

Load and initialize a multi-model pipeline.

Pipelines are pre-configured workflows that combine multiple models to solve complex computer vision tasks. For example, the "face-and-gaze-detection" pipeline combines a face detector with a gaze estimation model.

Each pipeline has default model configurations, but you can override them by providing custom models_parameters. All models in the pipeline are loaded using AutoModel.from_pretrained() with the same loading parameters.

Parameters:

• pipline_id

  (str) –

  Pipeline identifier. Use list_available_pipelines() to see available options. Examples: "face-and-gaze-detection".

• models_parameters

  (Optional[List[Optional[Union[str, dict, DependencyModelParameters]]]], default: None ) –

  Optional list of parameters for each model in the pipeline. Can be: - None (default): Use pipeline's default models - List of model IDs (strings): ["model1-id", "model2-id"] - List of parameter dicts: [{"model_id_or_path": "model1", "device": "cuda"}, ...] - List of DependencyModelParameters objects (advanced) - Mix of the above (None entries use defaults)

  The list length should match the number of models in the pipeline.

• weights_provider

  (str, default: 'roboflow' ) –

  Source for model weights. Default: "roboflow".

• api_key

  (Optional[str], default: None ) –

  Roboflow API key for accessing private models. If not provided, uses the ROBOFLOW_API_KEY environment variable.

• max_package_loading_attempts

  (Optional[int], default: None ) –

  Maximum number of model packages to try before failing. Default: Try all matching packages.

• verbose

  (bool, default: False ) –

  Enable detailed logging during pipeline and model loading. Default: False.

• model_download_file_lock_acquire_timeout

  (int, default: FILE_LOCK_ACQUIRE_TIMEOUT ) –

  Timeout in seconds for acquiring file locks during concurrent downloads. Default: 10.

• allow_untrusted_packages

  (bool, default: False ) –

  Allow loading model packages with custom code that haven't been verified. Security risk. Default: False.

• trt_engine_host_code_allowed

  (bool, default: True ) –

  Allow TensorRT engines to execute host code. Default: True.

• allow_local_code_packages

  (bool, default: True ) –

  Allow loading models with custom Python code from local directories. Default: True.

• verify_hash_while_download

  (bool, default: True ) –

  Verify file integrity using checksums during download. Default: True.

• download_files_without_hash

  (bool, default: False ) –

  Allow downloading files without checksums. Security risk. Default: False.

• use_auto_resolution_cache

  (bool, default: True ) –

  Enable caching of model resolution results. Default: True.

• auto_resolution_cache

  (Optional[AutoResolutionCache], default: None ) –

  Custom cache implementation. Advanced usage only.

• allow_direct_local_storage_loading

  (bool, default: True ) –

  Allow loading models directly from local paths. Default: True.

• model_access_manager

  (Optional[ModelAccessManager], default: None ) –

  Custom model access control manager. Advanced usage only.

• weights_provider_extra_query_params

  (Optional[List[Tuple[str, str]]], default: None ) –

  Extra query parameters to pass to the weights' provider. Advanced usage only.

• weights_provider_extra_headers

  (Optional[Dict[str, str]], default: None ) –

  Extra headers to pass to the weights' provider. Advanced usage only.

• **kwargs

  –

  Additional pipeline-specific parameters passed to the pipeline's with_models() method.

Returns:

• AnyModel –

  Initialized pipeline instance. The specific type depends on the pipeline.

Raises:

• ModelPipelineInitializationError –

  If pipeline initialization fails, models cannot be loaded, or required parameters are missing.

• UnauthorizedModelAccessError –

  If API key is invalid or model access is denied.

Examples:

Load pipeline with default models:

>>> from inference_models import AutoModelPipeline
>>> pipeline = AutoModelPipeline.from_pretrained("face-and-gaze-detection")
>>> results = pipeline(image)

Load pipeline with custom model parameters:

>>> pipeline = AutoModelPipeline.from_pretrained(
...     "face-and-gaze-detection",
...     models_parameters=[
...         "mediapipe/face-detector",  # Use specific face detector
...         {"model_id_or_path": "l2cs-net/rn50", "device": "cuda"}  # Custom gaze model
...     ]
... )

Load with verbose logging:

>>> pipeline = AutoModelPipeline.from_pretrained(
...     "face-and-gaze-detection",
...     verbose=True
... )

See Also

• AutoModelPipeline.list_available_pipelines(): List all available pipelines
• AutoModel.from_pretrained(): Load individual models

list_available_pipelines classmethod

list_available_pipelines()

Display all registered model pipelines available for loading.

Shows a tree view of all pipeline identifiers that can be used with AutoModelPipeline.from_pretrained(). Pipelines are multi-model workflows that combine multiple models to solve complex computer vision tasks.

Returns:

• None –

  None. Prints a formatted tree to the console showing all registered

• None –

  pipeline identifiers.

Examples:

List available pipelines:

>>> from inference_models import AutoModelPipeline
>>> AutoModelPipeline.list_available_pipelines()
# Displays:
# Available Model Pipelines:
# ├── face-and-gaze-detection
# └── ... (other registered pipelines)

See Also

• AutoModelPipeline.from_pretrained(): Load a pipeline