task-inference 0.1.1b1

Abstractions for AI tasks following the Open Inference Protocol v2, with a HuggingFace Transformers reference implementation

Task Inference

A Python library that provides task-oriented abstractions for AI inference, bridging HuggingFace Transformers pipeline tasks with the Open Inference Protocol v2 (OIP v2) tensor format.

Key Design

Each task is modelled in three layers:

Layer	Role
Protocol (protocol/v2.py)	Pydantic models for OIP v2 InferenceRequest / InferenceResponse
Task (tasks/)	Domain-specific input/output Pydantic schemas; process left abstract
Implementation (implementations/transformers/)	HuggingFace Transformers reference backend that fulfils process

This separation allows swapping backends (ONNX, TensorRT, REST endpoint, …) without changing the task schema or OIP conversion logic.

Input / Output schemas

Every task exposes a pair of Pydantic models (e.g. ImageClassificationInput / ImageClassificationOutput). These models own the conversion to and from OIP v2 tensors:

Method	Direction
XxxInput.to_inference_request()	Python input → InferenceRequest
XxxInput.from_inference_request(request)	InferenceRequest → Python input
XxxOutput.to_inference_response(model_name)	Python output → InferenceResponse
XxxOutput.from_inference_response(response)	InferenceResponse → Python output

The process(inputs: XxxInput) -> XxxOutput method on task classes works entirely with these domain objects — no raw OIP tensors required.

Supported Tasks

Vision

Task	Class	Default Model
Image Classification	TransformersImageClassificationTask	google/vit-base-patch16-224
Image Segmentation	TransformersImageSegmentationTask	facebook/mask2former-swin-large-coco-panoptic
Object Detection	TransformersObjectDetectionTask	facebook/detr-resnet-50
Depth Estimation	TransformersDepthEstimationTask	Intel/dpt-large
Mask Generation	TransformersMaskGenerationTask	facebook/sam-vit-base
Visual Question Answering	TransformersVQATask	dandelin/vilt-b32-finetuned-vqa
Image Anonymization	TransformersImageAnonymizationTask	hustvl/yolos-tiny

Audio

Task	Class	Default Model
Automatic Speech Recognition	TransformersASRTask	openai/whisper-base
Audio Classification	TransformersAudioClassificationTask	superb/wav2vec2-base-superb-ks

Installation

# Core (schemas + OIP v2 protocol only)
pip install task-inference

# With HuggingFace Transformers backend
pip install "task-inference[transformers]"

# With audio support
pip install "task-inference[all]"

Quick Start

Via the factory (recommended)

from task_inference import create_task

with open("cat.jpg", "rb") as f:
    image_bytes = f.read()

task = create_task(
    backend="transformers",
    task_name="image-classification",
    model_name="google/vit-base-patch16-224",
    model_params={"device": "cpu"},
)

from task_inference.tasks.vision.image_classification import (
    ImageClassificationInput,
    ImageClassificationOutput,
)

inp = ImageClassificationInput(image=image_bytes, top_k=3)
resp = task(inp.to_inference_request())
result = ImageClassificationOutput.from_inference_response(resp)
for r in result.results:
    print(r.label, r.score)

model_params is forwarded directly to the backend constructor, so any backend-specific keyword argument (e.g. device, chunk_length_s, points_per_batch) can be passed here.

To discover what backends and task names are available:

from task_inference import supported_tasks

print(supported_tasks())
# {'transformers': ['audio-classification', 'automatic-speech-recognition', ...]}

Direct instantiation

from task_inference.implementations.transformers.vision import (
    TransformersImageClassificationTask,
)
from task_inference.tasks.vision.image_classification import (
    ImageClassificationInput,
    ImageClassificationOutput,
)

task = TransformersImageClassificationTask(model_name="google/vit-base-patch16-224")

# Call via OIP v2 round-trip
inp = ImageClassificationInput(image=image_bytes, top_k=3)
resp = task(inp.to_inference_request())
result = ImageClassificationOutput.from_inference_response(resp)
for r in result.results:
    print(r.label, r.score)

# Convenience wrapper - build input from keyword arguments, returns InferenceResponse
resp = task.run(image=image_bytes, top_k=3)
result = ImageClassificationOutput.from_inference_response(resp)

OIP v2 round-trip

The input/output models handle all serialisation, so you can integrate with any OIP v2-compatible server without touching the task implementation:

from task_inference.tasks.vision.image_classification import (
    ImageClassificationInput,
    ImageClassificationOutput,
)

# --- Client side ---
inputs  = ImageClassificationInput(image=image_bytes, top_k=3)
request = inputs.to_inference_request()   # → InferenceRequest (send over HTTP)

# --- Server side ---
response = task(request)                  # returns InferenceResponse directly

# --- Client side (parse response) ---
output = ImageClassificationOutput.from_inference_response(response)
for r in output.results:
    print(r.label, r.score)

Factory reference

create_task(backend, task_name, model_name=None, model_params=None)

Parameter	Type	Description
backend	str	Backend name — currently "transformers"
task_name	str	Task identifier (see table below)
model_name	str | None	HuggingFace model id or local path. When None the backend's built-in default model is used.
model_params	dict | None	Extra keyword arguments passed to the constructor (e.g. device, chunk_length_s)

Raises ValueError for unknown backends or task names.

supported_tasks(backend=None)

Returns a dict[str, list[str]] mapping each backend to its supported task names. Pass a backend name to filter to a single backend.

Task name reference

Task name	Input class	Output class
image-classification	ImageClassificationInput	ImageClassificationOutput
object-detection	ObjectDetectionInput	ObjectDetectionOutput
depth-estimation	DepthEstimationInput	DepthEstimationOutput
image-segmentation	ImageSegmentationInput	ImageSegmentationOutput
image-anonymization	ImageAnonymizationInput	ImageAnonymizationOutput
mask-generation	MaskGenerationInput	MaskGenerationOutput
visual-question-answering	VQAInput	VQAOutput
image-text-to-text	ImageTextToTextInput	ImageTextToTextOutput
zero-shot-image-classification	ZeroShotImageClassificationInput	ZeroShotImageClassificationOutput
zero-shot-object-detection	ZeroShotObjectDetectionInput	ZeroShotObjectDetectionOutput
audio-classification	AudioClassificationInput	AudioClassificationOutput
automatic-speech-recognition	ASRInput	ASROutput

Project Structure

src/task_inference/
├── factory.py          # create_task() / supported_tasks() entry points
├── protocol/           # OIP v2 Pydantic models
│   └── v2.py
├── tasks/              # Abstract task definitions + input/output schemas
│   ├── base.py
│   ├── vision/         # Image-based tasks
│   └── audio/          # Audio-based tasks
├── implementations/
│   ├── transformers/   # HuggingFace reference backend
│   │   ├── base.py     # Shared image/audio helpers
│   │   ├── vision/
│   │   └── audio/
│   └── onnxruntime/    # ONNX Runtime backend
│       ├── base.py     # Shared ORT helpers
│       ├── vision/
│       ├── audio/
│       └── adapters/   # Dialect adapters (auto-detected from model I/O)
└── utils.py            # Image/audio encode-decode helpers

Extending

Implement a new backend by subclassing the relevant task and overriding process:

from task_inference.tasks.vision.image_classification import (
    ImageClassificationInput,
    ImageClassificationOutput,
    ImageClassificationTask,
)

class MyOnnxImageClassificationTask(ImageClassificationTask):
    def process(self, inputs: ImageClassificationInput) -> ImageClassificationOutput:
        # your ONNX / TensorRT / remote-endpoint logic here
        ...

ONNX Runtime adapters

The built-in onnxruntime backend supports multiple model families through a dialect-adapter layer that auto-detects the correct tensor contract from the model's I/O tensor names at load time. See docs/onnx-adapters.md for:

• The full dialect reference (tensor signatures, detection rules) for all 14 built-in dialects
• Model acquisition instructions (optimum-cli / torch.onnx.export) for each task
• Step-by-step guide for adding a custom adapter dialect

Security Policy

The current release is the supported version. Security fixes are released together with all other fixes in each new release.

If you discover a security vulnerability in this project, please do not open a public issue.

Instead, report it privately by emailing us at [e-mail]. Include as much detail as possible to help us understand and address the issue quickly and responsibly.

Contributing

To report a bug or request a feature, please first check the existing issues to avoid duplicates. If none exist, open a new issue with a clear title and a detailed description, including any steps to reproduce if it's a bug.

To contribute code, start by forking the repository. Clone your fork locally and create a new branch for your changes. Make sure your commits follow the Conventional Commits v1.0 specification to keep history readable and consistent.

Once your changes are ready, push your branch to your fork and open a pull request against the main branch. Be sure to include a summary of what you changed and why. If your pull request addresses an issue, mention it in the description (e.g., “Closes #123”).

Please note that new contributors may be asked to sign a Contributor License Agreement (CLA) before their pull requests can be merged. This helps us ensure compliance with open source licensing standards.

We appreciate contributions and help in improving the project!

Authors

This project is developed and maintained by DSLab – Fondazione Bruno Kessler, with contributions from the open source community. A complete list of contributors is available in the project’s commit history and pull requests.

For questions or inquiries, please contact: [e-mail]

Copyright and license

Copyright © 2025 DSLab – Fondazione Bruno Kessler and individual contributors.

This project is licensed under the Apache License, Version 2.0. You may not use this file except in compliance with the License. Ownership of contributions remains with the original authors and is governed by the terms of the Apache 2.0 License, including the requirement to grant a license to the project.