Skip to main content

Distill large foundational models into smaller, domain-specific models for deployment

Project description

Autodistill uses big, slower foundation models to train small, faster supervised models. Using autodistill, you can go from unlabeled images to inference on a custom model running at the edge with no human intervention in between.

Currently, autodistill supports vision tasks like object detection and instance segmentation, but in the future it can be expanded to support language (and other) models.

๐Ÿ”— Quicklinks

Tutorial Docs Supported Models Contribute

๐Ÿ‘€ Example Output

Here are example predictions of a Target Model detecting milk bottles and bottlecaps after being trained on an auto-labeled dataset using Autodistill (see the Autodistill YouTube video for a full walkthrough):

๐Ÿš€ Features

  • ๐Ÿ”Œ Pluggable interface to connect models together
  • ๐Ÿค– Automatically label datasets
  • ๐Ÿฐ Train fast supervised models
  • ๐Ÿ”’ Own your model
  • ๐Ÿš€ Deploy distilled models to the cloud or the edge

๐Ÿ“š Basic Concepts

To use autodistill, you input unlabeled data into a Base Model which uses an Ontology to label a Dataset that is used to train a Target Model which outputs a Distilled Model fine-tuned to perform a specific Task.

Autodistill defines several basic primitives:

  • Task - A Task defines what a Target Model will predict. The Task for each component (Base Model, Ontology, and Target Model) of an autodistill pipeline must match for them to be compatible with each other. Object Detection and Instance Segmentation are currently supported through the detection task. classification support will be added soon.
  • Base Model - A Base Model is a large foundation model that knows a lot about a lot. Base models are often multimodal and can perform many tasks. They're large, slow, and expensive. Examples of Base Models are GroundedSAM and GPT-4's upcoming multimodal variant. We use a Base Model (along with unlabeled input data and an Ontology) to create a Dataset.
  • Ontology - an Ontology defines how your Base Model is prompted, what your Dataset will describe, and what your Target Model will predict. A simple Ontology is the CaptionOntology which prompts a Base Model with text captions and maps them to class names. Other Ontologies may, for instance, use a CLIP vector or example images instead of a text caption.
  • Dataset - a Dataset is a set of auto-labeled data that can be used to train a Target Model. It is the output generated by a Base Model.
  • Target Model - a Target Model is a supervised model that consumes a Dataset and outputs a distilled model that is ready for deployment. Target Models are usually small, fast, and fine-tuned to perform a specific task very well (but they don't generalize well beyond the information described in their Dataset). Examples of Target Models are YOLOv8 and DETR.
  • Distilled Model - a Distilled Model is the final output of the autodistill process; it's a set of weights fine-tuned for your task that can be deployed to get predictions.

๐Ÿ’ก Theory and Limitations

Human labeling is one of the biggest barriers to broad adoption of computer vision. It can take thousands of hours to craft a dataset suitable for training a production model. The process of distillation for training supervised models is not new, in fact, traditional human labeling is just another form of distillation from an extremely capable Base Model (the human brain ๐Ÿง ).

Foundation models know a lot about a lot, but for production we need models that know a lot about a little.

As foundation models get better and better they will increasingly be able to augment or replace humans in the labeling process. We need tools for steering, utilizing, and comparing these models. Additionally, these foundation models are big, expensive, and often gated behind private APIs. For many production use-cases, we need models that can run cheaply and in realtime at the edge.

Autodistill's Base Models can already create datasets for many common use-cases (and through creative prompting and few-shotting we can expand their utility to many more), but they're not perfect yet. There's still a lot of work to do; this is just the beginning and we'd love your help testing and expanding the capabilities of the system!

๐Ÿ’ฟ Installation

Autodistill is modular. You'll need to install the autodistill package (which defines the interfaces for the above concepts) along with Base Model and Target Model plugins (which implement specific models).

By packaging these separately as plugins, dependency and licensing incompatibilities are minimized and new models can be implemented and maintained by anyone.


pip install autodistill autodistill-grounded-sam autodistill-yolov8
Install from source

You can also clone the project from GitHub for local development:

git clone
cd autodistill
pip install -e .

Additional Base and Target models are enumerated below.

๐Ÿš€ Quickstart

See the demo Notebook for a quick introduction to autodistill. This notebook walks through building a milk container detection model with no labeling.

Below, we have condensed key parts of the notebook for a quick introduction to autodistill.

You can also run Autodistill in one command. First, install autodistill:

pip install autodistill

Then, run:

autodistill images --base="grounding_dino" --target="yolov8" --ontology '{"prompt": "label"}' --output="./dataset"

This command will label all images in a directory called images with Grounding DINO and use the labeled images to train a YOLOv8 model. Grounding DINO will label all images with the "prompt" and save the label as the "label". You can specify as many prompts and labels as you want. The resulting dataset will be saved in a folder called dataset.

Install Packages

For this example, we'll show how to distill GroundedSAM into a small YOLOv8 model using autodistill-grounded-sam and autodistill-yolov8.

pip install autodistill autodistill-grounded-sam autodistill-yolov8

Distill a Model

from autodistill_grounded_sam import GroundedSAM
from autodistill.detection import CaptionOntology
from autodistill_yolov8 import YOLOv8

# define an ontology to map class names to our GroundingDINO prompt
# the ontology dictionary has the format {caption: class}
# where caption is the prompt sent to the base model, and class is the label that will
# be saved for that caption in the generated annotations
base_model = GroundedSAM(ontology=CaptionOntology({"shipping container": "container"}))

# label all images in a folder called `context_images`

target_model = YOLOv8("")
target_model.train("./dataset/data.yaml", epochs=200)

# run inference on the new model
pred = target_model.predict("./dataset/valid/your-image.jpg", confidence=0.5)

# optional: upload your model to Roboflow for deployment
from roboflow import Roboflow

rf = Roboflow(api_key="API_KEY")
project = rf.workspace().project("PROJECT_ID")
project.version(DATASET_VERSION).deploy(model_type="yolov8", model_path=f"./runs/detect/train/")
Visualize Predictions

To plot the annotations for a single image using autodistill, you can use the code below. This code is helpful to visualize the annotations generated by your base model (i.e. GroundedSAM) and the results from your target model (i.e. YOLOv8).

import supervision as sv
import cv2

img_path = "./images/your-image.jpeg"

image = cv2.imread(img_path)

detections = base_model.predict(img_path)
# annotate image with detections
box_annotator = sv.BoxAnnotator()

labels = [
    f"{base_model.ontology.classes()[class_id]} {confidence:0.2f}"
    for _, _, confidence, class_id, _ in detections

annotated_frame = box_annotator.annotate(
    scene=image.copy(), detections=detections, labels=labels

sv.plot_image(annotated_frame, (16, 16))

๐Ÿ“ Available Models

Our goal is for autodistill to support using all foundation models as Base Models and most SOTA supervised models as Target Models. We focused on object detection and segmentation tasks first but plan to launch classification support soon! In the future, we hope autodistill will also be used for models beyond computer vision.

  • โœ… - complete (click row/column header to go to repo)
  • ๐Ÿšง - work in progress

object detection

DETIC โœ… โœ… โœ… โœ… ๐Ÿšง
GroundedSAM โœ… โœ… โœ… โœ… ๐Ÿšง
GroundingDINO โœ… โœ… โœ… โœ… ๐Ÿšง
OWL-ViT โœ… โœ… โœ… โœ… ๐Ÿšง
SAM-CLIP โœ… โœ… โœ… โœ… ๐Ÿšง
LLaVA-1.5 โœ… โœ… โœ… โœ… ๐Ÿšง
Kosmos-2 โœ… โœ… โœ… โœ… ๐Ÿšง
OWLv2 โœ… โœ… โœ… โœ… ๐Ÿšง
Roboflow Universe Models (50k+ pre-trained models) โœ… โœ… โœ… โœ… ๐Ÿšง
CoDet โœ… โœ… โœ… โœ… ๐Ÿšง
VLPart โœ… โœ… โœ… โœ… ๐Ÿšง
Azure Custom Vision โœ… โœ… โœ… โœ… ๐Ÿšง
AWS Rekognition โœ… โœ… โœ… โœ… ๐Ÿšง
Google Vision โœ… โœ… โœ… โœ… ๐Ÿšง

instance segmentation

base / target YOLOv8 YOLO-NAS YOLOv5 YOLOv7 Segformer
GroundedSAM โœ… ๐Ÿšง ๐Ÿšง
SAM-CLIP โœ… ๐Ÿšง ๐Ÿšง
SegGPT โœ… ๐Ÿšง ๐Ÿšง
FastSAM ๐Ÿšง ๐Ÿšง ๐Ÿšง


base / target ViT YOLOv8 YOLOv5
CLIP โœ… โœ… ๐Ÿšง
MetaCLIP โœ… โœ… ๐Ÿšง
DINOv2 โœ… โœ… ๐Ÿšง
BLIP โœ… โœ… ๐Ÿšง
ALBEF โœ… โœ… ๐Ÿšง
FastViT โœ… โœ… ๐Ÿšง
AltCLIP โœ… โœ… ๐Ÿšง
Fuyu ๐Ÿšง ๐Ÿšง ๐Ÿšง
Open Flamingo ๐Ÿšง ๐Ÿšง ๐Ÿšง

Roboflow Model Deployment Support

You can optionally deploy some Target Models trained using Autodistill on Roboflow. Deploying on Roboflow allows you to use a range of concise SDKs for using your model on the edge, from roboflow.js for web deployment to NVIDIA Jetson devices.

The following Autodistill Target Models are supported by Roboflow for deployment:

model name Supported?
YOLOv8 Object Detection โœ…
YOLOv8 Instance Segmentation โœ…
YOLOv5 Object Detection โœ…
YOLOv5 Instance Segmentation โœ…
YOLOv8 Classification

๐ŸŽฌ Video Guides

Autodistill: Train YOLOv8 with ZERO Annotations Autodistill: Train YOLOv8 with ZERO Annotations

Published: 8 June 2023

In this video, we will show you how to use a new library to train a YOLOv8 model to detect bottles moving on a conveyor line. Yes, that's right - zero annotation hours are required! We dive deep into Autodistill's functionality, covering topics from setting up your Python environment and preparing your images, to the thrilling automatic annotation of images.

๐Ÿ’ก Community Resources

๐Ÿ—บ๏ธ Roadmap

Apart from adding new models, there are several areas we plan to explore with autodistill including:

  • ๐Ÿ’ก Ontology creation & prompt engineering
  • ๐Ÿ‘ฉโ€๐Ÿ’ป Human in the loop support
  • ๐Ÿค” Model evaluation
  • ๐Ÿ”„ Active learning
  • ๐Ÿ’ฌ Language tasks

๐Ÿ† Contributing

We love your input! Please see our contributing guide to get started. Thank you ๐Ÿ™ to all our contributors!

๐Ÿ‘ฉโ€โš–๏ธ License

The autodistill package is licensed under an Apache 2.0. Each Base or Target model plugin may use its own license corresponding with the license of its underlying model. Please refer to the license in each plugin repo for more information.

Frequently Asked Questions โ“

What causes the PytorchStreamReader failed reading zip archive: failed finding central directory error?

This error is caused when PyTorch cannot load the model weights for a model. Go into the ~/.cache/autodistill directory and delete the folder associated with the model you are trying to load. Then, run your code again. The model weights will be downloaded from scratch. Leave the installation process uninterrupted.

๐Ÿ’ป explore more Roboflow open source projects

Project Description
supervision General-purpose utilities for use in computer vision projects, from predictions filtering and display to object tracking to model evaluation.
Autodistill (this project) Automatically label images for use in training computer vision models.
Inference An easy-to-use, production-ready inference server for computer vision supporting deployment of many popular model architectures and fine-tuned models.
Notebooks Tutorials for computer vision tasks, from training state-of-the-art models to tracking objects to counting objects in a zone.
Collect Automated, intelligent data collection powered by CLIP.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

autodistill-0.1.26.tar.gz (32.1 kB view hashes)

Uploaded Source

Built Distribution

autodistill-0.1.26-py3-none-any.whl (30.1 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page