[AnnFlux] Annotation flux

A research tool for exploring and annotating large datasets with Active Learning

This standalone tool provides a basic interface for interacting with large datasets so that they can be explored and annotated efficiently.

The extensible design of the tool allows researchers from in- and outside Intel to contribute to the development of the functionality

License

Apache 2.0, see LICENSE.md

Contribute

See CONTRIBUTING.md

Install

1. Create and activate a new Conda environment:

   conda create -n AnnFlux_dev python=3.11
   conda activate AnnFlux_dev

2. Install the requirements:

   pip install -e .

Examples

StreetSurfaceVis

The public (Creative Commons Attribution Share Alike 4.0 International) road type dataset 'StreetSurfaceVis' (https://zenodo.org/records/11449977 / https://www.nature.com/articles/s41597-024-04295-9).

See StreetSurfaceVis

From images folder

Init project, compute features and embed

annflux go {PROJECT_FOLDER}

PROJECT_FOLDER should have at least a images folder

Example

Use an image dataset with a folder of images with a .jpg extension. A good size is 5,000 to 10,000 images.

Extract to ~/annflux/data/envdataset/images

On commandline

export HUGGINGFACE_CLIP_NAME={a hugging face CLIP model that supports the peft package}

annflux go ~/annflux/data/envdataset --start_labels Your_label_A Your_label_B Your_label_C`

Label some images, then

annflux train_then_features ~/annflux/data/envdataset

to perform parameter efficient fine-tuning of the (default) CLIP model, followed by computation of the adapted features.

Tests and code coverage

export HUGGINGFACE_CLIP_NAME={a hugging face CLIP model that supports the peft package}

export USER_DATASET_PATH={your project folder with an 'images' folder inside}

python run_coverage.py

Known issues

HIGH

• (None)

Changelog

1.0.1.0

• Basic UI contribution from Naturalis

1.0.0.0

• Final first release

0.9.3.0

• Improved test reporting

0.9.3.0

• Improved test reporting

0.9.2.0

• Preparing for open source