36g Rain Tagger

36g Rain Tagger is an automated image tagging software.

With it, you can:

• Run a tagging cli utility that crawls a file system, mapping a set of 10.8k tags to each image.
• Host a web UI to search tagged images. Searching can be done via tags, or images.

36g Rain Tagger uses timm and leverages the model SmilingWolf/wd-swinv2-tagger-v3.

It is named after 36g.

Set Up

Select your pytorch configuration by going through the menu at https://pytorch.org/get-started/locally/.

If you're using a CPU on Ubuntu, it should be something like python3.12 -m pip install torch --index-url https://download.pytorch.org/whl/cpu. You don't need torchvision or torchaudio.

If you're using a GPU on Ubuntu, python3.12 -m pip install torch will suffice.

git clone https://github.com/skwzrd/36g-Rain-Tagger
cd 36g-Rain-Tagger
python3.12 -m venv venv
source venv/bin/activate
python3.12 -m pip install -r requirements.txt
cd src/

Usage, The Tagger

An example,

python3.12 tagger.py --bsize 1 --gmin 0.2 --cmin 0.2 --exts jpg,jpeg,png,gif --printt --path /path/to/image.png

All the cli command options,

Image tagging utility for extracting and saving tags from images.

options:
  -h, --help             Show this help message and exit.
  --path                 Path to an image file or a directory containing images. Can also accept a list of paths.
  --gmin                 Minimum probability threshold for general tags. Range: [0.0, 1.0], where 1.0 means a very strong match. Default: 0.2
  --cmin                 Minimum probability threshold for character tags. Range: [0.0, 1.0], where 1.0 means a very strong match. Default: 0.2
  --exts                 Comma-separated list of valid image file extensions to process. Default: png,jpeg,jpg,gif
  --nmax                 Maximum number of images to tag. Set to 0 to process all images found in the specified path. Default: 0
  --bsize                Batch size for processing images. For faster processing, use a batch size of 1. Default: 1
  --db_name              Name of the SQLite database file to save results. Default: ./image.db
  --skip, --no-skip      Skip images that already have tags saved in the database. Use --no-skip to reprocess them. Default: --no-skip
  --idx, --no-idx        Enable index-to-probability mappings. Required to save results. Use --no-idx to disable. Default: --idx
  --save, --no-save      Save results to the SQLite database. Use --no-save to skip saving. Default: --save
  --printt, --no-printt  Print results. Use --no-printt to disable printing. Default: --no-print
  --cpu, --no-cpu        Run on CPU instead of GPU. Use --no-cpu to use GPU. Default: --no-cpu

Note: The tagger will automatically download the image tagging model and save it to ~/.cache/huggingface/hub.

Web UI

cd src/

cp   rename_to_configs.py   configs.py
# set variables in 'configs.py'

python3.12 web.py

# optional
celery -A web.celery_app worker --loglevel=info --pool solo
# ^ can be run in the background using --detach

Desktop, Info Mode

Desktop, Gallery Mode

Desktop, Uploaded Image Search

Mobile, Gallery Mode

Structures

The returned structures are determined by the --idx and --no-idx flags. With --idx, tag ids are gathered. With --no-idx, tags can be printed, but not saved.

Character: {}
Ratings: {'general': 0.944, 'sensitive': 0.053, 'questionable': 0.002, 'explicit': 0.0}
General tags: {'outdoors': 0.69, 'sky': 0.755, 'day': 0.68, 'cloud': 0.628, 'water': 0.767, 'blue_sky': 0.566, 'pokemon_(creature)': 0.558, 'no_humans': 0.851, 'ocean': 0.712, 'surfboard': 0.7}

Performance

Tagging

Benchmarked using https://github.com/sky-cake/image_data_set

From tagger.py,

main(
    path='/path/to/image_data_set',
    gmin=0.2,
    cmin=0.2,
    valid_extensions='png,jpeg,jpg,gif',
    bsize=1,
    nmax=0,
    db_name=make_path('image.db'),
    skip=False,
    idx=True,
    save=True,
    printt=False,
    cpu=?,
)

Device	Images	Total Time (s)	Time per Image (s)
4060 TI 16GB GPU	45	2.172	0.048
5700X x 8 CPU	45	21.277	0.473
i7 8665U x 8 CPU	45	76.273	1.695

Searching

0.1s - 0.4s results on hundreds of thousands of images.

Searched 238,302 in 0.313s and found 25 results.