tagwise

Tagwise is a command-line tool that automatically analyzes images with a CLIP-based ONNX model to find the most relevant keywords from a predefined vocabulary. It then renames each image file with the top predicted tags and its modification timestamp for easier search, organization, and sharing.

Models and Data

cd tagwise
git clone https://huggingface.co/mahyaret/tagwise ./model 

Build on Windows

Prerequisites:

• Visual Studio 2022 (Desktop development with C++)
• CMake
• vcpkg

Install dependencies with vcpkg:

git clone https://github.com/microsoft/vcpkg C:\vcpkg
cd C:\vcpkg
.\bootstrap-vcpkg.bat
.\vcpkg.exe install opencv4 nlohmann-json onnxruntime --triplet x64-windows

Configure and build:

cd C:\\Users\\home\\source\\repos\\tagwise

cmake -B build -S . -G "Visual Studio 17 2022" -A x64 -DVCPKG_INSTALLED_DIR="C:/vcpkg/installed/x64-windows"
cmake --build build --config Release

Build on macOS

Prerequisites (Homebrew):

brew install cmake opencv onnxruntime nlohmann-json

Build:

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j

How to run

.\tagwise.exe "[path_to_images_folder]" "[path_to_tags.txt]" "[path_to_model]"

If your model folder contains tags.txt, you can also run:

.\tagwise.exe "[path_to_images_folder]" "[path_to_model]"

If you see noisy ONNX schema registration logs on startup, TAGWISE_VERBOSE_INIT=1 disables Tagwise's stderr silencing during ONNXRuntime initialization.

If your image model has dynamic input size, Tagwise defaults to 224x224; override with TAGWISE_IMAGE_SIZE=224 or TAGWISE_IMAGE_SIZE=336x336.

Run on macOS

./build/bin/tagwise "[path_to_images_folder]" "[path_to_tags.txt]" "[path_to_model]"

Shorthand (uses [path_to_model]/tags.txt):

./build/bin/tagwise "[path_to_images_folder]" "[path_to_model]"

Demo