About the Project

This project is one of the components of the digital me (WIP), which aims to clone oneself to some extent.

This component burnxp is for cloning one's Aesthetics Preferences, also known as XP(性癖) in Chinese.

Based on burn, burnxp can train a ResNet model to score pictures according to owner's sesthetics preferences.

Based on ratatui, tagger can aid in tagging pictures by iteractively comparing the degree of compliance with sesthetics preferences.

Based on argmin, tagger divide can help in dividing the dataset into label-balanced train-set and valid-set.

Tagger

Features:

• tagger pick subcommand can help pick the images to be tagged (super fast image viewer in terminal)
• tagger tag subcommand to label pictures
• tagger divide subcommand can help divide the images into train-set and valid-set, an efficient solver to balance labels and compute weight
• tagger observe subcommand can help observe the distribution of labels

Trainer

multi-gpu training is supported.

Predictor

(outdated screenshot, will update soon)

Usage

Cuda 12.x should be installed for non-macOS users.

Torch version (recommended) depends on libtorch to accelerate, please set it up with provided dist/setup scripts. Instead, candle version can work independently (known issue: 1. cuda 12.6 unusable; 2. max_pool and avg_pool are not well-supported, which leads candle version actually unusable).

1. Use compiled release

You can download the burnxp and tagger in the release page.

Note that tagger is released independently. burnxp-xx with f16 is half-precision version. Candle version is currently not usable.

# torch-version only, candle version can skip this setup
# you can also setup manually like this script if you have libtorch else where
run/setup.xx
run/burnxp.xx
./tagger.xx
# xx is the suffix of executable file based on your OS

2. Compile yourself

git clone [email protected]:kingwingfly/burnxp.git
# torch-version only, candle version can skip this setup
scripts/setup_<your_os>.xx
# torch-version and macOS only, candle version can skip this
source .venv/bin/activate

cargo build -p tagger --release
cargo build -p burnxp --release
# or half-precision version
cargo build -p burnxp --release -F f16

Note

The tagger works well in kitty iTerm2 and wezterm while maybe not in other terminals (like Warp).

For Windows, I tried my best but failed to make it work perfectly. wezterm can be used and please just set font size to 18 to meet the preset.

A Tauri version may be under development.

TODO

• CLIP model instead of ResNet model
• Make candle version usable

Contributing

If you are working with feature tch:

Please setup all needed environment variables like scripts/setup before coding, or your IDE may not work well.

For macOS users, you need also activate python venv before coding.

Do not use nightly version of Rust.

If you are working with feature candle (ususable now due to pooling not supported well), need CUDA<=12.4.1, nothing else to configure, just enjoy coding.

License

MIT LICENSE

Acknowledgement

• CUDA_COMPUTE_CAP