tool for auto labeling, conversion and augmentation for darknet YOLO (and other annotation formats?).
kesa comes with a few binaries.
| name | explanation |
|---|---|
| kesa_al | for auto labeling, comes with onnx (ort) and torch (tch-rs) backends |
| kesa_l2y | for converting annotations to yolo txt format |
| kesa_split | for separating images/annotations to train, val, test batches. |
| kesa_aug | creates image augmentations from given labels and images |
currently kesa_al uses either torch(tch-rs) or onnxruntime(ort) to label images,
you can compile with either or both. you just need to download and link the libraries.
- build/download library onnxruntime
- add to your ~/.zshrc or ~/.bashrc:
export LD_LIBRARY_PATH=/path/to/onnxruntime-linux-x64-cuda-1.17.1/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
wip
- download libtorch from pytorch site
- add to your ~/.zshrc or ~/.bashrc:
export LIBTORCH=/media/hbpopos/penisf/libtorch-cxx11-abi-shared-with-deps-2.2.0+cu121/libtorch export LD_LIBRARY_PATH=${LIBTORCH}/lib:$LD_LIBRARY_PATH
- alternative options for installing
tch-rscan be found here
make sure you have the dependencies listed above
to build , run:
cargo build --releaseby default, kesa_al currently won't work without a backend specified, enable it by passing a feature flag.
// enable onnxruntime backend
cargo build --bin kesa_al --release --features onnxruntime
// enable torch backend
cargo build --bin kesa_al --release --features torch
// enable both
cargo build --bin kesa_al --release --features torch --features onnxruntime