#machine-learning #deep-learning #neural-networks #yolo


A Rust wrapper for Darknet, an open source neural network framework written in C and CUDA

13 releases

0.4.0 May 2, 2023
0.3.5 Dec 17, 2022
0.3.4 May 7, 2022
0.3.3 Oct 25, 2021
0.1.4 Mar 31, 2020

#31 in Machine learning

Download history 19/week @ 2023-06-12 34/week @ 2023-06-19 56/week @ 2023-06-26 27/week @ 2023-07-03 10/week @ 2023-07-10 11/week @ 2023-07-17 16/week @ 2023-07-24 8/week @ 2023-07-31 12/week @ 2023-08-07 25/week @ 2023-08-14 11/week @ 2023-08-21 35/week @ 2023-08-28 30/week @ 2023-09-04 7/week @ 2023-09-11 11/week @ 2023-09-18 25/week @ 2023-09-25

75 downloads per month

MIT license

985 lines

darknet-rust: A Rust bindings for AlexeyAB's Darknet


The crate is a Rust wrapper for AlexeyAB's Darknet.

It provides the following features:

  • Training and inference capabilities.
  • Load config files and model weights from upstream without modifications.
  • Safe type wrappers for C API. Includes network, detection and layer types.

Minimal rustc version: 1.43.0

Version 0.4 changes:

  • updated image to 0.24


The tiny_yolov3_inference example automatically downloads the YOLOv3 tiny weights, and produces inference results in output directory.

cargo run --release --example tiny_yolov3_inference

The run_inference example is an utility program that you can test a combination of model configs and weights on image files. For example, you can test the YOLOv4 mode.

cargo run --release --example run_inference -- \
    --label-file darknet/data/coco.names \
    --model-cfg darknet/cfg/yolov4.cfg \
    --weights yolov4.weights \

Read the example code in examples/ to understand the actual usage. More model configs and weights can be found here: (https://pjreddie.com/darknet/yolo/).


API documentation

If you are using version 0.1, consider migrating to 0.3 or newer as several critical bugs and memory leakages were fixed.


Terms used:

darknet-sys, darknet = Rust wrappers

libdarknet = C/C++ darknet implementation

By default, darknet will compile and link libdarknet statically. You can control the feature flags to change the behavior.

Before running tests:

git submodule init && git submodule update --recursive

Cargo Features

  • enable-cuda: Enable CUDA (expects CUDA 10.x and cuDNN 7.x).
  • enable-cudnn: Enable cuDNN
  • enable-opencv: Enable OpenCV.
  • enable-openmp: Enable OpenMP in darknet. Used for parallelization when running on the CPU. Enabled by default.
  • runtime: Link to libdarknet dynamic library. For example, libdark.so on Linux.
  • dylib: Build dynamic library instead of static
  • buildtime-bindgen: Generate bindings from libdarknet headers.

Method 1: Download and build from source (default)

darknet = "0.4"

You can optionally enable CUDA and OpenCV features. Please read Build with CUDA for more info.

darknet = {version = "0.4", features = ["enable-cuda", "enable-opencv"] }

Method 2: Build with custom source

If you want to build with custom libdarknet source, point DARKNET_SRC environment variable to your source path. It should contain CMakeLists.txt.

export DARKNET_SRC=/path/to/your/darknet/repo

With runtime feature, darknet-sys will not compile libdarknet source code and instead links to libdarknet dynamically. If you are using Linux, make sure libdark.so is installed on your system.

darknet = {version = "0.4", features = ["runtime"] }

Re-generate bindings

With buildtime-bindgen feature, darknet-sys re-generates bindings from headers. The option is necessary only when darkent is updated or modified.

darknet = {version = "0.4", features = ["buildtime-bindgen"] }

If you want to use your (possibly modified) header files, point DARKNET_INCLUDE_PATH environment variable to your header dir.

Build with CUDA

Please check that both CUDA 10.x and cuDNN 7.x are installed.

Darknet reads CUDA_PATH environment variable (which defaults to /opt/cuda if not set) and assumes it can find cuda libraries at ${CUDA_PATH}/lib64.

export CUDA_PATH=/usr/local/cuda-10.1
darknet = {version = "0.4", features = ["enable-cuda", "enable-opencv"] }

You can also set CUDA_ARCHITECTURES which is passed to libdarknet's cmake. It defaults to Auto, which auto-detects GPU architecture based on card present in the system during build.


The crate is licensed under MIT.


Huge thanks to all contributors!


~174K SLoC