darknet2onnx

README

darknet2onnx

Convert Darknet .cfg + .weights to ONNX format.

A standalone Go CLI tool that produces a single static binary with no Python or pip dependencies.

Table of Contents

• Installation
• Supported models
• Output format
• Build from source
• Usage
  • Flags
  • Example
• Validate output
• Supported layer types
• Protobuf source
• How it works
• License

Installation

From Go

go install github.com/LdDl/darknet2onnx@latest

Pre-built binaries

Download the latest release for your platform from Releases:

Platform	Archive
Linux amd64	linux-amd64-darknet2onnx.tar.gz
Linux arm64	linux-arm64-darknet2onnx.tar.gz
macOS amd64	darwin-amd64-darknet2onnx.tar.gz
macOS arm64	darwin-arm64-darknet2onnx.tar.gz
Windows amd64	windows-amd64-darknet2onnx.zip

Extract and place the binary somewhere in your $PATH.

Supported models

• YOLOv3, YOLOv3-tiny
• YOLOv4, YOLOv4-tiny
• YOLOv7, YOLOv7-tiny

Output format

The YOLO detection head decode logic (sigmoid, grid offsets, anchor application) is embedded into the ONNX graph. All heads are concatenated into a single output tensor.

Coordinates cx, cy, w, h are in absolute pixel units relative to the input image dimensions.

Two output formats are available via --format:

Format	Shape	Description
yolov5 (default)	[1, N, 5+C]	With objectness: cx, cy, w, h, obj, cls0..clsN
yolov8	[1, 4+C, N]	Without objectness: cx, cy, w, h, cls0..clsN (obj baked into cls scores)

where N is the total number of predictions and C is the number of classes.

Input tensor is named images, output tensor is named output0. This follows the Ultralytics naming convention, so the output ONNX should be compatible with most inference pipelines that expect this convention. Note that this is not "traditional" YOLO output format, but widly supported I believe.

This could not work for you, but in my case these formats are compatible with od_opencv:

• yolov5 -> Model::yolov5_ort()
• yolov8 -> Model::ort()

Build from source

Simple build for the current platform:

go build -o darknet2onnx .

Cross-compile for all platforms (linux/windows/macOS, amd64/arm64):

./build.sh

Usage

./darknet2onnx --cfg model.cfg --weights model.weights --output model.onnx

Flags

Flag	Default	Description
--cfg	(required)	Path to Darknet .cfg file
--weights	(required)	Path to Darknet .weights file
--output	model.onnx	Output ONNX file path
--opset	12	ONNX opset version
--format	yolov5	Output format: yolov5 or yolov8

Example

./darknet2onnx \
    --cfg pretrained/yolov3-tiny.cfg \
    --weights pretrained/yolov3-tiny.weights \
    --output pretrained/yolov3-tiny.onnx

Validate output

Install onnx in a Python virtual environment:

python3 -m venv .venv
.venv/bin/pip install onnx

Then validate:

.venv/bin/python3 -c "
import onnx
m = onnx.load('model.onnx')
onnx.checker.check_model(m)
print('Valid')
"

Clean up:

rm -rf .venv

Supported layer types

Darknet layer	ONNX op(s)
[convolutional]	Conv + BatchNormalization + activation
[maxpool]	MaxPool
[route]	Concat or passthrough (+ Slice for groups)
[shortcut]	Add + activation
[upsample]	Resize (nearest)
[yolo]	Decode subgraph (Reshape, Sigmoid, Add, Exp, Mul, Concat)

Activations: leaky (LeakyRelu), mish (Softplus + Tanh + Mul), swish (Sigmoid + Mul), logistic (Sigmoid), linear (none).

Protobuf source

The ONNX protobuf schema (proto/onnx.proto3) is downloaded from the official ONNX repository:

https://raw.githubusercontent.com/onnx/onnx/main/onnx/onnx.proto3

To regenerate Go bindings, you need protoc and protoc-gen-go:

go install google.golang.org/protobuf/cmd/protoc-gen-go@latest

Then run:

protoc -I proto proto/onnx.proto3 --go_out=./onnxpb --go_opt=paths=source_relative --experimental_allow_proto3_optional

How it works

@todo

License

Just MIT, see LICENSE.