Darknet::ONNXExport Class Reference

Everthing we need to convert .cfg and .weights to .onnx is contained within this class. More...

#include "darknet_onnx.hpp"

Collaboration diagram for Darknet::ONNXExport:

Public Member Functions
	ONNXExport (const std::filesystem::path &cfg_filename, const std::filesystem::path &weights_filename, const std::filesystem::path &onnx_filename)
	Constructor.
	~ONNXExport ()
	Destructor.
std::string	add_const_float_tensor (const std::string &stem, const float &f)
std::string	add_const_ints_tensor (const std::string &stem, const std::vector< int > &v)
ONNXExport &	add_node_activation (const size_t index, Darknet::CfgSection &section)
ONNXExport &	add_node_bn (const size_t index, Darknet::CfgSection &section)
ONNXExport &	add_node_conv (const size_t index, Darknet::CfgSection &section)
ONNXExport &	add_node_maxpool (const size_t index, Darknet::CfgSection &section)
ONNXExport &	add_node_resize (const size_t index, Darknet::CfgSection &section)
ONNXExport &	add_node_route_concat (const size_t index, Darknet::CfgSection &section)
ONNXExport &	add_node_route_identity (const size_t index, Darknet::CfgSection &section)
ONNXExport &	add_node_route_slice (const size_t index, Darknet::CfgSection &section)
ONNXExport &	add_node_shortcut (const size_t index, Darknet::CfgSection &section)
ONNXExport &	add_node_yolo (const size_t index, Darknet::CfgSection &section)
ONNXExport &	build_model ()
ONNXExport &	check_activation (const size_t index, Darknet::CfgSection &section)
ONNXExport &	display_summary ()
	Display some general information about the protocol buffer model.
ONNXExport &	initialize_model ()
	Initialize some of the simple protobuffer model fields.
ONNXExport &	load_network ()
	Use Darknet to load the neural network.
ONNXExport &	populate_graph_initializer (const float *f, const size_t n, const Darknet::Layer &l, const std::string &name, const bool simple=false)
ONNXExport &	populate_graph_input_frame ()
ONNXExport &	populate_graph_nodes ()
ONNXExport &	populate_graph_postprocess ()
ONNXExport &	populate_graph_YOLO_output ()
ONNXExport &	populate_input_output_dimensions (onnx::ValueInfoProto *proto, const std::string &name, const int v1, const int v2=-1, const int v3=-1, const int v4=-1, const size_t line_number=0)
Darknet::VStr	postprocess_yolo_boxes (const Darknet::VStr &output_tx_ty, const Darknet::VStr &output_tw_th)
ONNXExport &	postprocess_yolo_boxes (const Darknet::VStr &v)
ONNXExport &	postprocess_yolo_class (Darknet::CfgSection &section, const Darknet::VStr &v, VStr &output_names)
ONNXExport &	postprocess_yolo_confs (const Darknet::VStr &output_obj, const Darknet::VStr &output_class)
Darknet::VStr	postprocess_yolo_slice_and_concat (Darknet::CfgSection &section)
ONNXExport &	postprocess_yolo_to (Darknet::CfgSection &section, const Darknet::VStr &v, VStr &output_names)
ONNXExport &	postprocess_yolo_tw_th (Darknet::CfgSection &section, const Darknet::VStr &v, VStr &output_names)
ONNXExport &	postprocess_yolo_tx_ty (Darknet::CfgSection &section, const Darknet::VStr &v, VStr &output_names)
ONNXExport &	save_output_file ()
	Save the entire model as an .onnx file.

Public Attributes
size_t	bit_size
	Number of bits to use – 32, 16, and eventually 8.
Darknet::CfgFile	cfg
std::filesystem::path	cfg_fn
bool	fuse_batchnorm
	Whether or not we need to fuse batchnorm (fuse and dontfuse on the CLI).
onnx::GraphProto *	graph
std::string	input_string
	The dimensions used in populate_graph_input_frame().
onnx::ModelProto	model
std::map< std::string, size_t >	number_of_floats_exported
	The key is the last part of the string, and the value is the number of floats.
std::filesystem::path	onnx_fn
int	opset_version
	Which opset version to use (10, 18, ...)?
std::string	output_string
	The names of the output nodes for this neural network.
bool	postprocess_boxes
	Whether or not we need to output the post-processing nodes to deal with boxes.
std::filesystem::path	weights_fn

Detailed Description

Everthing we need to convert .cfg and .weights to .onnx is contained within this class.

Constructor & Destructor Documentation

ONNXExport()

Darknet::ONNXExport::ONNXExport	(	const std::filesystem::path &	cfg_filename,
		const std::filesystem::path &	weights_filename,
		const std::filesystem::path &	onnx_filename
	)

Constructor.

~ONNXExport()

Darknet::ONNXExport::~ONNXExport

(

)

Destructor.

Here is the call graph for this function:

Member Function Documentation

add_const_float_tensor()

std::string Darknet::ONNXExport::add_const_float_tensor	(	const std::string &	stem,
		const float &	f
	)

add_const_ints_tensor()

std::string Darknet::ONNXExport::add_const_ints_tensor	(	const std::string &	stem,
		const std::vector< int > &	v
	)

add_node_activation()

Darknet::ONNXExport & Darknet::ONNXExport::add_node_activation	(	const size_t	index,
		Darknet::CfgSection &	section
	)

Here is the call graph for this function:

add_node_bn()

Darknet::ONNXExport & Darknet::ONNXExport::add_node_bn	(	const size_t	index,
		Darknet::CfgSection &	section
	)

Here is the call graph for this function:

add_node_conv()

Darknet::ONNXExport & Darknet::ONNXExport::add_node_conv	(	const size_t	index,
		Darknet::CfgSection &	section
	)

Here is the call graph for this function:

add_node_maxpool()

Darknet::ONNXExport & Darknet::ONNXExport::add_node_maxpool	(	const size_t	index,
		Darknet::CfgSection &	section
	)

Here is the call graph for this function:

add_node_resize()

Darknet::ONNXExport & Darknet::ONNXExport::add_node_resize	(	const size_t	index,
		Darknet::CfgSection &	section
	)

Here is the call graph for this function:

add_node_route_concat()

Darknet::ONNXExport & Darknet::ONNXExport::add_node_route_concat	(	const size_t	index,
		Darknet::CfgSection &	section
	)

Here is the call graph for this function:

add_node_route_identity()

Darknet::ONNXExport & Darknet::ONNXExport::add_node_route_identity	(	const size_t	index,
		Darknet::CfgSection &	section
	)

Todo:

verify what ChatGPT has to say about this:

Many Darknet-to-ONNX converters mistakenly convert single-layer route layers into Split or Slice nodes due to:

1. Overgeneralized conversion logic (trying to handle all cases with the same code)
2. Misinterpretation of what route does
3. Incorrect assumptions about data layout
4. Legacy from older or broken conversion scripts

In reality, single-layer route is just a pass-through, but some tools try to “process” it unnecessarily.

...and:

In ONNX, the Identity operator is functionally equivalent to a no-op (no operation) on the tensor data. It simply passes the input through to the output unchanged.

Here is the call graph for this function:

add_node_route_slice()

Darknet::ONNXExport & Darknet::ONNXExport::add_node_route_slice	(	const size_t	index,
		Darknet::CfgSection &	section
	)

Here is the call graph for this function:

add_node_shortcut()

Darknet::ONNXExport & Darknet::ONNXExport::add_node_shortcut	(	const size_t	index,
		Darknet::CfgSection &	section
	)

Todo:

V5: unused? Do we have weights for shortcuts?

Here is the call graph for this function:

add_node_yolo()

Darknet::ONNXExport & Darknet::ONNXExport::add_node_yolo	(	const size_t	index,
		Darknet::CfgSection &	section
	)

Here is the call graph for this function:

build_model()

Darknet::ONNXExport & Darknet::ONNXExport::build_model

(

)

Here is the caller graph for this function:

check_activation()

Darknet::ONNXExport & Darknet::ONNXExport::check_activation	(	const size_t	index,
		Darknet::CfgSection &	section
	)

Here is the call graph for this function:

display_summary()

Darknet::ONNXExport & Darknet::ONNXExport::display_summary

(

)

Display some general information about the protocol buffer model.

Here is the call graph for this function:

Here is the caller graph for this function:

initialize_model()

Darknet::ONNXExport & Darknet::ONNXExport::initialize_model

(

)

Initialize some of the simple protobuffer model fields.

Todo:

We need a command-line parameter for this field.

Todo:

We need a command-line parameter for this field.

Here is the call graph for this function:

Here is the caller graph for this function:

load_network()

Darknet::ONNXExport & Darknet::ONNXExport::load_network

(

)

Use Darknet to load the neural network.

Here is the call graph for this function:

Here is the caller graph for this function:

populate_graph_initializer()

Darknet::ONNXExport & Darknet::ONNXExport::populate_graph_initializer	(	const float *	f,
		const size_t	n,
		const Darknet::Layer &	l,
		const std::string &	name,
		const bool	simple = false
	)

Todo:

V5 2025-08-13: This is black magic! I actually have no idea how the DIMS work. I saw some example Darknet/YOLO weights converted to ONNX and attempted to figure out the patern. While this seems to work for the few examples I have, I would be extremely happy if someone can point out to me exactly how this works so I can implement it correctly!

Here is the call graph for this function:

populate_graph_input_frame()

Darknet::ONNXExport & Darknet::ONNXExport::populate_graph_input_frame

(

)

populate_graph_nodes()

Darknet::ONNXExport & Darknet::ONNXExport::populate_graph_nodes

(

)

Here is the call graph for this function:

populate_graph_postprocess()

Darknet::ONNXExport & Darknet::ONNXExport::populate_graph_postprocess

(

)

populate_graph_YOLO_output()

Darknet::ONNXExport & Darknet::ONNXExport::populate_graph_YOLO_output

(

)

Here is the call graph for this function:

populate_input_output_dimensions()

Darknet::ONNXExport & Darknet::ONNXExport::populate_input_output_dimensions	(	onnx::ValueInfoProto *	proto,
		const std::string &	name,
		const int	v1,
		const int	v2 = -1,
		const int	v3 = -1,
		const int	v4 = -1,
		const size_t	line_number = 0
	)

Here is the call graph for this function:

postprocess_yolo_boxes() [1/2]

Darknet::VStr Darknet::ONNXExport::postprocess_yolo_boxes	(	const Darknet::VStr &	output_tx_ty,
		const Darknet::VStr &	output_tw_th
	)

Here is the call graph for this function:

postprocess_yolo_boxes() [2/2]

Darknet::ONNXExport & Darknet::ONNXExport::postprocess_yolo_boxes

(

const Darknet::VStr &

v

)

Here is the call graph for this function:

postprocess_yolo_class()

Darknet::ONNXExport & Darknet::ONNXExport::postprocess_yolo_class	(	Darknet::CfgSection &	section,
		const Darknet::VStr &	v,
		Darknet::VStr &	output_names
	)

Here is the call graph for this function:

postprocess_yolo_confs()

Darknet::ONNXExport & Darknet::ONNXExport::postprocess_yolo_confs	(	const Darknet::VStr &	output_obj,
		const Darknet::VStr &	output_class
	)

postprocess_yolo_slice_and_concat()

Darknet::VStr Darknet::ONNXExport::postprocess_yolo_slice_and_concat

(

Darknet::CfgSection &

section

)

Here is the call graph for this function:

postprocess_yolo_to()

Darknet::ONNXExport & Darknet::ONNXExport::postprocess_yolo_to	(	Darknet::CfgSection &	section,
		const Darknet::VStr &	v,
		Darknet::VStr &	output_names
	)

Here is the call graph for this function:

postprocess_yolo_tw_th()

Darknet::ONNXExport & Darknet::ONNXExport::postprocess_yolo_tw_th	(	Darknet::CfgSection &	section,
		const Darknet::VStr &	v,
		Darknet::VStr &	output_names
	)

postprocess_yolo_tx_ty()

Darknet::ONNXExport & Darknet::ONNXExport::postprocess_yolo_tx_ty	(	Darknet::CfgSection &	section,
		const Darknet::VStr &	v,
		Darknet::VStr &	output_names
	)

Here is the call graph for this function:

save_output_file()

Darknet::ONNXExport & Darknet::ONNXExport::save_output_file

(

)

Save the entire model as an .onnx file.

Here is the call graph for this function:

Here is the caller graph for this function:

Member Data Documentation

bit_size

size_t Darknet::ONNXExport::bit_size

Number of bits to use – 32, 16, and eventually 8.

cfg

Darknet::CfgFile Darknet::ONNXExport::cfg

cfg_fn

std::filesystem::path Darknet::ONNXExport::cfg_fn

fuse_batchnorm

bool Darknet::ONNXExport::fuse_batchnorm

Whether or not we need to fuse batchnorm (fuse and dontfuse on the CLI).

graph

onnx::GraphProto* Darknet::ONNXExport::graph

input_string

std::string Darknet::ONNXExport::input_string

The dimensions used in populate_graph_input_frame().

model

onnx::ModelProto Darknet::ONNXExport::model

number_of_floats_exported

std::map<std::string, size_t> Darknet::ONNXExport::number_of_floats_exported

The key is the last part of the string, and the value is the number of floats.

For example, for "000_conv_bias", we store the key as "bias".

onnx_fn

std::filesystem::path Darknet::ONNXExport::onnx_fn

opset_version

int Darknet::ONNXExport::opset_version

Which opset version to use (10, 18, ...)?

output_string

std::string Darknet::ONNXExport::output_string

The names of the output nodes for this neural network.

postprocess_boxes

bool Darknet::ONNXExport::postprocess_boxes

Whether or not we need to output the post-processing nodes to deal with boxes.

weights_fn

std::filesystem::path Darknet::ONNXExport::weights_fn

---

The documentation for this class was generated from the following files:

• darknet_onnx.hpp
• darknet_onnx.cpp