darknet_network.cpp File Reference

Include dependency graph for darknet_network.cpp:

Functions
void	backward_network (Darknet::Network &net, Darknet::NetworkState state)
void	calculate_binary_weights (DarknetNetworkPtr ptr)
	This is part of the original C API.
void	copy_cudnn_descriptors (const Darknet::Layer &src, Darknet::Layer *dst)
void	copy_weights_net (const Darknet::Network &net_train, Darknet::Network *net_map)
void	custom_get_region_detections (const Darknet::Layer &l, int w, int h, int net_w, int net_h, float thresh, int *map, float hier, int relative, Darknet::Detection *dets, int letter)
void	ema_apply (Darknet::Network &net)
void	ema_update (Darknet::Network &net, float ema_alpha)
void	fill_network_boxes (Darknet::Network *net, int w, int h, float thresh, float hier, int *map, int relative, Darknet::Detection *dets, int letter)
void	fill_network_boxes_batch (Darknet::Network *net, int w, int h, float thresh, float hier, int *map, int relative, Darknet::Detection *dets, int letter, int batch)
static void	fill_network_boxes_v3 (Darknet::Network *net, int w, int h, float thresh, float hier, int *map, int relative, Darknet::Detection *dets, int letter, Darknet::Output_Object_Cache &cache)
void	forward_blank_layer (Darknet::Layer &l, Darknet::NetworkState state)
void	forward_network (Darknet::Network &net, Darknet::NetworkState state)
void	free_batch_detections (det_num_pair *det_num_pairs, int n)
void	free_detections (detection *dets, int n)
	This is part of the original C API. Do not use in new code.
void	free_network (Darknet::Network &net)
	Free all memory allocations for the given neural network.
void	free_network_ptr (DarknetNetworkPtr ptr)
	This is part of the original C API.
void	free_network_recurrent_state (Darknet::Network &net)
void	fuse_conv_batchnorm (Darknet::Network &net)
	ChatGPT: The function fuse_conv_batchnorm() in the Darknet/YOLO codebase is responsible for fusing the convolutional layer and batch normalization layer into a single operation.
int	get_current_batch (const Darknet::Network &net)
int64_t	get_current_iteration (const Darknet::Network &net)
float	get_current_rate (const Darknet::Network &net)
float	get_current_seq_subdivisions (const Darknet::Network &net)
detection *	get_network_boxes (DarknetNetworkPtr ptr, int w, int h, float thresh, float hier, int *map, int relative, int *num, int letter)
	This is part of the original C API.
float	get_network_cost (const Darknet::Network &net)
Darknet::Image	get_network_image (Darknet::Network &net)
Darknet::Image	get_network_image_layer (Darknet::Network &net, int i)
int	get_network_input_size (Darknet::Network &net)
float *	get_network_output (Darknet::Network &net)
int	get_network_output_size (Darknet::Network &net)
int	get_sequence_value (const Darknet::Network &net)
int	is_ema_initialized (const Darknet::Network &net)
static float	lrelu (float src)
Darknet::Network	make_network (int n)
	Think of this as the constructor for the Darknet::Network object.
detection *	make_network_boxes (DarknetNetworkPtr ptr, float thresh, int *num)
	This is part of the original C API.
Darknet::Detection *	make_network_boxes_batch (Darknet::Network *net, float thresh, int *num, int batch)
Darknet::Detection *	make_network_boxes_v3 (Darknet::Network *net, const float thresh, int *num, Darknet::Output_Object_Cache &cache)
int	network_height (Darknet::Network *net)
float *	network_predict (Darknet::Network &net, float *input)
det_num_pair *	network_predict_batch (Darknet::Network *net, Darknet::Image im, int batch_size, int w, int h, float thresh, float hier, int *map, int relative, int letter)
matrix	network_predict_data (Darknet::Network &net, data test)
matrix	network_predict_data_multi (Darknet::Network &net, data test, int n)
float *	network_predict_image (DarknetNetworkPtr ptr, const DarknetImage im)
	This is part of the original C API.
float *	network_predict_image_letterbox (DarknetNetworkPtr ptr, DarknetImage im)
	This is part of the original C API.
float *	network_predict_ptr (DarknetNetworkPtr ptr, float *input)
	This is part of the original C API.
int	network_width (Darknet::Network *net)
int	num_detections (Darknet::Network *net, float thresh)
int	num_detections_batch (Darknet::Network *net, float thresh, int batch)
int	num_detections_v3 (Darknet::Network *net, float thresh, Darknet::Output_Object_Cache &cache)
	Basically a wrapper for yolo_num_detections_v3().
int	recalculate_workspace_size (Darknet::Network *net)
void	reject_similar_weights (Darknet::Network &net, float sim_threshold)
void	reset_network_state (Darknet::Network *net, int b)
void	reset_rnn (DarknetNetworkPtr ptr)
	This is part of the original C API.
int	resize_network (Darknet::Network *net, int w, int h)
void	restore_network_recurrent_state (Darknet::Network &net)
void	set_batch_network (Darknet::Network *net, int b)
float	train_network (Darknet::Network &net, data d)
float	train_network_batch (Darknet::Network net, data d, int n)
float	train_network_datum (Darknet::Network &net, float *x, float *y)
float	train_network_waitkey (Darknet::Network &net, data d, int wait_key)
void	update_network (Darknet::Network &net)
void	visualize_network (Darknet::Network &net)

Function Documentation

backward_network()

void backward_network	(	Darknet::Network &	net,
		Darknet::NetworkState	state
	)

Here is the caller graph for this function:

calculate_binary_weights()

void calculate_binary_weights

(

DarknetNetworkPtr

ptr

)

This is part of the original C API.

Do not use in new code.

See also

darknet_load_neural_network()

Darknet::load_neural_network()

If you were previously using calculate_binary_weights() from within C code, it used to pass the network by value. Starting with Darknet V3 in 2024-08-16, the network is now passed in as a pointer.

Here is the call graph for this function:

Here is the caller graph for this function:

copy_cudnn_descriptors()

void copy_cudnn_descriptors	(	const Darknet::Layer &	src,
		Darknet::Layer *	dst
	)

Here is the caller graph for this function:

copy_weights_net()

void copy_weights_net	(	const Darknet::Network &	net_train,
		Darknet::Network *	net_map
	)

Here is the call graph for this function:

Here is the caller graph for this function:

custom_get_region_detections()

void custom_get_region_detections	(	const Darknet::Layer &	l,
		int	w,
		int	h,
		int	net_w,
		int	net_h,
		float	thresh,
		int *	map,
		float	hier,
		int	relative,
		Darknet::Detection *	dets,
		int	letter
	)

Here is the call graph for this function:

Here is the caller graph for this function:

ema_apply()

void ema_apply

(

Darknet::Network &

net

)

Here is the call graph for this function:

Here is the caller graph for this function:

ema_update()

void ema_update	(	Darknet::Network &	net,
		float	ema_alpha
	)

Here is the call graph for this function:

Here is the caller graph for this function:

fill_network_boxes()

void fill_network_boxes	(	Darknet::Network *	net,
		int	w,
		int	h,
		float	thresh,
		float	hier,
		int *	map,
		int	relative,
		Darknet::Detection *	dets,
		int	letter
	)

Todo:

V3 JAZZ: most of the time is spent in this function

Here is the call graph for this function:

Here is the caller graph for this function:

fill_network_boxes_batch()

void fill_network_boxes_batch	(	Darknet::Network *	net,
		int	w,
		int	h,
		float	thresh,
		float	hier,
		int *	map,
		int	relative,
		Darknet::Detection *	dets,
		int	letter,
		int	batch
	)

Here is the call graph for this function:

Here is the caller graph for this function:

fill_network_boxes_v3()

static void fill_network_boxes_v3 ( Darknet::Network *  net, int  w, int  h, float  thresh, float  hier, int *  map, int  relative, Darknet::Detection *  dets, int  letter, Darknet::Output_Object_Cache &  cache  )

inlinestatic

Todo:

This assumes that "GAUSSIAN_YOLO", "REGION", and "DETECTION" layers don't exist, which is wrong. But they only exist in much older configurations which are hopefully not used anymore? Should we deprecate these?

Here is the call graph for this function:

Here is the caller graph for this function:

forward_blank_layer()

void forward_blank_layer	(	Darknet::Layer &	l,
		Darknet::NetworkState	state
	)

Here is the caller graph for this function:

forward_network()

void forward_network	(	Darknet::Network &	net,
		Darknet::NetworkState	state
	)

Here is the call graph for this function:

Here is the caller graph for this function:

free_batch_detections()

void free_batch_detections	(	det_num_pair *	det_num_pairs,
		int	n
	)

Here is the call graph for this function:

free_detections()

void free_detections	(	detection *	dets,
		int	n
	)

This is part of the original C API. Do not use in new code.

Here is the caller graph for this function:

free_network()

void free_network

(

Darknet::Network &

net

)

Free all memory allocations for the given neural network.

All of these functions perform the exact same task, so use the one that makes the most sense given your application. Some of these are C++ calls, some are C, some pass the network by refrence, some pass a pointer to the neural network.

See also

darknet_free_neural_network()

Darknet::free_neural_network()

free_network_ptr()

Here is the call graph for this function:

Here is the caller graph for this function:

free_network_ptr()

void free_network_ptr

(

DarknetNetworkPtr

ptr

)

This is part of the original C API.

Do not use in new code.

See also

darknet_free_neural_network()

Darknet::free_neural_network()

If you were previously using free_network() from within C code, please use free_network_ptr() instead by passing in the address of the network structure (pointer to the network).

Note

See the additional comments in free_network().

Here is the call graph for this function:

Here is the caller graph for this function:

free_network_recurrent_state()

void free_network_recurrent_state

(

Darknet::Network &

net

)

Here is the call graph for this function:

Here is the caller graph for this function:

fuse_conv_batchnorm()

void fuse_conv_batchnorm

(

Darknet::Network &

net

)

ChatGPT: The function fuse_conv_batchnorm() in the Darknet/YOLO codebase is responsible for fusing the convolutional layer and batch normalization layer into a single operation.

This process is essential for both model optimization and inference efficiency, as it reduces the number of operations needed during model execution.

In Darknet, when you have a convolutional layer followed by a batch normalization layer, the two layers can be fused into a single convolution-like operation that combines both the convolution and normalization. This reduces the need for the separate batch normalization step during inference, thus improving speed.

The fused operation allows the model to perform one pass of convolution instead of:

• Convolution operation.
• Batch normalization (normalization, scaling, shifting).
• Activation function (like ReLU).

Here is the call graph for this function:

Here is the caller graph for this function:

get_current_batch()

int get_current_batch

(

const Darknet::Network &

net

)

Here is the caller graph for this function:

get_current_iteration()

int64_t get_current_iteration

(

const Darknet::Network &

net

)

Here is the caller graph for this function:

get_current_rate()

float get_current_rate

(

const Darknet::Network &

net

)

Here is the call graph for this function:

Here is the caller graph for this function:

get_current_seq_subdivisions()

float get_current_seq_subdivisions

(

const Darknet::Network &

net

)

Here is the call graph for this function:

Here is the caller graph for this function:

get_network_boxes()

detection * get_network_boxes	(	DarknetNetworkPtr	ptr,
		int	w,
		int	h,
		float	thresh,
		float	hier,
		int *	map,
		int	relative,
		int *	num,
		int	letter
	)

This is part of the original C API.

Do not use in new code.

You must call free_detections() to free up memory once done with the detections.

See also

Darknet::predict()

free_detections()

Here is the call graph for this function:

Here is the caller graph for this function:

get_network_cost()

float get_network_cost

(

const Darknet::Network &

net

)

Here is the caller graph for this function:

get_network_image()

Darknet::Image get_network_image

(

Darknet::Network &

net

)

Here is the call graph for this function:

Here is the caller graph for this function:

get_network_image_layer()

Darknet::Image get_network_image_layer	(	Darknet::Network &	net,
		int	i
	)

Here is the call graph for this function:

Here is the caller graph for this function:

get_network_input_size()

int get_network_input_size

(

Darknet::Network &

net

)

Here is the caller graph for this function:

get_network_output()

float * get_network_output

(

Darknet::Network &

net

)

Here is the call graph for this function:

Here is the caller graph for this function:

get_network_output_size()

int get_network_output_size

(

Darknet::Network &

net

)

Here is the caller graph for this function:

get_sequence_value()

int get_sequence_value

(

const Darknet::Network &

net

)

Here is the caller graph for this function:

is_ema_initialized()

int is_ema_initialized

(

const Darknet::Network &

net

)

Here is the caller graph for this function:

lrelu()

static float lrelu ( float  src )

static

make_network()

Darknet::Network make_network

(

int

n

)

Think of this as the constructor for the Darknet::Network object.

Parameters

[in]

n

The number of network layers to initialize.

Here is the caller graph for this function:

make_network_boxes()

detection * make_network_boxes	(	DarknetNetworkPtr	ptr,
		float	thresh,
		int *	num
	)

This is part of the original C API.

See also

make_network_boxes_batch()

Todo:

V3 JAZZ: this is where we spend all our time

Here is the call graph for this function:

Here is the caller graph for this function:

make_network_boxes_batch()

Darknet::Detection * make_network_boxes_batch	(	Darknet::Network *	net,
		float	thresh,
		int *	num,
		int	batch
	)

See also

make_network_boxes()

Here is the call graph for this function:

Here is the caller graph for this function:

make_network_boxes_v3()

Darknet::Detection * make_network_boxes_v3	(	Darknet::Network *	net,
		const float	thresh,
		int *	num,
		Darknet::Output_Object_Cache &	cache
	)

Todo:

Is anything but YOLO still used as an output layer in a modern .cfg file? Should these be removed?

Todo:

V3 JAZZ: 97% of this function is spent in this next line

Here is the call graph for this function:

Here is the caller graph for this function:

network_height()

int network_height

(

Darknet::Network *

net

)

Here is the caller graph for this function:

network_predict()

float * network_predict	(	Darknet::Network &	net,
		float *	input
	)

Here is the call graph for this function:

Here is the caller graph for this function:

network_predict_batch()

det_num_pair * network_predict_batch	(	Darknet::Network *	net,
		Darknet::Image	im,
		int	batch_size,
		int	w,
		int	h,
		float	thresh,
		float	hier,
		int *	map,
		int	relative,
		int	letter
	)

Here is the call graph for this function:

network_predict_data()

matrix network_predict_data	(	Darknet::Network &	net,
		data	test
	)

Here is the call graph for this function:

network_predict_data_multi()

matrix network_predict_data_multi	(	Darknet::Network &	net,
		data	test,
		int	n
	)

Here is the call graph for this function:

network_predict_image()

float * network_predict_image	(	DarknetNetworkPtr	ptr,
		const DarknetImage	im
	)

This is part of the original C API.

Do not use in new code.

See also

network_predict_ptr()

Darknet::predict()

Here is the call graph for this function:

network_predict_image_letterbox()

float * network_predict_image_letterbox	(	DarknetNetworkPtr	ptr,
		DarknetImage	im
	)

This is part of the original C API.

Here is the call graph for this function:

network_predict_ptr()

float * network_predict_ptr	(	DarknetNetworkPtr	ptr,
		float *	input
	)

This is part of the original C API.

Do not use in new code.

See also

network_predict_image()

Darknet::predict()

If you were previously using network_predict() from within C code, please use network_predict_ptr() instead by passing in the address of the network structure (pointer to the network).

Here is the call graph for this function:

network_width()

int network_width

(

Darknet::Network *

net

)

Here is the caller graph for this function:

num_detections()

int num_detections	(	Darknet::Network *	net,
		float	thresh
	)

Todo:

V3 JAZZ: this is where we spend all our time

Here is the call graph for this function:

Here is the caller graph for this function:

num_detections_batch()

int num_detections_batch	(	Darknet::Network *	net,
		float	thresh,
		int	batch
	)

Here is the call graph for this function:

Here is the caller graph for this function:

num_detections_v3()

int num_detections_v3	(	Darknet::Network *	net,
		float	thresh,
		Darknet::Output_Object_Cache &	cache
	)

Basically a wrapper for yolo_num_detections_v3().

Returns

the number of objects found in the current image

Todo:

V3 JAZZ: this is where we spend all our time

Todo:

Is this still used in a modern .cfg file? Should it be removed?

Todo:

Is this still used in a modern .cfg file? Should it be removed?

Here is the call graph for this function:

Here is the caller graph for this function:

recalculate_workspace_size()

int recalculate_workspace_size

(

Darknet::Network *

net

)

Here is the call graph for this function:

Here is the caller graph for this function:

reject_similar_weights()

void reject_similar_weights	(	Darknet::Network &	net,
		float	sim_threshold
	)

Todo:

V5: would be nice if we could call rand_uniform_many() instead

Here is the call graph for this function:

Here is the caller graph for this function:

reset_network_state()

void reset_network_state	(	Darknet::Network *	net,
		int	b
	)

Here is the call graph for this function:

Here is the caller graph for this function:

reset_rnn()

void reset_rnn

(

DarknetNetworkPtr

ptr

)

This is part of the original C API.

Here is the call graph for this function:

resize_network()

int resize_network	(	Darknet::Network *	net,
		int	w,
		int	h
	)

Here is the call graph for this function:

Here is the caller graph for this function:

restore_network_recurrent_state()

void restore_network_recurrent_state

(

Darknet::Network &

net

)

Here is the call graph for this function:

Here is the caller graph for this function:

set_batch_network()

void set_batch_network	(	Darknet::Network *	net,
		int	b
	)

Here is the call graph for this function:

Here is the caller graph for this function:

train_network()

float train_network	(	Darknet::Network &	net,
		data	d
	)

Here is the call graph for this function:

Here is the caller graph for this function:

train_network_batch()

float train_network_batch	(	Darknet::Network	net,
		data	d,
		int	n
	)

Here is the call graph for this function:

train_network_datum()

float train_network_datum	(	Darknet::Network &	net,
		float *	x,
		float *	y
	)

Here is the call graph for this function:

Here is the caller graph for this function:

train_network_waitkey()

float train_network_waitkey	(	Darknet::Network &	net,
		data	d,
		int	wait_key
	)

Here is the call graph for this function:

Here is the caller graph for this function:

update_network()

void update_network

(

Darknet::Network &

net

)

Here is the call graph for this function:

Here is the caller graph for this function:

visualize_network()

void visualize_network

(

Darknet::Network &

net

)

Here is the call graph for this function:

Here is the caller graph for this function: