See also: https://www.ccoderun.ca/programming/darknet_faq/#configuration_template

Darknet comes with many sample .cfg files to use as templates.

To select which configuration file to use as a template, click on the configuration template button at the top of the darknet output window:

More than half of these configuration files are from many years ago, some dating from Joseph Redmon's original darknet repo. On this page the .cfg files have been split into 2 tables (new and old) loosely based on when a configuration file was last modified and/or if it is still considered relevant or outdated.

Note

In the tables below:

voc refers to PASCAL Visual Object Classes, a dataset with 20 classes (person, car, bicycle, bus, motorbike, train, aeroplane, chair, ...)
coco refers to COCO, a dataset with 80 classes (person, bicycle, car, motorcycle, airplane, bus, train, truck, ...)
spp refers to spatial pyramid pooling
lstm refers to long short-term memory

Getting Started

If uncertain where to start when training your own neural network, try one of the following templates:

Filename	Layers	Size	Release Date	Comments
`yolov4-tiny.cfg`	38	23.1 MiB	2020-06-25	Has 2 YOLO layers. Great configuration to use, quite fast. This is what I typically use when setting up new projects.
`yolov4-tiny-3l.cfg`	45	23.3 MiB	2020-07-11	Has 3 YOLO layers. Better than "tiny" at finding very small objects.
`yolov4.cfg`	162	245.7 MiB	2020-04-23	Has 3 YOLO layers. Best precision, but much larger and slower than the "tiny" variants.

New Configuration

The following configuration files are relatively new:

name	lines	layers	yolo layers	network size	weights size	related	last commit	commit name	notes	links
`cd53paspp-gamma.cfg`	1154	162	3	512x512		`yolov4.cfg`	2020-03-31	AlexeyAB	detector, the same as yolov4.cfg, but with leaky instead of mish
`csdarknet53-omega.cfg`	763	108		256x256		`yolov4.cfg`	2020-03-31	AlexeyAB	classifier, backbone for yolov4.cfg
`cspx-p7-mish-omega.cfg`	1459	212		320x320		`cspx-p7-mish.cfg`	2020-11-03	AlexeyAB	classifier, backbone for cspx-p7-mish.cfg
`cspx-p7-mish.cfg`	2604	375	5	1536x1536			2020-12-11	AlexeyAB	detector, yolov4-p7-large
`cspx-p7-mish_hp.cfg`	2638	375	5	896x896			2020-11-03	AlexeyAB	detector, experimental cfg file
`csresnext50-panet-spp-original-optimal.cfg`	1043	138	3	608x608	215.9 MiB	`csresnext50-panet-spp.cfg`	2020-03-31	AlexeyAB	"the best model for detection"	5092, 2859 (SPP)
`efficientnet-lite3.cfg`	1010	117		288x288			2020-03-23	AlexeyAB	classifier, EfficientNet-lite3	3380
`resnet152_trident.cfg`	2178	291	3	608x608			2020-01-22	AlexeyAB	"models are experimental"	5092
`yolov4-csp-s-mish.cfg`	914	112	3	640x640			2021-10-30	AlexeyAB
`yolov4-csp-swish.cfg`	1356	176	3	640x640			2021-07-12	AlexeyAB
`yolov4-csp-x-mish.cfg`	1554	204	3	640x640			2021-10-30	AlexeyAB
`yolov4-csp-x-swish-frozen.cfg`	1557	204	3	640x640			2021-07-12	AlexeyAB
`yolov4-csp-x-swish.cfg`	1557	204	3	640x640			2021-07-12	AlexeyAB
`yolov4-csp.cfg`	1280	175	3	512x512	202.1 MiB	`yolov4.cfg`	2020-12-15	AlexeyAB	cross-stage-partial; more accurate and faster than YOLOv4, but decreased detection for small objects	Scaled YOLO v4, YOLOv4-CSP whitepaper
`yolov4-custom.cfg`	1161	162	3	608x608	245.7 MiB	`yolov4.cfg`	2020-05-01	AlexeyAB	nearly identical; lower learning rate; one additional change to a convolutional layer	YOLOv4 whitepaper
`yolov4-p5-frozen.cfg`	1838	244	3	896x896			2021-06-29	AlexeyAB
`yolov4-p5.cfg`	1837	244	3	896x896		`yolov4.cfg`	2021-06-27	AlexeyAB		7414, 7087
`yolov4-p6.cfg`	2298	305	4	1280x1280		`yolov4.cfg`	2021-06-27	AlexeyAB		7414, 7087
`yolov4-sam-mish-csp-reorg-bfm.cfg`	1430	216	3	512x512			2021-05-11	AlexeyAB
`yolov4-tiny-3l.cfg`	333	45	3	608x608	23.3 MiB	`yolov4-tiny.cfg`	2020-07-11	AlexeyAB	better at finding small objects; "3l" refers to 3 YOLO layers vs the usual 2 in "tiny"
`yolov4-tiny-custom.cfg`	282	38	2	416x416	23.1 MiB	`yolov4-tiny.cfg`	2020-07-11	AlexeyAB	similar to yolov4-tiny.cfg, but contains 1 minor change to the first YOLO layer	5346
`yolov4-tiny.cfg`	295	38	2	416x416	23.1 MiB		2020-12-15	AlexeyAB	contains 2 YOLO layers	5346
`yolov4-tiny_contrastive.cfg`	453	55	1	416x416	27.5 MiB	`yolov4-tiny.cfg`	2020-10-20	AlexeyAB	"experimental"; "suitable for un-supervised learning and for multi-camera object tracking"	6892
`yolov4.cfg`	1159	162	3	608x608	245.7 MiB		2020-05-23	AlexeyAB	contains 3 YOLO layers	YOLOv4 whitepaper
`yolov4x-mish.cfg`	1437	203	3	640x640	380.9 MiB	`yolov4-csp.cfg`	2020-12-07	AlexeyAB	detector; something between yolov4-csp and yolov4-p5; more suitable for high resolutions 640x640 - 832x832 than yolov4.cfg; should be trained longer	7131
`yolov4_new.cfg`	1177	162	3	608x608			2021-10-30	AlexeyAB
`yolov7-tiny.cfg`	707	99	3	416x416			2022-07-07	AlexeyAB84
`yolov7.cfg`	1025	143	3	640x640			2022-08-11	AlexeyAB84
`yolov7x.cfg`	1153	159	3	640x640			2022-08-11	AlexeyAB84

Old Configuration

The following configuration files are from older versions of darknet:

name	lines	layers	yolo layers	network size	weights size	related	last commit	commit name	notes	links
`alexnet.cfg`	96	15		227x227			2016-09-12	Joseph Redmon	old cfg
`cifar.cfg`	127	19		32x32			2016-03-14	Joseph Redmon	"for Classification rather than Detection"	5092
`cifar.test.cfg`	120	18		32x32			2016-03-14	Joseph Redmon	old cfg
`crnn.train.cfg`	53	7				`rnn.train.cfg`	2019-03-18	AlexeyAB	old cfg	1624
`csresnext50-panet-spp.cfg`	1019	138	3	512x512			2019-12-18	AlexeyAB	old cfg	2859 (SPP)
`darknet.cfg`	112	17		224x224			2016-11-18	Joseph Redmon	old cfg
`darknet19.cfg`	195	27		224x224			2016-11-26	Joseph Redmon	old cfg
`darknet19_448.cfg`	203	27		448x448			2017-09-15	AlexeyAB	old cfg
`darknet53.cfg`	567	78		256x256			2018-10-01	AlexeyAB	"for Classification rather than Detection"	5092
`darknet53_448_xnor.cfg`	620	79		448x448			2019-06-22	AlexeyAB	"for Classification rather than Detection"	5092
`densenet201.cfg`	1955	306		256x256			2017-09-14	AlexeyAB	"for Classification rather than Detection"	5092
`efficientnet_b0.cfg`	1010	136		224x224			2019-12-04	AlexeyAB	"for Classification rather than Detection"	5092, 3380
`enet-coco.cfg`	1073	146	2	416x416			2019-09-04	AlexeyAB	"partial residual network"	Enriching Variety of Layer-wise Learning Information by Gradient Combination
`extraction.cfg`	207	28		224x224			2016-08-05	Joseph Redmon	old cfg
`extraction.conv.cfg`	180	27		256x256			2015-11-09	Joseph Redmon	old cfg
`extraction22k.cfg`	210	28		224x224			2016-09-01	Joseph Redmon	old cfg
`Gaussian_yolov3_BDD.cfg`	808	107		512x512			2019-11-16	AlexeyAB	"models are experimental"	5092
`go.test.cfg`	132	16		19x19			2016-09-01	Joseph Redmon	old cfg
`gru.cfg`	35	6					2016-06-06	Joseph Redmon	old cfg
`jnet-conv.cfg`	119	18		10x10			2015-07-20	Joseph Redmon	old cfg
`lstm.train.cfg`	36	6					2019-01-29	AlexeyAB	"long short-term memory"	3114
`resnet101.cfg`	991	138		256x256			2019-11-30	AlexeyAB	"for Classification rather than Detection"	5092
`resnet152.cfg`	1464	206		256x256			2018-02-15	AlexeyAB	"for Classification rather than Detection"	5092
`resnet50.cfg`	512	70		256x256			2017-09-14	AlexeyAB	"for Classification rather than Detection"	5092
`resnext152-32x4d.cfg`	1563	205		256x256			2018-10-01	AlexeyAB	"for Classification rather than Detection"	5092
`rnn.cfg`	41	6					2016-02-05	Joseph Redmon	old cfg
`rnn.train.cfg`	41	6				`crnn.train.cfg`	2019-01-28	AlexeyAB	old cfg	1624
`strided.cfg`	186	27		256x256			2015-11-03	Joseph Redmon	"for Classification rather than Detection"	5092
`t1.test.cfg`	118	16		224x224			2016-06-19	Joseph Redmon	old cfg
`tiny-coco.cfg`	126	16		448x448			2016-11-17	Joseph Redmon	old cfg
`tiny-yolo-voc.cfg`	135	16		416x416			2017-04-07	AlexeyAB	old cfg
`tiny-yolo.cfg`	135	16		416x416			2016-11-26	Joseph Redmon	old cfg
`tiny-yolo.cfg`	127	16		448x448			2016-11-17	Joseph Redmon	old cfg
`tiny-yolo_xnor.cfg`	149	16		416x416			2018-09-22	AlexeyAB	old cfg
`tiny.cfg`	173	23		224x224			2016-11-15	Joseph Redmon	old cfg
`vgg-16.cfg`	154	26		256x256			2015-08-17	Joseph Redmon	"for Classification rather than Detection"	5092
`vgg-conv.cfg`	122	18		224x224			2015-07-20	Joseph Redmon	"for Classification rather than Detection"	5092
`writing.cfg`	42	5		256x256			2015-09-23	Joseph Redmon	old cfg
`xyolo.test.cfg`	144	23		448x448			2016-11-17	Joseph Redmon	old cfg
`yolo-coco.cfg`	256	31		448x448			2016-11-17	Joseph Redmon	old cfg
`yolo-small.cfg`	240	34		448x448			2016-11-17	Joseph Redmon	old cfg
`yolo-voc.2.0.cfg`	245	31		416x416			2017-04-28	AlexeyAB	old cfg
`yolo-voc.cfg`	259	32		416x416			2017-04-07	AlexeyAB	old cfg
`yolo.2.0.cfg`	245	31		416x416			2017-04-28	AlexeyAB	old cfg
`yolo.cfg`	258	32		448x448			2016-11-17	Joseph Redmon	old cfg
`yolo.cfg`	259	32		416x416			2017-04-07	AlexeyAB	old cfg
`yolo.train.cfg`	258	32		448x448			2016-11-17	Joseph Redmon	old cfg
`yolo2.cfg`	252	31		448x448			2016-11-17	Joseph Redmon	old cfg
`yolo9000.cfg`	219	25		544x544			2017-07-28	AlexeyAB	old cfg
`yolov2-tiny-voc.cfg`	139	16		416x416			2018-03-28	AlexeyAB	old cfg
`yolov2-tiny.cfg`	140	16		416x416			2018-03-28	AlexeyAB	old cfg
`yolov2-voc.cfg`	259	32		416x416			2018-03-28	AlexeyAB	old cfg
`yolov2.cfg`	259	32		416x416			2018-03-28	AlexeyAB	old cfg
`yolov3-openimages.cfg`	790	107	3	608x608		`yolov3.cfg`	2018-10-15	AlexeyAB	YOLOv3 but already setup for 601 classes	OpenImage Dataset
`yolov3-spp.cfg`	823	114	3	608x608		`yolov3.cfg`	2018-08-03	AlexeyAB	YOLOv3 but with extra layers for spacial pyramid pooling	2859 (SPP)
`yolov3-tiny-prn.cfg`	200	28	2	416x416		`yolov3-tiny.cfg`	2019-09-04	AlexeyAB	"partial residual network"	Enriching Variety of Layer-wise Learning Information by Gradient Combination
`yolov3-tiny.cfg`	183	24	2	416x416	33.1 MiB		2018-05-07	Bartek Gąsiorzewski	contains 2 YOLO layers
`yolov3-tiny_3l.cfg`	227	31	3	608x608	34.4 MiB	`yolov3-tiny.cfg`	2019-01-04	AlexeyAB	better at finding small objects; "3l" refers to 3 YOLO layers vs the usual 2 in "tiny"
`yolov3-tiny_obj.cfg`	183	24	2	416x416	33.1 MiB	`yolov3-tiny.cfg`	2018-05-18	AlexeyAB	Exactly the same as yolov3-tiny.cfg.
`yolov3-tiny_occlusion_track.cfg`	219	27	2	416x416		`yolov3-tiny.cfg`	2019-03-02	AlexeyAB	"object Detection & Tracking using conv-rnn layer on frames from video"	2553
`yolov3-tiny_xnor.cfg`	198	24	2	416x416		`yolov3-tiny.cfg`	2018-09-22	AlexeyAB	"XNOR-net ~2x faster than cuDNN on CUDA"	3054
`yolov3-voc.cfg`	786	107	3	416x416		`yolov3.cfg`	2018-03-30	AlexeyAB	similar to the usual yolov3.cfg but pre-configured for 20 VOC classes
`yolov3-voc.yolov3-giou-40.cfg`	809	107	3	416x416		`yolov3-voc.cfg`	2019-06-01	AlexeyAB	"models are experimental"	5092, Generalized Intersection over Union (GIoU)
`yolov3.cfg`	790	107	3	416x416	234.9 MiB		2018-05-06	AlexeyAB	contains 3 YOLO layers
`yolov3.coco-giou-12.cfg`	807	107	3	608x608		`yolov3.cfg`	2019-06-01	AlexeyAB	"models are experimental"	5092, Generalized Intersection over Union (GIoU)
`yolov3_5l.cfg`	969	131	5	416x416	236.2 MiB	`yolov3.cfg`	2018-12-11	AlexeyAB	"5l" refers to 5 YOLO layers; "...or very small objects, or if you want to set high network resolution"	5092