Awesome Open Source
Awesome Open Source


lightnet is a turnkey solution to real world problems accelerated with deep learning AI technology, including but not limited to object detection, image classification and human pose estimation.

How to read the source code

This project is dependent on a few open-source projects:

  • modules/darknet - the main engine for training & inferencing.
  • modules/Yolo_mark - the toolkit to prepare training data for object detection.
  • modules/yolo2_light - lightweighted inferencing engine [optional].
  • modules/cvui - lightweighted GUI based purely on OpenCV.
  • moudles/pytorch-caffe-darknet-convert - DL framework model converter
  • modules/minitrace - library to generate tracing logs for Chrome "about:tracing"
  • modules/readerwriterqueue - single-producer, single-consumer lock-free queue for C++
  • modules/bhtsne - Barnes-Hut implementation of the t-SNE algorithm

How to build from Visual Studio 2015


Install OpenCV

Build it

Execute the batch file


Object Detection - inference w/ pre-trained weights

First you need to download the weights. You can read more details on darknet website.

cfg weights names
cfg/yolov2.cfg data/coco.names
cfg/yolov2-tiny.cfg coco.names
cfg/yolo9000.cfg cfg/9k.names
cfg/yolov3.cfg cfg/coco.names
cfg/yolov3-openimages.cfg data/openimages.names
cfg/yolov3-tiny.cfg cfg/coco.names
cfg/yolov2_shoe.cfg yolov2_shoe.weights obj.names

Syntax for object detection

darknet.exe detector demo <data> <cfg> <weights> -c <camera_idx> -i <gpu_idx>
darknet.exe detector demo <data> <cfg> <weights> <video_filename> -i <gpu_idx>
darknet.exe detector test <data> <cfg> <weights> <img_filename> -i <gpu_idx>

Default launch device combination is -i 0 -c 0.

Run yolov3 on camera #0

darknet.exe detector demo cfg/ cfg/yolov3.cfg yolov3.weights

Run yolo9000 on camera #0

darknet.exe detector demo cfg/ cfg/yolo9000.cfg yolo9000.weights

Run yolo9000 on images

darknet.exe detector test cfg/ cfg/yolo9000.cfg yolo9000.weights

Run yolo9000 CPU on camera #0

darknet_no_gpu.exe detector demo cfg/ cfg/yolo9000.cfg yolo9000.weights

Object Detection - label images manually

Object Detection - train yolo v2 network

  1. Fork __template-yolov2 to my-yolo-net

  2. Download pre-trained weights for the convolutional layers: to bin/darknet19_448.conv.23

  3. To training for your custom objects, you should change 2 lines in file obj.cfg:

  • change classes in
  • set number of classes (objects) in obj.cfg#L230
  • set filter-value equal to (classes + 5)*5 in obj.cfg#L224
  1. Run my-yolo-net/train.cmd

Object Detection - train yolo v3 network

  1. Fork __template-yolov3 to my-yolo-net

  2. Download pre-trained weights for the convolutional layers: to bin/darknet53.conv.74

  3. Create file obj.cfg with the same content as in yolov3.cfg (or copy yolov3.cfg to obj.cfg) and:

  • change line batch to batch=64
  • change line subdivisions to subdivisions=8
  • change line classes=80 to your number of objects in each of 3 [yolo]-layers:
    • obj.cfg#L610
    • obj.cfg#L696
    • obj.cfg#L783
  • change [filters=255] to filters=(classes + 5)x3 in the 3 [convolutional] before each [yolo] layer
    • obj.cfg#L603
    • obj.cfg#L689
    • obj.cfg#L776

So if classes=1 then should be filters=18. If classes=2 then write filters=21.

(Do not write in the cfg-file: filters=(classes + 5)x3)

(Generally filters depends on the classes, coords and number of masks, i.e. filters=(classes + coords + 1)*<number of mask>, where mask is indices of anchors. If mask is absence, then filters=(classes + coords + 1)*num)

So for example, for 2 objects, your file obj.cfg should differ from yolov3.cfg in such lines in each of 3 [yolo]-layers:



Image Classification - inference w/ pre-trained weights

Again, you need download weights first. You can read more details on darknet website.

cfg weights

Image Classification - train darknet19_448 network

  1. Fork __template-darknet19_448 to my-darknet19-net

  2. Download pre-trained weights for the convolutional layers: to bin/darknet19_448.conv.23

  3. Create file obj.cfg with the same content as in darknet19_448.cfg (or copy darknet19_448.cfg to obj.cfg) and:

  • set batch to 128 or 64 or 32 depends on your GPU memory in darknet19-classify.cfg#L4
  • change line to subdivisions=4
  • set filter-value equal to classes in darknet19-classify.cfg#L189

Human Pose Estimation - inference w/ pre-trained weights

This project lives in DancingGaga

For more details, please check the README there.

[Weight file] (darknet version openpose.weight)


blobFromImage() vs letterbox_image() vs resize_image()

How to fix CUDA Error: no kernel image is available for execution on the device?

# Tesla V100
# ARCH= -gencode arch=compute_70,code=[sm_70,compute_70]

# GeForce RTX 2080 Ti, RTX 2080, RTX 2070, Quadro RTX 8000, Quadro RTX 6000, Quadro RTX 5000, Tesla T4, XNOR Tensor Cores
# ARCH= -gencode arch=compute_75,code=[sm_75,compute_75]

# Jetson XAVIER
# ARCH= -gencode arch=compute_72,code=[sm_72,compute_72]

# GTX 1080, GTX 1070, GTX 1060, GTX 1050, GTX 1030, Titan Xp, Tesla P40, Tesla P4
# ARCH= -gencode arch=compute_61,code=sm_61 -gencode arch=compute_61,code=compute_61

# GP100/Tesla P100 - DGX-1
# ARCH= -gencode arch=compute_60,code=sm_60

# For Jetson TX1, Tegra X1, DRIVE CX, DRIVE PX - uncomment:
# ARCH= -gencode arch=compute_53,code=[sm_53,compute_53]

# For Jetson Tx2 or Drive-PX2 uncomment:
# ARCH= -gencode arch=compute_62,code=[sm_62,compute_62]

How to fine tune a existing network?

darknet.exe partial cfg/darknet.cfg darknet.weights darknet.conv.13 13 darknet.exe partial cfg/extraction.cfg extraction.weights extraction.conv.23 23 darknet.exe partial cfg/darknet19.cfg darknet19.weights darknet19.conv.23 23 darknet.exe partial cfg/darknet19_448.cfg darknet19_448.weights darknet19_448.conv.23 23 darknet.exe partial cfg/darknet53.cfg darknet53.weights darknet53.conv.74 74 darknet.exe partial cfg/resnet50.cfg resnet50.weights resnet50.conv.66 66

Explanation of yolo training output

CFG Parameters

How to use scripts folder?

pip install -r scripts/requirements.txt

Get A Weekly Email With Trending Projects For These Topics
No Spam. Unsubscribe easily at any time.
Python (1,147,256
Deep Learning (23,829
Neural Network (8,646
Computer Vision (8,257
Opencv (6,350
Ai (4,833
Cuda (3,977
Classification (3,163
Deep Neural Networks (2,872
Gpu (1,712
Detection (1,222
Yolo (648
Yolov3 (639
Dnn (463
Pose Estimation (445
Darknet (311
Yolov2 (155
Cudnn (106
Yolo2 (34
Related Projects