1. Setup

Install darknet framework writing in C on Jetson Nano

cd ~
git clone https://github.com/SokPhanith/image_classification_darknet_jetson_nano.git
cd image_classification_darknet_jetson_nano
git clone https://github.com/AlexeyAB/darknet
rm darknet/Makefile
cp Makefile darknet/
cd darknet
make

2. Download model

Download Image Classification pretrain model 1000 classes from ImgaeNet dataset training in darknet framework. Source

cd ~
cd image_classification_darknet_jetson_nano
./download_model.sh

3. Testing with darknet application build from C

cd ~
cd image_classification_darknet_jetson_nano/darknet
./darknet classifier predict cfg/imagenet1k.data cfg/darknet19.cfg ../weight/darknet19.weights data/dog.jpg
./darknet classifier predict cfg/imagenet1k.data cfg/darknet53.cfg ../weight/darknet53.weights

4. Testing with python by libdarknet.so librery

cd ~
cd image_classification_darknet_jetson_nano
cp *.py darknet/
cd darknet/
python3 runtime_gpu.py --weights ../weight/darknet53.weights --config_file cfg/darknet53.cfg --data_file cfg/imagenet1k.data --csi
python3 runtime_cpu.py --weights ../weight/darknet53.weights --config_file cfg/darknet53.cfg --data_file cfg/imagenet1k.data --image data/dog.jpg

5. Training with cifar dataset

cd image_classification_darknet_jetson_nano/darknet/data
wget https://pjreddie.com/media/files/cifar.tgz
tar xzf cifar.tgz
cd cifar
find `pwd`/train -name \*.png > train.list
find `pwd`/test -name \*.png > test.list
cd ../..
cp ../cifar.data cfg/
cp ../cifar_small.cfg cfg/
./darknet classifier train cfg/cifar.data cfg/cifar_small.cfg
./darknet classifier valid cfg/cifar.data cfg/cifar_small.cfg backup/cifar_small_final.weights
python3 runtime_gpu.py --weights backup/cifar_small_final.weights --config_file cfg/cifar_small.cfg --data_file cfg/cifar.data --image data/cifar/test/0_cat.png

6. Training with custom dataset darknet19 model

My custom dataset has 4 classes arduino, cnc, esp8266 and pyboard. In data/custom have test/ and train/. In a test/ and train/ have classes arduino, cnc, esp8266 and pyboard folder. Imgae was put in by there folder name. If you have a different classes or more then this classes must edit filters=4 line 181 by count same your dataset.

cd ~
cd image_classification_darknet_jetson_nano/darknet
python3 convert_dataset.py --path_dataset data/custom --new_path_dataset data/board
cp ../darknet19.cfg data/board
./darknet classifier train cfg/board/custom.data cfg/board/darknet19.cfg
./darknet classifier valid cfg/board/custom.data cfg/board/darknet19.cfg cfg/board/darknet19_last.weights
python3 runtime_gpu.py --weights cfg/board/darknet19_last.weights --config_file cfg/board/darknet19.cfg --data_file cfg/board/custom.data --image data/board/test/0_esp8266.jpg