● Note that apt install with deb is preferred to run file and source file build for both of CUDA and cuDNN

• Download and install CUDA following instructions at here - https://developer.nvidia.com/cuda-downloads
• Download and install cuDNN following instructions at here - https://developer.nvidia.com/cudnn-downloads
  • If you want, also refer to here - https://docs.nvidia.com/deeplearning/cudnn/installation/linux.html#
• Set up environmental paths

gedit ~/.bashrc
*** Type and save below, CUDA_PATH should be like /usr/local/cuda-11.5, depending on your version ***
export PATH=CUDA_PATH/bin:$PATH 
export LD_LIBRARY_PATH=CUDA_PATH/lib64:$LD_LIBRARY_PATH

. ~/.bashrc

gedit ~/.profile
*** Type and save below, CUDA_PATH should be like /usr/local/cuda-11.5, depending on your version ***
export PATH=CUDA_PATH/bin:$PATH 
export LD_LIBRARY_PATH=CUDA_PATH/lib64:$LD_LIBRARY_PATH

. ~/.profile

• Verify, if installed properly

# Verify
dpkg -l | grep cuda
dpkg -l | grep cudnn
nvcc --version

● Note that apt install with deb is preferred to other methods for TensorRT

• Download TensorRT at here - https://developer.nvidia.com/tensorrt-download
• Follow the instructions at here - https://docs.nvidia.com/deeplearning/tensorrt/install-guide/index.html#installing-debian
  • Installing full packages is recommended, which means:

  sudo apt install tensorrt
  sudo apt install python3-libnvinfer-dev
  sudo apt install onnx-graphsurgeon
  

● Common step for training / engine file

0. Make sure that you have installed all dependencies properly.

• Particularly, you should install full packages of TensorRT: tensorrt, python3-libnvinfer-dev, onnx-graphsurgeon

1. Install and make Python3 virtual env

python3 -m pip install virtualenv virtualenvwrapper
cd <PATH YOU WANT TO SAVE VIRTUAL ENVIRONMENT>
virtualenv -p python3 <NAME YOU WANT>

*** Now you can activate with
source <PATH YOU SAVED>/<NAME YOU WANT>/bin/activate

*** Deactivate with
deactivate

2. (While virtual env being activated), clone YOLOv9 repo and install requirements

git clone https://github.com/WongKinYiu/yolov9
cd yolov9
pip install -r requirements.txt

● Converting .pt to .onnx, and then .engine

0. (While virtual env being activated)
1. Get trained YOLOv9 weight file as .pt by training your own data or downloading the pre-trained model at here - https://github.com/WongKinYiu/yolov9/releases
2. Reparameterize the .pt file (saving computation, memory, and size by trimming unnecessary parts for inference but necessary only for training)

cd yolov9 # cloned at above step
wget https://raw.githubusercontent.com/engcang/TensorRT_YOLOv9_ROS/main/reparameterize.py

*** Change the number of classes in the reparameterize.py in line 8 (nc=80)
python reparameterize.py yolov9-c.pt yolov9-c-reparameterized.pt # input.pt output.put

3. Export .pt file as .onnx

python export.py --weights yolov9-c-reparameterized.pt --include onnx

4. Then .onnx to .engine

/usr/src/tensorrt/bin/trtexec --onnx=yolov9-c-reparameterized.onnx --saveEngine=yolov9-c.engine
#for faster, less accurate
/usr/src/tensorrt/bin/trtexec --onnx=yolov9-c-reparameterized.onnx --saveEngine=yolov9-c-fp16.engine --fp16
#not recommended - much faster, much less accurate
/usr/src/tensorrt/bin/trtexec --onnx=yolov9-c-reparameterized.onnx --saveEngine=yolov9-c-int8.engine --int8

● Training your own data

0. (While virtual env being activated) + YOLOv9 is cloned already, requirements are installed already
1. Prepare data and labels in YOLO format.

• You may want to use this - https://github.com/AlexeyAB/Yolo_mark
• Or roboflow - https://docs.ultralytics.com/yolov5/tutorials/roboflow_datasets_integration/

2. Make proper data.yaml file by copying and editing yolov9/data/coco.yaml as follows:

path: training  # dataset root dir (relative from train.py file)
train: train    # train images folder (relative to 'path')
val: val        # val images folder (relative to 'path')
test: test      # test images folder (relative to 'path')

# Classes
names:
  0: Transmission tower
  1: Insulator

3. Make proper yolov9.yaml file by copying and editing yolov9/models/detect/yolov9.yaml or yolov9-c, yolov9-e, etc.

# parameters
nc: 2  # number of classes
depth_multiple: 1.0  # model depth multiple
width_multiple: 1.0  # layer channel multiple
#activation: nn.LeakyReLU(0.1)
#activation: nn.ReLU()

# anchors
anchors: 3

# YOLOv9 backbone
backbone:
  [
   [-1, 1, Silence, []],  
   
   # conv down
   [-1, 1, Conv, [64, 3, 2]],  # 1-P1/2
   
   ...
  ]

4. Edit learning parameters by editing yolov9/data/hyps/hyp.scratch-high.yaml
5. Put all of files properly in the yolov9 folder. If outside the yolov9 folder, error occurs!

yolov9
│  ...
├─ data # Reference folder
│  ├─ coco.yaml
│  └─ hyps
│     └─ hyp.scratch-high.yaml
├─ models # Reference folder
│  ...
│  ├─ detect
│  ...
│  │  ├─ yolov9-c.yaml
│  │  ├─ yolov9-e.yaml
│  │  └─ yolov9.yaml
├─ runs # Output saved folder
│  ...
├─ train.py # Using this file for GELAN
├─ train_dual.py # Using this file for YOLOv9
├─ training # Using this folder
│  ├─ yolov9-c.pt
│  ├─ data.yaml
│  ├─ yolov9.yaml
│  ├─ test
│  │  ├─ 02001.jpg
│  │  ├─ 02001.txt
│  │  └─ ...
│  ├─ train
│  │  ├─ 00001.jpg
│  │  ├─ 00001.txt
│  │  └─ ...
│  ├─ val
│  │  ├─ 04000.jpg
│  │  ├─ 04000.txt
│  │  └─ ...
└─ └─ ...

6. Train

cd yolov9

*** Using pretrained model (yolov9-c.pt here), fine-tuning:
python train_dual.py --batch-size 4 --epochs 100 --img 640 --device 0 --close-mosaic 15 \
--data training/data.yaml --weights training/yolov9-c.pt --cfg training/yolov9.yaml --hyp data/hyps/hyp.scratch-high.yaml

*** From the scratch:
python train_dual.py --batch-size 4 --epochs 100 --img 640 --device 0 --close-mosaic 15 \
--data training/data.yaml --weights '' --cfg training/yolov9.yaml --hyp data/hyps/hyp.scratch-high.yaml

● Trouble shooting for training

0. (While virtual env being activated)
1. AttributeError: 'FreeTypeFont' object has no attribute 'getsize'

• This is because installed Pillow version is too recent.
• Solve with pip install Pillow==9.5.0

2. Getting Killed and does not train

• Lack of memory, reduce batch-size a lot

3. AssertionError: Invalid CUDA '--device 0' requested, use '--device cpu' or pass valid CUDA device(s)

• This is because installed torch and torchvision are not CUDA versions.
• Solve as:

*** Check the version at https://download.pytorch.org/whl/torch_stable.html
*** torch >= 1.7.0, torchvision>=0.8.1

pip install torch==1.11.0+cu115 torchvision==0.12.0+cu115 -f https://download.pytorch.org/whl/torch_stable.html

4. RuntimeError: CUDA out of memory. Tried to allocate 50.00 MiB (GPU 0; 9.76 GiB total capacity; 6.68 GiB already allocated; 45.00 MiB free; 6.82 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

• Lack of memory, reduce batch-size a lot