Readme.Windows.gpu.md

Windows 11

Prerequisite

CUDA GPU is needed for GPU training. Tested with RTX 2080 super(8GB), windows 11.

Create environment

conda create -n yolov8_gpu python=3.9
conda activate yolov8_gpu
pip install ultralytics==8.0.221
pip install --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install tensorflow==2.13.1

Export yolov8n to tflite and onnx format

python export_models.py
python export_models.py --format onnx

Note, It seems like there is a bug when I export tflite and onnx at the same time. So for now export them separately.

Run

Run yolov8n.pt

• --debug option show debug window with annotation, good for debugging but slows down the fps
• --print_fps option prints fps every 1 sec.

python main.py --debug
python main.py --print_fps

Run exported models

python main.py --model=./models/yolov8n.onnx --debug
python main.py --model=./models/yolov8n_saved_model/yolov8n_integer_quant.tflite --debug

Train

Training yolov8n - low resolution(320) with coco dataset.

python train.py
# or
python train.py --model=yolov8n.yaml --imgsz=320 --batch 128