AI PIPELINE🚀

See the YOLOv5 Docs for full documentation on training, testing and deployment. See below for quickstart examples.

How to Generate Yolov5 model for VA8801?

1. Prepare Dataset: Use example data at data/dataset or Use your custom dataset

2. Inference: Inference testing data with a TFLite pretrained model, which can be downloaded from the model zoo for the Human model (input=96x96x1)

• Please check your local model path -w "pretrained pytorch model path"

  Example of your local model folder

  path: VA8801_Model_Zoo/ObjectDetection/Human_Detection/Yolo

bash python tflite_runtime.py -s data/dataset/test/human_001.jpg -w path/HUMAN_DET_6_001_001.tflite --img_ch 1 2. Infernece: Inference testing data with a TFLite pretrained model, which can be downloaded from the model zoo for the Human model (input=320x320x3)

bash python tflite_runtime.py -s data/dataset/test/human_002.jpg -w path/HUMAN_DET_7_002_002.tflite --img_ch 3

1. Train model: Transfer learning with a PyTorch pretrained model, which can be downloaded from the model zoo for the Human model (input=96x96x1)

• Please check your local model path --weights "pretrained pytorch model path"

  Example of your local model folder

  path: VA8801_Model_Zoo/ObjectDetection/Human_Detection/Yolo

• Please check your PC device --device "cuda device, i.e. 0 or 0,1,2,3 or cpu"

bash python train.py --device 0 --data data/training_cfg/data_config.yaml --weights path/HUMAN_DET_6_001_001.pt --imgsz 96 --imgch 1 --cfg models/yolov5n_WM005_DM033.yaml

1. Train model: Transfer learning with a PyTorch pretrained model, which can be downloaded from the model zoo for the Human model (input=320x320x3)

bash python train.py --device 0 --data data/training_cfg/data_config.yaml --weights path/HUMAN_DET_7_002_002.pt --imgsz 320 --imgch 3 --cfg models/2_head_yolov5n_WM022.yaml

1. Export int8 tflite model
2. Please check your local model path --weights "your pytorch model path"
   • After training, your trained model will be saved at results/yyyy_mm_dd/trialx/weights/best.pt
3. Please check the image size for export to the TFLite model --imgsz_tflite "image size".

4. Please check your PC device --device "cuda device, i.e. 0 or 0,1,2,3 or cpu"

   If model input=96x96x1 ```bash python aipipeline.py --data data/trainingcfg/dataconfig.yaml --weights path/HUMANDET6001001.pt --batch-size 1 --imgch 1 --imgsz 96 --imgsztflite 96 --device 0 --include tflite --int8 --run export

   ```

   If model input=320x320x3 ```bash python aipipeline.py --data data/trainingcfg/dataconfig.yaml --weights path/HUMANDET7002002.pt --batch-size 1 --imgch 3 --imgsz 320 --imgsztflite 320 --device 0 --include tflite --int8 --run export

   ```

Example for train from scatch

run training only ```bash python ai_pipeline.py --data --cfg --epochs 10 --batch-size 64 --imgch 1 --imgsz 320 --patience 0 --device 0 --run train ``` run export only ```bash python ai_pipeline.py --data --weights --batch-size 1 --imgch 1 --imgsz 192 --device 0 --include tflite --int8 --run export ``` run inference only ```bash python ai_pipeline.py --data --conf-thres-test 0 --device 0 --tflite_model_path --save_dir --run inference ``` run tflite inference for va8801 results only ```bash python tflite_runtime.py -s -w ```