Trained to detect tabla (Indian classical musical instrument)
git@github.com:walnashgit/S15YoloV9.git
The custom data set is in the folder tabla. If using a different dataset, follow the folder structue same as the one here and update the data.yaml file accordingly.
%cd S15YoloV9
!pip install -r requirements.txt -q
#!wget https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-c.pt
!wget https://github.com/WongKinYiu/yolov9/releases/download/v0.1/yolov9-e.pt
# !wget https://github.com/WongKinYiu/yolov9/releases/download/v0.1/gelan-c.pt
#!wget https://github.com/WongKinYiu/yolov9/releases/download/v0.1/gelan-e.pt
# train yolov9 models
!python train_dual.py \
--batch 8 --epochs 25 --img 640 --device 0 --min-items 0 --close-mosaic 15 \
--data /content/S15YoloV9/tabla/data.yaml \
--weights /content/S15YoloV9/yolov9-e.pt \
--cfg /content/S15YoloV9/models/detect/yolov9-e.yaml \
--hyp hyp.scratch-high.yaml
# train gelan models
!python train.py \
--batch 8 --epochs 25 --img 640 --device 0 --min-items 0 --close-mosaic 15 \
--data /content/S15YoloV9/tabla/data.yaml \
--weights /content/S15YoloV9/yolov9-e.pt \
--cfg /content/S15YoloV9/models/detect/gelan-c.yaml \
--hyp hyp.scratch-high.yaml
Note the cfg file should be based on the model weight being used. Adjust batch size to suit the available memory.
The results of trainig are stored in runs/train/exp. The trained model weight is in runs/train/exp/weights/best.pt.
from IPython.display import Image
Image(filename=f"/content/S15YoloV9/runs/train/exp/val_batch0_pred.jpg", width=1000)
!python detect.py \
--img 1280 --conf 0.1 --device 0 \
--weights /content/S15YoloV9/runs/train/exp/weights/best.pt \
--source /content/S15YoloV9/tabla/valid/images
The training results are stored in runs/detect/exp
Ref video for training.
Read the Readme from the original repo for details about the model
Check out this video to understand the architecture.