Welding Defect Detection using improved YOLOv7 model

Atanu Shuvam Roy

Last updated on Jan 13, 2024

Training Section

Training model on YOLOV7 based Architecture

Browse to YOLO Folder:

cd yolov7

Basic Training Command:

python train.py --data 'data/custom.yaml' --weights 'yolov7-tiny.pt'

To train on the yolov7-tiny architecture [Leaky ReLu] :

 python train.py --weights yolov7-tiny.pt --data data/custom.yaml --cfg cfg/final/OG/yolov7-tiny.yaml --device 0 --epochs 200 --name yolov7-tiny-pipe --hyp data/hyp.scratch.tiny.yaml

To train on the yolov7-tiny architecture [Parametric ReLu] :

python train.py --weights yolov7-tiny.pt --data data/custom.yaml --cfg cfg/final/yolov7-tiny-PRelu.yaml --device 0 --epochs 200 --name yolov7-tiny-prelu-pipe --hyp data/hyp.scratch.tiny.yaml

To train on the yolov7 + ecanet architecture :

python train.py --weights yolov7-tiny.pt --data data/custom.yaml --cfg cfg/final/yolov7-tiny-ecanet-NO-SPPCSPC.yaml --device 0 --epochs 200 --name yolov7-ecanet-pipe-NS --hyp data/hyp.scratch.tiny.yaml

To train on the yolov7 + hornet architecture :

python train.py --weights yolov7-tiny.pt --data data/custom.yaml --cfg cfg/final/yolov7-tiny-hornet2b-pipe.yaml --device 0 --epochs 200 --name yolov7-hornet-pipe --hyp data/hyp.scratch.tiny.yaml

To train on the yolov7 + ecanet architecture :

python train.py --weights yolov7-tiny.pt --data data/custom.yaml --cfg cfg/final/yolov7-tiny-ecanet-NO-SPPCSPC.yaml --device 1 --epochs 200 --name yolov7-ecanet-pipe --hyp data/hyp.scratch.p5.yaml

Training model on YOLOV5 based Architecture

Browse to YOLO Folder:

cd yolov5

To train on the yolov5s:

python train.py --weights yolov5s.pt --data data/custom.yaml --name yolov5s-pipe

To train on the yolov5n:

python train.py --weights yolov5n.pt --data data/custom.yaml --name yolov5n-pipe

Troubleshooting Training

To resume training (on failure):

python train.py --weights runs/train/exp14/last.pt --resume

Testing Section

Detecting model on YOLOV7 based Architecture

To detect on video (Source: VIDEO):

python detect.py --weights '../finalruns/yolov7-nano-hornet-ecanet-pipe/weights/best.pt' --img-size 640 --source ../testdata/test.mp4

To detect on video (Source: WEBCAM):

python detect.py --weights '../finalruns/yolov7-nano-hornet-ecanet-pipe/weights/best.pt' --img-size 640 --source 0

Model Section

Final Declared Models

FINAL NANO MODEL:

../finalruns/yolov7-nano-hornet-ecanet-pipe2-1MB/weights/best.pt

FINAL TINY MODEL:

../finalruns/yolov7-tiny-hornet-ecanet-pipe2/weights/best.pt

Model Architecture

Original YOLOV7 Architecture

Modded YOLOV7 Architecture

Testing Architecture

Model Results

Modded YOLOV7 Architecture Comparison

Modded YOLOV7 Architectiure Inference Times on Raspberry Pi

iot