YOLOV3SlimFlowChart

clone YOLOV3 
> git clone 

> git checkout ktian08-hyp & git pull



Step 1 :

Trainin Darknet with 4 classes using AlexeyAB repo normally and get last weights. 

Step 2 :

Training with sparstiy using https://github.com/erikguo/yolov3
> cd yolov3_erikguo
> python3 train_boat.py  --weights weights/yolov3-boat_last.weights --sparsity 0.0001
add continue parameter to continue from last stored checkpoint later if stoped training. 
Test using 
>python3 test.py --cfg cfg/yolov3-boat.cfg  --data boat2019/boat.data --weights weights/yolov3-boat_last.weights  

check tensorboard

> tensorboard --logdir=runs
Can also detect for the detection.

After finish training convert .pt weights to .weights  using https://github.com/erikguo/yolov3
> python3 convert_pt_weights.py cfg/yolov3-boat.cfg  weights/best.pt 

or convert .pt weights to .weights using https://github.com/ultralytics/yolov3/

> python3  -c "from models import *; convert('cfg/prune5_boat.cfg', 'weights/prune5_boat.pt')"



Step 3: Prune the model 

> cd SlimYOLOv3
> python3 yolov3/prune.py --cfg boat/yolov3-boat.cfg    --weights boat/best_converted.weights  --overall_ratio 0.5 --perlayer_ratio 0.1

Step 4 :

Train again using Alexey Ab darknet repo using generated prune.cfg and prune.weights with step 3.
> cd darknet
> ./darknet detector train data/boat.data cfg/prune.cfg  backup/prune.weights 


Additional 
To run on video using https://github.com/ultralytics/yolov3/
> conda activate tracking
> cd yolov3_ultralytics
to run on single image or video
> python detect_boat.py  --source vidoe.mp4     //output will be saved in  output folder

to run on all images and videos in a folder 
>python detect_boat.py --source /folder name   //output will be saved in  output folder