This repo extend the work from paper 'Redundant Feature Pruning for Accelerated Inference in Deep Neural Networks' Babajide O. Ayindea, Tamer Inanca, Jacek M. Zuradaa.
Contribution of this work:

• apply proposed pruning method from paper for all yolov5 layer
• fine-tuned pruned model to regain accuracy (in my case, pruned half parameters of yolov5n without losing in accuracy)
• custom Tensorflowjs class to deploy on web: seprate nms module from model graph and execute in cpu for faster inference
• end2end pipeline from train - prune - finetune - convert and post training quantization tfjs - deploy tfjs

python channel_prune.py --weight_path path/to/checkpoint --pruned_weight_path path/to/output/checkpoint --thres 0.5

• Specify model checkpoint, output of purned model in weigth_path and pruned_weight_path
• Specify cosin similarity distance threshold. The higher threshold the more filter weights are pruned

python fine-tune.py --batch-size 144 --img 320 --data data/coco-old-cp.yaml --hyp data/hyps/hyp.scratch-high.yaml --weights runs/train/carpart-add-car-v2-anchor-tune-3-s-model-od/weights/best-prune-066.pt --sync-bn --device 2 --name prune-66-finetune-carpart-add-car-v2-anchor-tune-3-od-od

• Load pruned model checkpoint as pretrain and retrain model on same dataset

python export.py --weight path/to/pruned/checkpoint --int8 --includes tfjs

• Move NMS separate from model graph, nms will excecute by cpu for faster inference (detail)
• Link demo : here