pse-lite-pytorch

---

data format

follow icdar15 dataset format, x1,y1,x2,y2,x3,y3,x4,y4,label

image
│   1.jpg
│   2.jpg   
│		...
label
│   gt_1.txt
│   gt_2.txt
|		...

Compression model mode one,use lite basemodel

---

Support switching basemodel,(mobilenet,squeezenet,shufflenet,resnet)

train

python3 train.py --backbone mobile 

test

python3 inference.py

---

Compression model mode two,Channel clipping

Sparse training

python3 train.py --backbone resnet --sr_lr 0.00001

prune model

python3 prune.py 

fintune

python3 train_prune_finetune.py 

prune test

python3 inference_prune.py 

performance

Method	precision	recall	hmean	prune ratio	modelsize(M)	infer time(v100)(ms)
PSENet-1s (ResNet50)				0	114.5	12
PSENet-1s (ResNet50)	0.8179	0.7958	0.8067	0.8	25.1	9
PSENet-1s (ResNet50)	0.8124	0.7862	0.7991	0.9	16.6	7

---

Compression model mode three, Model distillation

reference

1. https://github.com/whai362/PSENet
2. https://github.com/xiaolai-sqlai/mobilenetv3
3. https://github.com/MhLiao/DB
4. https://github.com/tanluren/yolov3-channel-and-layer-pruning