Skip to content

Latest commit

 

History

History
126 lines (90 loc) · 2.31 KB

README_en.md

File metadata and controls

126 lines (90 loc) · 2.31 KB

English| 简体中文

DBNet-lite-pytorch

setup

pip install -r requirement.txt
cd models/dcn/
sh make.sh

data format for Horizontal or slanted text

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
|		...

data format for curved text

dataset format, x1,y1,x2,y2,x3,y3,x4,y4 ...xn,yn,label

The number of N can be inconsistent,The arrangement of points is clockwise or counterclockwise

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

train

Go to configure config.yaml in the root directory

python3 train.py

test

set is_poly = True in config.yaml for curved text , others set is_poly = False

python3 inference.py

Channel clipping for model compression

Training section

  1. sparse training is performed first. firstly, modify config.yaml to set use_sr to True, and set sr_lr. the larger this setting is, the more pressure it will have. pay attention to the fact that it may not converge if it is too large.
python3 train.py
  1. Compression model Set the parameters of pruned part in config.yaml and run it
python3 ./pruned/prune.py
  1. Re-finetune model Here, the accuracy will pick up quickly. Generally, you can train 50-100epoch and do your own experiments
python3 ./pruned/train_fintune.py

test section

python3 ./pruned/prune_inference.py

performance in icdar2015

Method head extra data prune ratio model size(M) precision(%) recall(%) hmean(%) model_file
Resnet18 FPN no 0 62.6 86.11 76.45 80.99 baiduyun (extract code: p0bk)
Resnet18 DB no 0.8 20.1 85.55 76.40 80.72
mobilev3 DB no 0 2.5 85.55 76.40 74.71

some result

ToDoList

  • tranform DB code format from MhLiao/DB
  • add some performance
  • add light backbone
  • pruned big model by channel clipping
  • Model distillation

reference

  1. https://github.com/whai362/PSENet
  2. https://github.com/MhLiao/DB
  3. https://github.com/Jzz24/pytorch_quantization