zonghaofan / dbnet_torch Goto Github PK

my chinese blog:https://blog.csdn.net/fanzonghao/article/details/107199538

Dbnet is usually used to detect word, in fact barcode can be detected.

This project also provide word detect model.

model:

​ follow icdar15 dataset format, x1,y1,x2,y2,x3,y3,x4,y4,label,(x1,y1) is left top,(x2,y2) is right top.

where config/icdar2015_resnet18_FPN_DBhead_polyLR_code_phone.yaml you can change learning rate,train_path and so on.

single gpu train: python train_code_phone.py

multi gpus train:sh multi_gpu_train.sh , nedd notice os.environ['CUDA_VISIBLE_DEVICES'] is match nproc_per_node.

​ python predict_code_phone.py

First python model_to_onnx.py to get onnx model. Then where onnx_project you can python dbcode_tensorrt_predict.py.

notice:change model path

python train_word_industry_res50.py　train teacher(res50) model;

python train_word_industry_res18_kd.py train student(res18)model;

​ --change you own path in labelme_txt_box.py
​ python labelme_txt_box.py

pytorch1.5

torchvision0.6

cuda9.0+

tensorrt 7.0

1.word:https://github.com/zonghaofan/dbnet_torch/tree/master/phone_word_model 2.code:https://github.com/zonghaofan/dbnet_torch/tree/master/phone_code_model

1. learning rate　show

   

   ２.some test examples

   

3.train loss

1. https://github.com/WenmuZhou/DBNet.pytorch

More tensortrt inference.