pytorch版本PSEnet训练并部署方式

概述

源码地址

torch版本

训练环境没有按照torch的readme一样的环境，自己部署环境为：

torch==1.9.1
torchvision==0.10.1
python==3.8.0
cuda==10.2
mmcv==0.2.12
editdistance==0.5.3
Polygon3==3.0.9.1
pyclipper==1.3.0
opencv-python==3.4.2.17
Cython==0.29.24

./compile.sh

制作数据集

1、训练的数据集

采用的是rolabelimg进行标注，需要转换为ic2015格式的数据。

转换代码：

import os
from lxml import etree
import numpy as np
import math
src_xml = "ANN"
txt_dir = "gt"
xml_listdir = os.listdir(src_xml)
xml_listpath = [os.path.join(src_xml,xml_listdir1) for xml_listdir1 in xml_listdir]
def xml_out(xml_path):
   gt_lines = []
   ET = etree.parse(xml_path)
   objs = ET.findall("object")
   for ix,obj in enumerate(objs):
       name = obj.find("name").text
       robox = obj.find("robndbox")
       cx = int(float(robox.find("cx").text))
       cy = int(float(robox.find("cy").text))
       w = int(float(robox.find("w").text))
       h = int(float(robox.find("h").text))
       angle = float(robox.find("angle").text)
       # angle = math.degrees(angle1)
       wx1 = cx - int(0.5 * w)
       wy1 = cy - int(0.5 * h)
       wx2 = cx + int(0.5 * w)
       wy2 = cy - int(0.5 * h)
       wx3 = cx - int(0.5 * w)
       wy3 = cy + int(0.5 * h)
       wx4 = cx + int(0.5 * w)
       wy4 = cy + int(0.5 * h)
       x1 = int((wx1 - cx) * np.cos(angle) - (wy1 - cy) * np.sin(angle) + cx)
       y1 = int((wx1 - cx) * np.sin(angle) - (wy1 - cy) * np.cos(angle) + cy)
       x2 = int((wx2 - cx) * np.cos(angle) - (wy2 - cy) * np.sin(angle) + cx)
       y2 = int((wx2 - cx) * np.sin(angle) - (wy2 - cy) * np.cos(angle) + cy)
       x3 = int((wx3 - cx) * np.cos(angle) - (wy3 - cy) * np.sin(angle) + cx)
       y3 = int((wx3 - cx) * np.sin(angle) - (wy3 - cy) * np.cos(angle) + cy)
       x4 = int((wx4 - cx) * np.cos(angle) - (wy4 - cy) * np.sin(angle) + cx)
       y4 = int((wx4 - cx) * np.sin(angle) - (wy4 - cy) * np.cos(angle) + cy)
       lines = str(x1)+","+str(y1)+","+str(x2)+","+str(y2)+","+\
               str(x3)+","+str(y3)+","+str(x4)+","+str(y4)+","+str(name)+"\n"
       gt_lines.append(lines)
       return gt_lines
def main():
   count = 0
   for xml_dir in xml_listdir:
       gt_lines = xml_out(os.path.join(src_xml,xml_dir))
       txt_path = "gt_" + xml_dir[:-4] + ".txt"
       with open(os.path.join(txt_dir,txt_path),"a+") as fd:
           fd.writelines(gt_lines)
       count +=1
       print("Write file ％s" ％ str(count))
if __name__ == "__main__":
   main()

rolabelimg标注后的xml文件和labelimg的xml有些区别，根据不同的标注软件，转换代码略有区别。

转换后的格式为x1,y1,x2,y2,x3,y3,x4,y4,"classes",此处classes为检测的类别，如果是模糊训练的话，classes为“###”。

但是重点，这个源代码对于模糊训练，loss一直为1。

2、将数据集分成训练集和测试集

这里可以按照源码路径存放数据集，也可以修改源码存放位置。

PSENet-python3\dataset\psenet\psenet_ic15.py

修改下述代码为自己文件夹

3、训练

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py config/psenet/psenet_r50_ic15_736.py

其中根据源码中的readme，

可以根据自己的需要，自行选择配置文件。

4、部署测试

import torch
import numpy as np
import argparse
import os
import os.path as osp
import sys
import time
import json
from mmcv import Config
import cv2
from torchvision import transforms
from dataset import build_data_loader
from models import build_model
from models.utils import fuse_module
from utils import ResultFormat, AverageMeter
def prepare_image(image, target_size):
   """Do image preprocessing before prediction on any data.
   :param image:       original image
   :param target_size: target image size
   :return:
                       preprocessed image
   """
   #assert os.path.exists(img), 'file is not exists'
   #img = cv2.imread(img)
   img = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
   # h, w = image.shape[:2]
   # scale = long_size / max(h, w)
   img = cv2.resize(img, target_size)
   # 将图片由(w,h)变为(1,img_channel,h,w)
   tensor = transforms.ToTensor()(img)
   tensor = tensor.unsqueeze_(0)
   tensor = tensor.to(torch.device("cuda:0"))
   return tensor
def report_speed(outputs, speed_meters):
   total_time = 0
   for key in outputs:
       if 'time' in key:
           total_time += outputs[key]
           speed_meters[key].update(outputs[key])
           print('％s: ％.4f' ％ (key, speed_meters[key].avg))
   speed_meters['total_time'].update(total_time)
   print('FPS: ％.1f' ％ (1.0 / speed_meters['total_time'].avg))
def load_model(cfg):
   model = build_model(cfg.model)
   model = model.cuda()
   model.eval()
   checkpoint = "psenet_r50_ic15_1024_finetune/checkpoint_580ep.pth.tar"
   if checkpoint is not None:
       if os.path.isfile(checkpoint):
           print("Loading model and optimizer from checkpoint '{}'".format(checkpoint))
           sys.stdout.flush()
           checkpoint = torch.load(checkpoint)
           d = dict()
           for key, value in checkpoint['state_dict'].items():
               tmp = key[7:]
               d[tmp] = value
           model.load_state_dict(d)
       else:
           print("No checkpoint found at")
           raise
       # fuse conv and bn
   model = fuse_module(model)
   return model
if __name__ == '__main__':
   src_dir = "testimg/"
   save_dir = "test_save/"
   if not os.path.exists(save_dir):
       os.makedirs(save_dir)
   cfg = Config.fromfile("PSENet/config/psenet/psenet_r50_ic15_1024_finetune.py")
   for d in [cfg, cfg.data.test]:
       d.update(dict(
           report_speed=False
       ))
   if cfg.report_speed:
       speed_meters = dict(
           backbone_time=AverageMeter(500),
           neck_time=AverageMeter(500),
           det_head_time=AverageMeter(500),
           det_pse_time=AverageMeter(500),
           rec_time=AverageMeter(500),
           total_time=AverageMeter(500)
       )
   model = load_model(cfg)
   model.eval()
   count = 0
   for img_name in os.listdir(src_dir):
       img = cv2.imread(src_dir + img_name)
       tensor = prepare_image(img, target_size=(1376, 1024))
       data = dict()
       img_metas = dict()
       data['imgs'] = tensor
       img_metas['org_img_size'] = torch.tensor([[img.shape[0], img.shape[1]]])
       img_metas['img_size'] = torch.tensor([[1376, 1024]])
       data['img_metas'] = img_metas
       data.update(dict(
           cfg=cfg
       ))
       with torch.no_grad():
           outputs = model(**data)
       if cfg.report_speed:
           report_speed(outputs, speed_meters)
       for bboxes in outputs['bboxes']:
           x1 = bboxes[0]
           y1 = bboxes[1]
           x2 = bboxes[4]
           y2 = bboxes[5]
           cv2.rectangle(img, (x1, y1), (x2, y2), (0, 0, 255), 3)
       count = count + 1
       cv2.imwrite(save_dir + img_name, img)
       print("img test:", count)

from dataset import build_data_loader
from models import build_model
from models.utils import fuse_module
from utils import ResultFormat, AverageMeter

训练代码里含有。

来源：https://blog.csdn.net/WSNjiang/article/details/120821227