PaddleDetection使用VOC数据集训练

当前PaddleDetection版本: release/2.4

以下采用PP-YOLOE为例

1. 数据准备

   1. 将图片和标注文件保存至如下文件格式
      

   2. 使用脚本切分训练集

      pip install matplotlib
      python3 make_voc.py --dataset datasets/测试数据集
      

      import argparse
      import random
      import re
      from pathlib import Path
      from matplotlib import pyplot as plt
      
      annotations_name = 'Annotations'
      jpegimages_name = 'JPEGImages'
      trainval_name = 'trainval.txt'
      text_file_name = 'test.txt'
      val_file_name = 'val.txt'
      
      def write_file(datas, file):
          results = []
          for data in datas:
              annotation = Path(annotations_name, f'{data}.xml')
              image =  Path(jpegimages_name, f'{data}.jpg')
              result = f'{image} {annotation}'
              results.append(result)
          with file.open('w')as f:
              f.write("\n".join(results))
          print(file, "success")
      
      def make_paddle_voc(dataset):
          dataset = Path(dataset)
          assert dataset.is_dir(), "Error dataset path"
          annotations_path = dataset / annotations_name
          images_path = dataset / jpegimages_name
          if not (annotations_path.is_dir() and images_path.is_dir()):
              print("Please use the following structure:")
              print(f"{dataset.name}")
              print(f'\t{annotations_name}\n\t\t|--00001.xml\n\t\t|--00002.xml')
              print(f'\t{jpegimages_name}\n\t\t|--00001.jpg\n\t\t|--00002.jpg')
              raise RuntimeError(f'Error {jpegimages_name} or {annotations_name} dir')
          paddle_path = dataset / 'paddle'
          paddle_path.mkdir(exist_ok=True, parents=True)
          trainval_path = paddle_path / trainval_name
          test_path = paddle_path / text_file_name
          val_path = paddle_path / val_file_name
          annotations = annotations_path.glob('*.xml')
          images = images_path.glob('*.jpg')
          images_list = [image.name for image in images]
          annotations_list = []
          for annotation in annotations:
              image = annotation.with_suffix('.jpg').name
              assert image in images_list, f"{annotation.name} has no image file"
              annotations_list.append(annotation.stem)
          random.shuffle(annotations_list)
          per = round(len(annotations_list)/10)
          trainval = annotations_list[:per*7] # 此处切分比例为8:1:1 可行修改比例
          test = annotations_list[per*7:per*9]
          val = annotations_list[per*9:]
          write_file(trainval, trainval_path)
          write_file(test, test_path)
          write_file(val, val_path)
      
      def check_labels(dataset):
          dataset = Path(dataset)
          annotations_path = dataset / annotations_name
        	class_list = {}
        	for annotation_path in annotations_path.glob('*.xml'):
        	    with annotation_path.open('r')as f:
        	        names = re.findall(r'<name>(.+?)</name>', f.read())
        	        for name in names:
        	            if name not in class_list.keys():
        	                class_list[name] = 0
        	            class_list[name] += 1
        	class_list = dict(sorted(class_list.items()))
        	label_list = dataset / 'label_list.txt'
        	with label_list.open('w')as f:
        	    f.write('\n'.join(class_list.keys()))
        	fig, ax = plt.subplots()
        	[ax.text(a, b + 1, b, ha='center', va='bottom') for a, b in class_list.items()]
        	plt.bar(*zip(*class_list.items()))
        	plt.xticks(rotation=270)
        	plt.title('Detection category distribution')
        	plt.xlabel('Class')
        	plt.ylabel('Number')
        	plt.tight_layout()
        	plt.savefig(f"{dataset / '类别分布.jpg'}", format="jpg")
        	print(f"Save as {dataset / '类别分布.jpg'}")
        	print(f"共{len(class_list)}类")
        	
      if __name__=='__main__':
          parser = argparse.ArgumentParser()
          parser.add_argument('--dataset', required=True, help='input dataset path, necessary parameter')
          
          opt = parser.parse_args()
          check_dataset(opt.dataset)
          make_paddle_voc(opt.dataset)
          check_labels(opt.dataset)
      

      

   3. 创建数据集配置文件

      1. 将configs/datasets/voc.yml复制为custom.yml
         
      2. 修改配置，红色标记处，根据自己的数据集修改
         

2. 修改参数

   1. 在configs中选择你要训练的配置，复制一份并将coco后缀改为voc
      

   2. 修改配置

      第一处标记为自定义VOC数据集，第二、三处根据自己电脑配置，参照官方说明修改。
      

3. 开启训练

   python3 tools/train.py -c configs/ppyoloe/ppyoloe_crn_s_300e_voc.yml
   

4. 模型评估

   python3 -u tools/eval.py -c configs/ppyoloe/ppyoloe_crn_s_300e_voc.yml
   

5. 模型测试

   python3 -u tools/infer.py -c configs/ppyoloe/ppyoloe_crn_s_300e_voc.yml --infer_img=test.jpg --draw_threshold=0.5
   

6. 模型导出

   python3 tools/export_model.py -c configs/ppyoloe/ppyoloe_crn_s_300e_voc.yml --output_dir=output/ppyoloe_crn_s_300e_voc/ -o weights=output/ppyoloe_crn_s_300e_voc/best_model