CenterNet：Objects as Points训练相关技巧

目录

1.训练自己数据集【以VOC数据集为样例】

1.准备数据：准备好自己要用的JFPGImages+Annotations

2.xml转json：【我比较熟悉的是.xml文件，但作者用的是coco类似的.json文件】

3.把文件夹放置的与作者生成的VOC数据集相同：

4.修改一些文件：

5.试着训练一下叭：

6.开始训练：【能够正常开始训练】

7.测试前修改文件：

7.mAP测试结果：

8.单张图片检测结果：

2.关于预训练模型

1.这里是如何将自己生成的.pth模型，转换为可以用来训练的模型

2.检验下生成的预训练模型是否有效：

---

1.训练自己数据集【以VOC数据集为样例】

【因为更熟悉voc，所以还是以vocal数据集为例】之前的环境配置和VOC训练：https://blog.csdn.net/weixin_38715903/article/details/98039181

1.准备数据：准备好自己要用的JFPGImages+Annotations

2.xml转json：【我比较熟悉的是.xml文件，但作者用的是coco类似的.json文件】

感恩博主：https://blog.csdn.net/u010397980/article/details/90341223

更改了一下代码，最终代码如下：

#coding:utf-8
# pip install lxml
 
import os
import glob
import json
import shutil
import numpy as np
import xml.etree.ElementTree as ET
 
path2 = "./INF"
 
START_BOUNDING_BOX_ID = 1
 
def get(root, name):
    return root.findall(name)
 
def get_and_check(root, name, length):
    vars = root.findall(name)
    if len(vars) == 0:
        raise NotImplementedError('Can not find %s in %s.'%(name, root.tag))
    if length > 0 and len(vars) != length:
        raise NotImplementedError('The size of %s is supposed to be %d, but is %d.'%(name, length, len(vars)))
    if length == 1:
        vars = vars[0]
    return vars
 
def convert(xml_list, json_file):
    json_dict = {"images": [], "type": "instances", "annotations": [], "categories": []}
    categories = pre_define_categories.copy()
    bnd_id = START_BOUNDING_BOX_ID
    all_categories = {}
    for index, line in enumerate(xml_list):
        # print("Processing %s"%(line))
        xml_f = line
        tree = ET.parse(xml_f)
        root = tree.getroot()
        
        filename = os.path.basename(xml_f)[:-4] + ".jpg"
        image_id = 1 + index
        size = get_and_check(root, 'size', 1)
        width = int(get_and_check(size, 'width', 1).text)
        height = int(get_and_check(size, 'height', 1).text)
        image = {'file_name': filename, 'height': height, 'width': width, 'id':image_id}
        json_dict['images'].append(image)
        ## Cruuently we do not support segmentation
        #  segmented = get_and_check(root, 'segmented', 1).text
        #  assert segmented == '0'
        for obj in get(root, 'object'):
            category = get_and_check(obj, 'name', 1).text
            if category in all_categories:
                all_categories[category] += 1
            else:
                all_categories[category] = 1
            if category not in categories:
                if only_care_pre_define_categories:
                    continue
                new_id = len(categories) + 1
                print("[warning] category '{}' not in 'pre_define_categories'({}), create new id: {} automatically".format(category, pre_define_categories, new_id))
                categories[category] = new_id
            category_id = categories[category]
            bndbox = get_and_check(obj, 'bndbox', 1)
            xmin = int(float(get_and_check(bndbox, 'xmin', 1).text))
            ymin = int(float(get_and_check(bndbox, 'ymin', 1).text))
            xmax = int(float(get_and_check(bndbox, 'xmax', 1).text))
            ymax = int(float(get_and_check(bndbox, 'ymax', 1).text))
            assert(xmax > xmin), "xmax <= xmin, {}".format(line)
            assert(ymax > ymin), "ymax <= ymin, {}".format(line)
            o_width = abs(xmax - xmin)
            o_height = abs(ymax - ymin)
            ann = {'area': o_width*o_height, 'iscrowd': 0, 'image_id':
                   image_id, 'bbox':[xmin, ymin, o_width, o_height],
                   'category_id': category_id, 'id': bnd_id, 'ignore': 0,
                   'segmentation': []}
            json_dict['annotations'].append(ann)
            bnd_id = bnd_id + 1
 
    for cate, cid in categories.items():
        cat = {'supercategory': 'none', 'id': cid, 'name': cate}
        json_dict['categories'].append(cat)
    json_fp = open(json_file, 'w')
    json_str = json.dumps(json_dict)
    json_fp.write(json_str)
    json_fp.close()
    print("------------create {} done--------------".format(json_file))
    print("find {} categories: {} -->>> your pre_define_categories {}: {}".format(len(all_categories), all_categories.keys(), len(pre_define_categories), pre_define_categories.keys()))
    print("category: id --> {}".format(categories))
    print(categories.keys())
    print(categories.values())
 
if __name__ == '__main__':
    classes = ['car', 'person', 'bicycle']
    pre_define_categories = {}
    for i, cls in enumerate(classes):
        pre_define_categories[cls] = i + 1
    # pre_define_categories = {'a1': 1, 'a3': 2, 'a6': 3, 'a9': 4, "a10": 5}
    only_care_pre_define_categories = True
    # only_care_pre_define_categories = False
 
    train_ratio = 0.9
    save_json_train = './INF/annotations/INF_train.json'
    save_json_val = './INF/annotations/INF_test.json'
    xml_dir = "./INF/Annotations"
    img_dir="./INF/JFPGImages"
 
    xml_list = glob.glob(xml_dir + "/*.xml")#返回所有匹配的.xml文件路径列表。
    xml_list = np.sort(xml_list)
    np.random.seed(100)
    np.random.shuffle(xml_list)
    #print(xml_list[:100])
    train_num = int(len(xml_list)*train_ratio)
    xml_list_train = xml_list[:train_num]
    xml_list_val = xml_list[train_num:]
 
    if os.path.exists(path2 + "/annotations"):
        shutil.rmtree(path2 + "/annotations")
    os.makedirs(path2 + "/annotations")
    if os.path.exists(path2 + "/images/train2019"):
        shutil.rmtree(path2 + "/images/train2019")
    os.makedirs(path2 + "/images/train2019")
    if os.path.exists(path2 + "/images/val2019"):
        shutil.rmtree(path2 +"/images/val2019")
    os.makedirs(path2 + "/images/val2019")
    
    convert(xml_list_train, save_json_train)
    convert(xml_list_val, save_json_val)
    
    f1 = open("./INF/train.txt", "w")
    for xml in xml_list_train:
        img1 = img_dir+xml[17:-4] + ".jpg"#'这里的17其实是'./INF/Annotations'的长度'
        #print(img1)
        f1.write(os.path.basename(xml)[:-4] + "\n")
        shutil.copyfile(img1, path2 + "/images/train2019/" + os.path.basename(img1))
 
    f2 = open("./INF/test.txt", "w")
    for xml in xml_list_val:
        img2 = img_dir+xml[17:-4] + ".jpg"#'这里的17其实是'./INF/Annotations'的长度'
        f2.write(os.path.basename(xml)[:-4] + "\n") 
        shutil.copyfile(img2, path2 + "/images/val2019/" + os.path.basename(img2))
        
    f1.close()
    f2.close()
    print("-------------------------------")
    print("train number:", len(xml_list_train))
    print("val number:", len(xml_list_val))

3.把文件夹放置的与作者生成的VOC数据集相同：

voc_INF|--annotations|--INF_test.json
       |             |--INF_train.json
       |--images
       |--VOCdevkit

PS:
1.'annotations'存放.json文件，如果你分了train,val,test三个部分，还要运行merge_pascal_json.py将train和val放在一个.json文件里
2.'image'存放所有的图片
3.'VOCdevkit'存放普通的VOC数据集包括JFPGImages、Annotations、ImageSets
4.上述存放的文件与后续修改路径有关

4.修改一些文件：

• ~/CenterNet/src/lib/datasets/dataset文件夹中

将pascal.py复制为pascal_INF.py，修改部分路径代码如下：

#从13行开始
"""类名与文件名一致"""
class PascalINF(data.Dataset):
  """类别数目：20"""
  num_classes = 3
  default_resolution = [384, 384]
  mean = np.array([0.485, 0.456, 0.406],
                   dtype=np.float32).reshape(1, 1, 3)
  std  = np.array([0.229, 0.224, 0.225],
                   dtype=np.float32).reshape(1, 1, 3)
  
  def __init__(self, opt, split):
    super(PascalINF, self).__init__()
    """data_dir：是存放你数据的文件名,我的是~/CenterNet/data/voc_INF/"""
    self.data_dir = os.path.join(opt.data_dir, 'voc_INF')
    self.img_dir = os.path.join(self.data_dir, 'images')
    """这里照着annotations中的文件名修改,我只有train和test"""
    _ann_name = {'train': 'train', 'val': 'test'}
    """这里照着annotations中的文件名修改,我的json文件命名规则是INF_test和INF_train"""
    self.annot_path = os.path.join(
      self.data_dir, 'annotations', 
      'INF_{}.json').format(_ann_name[split])
    self.max_objs = 50
    """修改你的类别,记得与生成json文件时顺序一致,涉及到class_id匹配问题"""
    self.class_name = ['__background__', 'car', 'person', 'bicycle']
    """4=class_number+1(blackground)"""
    self._valid_ids = np.arange(1, 4, dtype=np.int32)
    self.cat_ids = {v: i for i, v in enumerate(self._valid_ids)}
    self._data_rng = np.random.RandomState(123)
    self._eig_val = np.array([0.2141788, 0.01817699, 0.00341571],
                             dtype=np.float32)
    self._eig_vec = np.array([
        [-0.58752847, -0.69563484, 0.41340352],
        [-0.5832747, 0.00994535, -0.81221408],
        [-0.56089297, 0.71832671, 0.41158938]
    ], dtype=np.float32)
    self.split = split
    self.opt = opt

• ~/CenterNet/src/lib/datasets文件夹中，修改dataset_factory.py

#第10行添加
from .dataset.coco import COCO
from .dataset.pascal import PascalVOC
"""添加自己的数据集"""
from .dataset.pascal_INF import PascalINF
from .dataset.kitti import KITTI
from .dataset.coco_hp import COCOHP

#第17行添加
dataset_factory = {
  'coco': COCO,
  'pascal': PascalVOC,
  'kitti': KITTI,
  'coco_hp': COCOHP,
  """添加自己的数据集"""
  'inf': PascalINF
}

5.试着训练一下叭：

cd src
# train
sudo python3.6 main.py ctdet --exp_id INF --dataset inf --num_epochs 70 --lr_step 45,60 --batch_size 32 --master_batch 16 --lr 1.25e-4 --gpus 0,1 

"""
PS:
1.--exp_id INF 存放日志的文件名
2.--dataset inf 你的数据类型：刚刚更改的部分
"""

6.开始训练：【能够正常开始训练】

明天测试一下模型结果，就酱

7.测试前修改文件：

• 需要更改的文件：\CenterNet\src\lib\debugger.py

'第44-50行左右，添加elif dataset == 'inf'，如下：'
elif num_classes == 80 or dataset == 'coco':
      self.names = coco_class_name
elif num_classes == 20 or dataset == 'pascal':
      self.names = pascal_class_name
elif dataset == 'inf':
      self.names = inf_class_name

'第440-447行左右，添加inf_class_name，如下：'
gta_class_name = [
  'p', 'v'
]
inf_class_name = ["car", "person", "bicycle"]
  
pascal_class_name = ["aeroplane", "bicycle", "bird", "boat", "bottle", "bus", 
  "car", "cat", "chair", "cow", "diningtable", "dog", "horse", "motorbike", 
  "person", "pottedplant", "sheep", "sofa", "train", "tvmonitor"]

• \CenterNet\src\tools\reval.py-->\CenterNet\src\tools\reval_inf.py

1.复制reval.py,并改名为reval_inf.py
2.根据需求修改代码：【3处】
  '修改引入的库'
  from datasets.pascal_inf import pascal_inf

  '#第33到37行左右，修改imdb的默认名称'
  parser.add_argument('--imdb', dest='imdb_name',
                      help='dataset to re-evaluate',
                      default='INF_test', type=str)

   def from_dets(imdb_name, detection_file, args):
   '修改函数名称'
      imdb = pascal_inf('test')
      imdb.competition_mode(args.comp_mode)
      imdb.config['matlab_eval'] = args.matlab_eval
      with open(os.path.join(detection_file), 'rb') as f:
        if 'json' in detection_file:
          dets = json.load(f)
        else:
          dets = pickle.load(f, encoding='latin1')
      # import pdb; pdb.set_trace()
      if args.apply_nms:
        print('Applying NMS to all detections')
        test_nms = 0.3
        nms_dets = apply_nms(dets, test_nms)
      else:
        nms_dets = dets

      print('Evaluating detections')
      imdb.evaluate_detections(nms_dets)

• \CenterNet\src\tools\voc_eval_lib\datasets\pascal_voc.py-->\CenterNet\src\tools\voc_eval_lib\datasets\pascal_inf.py

1.将pascal_voc.py复制,并改名为pascal_inf.py
2.按需求修改代码：
    1). #-*-coding:utf-8-*-
    2).'初始化部分'【5处】
    '类名修改'
    class pascal_inf(imdb):
    '输入参数修改'
      def __init__(self, image_set, use_diff=False):
    'name命名修改：INF_test'
        name = 'INF_'+ image_set
        if use_diff:
          name += '_diff'
        imdb.__init__(self, name)
        self._image_set = image_set
        self._devkit_path = self._get_default_path()
    'Data地址修改：cfg.DATA_DIR+'voc_INF'+ 'VOCdevkit'[==self._devkit_path]'
        self._data_path = os.path.join(self._devkit_path)
    '按自己类别修改：'
        self._classes = ('__background__',  # always index 0
                     'car', 'person', 'bicycle')
        self._class_to_ind = dict(list(zip(self.classes,
                     list(range(self.num_classes)))))
        self._image_ext = '.jpg'
        self._image_index = self._load_image_set_index()
        # Default to roidb handler
        self._roidb_handler = self.gt_roidb
        self._salt = str(uuid.uuid4())
        self._comp_id = 'comp4'

    3). '数据地址修改：~/CenterNet/data/voc_INF/VOCdekit/'
      def _get_default_path(self):【1处】
        """
        Return the default path where PASCAL VOC is expected to be installed.
        """
        return os.path.join(cfg.DATA_DIR, 'voc_INF', 'VOCdevkit')

    4).修改模式，因为我不需要年份【1处】
        def rpn_roidb(self):
        '这里修改：删去年份判断'
            if self._image_set != 'test':
              gt_roidb = self.gt_roidb()
              rpn_roidb = self._load_rpn_roidb(gt_roidb)
              roidb = imdb.merge_roidbs(gt_roidb, rpn_roidb)
            else:
              roidb = self._load_rpn_roidb(None)
            return roidb

    5).result的存储地址修改【1处】
        def _get_voc_results_file_template(self):
            # VOCdevkit/results/VOC2007/Main/<comp_id>_det_test_aeroplane.txt
            filename = self._get_comp_id() + '_det_' + self._image_set + '_{:s}.txt'
            'result的存储地址：~/CenterNet/data/voc_INF/VOCdekit/results/'
            path = os.path.join(
              self._devkit_path,
              'results',
              filename)
            return path

    6)._do_python_eval参数修改【1处】
    def _do_python_eval(self, output_dir=None):
        annopath = os.path.join(
          self._devkit_path,
          'Annotations',
          '{:s}.xml')
        imagesetfile = os.path.join(
          self._devkit_path,
          'ImageSets',
          'Main',
          self._image_set + '.txt')
        cachedir = os.path.join(self._devkit_path, 'annotations_cache')
        aps = []
        # The PASCAL VOC metric changed in 2010
       '我的数据集中没有年份判断,所以做了删改,这里选择直接使用07_metric的方式计算AP,也可以选择其他模式'
        use_07_metric = True 
        if output_dir is not None and not os.path.isdir(output_dir):
          os.mkdir(output_dir)
        for i, cls in enumerate(self._classes):
          if cls == '__background__':
            continue
          filename = self._get_voc_results_file_template().format(cls)
          rec, prec, ap = voc_eval(
                filename, annopath, imagesetfile, cls, cachedir, ovthresh=0.5,
                use_07_metric=use_07_metric, use_diff=self.config['use_diff'])
          aps += [ap]
          print(('AP for {} = {:.4f}'.format(cls, ap)))
          if output_dir is not None:
            with open(os.path.join(output_dir, cls + '_pr.pkl'), 'wb') as f:
                  pickle.dump({'rec': rec, 'prec': prec, 'ap': ap}, f)
        print(('Mean AP = {:.4f}'.format(np.mean(aps))))
        print('~~~~~~~~')
    7).更改库名啦：【1处】
    if __name__ == '__main__':
      from datasets.pascal_inf import pascal_inf

      d = pascal_voc('trainval')
      res = d.roidb
      from IPython import embed;

      embed()

• /CenterNet/src/lib/datasets/dataset/pascal_INF.py【训练时我们更改的文件】

拉到文件末尾,修改使用的reval文件名：
    os.system('python tools/reval_inf.py ' + \
              '{}/results.json'.format(save_dir))

7.mAP测试结果：

sudo python3.6 test.py ctdet --exp_id INF --dataset inf --load_model ../exp/ctdet/INF/model_last.pth --flip_test

8.单张图片检测结果：

同样地：修改demo.py和opts.py文件【将其复制,并改名】

1).opts_inf:
    default_dataset_info = {
      'ctdet': {'default_resolution': [512, 512], 'num_classes': 80, 
                'mean': [0.408, 0.447, 0.470], 'std': [0.289, 0.274, 0.278],
                'dataset': 'coco'},   
 
'改为'
 
     default_dataset_info = {
      'ctdet': {'default_resolution': [512, 512], 'num_classes': 3, 
                'mean': [0.485, 0.456, 0.406], 'std': [0.229, 0.224, 0.225],
                'dataset': 'inf'},
2).demo_inf.py
    from opts import opts
'改为'
    from opts_inf import opts

【我的数据集要保密，所以随便从百度上拿了几张图片给一下结果，以下图片来源于百度】

2.关于预训练模型

1.这里是如何将自己生成的.pth模型，转换为可以用来训练的模型

• 先查看一下作者提供的预训练模型结构：

model_pre=torch.load('dla34-ba72cf86.pth')
for name in model_pre:
    print(name)

输出：

base_layer.0.weight
base_layer.1.weight
base_layer.1.bias
base_layer.1.running_mean
base_layer.1.running_var
level0.0.weight
level0.1.weight
level0.1.bias
level0.1.running_mean
level0.1.running_var
level1.0.weight
level1.1.weight
level1.1.bias
level1.1.running_mean
level1.1.running_var
level2.tree1.conv1.weight
level2.tree1.bn1.weight
level2.tree1.bn1.bias
level2.tree1.bn1.running_mean
level2.tree1.bn1.running_var
level2.tree1.conv2.weight
level2.tree1.bn2.weight
level2.tree1.bn2.bias
level2.tree1.bn2.running_mean
level2.tree1.bn2.running_var
level2.tree2.conv1.weight
level2.tree2.bn1.weight
level2.tree2.bn1.bias
level2.tree2.bn1.running_mean
level2.tree2.bn1.running_var
level2.tree2.conv2.weight
level2.tree2.bn2.weight
level2.tree2.bn2.bias
level2.tree2.bn2.running_mean
level2.tree2.bn2.running_var
level2.root.conv.weight
level2.root.bn.weight
level2.root.bn.bias
level2.root.bn.running_mean
level2.root.bn.running_var
level2.project.0.weight
level2.project.1.weight
level2.project.1.bias
level2.project.1.running_mean
level2.project.1.running_var
level3.tree1.tree1.conv1.weight
level3.tree1.tree1.bn1.weight
level3.tree1.tree1.bn1.bias
level3.tree1.tree1.bn1.running_mean
level3.tree1.tree1.bn1.running_var
level3.tree1.tree1.conv2.weight
level3.tree1.tree1.bn2.weight
level3.tree1.tree1.bn2.bias
level3.tree1.tree1.bn2.running_mean
level3.tree1.tree1.bn2.running_var
level3.tree1.tree2.conv1.weight
level3.tree1.tree2.bn1.weight
level3.tree1.tree2.bn1.bias
level3.tree1.tree2.bn1.running_mean
level3.tree1.tree2.bn1.running_var
level3.tree1.tree2.conv2.weight
level3.tree1.tree2.bn2.weight
level3.tree1.tree2.bn2.bias
level3.tree1.tree2.bn2.running_mean
level3.tree1.tree2.bn2.running_var
level3.tree1.root.conv.weight
level3.tree1.root.bn.weight
level3.tree1.root.bn.bias
level3.tree1.root.bn.running_mean
level3.tree1.root.bn.running_var
level3.tree1.project.0.weight
level3.tree1.project.1.weight
level3.tree1.project.1.bias
level3.tree1.project.1.running_mean
level3.tree1.project.1.running_var
level3.tree2.tree1.conv1.weight
level3.tree2.tree1.bn1.weight
level3.tree2.tree1.bn1.bias
level3.tree2.tree1.bn1.running_mean
level3.tree2.tree1.bn1.running_var
level3.tree2.tree1.conv2.weight
level3.tree2.tree1.bn2.weight
level3.tree2.tree1.bn2.bias
level3.tree2.tree1.bn2.running_mean
level3.tree2.tree1.bn2.running_var
level3.tree2.tree2.conv1.weight
level3.tree2.tree2.bn1.weight
level3.tree2.tree2.bn1.bias
level3.tree2.tree2.bn1.running_mean
level3.tree2.tree2.bn1.running_var
level3.tree2.tree2.conv2.weight
level3.tree2.tree2.bn2.weight
level3.tree2.tree2.bn2.bias
level3.tree2.tree2.bn2.running_mean
level3.tree2.tree2.bn2.running_var
level3.tree2.root.conv.weight
level3.tree2.root.bn.weight
level3.tree2.root.bn.bias
level3.tree2.root.bn.running_mean
level3.tree2.root.bn.running_var
level3.project.0.weight
level3.project.1.weight
level3.project.1.bias
level3.project.1.running_mean
level3.project.1.running_var
level4.tree1.tree1.conv1.weight
level4.tree1.tree1.bn1.weight
level4.tree1.tree1.bn1.bias
level4.tree1.tree1.bn1.running_mean
level4.tree1.tree1.bn1.running_var
level4.tree1.tree1.conv2.weight
level4.tree1.tree1.bn2.weight
level4.tree1.tree1.bn2.bias
level4.tree1.tree1.bn2.running_mean
level4.tree1.tree1.bn2.running_var
level4.tree1.tree2.conv1.weight
level4.tree1.tree2.bn1.weight
level4.tree1.tree2.bn1.bias
level4.tree1.tree2.bn1.running_mean
level4.tree1.tree2.bn1.running_var
level4.tree1.tree2.conv2.weight
level4.tree1.tree2.bn2.weight
level4.tree1.tree2.bn2.bias
level4.tree1.tree2.bn2.running_mean
level4.tree1.tree2.bn2.running_var
level4.tree1.root.conv.weight
level4.tree1.root.bn.weight
level4.tree1.root.bn.bias
level4.tree1.root.bn.running_mean
level4.tree1.root.bn.running_var
level4.tree1.project.0.weight
level4.tree1.project.1.weight
level4.tree1.project.1.bias
level4.tree1.project.1.running_mean
level4.tree1.project.1.running_var
level4.tree2.tree1.conv1.weight
level4.tree2.tree1.bn1.weight
level4.tree2.tree1.bn1.bias
level4.tree2.tree1.bn1.running_mean
level4.tree2.tree1.bn1.running_var
level4.tree2.tree1.conv2.weight
level4.tree2.tree1.bn2.weight
level4.tree2.tree1.bn2.bias
level4.tree2.tree1.bn2.running_mean
level4.tree2.tree1.bn2.running_var
level4.tree2.tree2.conv1.weight
level4.tree2.tree2.bn1.weight
level4.tree2.tree2.bn1.bias
level4.tree2.tree2.bn1.running_mean
level4.tree2.tree2.bn1.running_var
level4.tree2.tree2.conv2.weight
level4.tree2.tree2.bn2.weight
level4.tree2.tree2.bn2.bias
level4.tree2.tree2.bn2.running_mean
level4.tree2.tree2.bn2.running_var
level4.tree2.root.conv.weight
level4.tree2.root.bn.weight
level4.tree2.root.bn.bias
level4.tree2.root.bn.running_mean
level4.tree2.root.bn.running_var
level4.project.0.weight
level4.project.1.weight
level4.project.1.bias
level4.project.1.running_mean
level4.project.1.running_var
level5.tree1.conv1.weight
level5.tree1.bn1.weight
level5.tree1.bn1.bias
level5.tree1.bn1.running_mean
level5.tree1.bn1.running_var
level5.tree1.conv2.weight
level5.tree1.bn2.weight
level5.tree1.bn2.bias
level5.tree1.bn2.running_mean
level5.tree1.bn2.running_var
level5.tree2.conv1.weight
level5.tree2.bn1.weight
level5.tree2.bn1.bias
level5.tree2.bn1.running_mean
level5.tree2.bn1.running_var
level5.tree2.conv2.weight
level5.tree2.bn2.weight
level5.tree2.bn2.bias
level5.tree2.bn2.running_mean
level5.tree2.bn2.running_var
level5.root.conv.weight
level5.root.bn.weight
level5.root.bn.bias
level5.root.bn.running_mean
level5.root.bn.running_var
level5.project.0.weight
level5.project.1.weight
level5.project.1.bias
level5.project.1.running_mean
level5.project.1.running_var
fc.weight
fc.bias

• 我们训练后的模型如下结构：

models_weights|--epoch
              |--state_dict|--'与预训练模型类似的参数'

model_weights['state_dict']的内容如下：

model_weights=torch.load('ctdet_pascal_dla_384.pth')
for name in model_weights['state_dict']:
    print(name)

可以看到输出的name与预训练模型相比，多了'base.',且多了后续层的参数比如'dla_up.'、'ida_up.'等开头以及'tracked'结尾的name参数：

base.base_layer.0.weight
base.base_layer.1.weight
base.base_layer.1.bias
base.base_layer.1.running_mean
base.base_layer.1.running_var
base.base_layer.1.num_batches_tracked
base.level0.0.weight
base.level0.1.weight
base.level0.1.bias
base.level0.1.running_mean
base.level0.1.running_var
base.level0.1.num_batches_tracked
base.level1.0.weight
base.level1.1.weight
base.level1.1.bias
base.level1.1.running_mean
base.level1.1.running_var
base.level1.1.num_batches_tracked
base.level2.tree1.conv1.weight
base.level2.tree1.bn1.weight
base.level2.tree1.bn1.bias
base.level2.tree1.bn1.running_mean
base.level2.tree1.bn1.running_var
base.level2.tree1.bn1.num_batches_tracked
base.level2.tree1.conv2.weight
base.level2.tree1.bn2.weight
base.level2.tree1.bn2.bias
base.level2.tree1.bn2.running_mean
base.level2.tree1.bn2.running_var
base.level2.tree1.bn2.num_batches_tracked
base.level2.tree2.conv1.weight
base.level2.tree2.bn1.weight
base.level2.tree2.bn1.bias
base.level2.tree2.bn1.running_mean
base.level2.tree2.bn1.running_var
base.level2.tree2.bn1.num_batches_tracked
base.level2.tree2.conv2.weight
base.level2.tree2.bn2.weight
base.level2.tree2.bn2.bias
base.level2.tree2.bn2.running_mean
base.level2.tree2.bn2.running_var
base.level2.tree2.bn2.num_batches_tracked
base.level2.root.conv.weight
base.level2.root.bn.weight
base.level2.root.bn.bias
base.level2.root.bn.running_mean
base.level2.root.bn.running_var
base.level2.root.bn.num_batches_tracked
base.level2.project.0.weight
base.level2.project.1.weight
base.level2.project.1.bias
base.level2.project.1.running_mean
base.level2.project.1.running_var
base.level2.project.1.num_batches_tracked
base.level3.tree1.tree1.conv1.weight
base.level3.tree1.tree1.bn1.weight
base.level3.tree1.tree1.bn1.bias
base.level3.tree1.tree1.bn1.running_mean
base.level3.tree1.tree1.bn1.running_var
base.level3.tree1.tree1.bn1.num_batches_tracked
base.level3.tree1.tree1.conv2.weight
base.level3.tree1.tree1.bn2.weight
base.level3.tree1.tree1.bn2.bias
base.level3.tree1.tree1.bn2.running_mean
base.level3.tree1.tree1.bn2.running_var
base.level3.tree1.tree1.bn2.num_batches_tracked
base.level3.tree1.tree2.conv1.weight
base.level3.tree1.tree2.bn1.weight
base.level3.tree1.tree2.bn1.bias
base.level3.tree1.tree2.bn1.running_mean
base.level3.tree1.tree2.bn1.running_var
base.level3.tree1.tree2.bn1.num_batches_tracked
base.level3.tree1.tree2.conv2.weight
base.level3.tree1.tree2.bn2.weight
base.level3.tree1.tree2.bn2.bias
base.level3.tree1.tree2.bn2.running_mean
base.level3.tree1.tree2.bn2.running_var
base.level3.tree1.tree2.bn2.num_batches_tracked
base.level3.tree1.root.conv.weight
base.level3.tree1.root.bn.weight
base.level3.tree1.root.bn.bias
base.level3.tree1.root.bn.running_mean
base.level3.tree1.root.bn.running_var
base.level3.tree1.root.bn.num_batches_tracked
base.level3.tree1.project.0.weight
base.level3.tree1.project.1.weight
base.level3.tree1.project.1.bias
base.level3.tree1.project.1.running_mean
base.level3.tree1.project.1.running_var
base.level3.tree1.project.1.num_batches_tracked
base.level3.tree2.tree1.conv1.weight
base.level3.tree2.tree1.bn1.weight
base.level3.tree2.tree1.bn1.bias
base.level3.tree2.tree1.bn1.running_mean
base.level3.tree2.tree1.bn1.running_var
base.level3.tree2.tree1.bn1.num_batches_tracked
base.level3.tree2.tree1.conv2.weight
base.level3.tree2.tree1.bn2.weight
base.level3.tree2.tree1.bn2.bias
base.level3.tree2.tree1.bn2.running_mean
base.level3.tree2.tree1.bn2.running_var
base.level3.tree2.tree1.bn2.num_batches_tracked
base.level3.tree2.tree2.conv1.weight
base.level3.tree2.tree2.bn1.weight
base.level3.tree2.tree2.bn1.bias
base.level3.tree2.tree2.bn1.running_mean
base.level3.tree2.tree2.bn1.running_var
base.level3.tree2.tree2.bn1.num_batches_tracked
base.level3.tree2.tree2.conv2.weight
base.level3.tree2.tree2.bn2.weight
base.level3.tree2.tree2.bn2.bias
base.level3.tree2.tree2.bn2.running_mean
base.level3.tree2.tree2.bn2.running_var
base.level3.tree2.tree2.bn2.num_batches_tracked
base.level3.tree2.root.conv.weight
base.level3.tree2.root.bn.weight
base.level3.tree2.root.bn.bias
base.level3.tree2.root.bn.running_mean
base.level3.tree2.root.bn.running_var
base.level3.tree2.root.bn.num_batches_tracked
base.level3.project.0.weight
base.level3.project.1.weight
base.level3.project.1.bias
base.level3.project.1.running_mean
base.level3.project.1.running_var
base.level3.project.1.num_batches_tracked
base.level4.tree1.tree1.conv1.weight
base.level4.tree1.tree1.bn1.weight
base.level4.tree1.tree1.bn1.bias
base.level4.tree1.tree1.bn1.running_mean
base.level4.tree1.tree1.bn1.running_var
base.level4.tree1.tree1.bn1.num_batches_tracked
base.level4.tree1.tree1.conv2.weight
base.level4.tree1.tree1.bn2.weight
base.level4.tree1.tree1.bn2.bias
base.level4.tree1.tree1.bn2.running_mean
base.level4.tree1.tree1.bn2.running_var
base.level4.tree1.tree1.bn2.num_batches_tracked
base.level4.tree1.tree2.conv1.weight
base.level4.tree1.tree2.bn1.weight
base.level4.tree1.tree2.bn1.bias
base.level4.tree1.tree2.bn1.running_mean
base.level4.tree1.tree2.bn1.running_var
base.level4.tree1.tree2.bn1.num_batches_tracked
base.level4.tree1.tree2.conv2.weight
base.level4.tree1.tree2.bn2.weight
base.level4.tree1.tree2.bn2.bias
base.level4.tree1.tree2.bn2.running_mean
base.level4.tree1.tree2.bn2.running_var
base.level4.tree1.tree2.bn2.num_batches_tracked
base.level4.tree1.root.conv.weight
base.level4.tree1.root.bn.weight
base.level4.tree1.root.bn.bias
base.level4.tree1.root.bn.running_mean
base.level4.tree1.root.bn.running_var
base.level4.tree1.root.bn.num_batches_tracked
base.level4.tree1.project.0.weight
base.level4.tree1.project.1.weight
base.level4.tree1.project.1.bias
base.level4.tree1.project.1.running_mean
base.level4.tree1.project.1.running_var
base.level4.tree1.project.1.num_batches_tracked
base.level4.tree2.tree1.conv1.weight
base.level4.tree2.tree1.bn1.weight
base.level4.tree2.tree1.bn1.bias
base.level4.tree2.tree1.bn1.running_mean
base.level4.tree2.tree1.bn1.running_var
base.level4.tree2.tree1.bn1.num_batches_tracked
base.level4.tree2.tree1.conv2.weight
base.level4.tree2.tree1.bn2.weight
base.level4.tree2.tree1.bn2.bias
base.level4.tree2.tree1.bn2.running_mean
base.level4.tree2.tree1.bn2.running_var
base.level4.tree2.tree1.bn2.num_batches_tracked
base.level4.tree2.tree2.conv1.weight
base.level4.tree2.tree2.bn1.weight
base.level4.tree2.tree2.bn1.bias
base.level4.tree2.tree2.bn1.running_mean
base.level4.tree2.tree2.bn1.running_var
base.level4.tree2.tree2.bn1.num_batches_tracked
base.level4.tree2.tree2.conv2.weight
base.level4.tree2.tree2.bn2.weight
base.level4.tree2.tree2.bn2.bias
base.level4.tree2.tree2.bn2.running_mean
base.level4.tree2.tree2.bn2.running_var
base.level4.tree2.tree2.bn2.num_batches_tracked
base.level4.tree2.root.conv.weight
base.level4.tree2.root.bn.weight
base.level4.tree2.root.bn.bias
base.level4.tree2.root.bn.running_mean
base.level4.tree2.root.bn.running_var
base.level4.tree2.root.bn.num_batches_tracked
base.level4.project.0.weight
base.level4.project.1.weight
base.level4.project.1.bias
base.level4.project.1.running_mean
base.level4.project.1.running_var
base.level4.project.1.num_batches_tracked
base.level5.tree1.conv1.weight
base.level5.tree1.bn1.weight
base.level5.tree1.bn1.bias
base.level5.tree1.bn1.running_mean
base.level5.tree1.bn1.running_var
base.level5.tree1.bn1.num_batches_tracked
base.level5.tree1.conv2.weight
base.level5.tree1.bn2.weight
base.level5.tree1.bn2.bias
base.level5.tree1.bn2.running_mean
base.level5.tree1.bn2.running_var
base.level5.tree1.bn2.num_batches_tracked
base.level5.tree2.conv1.weight
base.level5.tree2.bn1.weight
base.level5.tree2.bn1.bias
base.level5.tree2.bn1.running_mean
base.level5.tree2.bn1.running_var
base.level5.tree2.bn1.num_batches_tracked
base.level5.tree2.conv2.weight
base.level5.tree2.bn2.weight
base.level5.tree2.bn2.bias
base.level5.tree2.bn2.running_mean
base.level5.tree2.bn2.running_var
base.level5.tree2.bn2.num_batches_tracked
base.level5.root.conv.weight
base.level5.root.bn.weight
base.level5.root.bn.bias
base.level5.root.bn.running_mean
base.level5.root.bn.running_var
base.level5.root.bn.num_batches_tracked
base.level5.project.0.weight
base.level5.project.1.weight
base.level5.project.1.bias
base.level5.project.1.running_mean
base.level5.project.1.running_var
base.level5.project.1.num_batches_tracked
base.fc.weight
base.fc.bias
dla_up.ida_0.proj_1.actf.0.weight
dla_up.ida_0.proj_1.actf.0.bias
dla_up.ida_0.proj_1.actf.0.running_mean
dla_up.ida_0.proj_1.actf.0.running_var
dla_up.ida_0.proj_1.actf.0.num_batches_tracked
dla_up.ida_0.proj_1.conv.weight
dla_up.ida_0.proj_1.conv.bias
dla_up.ida_0.proj_1.conv.conv_offset_mask.weight
dla_up.ida_0.proj_1.conv.conv_offset_mask.bias
dla_up.ida_0.up_1.weight
dla_up.ida_0.node_1.actf.0.weight
dla_up.ida_0.node_1.actf.0.bias
dla_up.ida_0.node_1.actf.0.running_mean
dla_up.ida_0.node_1.actf.0.running_var
dla_up.ida_0.node_1.actf.0.num_batches_tracked
dla_up.ida_0.node_1.conv.weight
dla_up.ida_0.node_1.conv.bias
dla_up.ida_0.node_1.conv.conv_offset_mask.weight
dla_up.ida_0.node_1.conv.conv_offset_mask.bias
dla_up.ida_1.proj_1.actf.0.weight
dla_up.ida_1.proj_1.actf.0.bias
dla_up.ida_1.proj_1.actf.0.running_mean
dla_up.ida_1.proj_1.actf.0.running_var
dla_up.ida_1.proj_1.actf.0.num_batches_tracked
dla_up.ida_1.proj_1.conv.weight
dla_up.ida_1.proj_1.conv.bias
dla_up.ida_1.proj_1.conv.conv_offset_mask.weight
dla_up.ida_1.proj_1.conv.conv_offset_mask.bias
dla_up.ida_1.up_1.weight
dla_up.ida_1.node_1.actf.0.weight
dla_up.ida_1.node_1.actf.0.bias
dla_up.ida_1.node_1.actf.0.running_mean
dla_up.ida_1.node_1.actf.0.running_var
dla_up.ida_1.node_1.actf.0.num_batches_tracked
dla_up.ida_1.node_1.conv.weight
dla_up.ida_1.node_1.conv.bias
dla_up.ida_1.node_1.conv.conv_offset_mask.weight
dla_up.ida_1.node_1.conv.conv_offset_mask.bias
dla_up.ida_1.proj_2.actf.0.weight
dla_up.ida_1.proj_2.actf.0.bias
dla_up.ida_1.proj_2.actf.0.running_mean
dla_up.ida_1.proj_2.actf.0.running_var
dla_up.ida_1.proj_2.actf.0.num_batches_tracked
dla_up.ida_1.proj_2.conv.weight
dla_up.ida_1.proj_2.conv.bias
dla_up.ida_1.proj_2.conv.conv_offset_mask.weight
dla_up.ida_1.proj_2.conv.conv_offset_mask.bias
dla_up.ida_1.up_2.weight
dla_up.ida_1.node_2.actf.0.weight
dla_up.ida_1.node_2.actf.0.bias
dla_up.ida_1.node_2.actf.0.running_mean
dla_up.ida_1.node_2.actf.0.running_var
dla_up.ida_1.node_2.actf.0.num_batches_tracked
dla_up.ida_1.node_2.conv.weight
dla_up.ida_1.node_2.conv.bias
dla_up.ida_1.node_2.conv.conv_offset_mask.weight
dla_up.ida_1.node_2.conv.conv_offset_mask.bias
dla_up.ida_2.proj_1.actf.0.weight
dla_up.ida_2.proj_1.actf.0.bias
dla_up.ida_2.proj_1.actf.0.running_mean
dla_up.ida_2.proj_1.actf.0.running_var
dla_up.ida_2.proj_1.actf.0.num_batches_tracked
dla_up.ida_2.proj_1.conv.weight
dla_up.ida_2.proj_1.conv.bias
dla_up.ida_2.proj_1.conv.conv_offset_mask.weight
dla_up.ida_2.proj_1.conv.conv_offset_mask.bias
dla_up.ida_2.up_1.weight
dla_up.ida_2.node_1.actf.0.weight
dla_up.ida_2.node_1.actf.0.bias
dla_up.ida_2.node_1.actf.0.running_mean
dla_up.ida_2.node_1.actf.0.running_var
dla_up.ida_2.node_1.actf.0.num_batches_tracked
dla_up.ida_2.node_1.conv.weight
dla_up.ida_2.node_1.conv.bias
dla_up.ida_2.node_1.conv.conv_offset_mask.weight
dla_up.ida_2.node_1.conv.conv_offset_mask.bias
dla_up.ida_2.proj_2.actf.0.weight
dla_up.ida_2.proj_2.actf.0.bias
dla_up.ida_2.proj_2.actf.0.running_mean
dla_up.ida_2.proj_2.actf.0.running_var
dla_up.ida_2.proj_2.actf.0.num_batches_tracked
dla_up.ida_2.proj_2.conv.weight
dla_up.ida_2.proj_2.conv.bias
dla_up.ida_2.proj_2.conv.conv_offset_mask.weight
dla_up.ida_2.proj_2.conv.conv_offset_mask.bias
dla_up.ida_2.up_2.weight
dla_up.ida_2.node_2.actf.0.weight
dla_up.ida_2.node_2.actf.0.bias
dla_up.ida_2.node_2.actf.0.running_mean
dla_up.ida_2.node_2.actf.0.running_var
dla_up.ida_2.node_2.actf.0.num_batches_tracked
dla_up.ida_2.node_2.conv.weight
dla_up.ida_2.node_2.conv.bias
dla_up.ida_2.node_2.conv.conv_offset_mask.weight
dla_up.ida_2.node_2.conv.conv_offset_mask.bias
dla_up.ida_2.proj_3.actf.0.weight
dla_up.ida_2.proj_3.actf.0.bias
dla_up.ida_2.proj_3.actf.0.running_mean
dla_up.ida_2.proj_3.actf.0.running_var
dla_up.ida_2.proj_3.actf.0.num_batches_tracked
dla_up.ida_2.proj_3.conv.weight
dla_up.ida_2.proj_3.conv.bias
dla_up.ida_2.proj_3.conv.conv_offset_mask.weight
dla_up.ida_2.proj_3.conv.conv_offset_mask.bias
dla_up.ida_2.up_3.weight
dla_up.ida_2.node_3.actf.0.weight
dla_up.ida_2.node_3.actf.0.bias
dla_up.ida_2.node_3.actf.0.running_mean
dla_up.ida_2.node_3.actf.0.running_var
dla_up.ida_2.node_3.actf.0.num_batches_tracked
dla_up.ida_2.node_3.conv.weight
dla_up.ida_2.node_3.conv.bias
dla_up.ida_2.node_3.conv.conv_offset_mask.weight
dla_up.ida_2.node_3.conv.conv_offset_mask.bias
ida_up.proj_1.actf.0.weight
ida_up.proj_1.actf.0.bias
ida_up.proj_1.actf.0.running_mean
ida_up.proj_1.actf.0.running_var
ida_up.proj_1.actf.0.num_batches_tracked
ida_up.proj_1.conv.weight
ida_up.proj_1.conv.bias
ida_up.proj_1.conv.conv_offset_mask.weight
ida_up.proj_1.conv.conv_offset_mask.bias
ida_up.up_1.weight
ida_up.node_1.actf.0.weight
ida_up.node_1.actf.0.bias
ida_up.node_1.actf.0.running_mean
ida_up.node_1.actf.0.running_var
ida_up.node_1.actf.0.num_batches_tracked
ida_up.node_1.conv.weight
ida_up.node_1.conv.bias
ida_up.node_1.conv.conv_offset_mask.weight
ida_up.node_1.conv.conv_offset_mask.bias
ida_up.proj_2.actf.0.weight
ida_up.proj_2.actf.0.bias
ida_up.proj_2.actf.0.running_mean
ida_up.proj_2.actf.0.running_var
ida_up.proj_2.actf.0.num_batches_tracked
ida_up.proj_2.conv.weight
ida_up.proj_2.conv.bias
ida_up.proj_2.conv.conv_offset_mask.weight
ida_up.proj_2.conv.conv_offset_mask.bias
ida_up.up_2.weight
ida_up.node_2.actf.0.weight
ida_up.node_2.actf.0.bias
ida_up.node_2.actf.0.running_mean
ida_up.node_2.actf.0.running_var
ida_up.node_2.actf.0.num_batches_tracked
ida_up.node_2.conv.weight
ida_up.node_2.conv.bias
ida_up.node_2.conv.conv_offset_mask.weight
ida_up.node_2.conv.conv_offset_mask.bias
hm.0.weight
hm.0.bias
hm.2.weight
hm.2.bias
wh.0.weight
wh.0.bias
wh.2.weight
wh.2.bias
reg.0.weight
reg.0.bias
reg.2.weight
reg.2.bias

'所以要删除的名字有如下特点：'
   name[-7:]=='tracked'
   name[:6]=='dla_up'
   name[:6]=='ida_up'
   name[:2]=='hm'
   name[:2]=='wh'
   name[:3]=='reg'

最终生成与预训练模型结构相似的结构模型代码如下：

import torch
from torch import nn
new_model={}
model_weights=torch.load('ctdet_pascal_dla_384.pth')
for name in model_weights['state_dict']:
    print(name)
    n=0
    if(name[-7:]=='tracked'):
        n=1
    elif(name[:6]=='dla_up' or  name[:6]=='ida_up'):
        n=1
    elif(name[:2]=='hm' or name[:2]=='wh' or name[:3]=='reg'):
        n=1
    if(n==0):
        new_model[name[5:]]=model_weights['state_dict'][name]
for name in new_model:
    print(name)
torch.save(new_model,'pre_pascal_dla_384.pth')

2.检验下生成的预训练模型是否有效：