DEtection TRansformers(DETR)
DEtection TRansformer(DETR)是Facebook AI的研究者提出的Transformer的视觉版本,用于目标检测和全景分割。这是第一个将Transformer成功整合为检测pipeline中心构建块的目标检测框架。(原文地址)
本文总结了网上多位博主的复现工作,并就本人复现经历加以说明,是较为详尽的可完全按照此过程操作完成程序复现。
1、代码地址Github:https://github.com/facebookresearch/detr
2、论文地址:End-to-End Object Detection with Transformers
第一步,先将代码下载下来,然后在pycharm中打开,运行terminal,输入
pip install -r requirements.txt
在安装第二个库cocoapi时出现了安装失败。
解决方法:–>本地安装加换其它大神编译好的包,下载地址:https://github.com/philferriere/cocoapi
打开anconda的命令界面输入activate your_env_name (激活anaconda虚拟环境,在pycharm的端口输入);
进入刚才下载的源码setup.py所在目录 cocoapi-master\PythonAPI;
运行python setup.py build_ext install即可安装完成。
解决了这个问题,安装其他的库就没得什么问题了,如果遇到vc++2015之类的问题建议在vs studio中把c++那一项也安装好。
第二步,将源程序的pth文件改一下,因为他是用的coco数据集,而我们只需要训练自己的数据集,就是下图这个文件
(该文件通过github上下载的包中README.md文件有网址直接下载)
之后运行change.py(内容如下程序),改变当前pth文件的class,修改num_classes为自己的类别数+1,以适应自己训练的数据集。
import torch pretrained_weights = torch.load('detr-r50-e632da11.pth') num_class = 3 #这里是你的物体数+1,因为背景也算一个 pretrained_weights["model"]["class_embed.weight"].resize_(num_class+1, 256) pretrained_weights["model"]["class_embed.bias"].resize_(num_class+1) torch.save(pretrained_weights, "detr-r50_%d.pth"%num_class
运行完后会生成下图的文件:
,将该文件保存到当前工程路径下。
第三步,准备自己的训练集,首先你要用自己的数据标记一个VOC类型的数据集,这里就不多赘述了,可以自己百度下载labelimg来打标,然后将文件放入这几个文件夹,就可以了。
DETR需要的数据集格式为coco格式,图片和标签文件保存于训练集、测试集、验证集、标签文件四个文件夹中,其中annotations中存放json格式的标签文件
(数据集按此文件结构放置,命名也不要变)
第四步,转换成json格式,生成的文件夹记得改为instances_train2017,json这种样子,如下图
下面的代码包含了几种数据集RSOD、NWPU、DIOR、YOLO数据集标签文件转换json功能。新建py文件tojson.py,使用如下代码生成需要的json文件。
生成instances_train2017.json
(a)修改29行image_path默认路径为train2017的路径;
(b)修改31行annotation_path默认路径为标签文件路径(train和val的标签都放在这个文件夹下,所以生成instances_val2017.json时就不需要再修改这个路径了);
(c)修改33行dataset为自己的数据集名称NWPU
(d).修改34行save的默认路径为json文件的保存路径…/NWPUVHR-10/annotations/instances_train2017.json
import os
import cv2
import json
import argparse
from tqdm import tqdm
import xml.etree.ElementTree as ET
COCO_DICT=['images','annotations','categories']
IMAGES_DICT=['file_name','height','width','id']
ANNOTATIONS_DICT=['image_id','iscrowd','area','bbox','category_id','id']
CATEGORIES_DICT=['id','name']
## {'supercategory': 'person', 'id': 1, 'name': 'person'}
## {'supercategory': 'vehicle', 'id': 2, 'name': 'bicycle'}
YOLO_CATEGORIES=['person']
RSOD_CATEGORIES=['aircraft','playground','overpass','oiltank']
NWPU_CATEGORIES=['airplane','ship','storage tank','baseball diamond','tennis court',\
'basketball court','ground track field','harbor','bridge','vehicle']
VOC_CATEGORIES=['aeroplane','bicycle','bird','boat','bottle','bus','car','cat','chair','cow',\ 'diningtable','dog','horse','motorbike','person','pottedplant','sheep','sofa','train','tvmonitor']
DIOR_CATEGORIES=['golffield','Expressway-toll-station','vehicle','trainstation','chimney','storagetank',\
'ship','harbor','airplane','groundtrackfield','tenniscourt','dam','basketballcourt',\
'Expressway-Service-area','stadium','airport','baseballfield','bridge','windmill','overpass']
parser=argparse.ArgumentParser(description='2COCO')
#parser.add_argument('--image_path',type=str,default=r'T:/shujuji/DIOR/JPEGImages-trainval/',help='config file')
parser.add_argument('--image_path',type=str,default=r'G:/NWPU VHR-10 dataset/positive image set/',help='config file')
#parser.add_argument('--annotation_path',type=str,default=r'T:/shujuji/DIOR/Annotations/',help='config file')
parser.add_argument('--annotation_path',type=str,default=r'G:/NWPU VHR-10 dataset/ground truth/',help='config file')
parser.add_argument('--dataset',type=str,default='NWPU',help='config file')
parser.add_argument('--save',type=str,default='G:/NWPU VHR-10 dataset/instances_train2017.json',help='config file')
args=parser.parse_args()
def load_json(path):
with open(path,'r') as f:
json_dict=json.load(f)
for i in json_dict:
print(i)
print(json_dict['annotations'])
def save_json(dict,path):
print('SAVE_JSON...')
with open(path,'w') as f:
json.dump(dict,f)
print('SUCCESSFUL_SAVE_JSON:',path)
def load_image(path):
img=cv2.imread(path)
return img.shape[0],img.shape[1]
def generate_categories_dict(category): #ANNOTATIONS_DICT=['image_id','iscrowd','area','bbox','category_id','id']
print('GENERATE_CATEGORIES_DICT...')
return [{CATEGORIES_DICT[0]:category.index(x)+1,CATEGORIES_DICT[1]:x} for x in category] #CATEGORIES_DICT=['id','name']
def generate_images_dict(imagelist,image_path,start_image_id=11725): #IMAGES_DICT=['file_name','height','width','id']
print('GENERATE_IMAGES_DICT...')
images_dict=[]
with tqdm(total=len(imagelist)) as load_bar:
for x in imagelist: #x就是图片的名称
#print(start_image_id)
dict={IMAGES_DICT[0]:x,IMAGES_DICT[1]:load_image(image_path+x)[0],\
IMAGES_DICT[2]:load_image(image_path+x)[1],IMAGES_DICT[3]:imagelist.index(x)+start_image_id}
load_bar.update(1)
images_dict.append(dict)
return images_dict
def DIOR_Dataset(image_path,annotation_path,start_image_id=11725,start_id=0):
categories_dict=generate_categories_dict(DIOR_CATEGORIES) #CATEGORIES_DICT=['id':,1'name':golffield......] id从1开始
imgname=os.listdir(image_path)
images_dict=generate_images_dict(imgname,image_path,start_image_id) #IMAGES_DICT=['file_name','height','width','id'] id从0开始的
print('GENERATE_ANNOTATIONS_DICT...') #生成cooc的注记 ANNOTATIONS_DICT=['image_id','iscrowd','area','bbox','category_id','id']
annotations_dict=[]
id=start_id
for i in images_dict:
image_id=i['id']
print(image_id)
image_name=i['file_name']
annotation_xml=annotation_path+image_name.split('.')[0]+'.xml'
tree=ET.parse(annotation_xml)
root=tree.getroot()
for j in root.findall('object'):
category=j.find('name').text
category_id=DIOR_CATEGORIES.index(category) #字典的索引,是从1开始的
x_min=float(j.find('bndbox').find('xmin').text)
y_min=float(j.find('bndbox').find('ymin').text)
w=float(j.find('bndbox').find('xmax').text)-x_min
h=float(j.find('bndbox').find('ymax').text)-y_min
area = w * h
bbox = [x_min, y_min, w, h]
dict = {'image_id': image_id, 'iscrowd': 0, 'area': area, 'bbox': bbox, 'category_id': category_id,
'id': id}
annotations_dict.append(dict)
id=id+1
print('SUCCESSFUL_GENERATE_DIOR_JSON')
return {COCO_DICT[0]:images_dict,COCO_DICT[1]:annotations_dict,COCO_DICT[2]:categories_dict}
def NWPU_Dataset(image_path,annotation_path,start_image_id=0,start_id=0):
categories_dict=generate_categories_dict(NWPU_CATEGORIES)
imgname=os.listdir(image_path)
images_dict=generate_images_dict(imgname,image_path,start_image_id)
print('GENERATE_ANNOTATIONS_DICT...')
annotations_dict=[]
id=start_id
for i in images_dict:
image_id=i['id']
image_name=i['file_name']
annotation_txt=annotation_path+image_name.split('.')[0]+'.txt'
txt=open(annotation_txt,'r')
lines=txt.readlines()
for j in lines:
if j=='\n':
continue
category_id=int(j.split(',')[4])
category=NWPU_CATEGORIES[category_id-1]
print(category_id,' ',category)
x_min=float(j.split(',')[0].split('(')[1])
y_min=float(j.split(',')[1].split(')')[0])
w=float(j.split(',')[2].split('(')[1])-x_min
h=float(j.split(',')[3].split(')')[0])-y_min
area=w*h
bbox=[x_min,y_min,w,h]
dict = {'image_id': image_id, 'iscrowd': 0, 'area': area, 'bbox': bbox, 'category_id': category_id,
'id': id}
id=id+1
annotations_dict.append(dict)
print('SUCCESSFUL_GENERATE_NWPU_JSON')
return {COCO_DICT[0]:images_dict,COCO_DICT[1]:annotations_dict,COCO_DICT[2]:categories_dict}
def YOLO_Dataset(image_path,annotation_path,start_image_id=0,start_id=0):
categories_dict=generate_categories_dict(YOLO_CATEGORIES)
imgname=os.listdir(image_path)
images_dict=generate_images_dict(imgname,image_path)
print('GENERATE_ANNOTATIONS_DICT...')
annotations_dict=[]
id=start_id
for i in images_dict:
image_id=i['id']
image_name=i['file_name']
W,H=i['width'],i['height']
annotation_txt=annotation_path+image_name.split('.')[0]+'.txt'
txt=open(annotation_txt,'r')
lines=txt.readlines()
for j in lines:
category_id=int(j.split(' ')[0])+1
category=YOLO_CATEGORIES
x=float(j.split(' ')[1])
y=float(j.split(' ')[2])
w=float(j.split(' ')[3])
h=float(j.split(' ')[4])
x_min=(x-w/2)*W
y_min=(y-h/2)*H
w=w*W
h=h*H
area=w*h
bbox=[x_min,y_min,w,h]
dict={'image_id':image_id,'iscrowd':0,'area':area,'bbox':bbox,'category_id':category_id,'id':id}
annotations_dict.append(dict)
id=id+1
print('SUCCESSFUL_GENERATE_YOLO_JSON')
return {COCO_DICT[0]:images_dict,COCO_DICT[1]:annotations_dict,COCO_DICT[2]:categories_dict}
def RSOD_Dataset(image_path,annotation_path,start_image_id=0,start_id=0):
categories_dict=generate_categories_dict(RSOD_CATEGORIES)
imgname=os.listdir(image_path)
images_dict=generate_images_dict(imgname,image_path,start_image_id)
print('GENERATE_ANNOTATIONS_DICT...')
annotations_dict=[]
id=start_id
for i in images_dict:
image_id=i['id']
image_name=i['file_name']
annotation_txt=annotation_path+image_name.split('.')[0]+'.txt'
txt=open(annotation_txt,'r')
lines=txt.readlines()
for j in lines:
category=j.split('\t')[1]
category_id=RSOD_CATEGORIES.index(category)+1
x_min=float(j.split('\t')[2])
y_min=float(j.split('\t')[3])
w=float(j.split('\t')[4])-x_min
h=float(j.split('\t')[5])-y_min
area = w * h
bbox = [x_min, y_min, w, h]
dict = {'image_id': image_id, 'iscrowd': 0, 'area': area, 'bbox': bbox, 'category_id': category_id,
'id': id}
annotations_dict.append(dict)
id=id+1
print('SUCCESSFUL_GENERATE_RSOD_JSON')
return {COCO_DICT[0]:images_dict,COCO_DICT[1]:annotations_dict,COCO_DICT[2]:categories_dict}
if __name__=='__main__':
dataset=args.dataset #数据集名字
save=args.save #json的保存路径
image_path=args.image_path #对于coco是图片的路径
annotation_path=args.annotation_path #coco的annotation路径
if dataset=='RSOD':
json_dict=RSOD_Dataset(image_path,annotation_path,0)
if dataset=='NWPU':
json_dict=NWPU_Dataset(image_path,annotation_path,0)
if dataset=='DIOR':
json_dict=DIOR_Dataset(image_path,annotation_path,11725)
if dataset=='YOLO':
json_dict=YOLO_Dataset(image_path,annotation_path,0)
save_json(json_dict,save)
数据集到此就全部准备好了。
第五步,修改detr.py中305行的num_classes改成你的物体种类的数目
第六步,运行main.py进行训练
修改main.py文件的epochs、lr、batch_size等训练参数。两种方法:修改默认参数;用命令行运行。
1、修改默认参数,在main.py中,修改下面参数。
2、命令行执行
python main.py --dataset_file "coco" --coco_path data/coco --epochs 100 --lr=1e-4 --batch_size=2 --num_workers=4 --output_dir="outputs" --resume="detr-r50_3.pth"
训练完后会在outputs生成下图的文件,log文件是记录每一个epoch的一些信息。
到这里就完成了整个训练。
其实训练过程中,DETR在每个epoch都会自动进行一次精度评估,评估结果查看直接输出或生成的中间文件就行,也可以再用main.py评估一下精度,将下列代码中的–resume改为最后一个epoch输出的模型路径,–coco_path改为自己的数据集路径即可。
python main.py --batch_size 6 --no_aux_loss --eval --resume /home/detr-main/outputs/checkpoint0299.pth --coco_path /home/NWPUVHR-10
最后,就是拿自己的训练的模型进行测试,更改这里的图片路径为要测试的图片的路径,还有第19行的CLASSES=[],记得改成自己的类别!
将要预测的图片保存在一个文件夹下,预测时一次输出所有图片的预测结果
需要修改的参数有:
backbone,我当时下载的backbone是resnet50,修改(训练时已经下载好了主干特征网络是Resnet50的DETR权重文件,放在主文件夹下)
数据集有关参数
–coco_path 修改为自己的数据集路径
–outputdir 修改为建立的预测图片的保存文件夹
–resume 修改为训练好的模型文件路径
修改待预测的图片文件夹路径image_file_path和image_path
放一张检测后的图像。
预测程序:
import argparse
import random
import time
from pathlib import Path
import numpy as np
import torch
from models import build_model
from PIL import Image
import os
import torchvision
from torchvision.ops.boxes import batched_nms
import cv2
#-------------------------------------------------------------------------设置参数
def get_args_parser():
parser = argparse.ArgumentParser('Set transformer detector', add_help=False)
parser.add_argument('--lr', default=1e-4, type=float)
parser.add_argument('--lr_backbone', default=1e-5, type=float)
parser.add_argument('--batch_size', default=2, type=int)
parser.add_argument('--weight_decay', default=1e-4, type=float)
parser.add_argument('--epochs', default=300, type=int)
parser.add_argument('--lr_drop', default=200, type=int)
parser.add_argument('--clip_max_norm', default=0.1, type=float,
help='gradient clipping max norm')
# Model parameters
parser.add_argument('--frozen_weights', type=str, default=None,
help="Path to the pretrained model. If set, only the mask head will be trained")
# * Backbone
parser.add_argument('--backbone', default='resnet101', type=str,
help="Name of the convolutional backbone to use")
parser.add_argument('--dilation', action='store_true',
help="If true, we replace stride with dilation in the last convolutional block (DC5)")
parser.add_argument('--position_embedding', default='sine', type=str, choices=('sine', 'learned'),
help="Type of positional embedding to use on top of the image features")
# * Transformer
parser.add_argument('--enc_layers', default=6, type=int,
help="Number of encoding layers in the transformer")
parser.add_argument('--dec_layers', default=6, type=int,
help="Number of decoding layers in the transformer")
parser.add_argument('--dim_feedforward', default=2048, type=int,
help="Intermediate size of the feedforward layers in the transformer blocks")
parser.add_argument('--hidden_dim', default=256, type=int,
help="Size of the embeddings (dimension of the transformer)")
parser.add_argument('--dropout', default=0.1, type=float,
help="Dropout applied in the transformer")
parser.add_argument('--nheads', default=8, type=int,
help="Number of attention heads inside the transformer's attentions")
parser.add_argument('--num_queries', default=100, type=int,
help="Number of query slots")
parser.add_argument('--pre_norm', action='store_true')
# * Segmentation
parser.add_argument('--masks', action='store_true',
help="Train segmentation head if the flag is provided")
# Loss
parser.add_argument('--no_aux_loss', dest='aux_loss', default='False',
help="Disables auxiliary decoding losses (loss at each layer)")
# * Matcher
parser.add_argument('--set_cost_class', default=1, type=float,
help="Class coefficient in the matching cost")
parser.add_argument('--set_cost_bbox', default=5, type=float,
help="L1 box coefficient in the matching cost")
parser.add_argument('--set_cost_giou', default=2, type=float,
help="giou box coefficient in the matching cost")
# * Loss coefficients
parser.add_argument('--mask_loss_coef', default=1, type=float)
parser.add_argument('--dice_loss_coef', default=1, type=float)
parser.add_argument('--bbox_loss_coef', default=5, type=float)
parser.add_argument('--giou_loss_coef', default=2, type=float)
parser.add_argument('--eos_coef', default=0.1, type=float,
help="Relative classification weight of the no-object class")
# dataset parameters
parser.add_argument('--dataset_file', default='coco')
parser.add_argument('--coco_path', type=str,default="coco")
parser.add_argument('--coco_panoptic_path', type=str)
parser.add_argument('--remove_difficult', action='store_true')
parser.add_argument('--output_dir', default='inference_demo/inference_output',
help='path where to save, empty for no saving')
parser.add_argument('--device', default='cuda',
help='device to use for training / testing')
parser.add_argument('--seed', default=42, type=int)
parser.add_argument('--resume', default='inference_demo/weights/detr-r101-2c7b67e5.pth', help='resume from checkpoint')
parser.add_argument('--start_epoch', default=0, type=int, metavar='N',
help='start epoch')
parser.add_argument('--eval', default="True")
parser.add_argument('--num_workers', default=2, type=int)
# distributed training parameters
parser.add_argument('--world_size', default=1, type=int,
help='number of distributed processes')
parser.add_argument('--dist_url', default='env://', help='url used to set up distributed training')
return parser
def box_cxcywh_to_xyxy(x):
#将DETR的检测框坐标(x_center,y_cengter,w,h)转化成coco数据集的检测框坐标(x0,y0,x1,y1)
x_c, y_c, w, h = x.unbind(1)
b = [(x_c - 0.5 * w), (y_c - 0.5 * h),
(x_c + 0.5 * w), (y_c + 0.5 * h)]
return torch.stack(b, dim=1)
def rescale_bboxes(out_bbox, size):
#把比例坐标乘以图像的宽和高,变成真实坐标
img_w, img_h = size
b = box_cxcywh_to_xyxy(out_bbox)
b = b * torch.tensor([img_w, img_h, img_w, img_h], dtype=torch.float32)
return b
def filter_boxes(scores, boxes, confidence=0.7, apply_nms=True, iou=0.5):
#筛选出真正的置信度高的框
keep = scores.max(-1).values > confidence
scores, boxes = scores[keep], boxes[keep]
if apply_nms:
top_scores, labels = scores.max(-1)
keep = batched_nms(boxes, top_scores, labels, iou)
scores, boxes = scores[keep], boxes[keep]
return scores, boxes
# COCO classes
CLASSES = [
'N/A', 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus',
'train', 'truck', 'boat', 'traffic light', 'fire hydrant', 'N/A',
'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse',
'sheep', 'cow', 'elephant', 'bear', 'zebra', 'giraffe', 'N/A', 'backpack',
'umbrella', 'N/A', 'N/A', 'handbag', 'tie', 'suitcase', 'frisbee', 'skis',
'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove',
'skateboard', 'surfboard', 'tennis racket', 'bottle', 'N/A', 'wine glass',
'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich',
'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake',
'chair', 'couch', 'potted plant', 'bed', 'N/A', 'dining table', 'N/A',
'N/A', 'toilet', 'N/A', 'tv', 'laptop', 'mouse', 'remote', 'keyboard',
'cell phone', 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'N/A',
'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier',
'toothbrush'
]
def plot_one_box(x, img, color=None, label=None, line_thickness=1):
#把检测框画到图片上
tl = line_thickness or round(0.002 * (img.shape[0] + img.shape[1]) / 2) + 1 # line/font thickness
color = color or [random.randint(0, 255) for _ in range(3)]
c1, c2 = (int(x[0]), int(x[1])), (int(x[2]), int(x[3]))
cv2.rectangle(img, c1, c2, color, thickness=tl, lineType=cv2.LINE_AA)
if label:
tf = max(tl - 1, 1) # font thickness
t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0]
c2 = c1[0] + t_size[0], c1[1] - t_size[1] - 3
cv2.rectangle(img, c1, c2, color, -1, cv2.LINE_AA) # filled
cv2.putText(img, label, (c1[0], c1[1] - 2), 0, tl / 3, [225, 255, 255], thickness=tf, lineType=cv2.LINE_AA)
def main(args):
print(args)
device = torch.device(args.device)
#------------------------------------导入网络
#下面的criterion是算损失函数要用的,推理用不到,postprocessors是解码用的,这里也没有用,用的是自己的。
model, criterion, postprocessors = build_model(args)
#------------------------------------加载权重
checkpoint = torch.load(args.resume, map_location='cuda')
model.load_state_dict(checkpoint['model'])
#------------------------------------把权重加载到gpu或cpu上
model.to(device)
#------------------------------------打印出网络的参数大小
n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad)
print("parameters:",n_parameters)
#------------------------------------设置好存储输出结果的文件夹
output_dir = Path(args.output_dir)
#-----------------------------------读取数据集,进行推理
image_Totensor=torchvision.transforms.ToTensor()
image_file_path = os.listdir("inference_demo/detect_demo")
image_set = []
for image_item in image_file_path:
print("inference_image:",image_item)
image_path = os.path.join("inference_demo/detect_demo",image_item)
image = Image.open(image_path)
image_tensor = image_Totensor(image)
image_tensor = torch.reshape(image_tensor,[-1,image_tensor.shape[0],image_tensor.shape[1],image_tensor.shape[2]])
image_tensor=image_tensor.to(device)
time1 = time.time()
inference_result = model(image_tensor)
time2 = time.time()
print("inference_time:",time2-time1)
probas = inference_result['pred_logits'].softmax(-1)[0, :, :-1].cpu()
bboxes_scaled = rescale_bboxes(inference_result['pred_boxes'][0, ].cpu(),(image_tensor.shape[3],image_tensor.shape[2]))
scores, boxes = filter_boxes(probas,bboxes_scaled)
scores = scores.data.numpy()
boxes = boxes.data.numpy()
for i in range(boxes.shape[0]):
class_id = scores[i].argmax()
label = CLASSES[class_id]
confidence = scores[i].max()
text = f"{label} {confidence:.3f}"
image = np.array(image)
plot_one_box(boxes[i],image,label=text)
cv2.imshow("images",image)
cv2.waitKey(1)
image=Image.fromarray(image)
image.save(os.path.join(args.output_dir,image_item))
if __name__ == '__main__':
parser = argparse.ArgumentParser('DETR training and evaluation script', parents=[get_args_parser()])
args = parser.parse_args()
if args.output_dir:
Path(args.output_dir).mkdir(parents=True, exist_ok=True)
main(args)
到此就将复现工作完全结束了,感谢博主w1520039381_mmdetection,colab,detr-CSDN博客和暮已深_数据结构,数字图像处理,其他学习笔记-CSDN博客。
参考博客地址:【DETR】训练自己的数据集-实践笔记;pytorch实现DETR的推理程序;windows10复现DEtection TRansformers(DETR)并实现自己的数据集。