p2p train install
conda create -n p2pnet python=3.7 -y
conda activate p2pnet
conda install pytorch torchvision torchaudio pytorch-cuda=12.4 -c pytorch -c nvidia -y
pip install -r requirements.txt结果我的python解释器不支持3.8以下的
conda create -n p2pnet38 python=3.8 -y
conda activate p2pnet38
conda install pytorch torchvision torchaudio pytorch-cuda=12.4 -c pytorch -c nvidia -y
pip install -r requirements.txtDataset creation needs to be combined with the README.md.
我是在xml文件中直接做成点的信息的，直接转换成 x y format的文件
import xml.etree.ElementTree as ET
import os
# Convert an XML file of a document into a format with x and y coordinates.
def convert_to_center_points(xml_file_path, txt_file_path):
xml_file_path.endswith('.xml')
tree = ET.parse(xml_file_path)
root = tree.getroot()
with open(txt_file_path, 'w') as txt_file:
for obj in root.findall('object'):
xmin = int(obj.find('bndbox').find('xmin').text)
ymin = int(obj.find('bndbox').find('ymin').text)
xmax = int(obj.find('bndbox').find('xmax').text)
ymax = int(obj.find('bndbox').find('ymax').text)
x = (xmin + xmax) / 2
y = (ymin + ymax) / 2
txt_file.write(f"{x} {y}\n")
if __name__ == "__main__":
xml_file_floder = r'G:\CASIA_Windows\Code_Repository\yolox-pytorch-main\yolox-pytorch-main\VOCdevkit\VOC2007\Annotations'
txt_file_floder = r'G:\CASIA_Windows\Code_Repository\yolox-pytorch-main\yolox-pytorch-main\VOCdevkit\VOC2007\txt'
xml_file_list = os.listdir(xml_file_floder)
sample_num = len(xml_file_list)
print("len(xml_file_list) : ",sample_num)
for xml_file in xml_file_list:
xml_file_path = os.path.join(xml_file_floder,xml_file)
xml_name=os.path.basename(xml_file_path).split(".")[0]
# print("xml_name : ",xml_name)
txt_file_path = os.path.join(txt_file_floder,xml_name+".txt")
#print("txt_file_path : ",txt_file_path)
convert_to_center_points(xml_file_path, txt_file_path)检测一下转化的有没有问题 可视化
from PIL import Image, ImageDraw
import os
# Visualize the point annotation information to ensure the processed data is correct.
def draw_points_on_image(txt_file_path, input_image_path, output_folder_path):
with open(txt_file_path, 'r') as file:
lines = file.readlines()
points = [tuple(map(float, line.strip().split(' '))) for line in lines]
try:
image = Image.open(input_image_path)
except FileNotFoundError:
print(f"Image file {input_image_path} not found. Skipping.")
for point in points:
x, y = point
draw = ImageDraw.Draw(image)
draw.ellipse((x-2, y-2, x+2, y+2), fill='red', outline='red')
image.save(output_folder_path)
print(f"Point ({x} {y}) drawn on {input_image_path} and saved to {output_folder_path}")
if __name__ == "__main__":
txt_file = r'G:\CASIA_Windows\Code_Repository\yolox-pytorch-main\yolox-pytorch-main\VOCdevkit\VOC2007\txt'
image_folder = r'G:\CASIA_Windows\Code_Repository\yolox-pytorch-main\yolox-pytorch-main\VOCdevkit\VOC2007\JPEGImages'
output_folder = r'G:\CASIA_Windows\Code_Repository\yolox-pytorch-main\yolox-pytorch-main\VOCdevkit\VOC2007\output_images'
if not os.path.exists(output_folder):
os.makedirs(output_folder)
txt_file_list = os.listdir(txt_file)
txt_num = len(txt_file)
for i in txt_file_list:
name = i.split(".")[0]
txt_file_path = os.path.join(txt_file,i)
image_folder_path = os.path.join(image_folder,name+".jpg")
output_folder_path=os.path.join(output_folder,name+".jpg")
draw_points_on_image(txt_file_path, image_folder_path, output_folder_path)将已将处理好的 image 和 txt文件移动到对应的位置 分别为train test文件夹
我事先生成了对应比例的train.txt val.test test.txt 所以我直接按照这个比例的名称直接放在对应的文件夹了
import os
import shutil
# Move the processed files to the corresponding folders.
def move_images(source_folder, destination_folder, file_list_path):
if not os.path.exists(destination_folder):
os.makedirs(destination_folder)
with open(file_list_path, 'r') as file:
for line in file:
image_name = line.strip()
image_path = os.path.join(source_folder, image_name+".txt")
if os.path.isfile(image_path):
shutil.move(image_path, destination_folder)
print(f"Moved: {image_name}")
else:
print(f"File not found: {image_name}")
if __name__ == "__main__":
source_folder = r'G:\CASIA_Windows\Code_Repository\yolox-pytorch-main\yolox-pytorch-main\VOCdevkit\VOC2007\txt'
destination_folder = r"crowd_datasets\SHHA\train"
file_list_path = r'crowd_datasets\SHHA\train_val_test\train.txt' # floder train <--- train.txt | floder test <--- val.txt
move_images(source_folder, destination_folder, file_list_path)下面就是按照数据集训练对应的格式生成对应的。list文件
def generate_file_paths(input_txt_path, output_txt_path, img_folder, txt_folder):
"""
生成包含.jpg和.txt文件路径的列表，并将它们写入到一个新的文件中。
:param input_txt_path: 输入的train.txt文件路径
:param output_txt_path: 输出文件路径
:param img_folder: 存放.jpg文件的文件夹路径
:param txt_folder: 存放.txt文件的文件夹路径
"""
with open(input_txt_path, 'r') as input_file, open(output_txt_path, 'w') as output_file:
for line in input_file:
img_name = line.strip()
img_path = f"{img_folder}/{img_name}.jpg"
txt_path = f"{txt_folder}/{img_name}.txt"
output_file.write(f"{img_path} {txt_path}\n")
if __name__ == "__main__":
input_txt_path = r"crowd_datasets\SHHA\train_val_test\val.txt" # floder train <--- train.txt | floder test <--- val.txt
output_txt_path = r"crowd_datasets\SHHA\test.txt"
img_folder = r'crowd_datasets\SHHA\train'
txt_folder = r'crowd_datasets\SHHA\train'
generate_file_paths(input_txt_path, output_txt_path, img_folder, txt_folder)问题1： util\misc.py
cannot import name ‘_new_empty_tensor‘ from ‘torchvision.ops
Python问题: cannot import name ‘_new_empty_tensor‘ from ‘torchvision.ops
'''
if float(torchvision.__version__[:3]) < 0.7:
from torchvision.ops import _new_empty_tensor
from torchvision.ops.misc import _output_size
'''问题2：train.py
注释掉即可 assert args.masks这个参数没找到
if args.frozen_weights is not None:
pass
\# assert args.masks, "Frozen training is meant for segmentation only"
\# backup the arguments问题3：models\vgg_.py
修改要用到的路径
model_paths = {
'vgg16_bn': r"weights\vgg16_bn-6c64b313.pth" , # '/apdcephfs/private_changanwang/checkpoints/vgg16_bn-6c64b313.pth',
'vgg16': '/apdcephfs/private_changanwang/checkpoints/vgg16-397923af.pth',
}问题4：crowd_datasets\SHHA\SHHA.py
[Errno 2] No such file or directory: 'crowd_datasets\SHHA\shanghai_tech_part_a_train.list'
修改
self.train_lists = "shanghai_tech_part_a_train.list"
self.eval_list = "shanghai_tech_part_a_test.list"修改到自己正确的位置
比如我的是
self.train_lists = r"crowd_datasets\SHHA\train.list"
self.eval_list = r"crowd_datasets\SHHA\test.list"注意
parser.add_argument('--data_root', default=r'./',
help='path where the dataset is')crowd_datasets\SHHA这个是默认地址
问题5：train.py
'P2PNet' object has no attribute 'detr'
if args.frozen_weights is not None:
checkpoint = torch.load(args.frozen_weights, map_location='cpu')
model_without_ddp.detr.load_state_dict(checkpoint['model'])改为
model_without_ddp.load_state_dict(checkpoint['model'])问题6：训练过程中
engine.py
Caught error in DataLoader worker process 0.
迭代DataLoader时出现TypeError: Caught TypeError in DataLoader worker process 0.TypeError: 'NoneType' obj。
train，py
parser.add_argument('--num_workers', default=8, type=int)改为0
parser.add_argument('--num_workers', default=0, type=int)问题7：crowd_datasets\SHHA\SHHA.py
def load_data(img_gt_path, train):
img_path, gt_path = img_gt_path
# load the images
img = cv2.imread(img_path)
img = Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))OpenCV(4.10.0) D:\a\opencv-python\opencv-python\opencv\modules\imgproc\src\color.cpp:196: error: (-215:Assertion failed) !_src.empty() in function 'cv::cvtColor'注意train.list和test.list里面的路径转义字符
参数记录
Namespace(backbone='vgg16_bn', batch_size=32, checkpoints_dir='./ckpt', clip_max_norm=0.1, data_root='./', dataset_file='SHHA', eos_coef=0.5, epochs=3500, eval=False, eval_freq=5, frozen_weights='weights\\SHTechA.pth', gpu_id=0, line=2, lr=0.0001, lr_backbone=1e-05, lr_drop=3500, num_workers=0, output_dir='./log', point_loss_coef=0.0002, resume='', row=2, seed=42, set_cost_class=1, set_cost_point=0.05, start_epoch=0, tensorboard_dir='./runs', weight_decay=0.0001)
number of params: 21579344eval可能会有问题 先将频率改为1
运行run_test.py文件，但是这是对单个图片进行验证
如果是对文件进行检测 运行run_test_processfor_folder.py文件。
一定要注意,这种插值方法会出问题
# img_raw = img_raw.resize((new_width, new_height), Image.ANTIALIAS)
img_raw = img_raw.resize((new_width, new_height))要确保// 128 * 128，确定是整数，
new_width = width // 128 * 128
new_height = height // 128 * 128
不然的话后面经过处理，会导致维度不匹配
P4_x = self.P4_1(C4) # P4_x--->torch.Size([1, 256, 189, 240])
P4_x = P5_upsampled_x + P4_x # P5_upsampled_x--->torch.Size([1, 256, 188, 240])
[Errno 13] Permission denied: 'crowd_datasets\SHHA\val_real_test'
f = open(args.test_text, 'w', encoding='utf-8')
当你打开的是一个文件夹而不是一个文件时经常会遇到这样的情况，确保要打开的是一个文件（txt jpg等）
输出pre_gd_cnt.txt
result_anlysis_tools\calculate_mae.py 计算mae