借用Ultralytics Yolo快速训练一个物体检测器
借用Ultralytics Yolo快速训练一个物体检测器
同步发表于 https://www.codebonobo.tech/post/14
大约在16/17年, 深度学习刚刚流行时, Object Detection 还是相当高端的技术, 各大高校还很流行水Fast RCNN / Faster RCNN之类的论文, 干着安全帽/行人/车辆检测之类的横项. 但到了2024年, 随着技术成熟, 物体检测几乎已经是个死方向了, 现在的学校应该忙着把别人好不容易训练的通用大模型退化成各领域的专用大模型吧...哈哈
回到主题, 目前物体检测模型的训练已经流程化了, 不再需要费脑去写训练代码, 使用Ultralytics Yolo就可以快速训练一个特定物品的检测器, 比如安全帽检测.
https://github.com/ultralytics/ultralytics
Step-1 准备数据集
你需要一些待检测物体比如安全帽, 把它从各个角度拍摄一下. 再找一些不相关的背景图片. 然后把安全帽给放大缩小旋转等等贴到背景图片上去, 生成一堆训练数据.
配置文件:
extract_cfg:
output_dir: '/datasets/images'
fps: 0.25
screen_images_path: '/datasets/待检测图片'
max_scale: 1.0
min_scale: 0.1
manual_scale: [ {name: 'logo', min_scale: 0.05, max_scale: 0.3},
{name: 'logo', min_scale: 0.1, max_scale: 0.5},
{name: '箭头', min_scale: 0.1, max_scale: 0.5}
]
data_cfgs: [ {id: 0, name: 'logo', min_scale: 0.05, max_scale: 0.3, gen_num: 2},
{id: 1, name: '截屏', min_scale: 0.1, max_scale: 1.0, gen_num: 3, need_full_screen: true},
{id: 2, name: '红包', min_scale: 0.1, max_scale: 0.5, gen_num: 2},
{id: 3, name: '箭头', min_scale: 0.1, max_scale: 0.5, gen_num: 2, rotate_aug: true},
]
save_oss_dir: /datasets/gen_datasets/
gen_num_per_image: 2
max_bg_img_sample:
数据集生成:
from pathlib import Path
import io
import random
import cv2
import numpy as np
from PIL import Image
import hydra
from omegaconf import DictConfig
import json
from tqdm import tqdm
# 加载图片
def load_images(background_path, overlay_path):
background = cv2.imread(background_path)
overlay = cv2.imread(overlay_path, cv2.IMREAD_UNCHANGED)
return background, overlay
# 随机缩放和位置
def random_scale_and_position(bg_shape, overlay_shape, max_scale=1.0, min_scale=0.1):
max_height, max_width = bg_shape[:2]
overlay_height, overlay_width = overlay_shape[:2]
base_scale = min(max_height / overlay_height, max_width / overlay_width)
# 随机缩放
scale_factor = random.uniform(
min_scale * base_scale, max_scale * base_scale)
new_height, new_width = int(
overlay_height * scale_factor), int(overlay_width * scale_factor)
# 随机位置
max_x = max_width - new_width - 1
max_y = max_height - new_height - 1
position_x = random.randint(0, max_x)
position_y = random.randint(0, max_y)
return scale_factor, (position_x, position_y)
def get_resized_overlay(overlay, scale):
overlay_resized = cv2.resize(overlay, (0, 0), fx=scale, fy=scale)
return overlay_resized
def rotate_image(img, angle):
if isinstance(img, np.ndarray):
img = Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGRA2RGBA))
# 确保图像具有alpha通道（透明度）
img = img.convert("RGBA")
# 旋转原始图像并粘贴到新的透明图像框架中
rotated_img = img.rotate(angle, resample=Image.BICUBIC, expand=True)
rotated_img = np.asarray(rotated_img)
return cv2.cvtColor(rotated_img, cv2.COLOR_RGBA2BGRA)
# 合成图片
def overlay_image(background, overlay_resized, position, scale):
h, w = overlay_resized.shape[:2]
x, y = position
# 透明度处理
alpha_s = overlay_resized[:, :, 3] / 255.0
alpha_l = 1.0 - alpha_s
for c in range(0, 3):
background[y:y + h, x:x + w, c] = (alpha_s * overlay_resized[:, :, c] +
alpha_l * background[y:y + h, x:x + w, c])
# 画出位置，调试使用
# print("position", x, y, w, h)
# cv2.rectangle(background, (x, y), (x + w, y + h), (0, 255, 0), 2)
background = cv2.cvtColor(background, cv2.COLOR_BGR2RGB)
return Image.fromarray(background)
class Box:
def __init__(self, x, y, width, height, category_id, image_width, image_height):
self.x = x
self.y = y
self.width = width
self.height = height
self.image_width = image_width
self.image_height = image_height
self.category_id = category_id
def to_yolo_format(self):
x_center = (self.x + self.width / 2) / self.image_width
y_center = (self.y + self.height / 2) / self.image_height
width = self.width / self.image_width
height = self.height / self.image_height
box_line = f"{self.category_id} {x_center} {y_center} {width} {height}"
return box_line
class SingleCategoryGen:
def __init__(self, cfg, data_cfg, output_dir):
self.output_dir = output_dir
self.screen_png_images = []
self.coco_images = []
self.coco_annotations = []
screen_images_path = Path(
cfg.screen_images_path.format(user_root=user_root))
self.manual_scale = {}
self.data_cfg = data_cfg
self.category_id = data_cfg.id
self.category_name = self.data_cfg.name
self.max_scale = self.data_cfg.max_scale
self.min_scale = self.data_cfg.min_scale
self.gen_num = self.data_cfg.gen_num
self.rotate_aug = self.data_cfg.get("rotate_aug", False)
self.need_full_screen = self.data_cfg.get("need_full_screen", False)
self.category_num = 0
self.category_names = {}
self.butcket = get_oss_bucket(cfg.bucket_name)
output_dir = Path(output_dir)
save_oss_dir = f"{cfg.save_oss_dir}/{output_dir.parent.name}/{output_dir.name}"
self.save_oss_dir = save_oss_dir
self.images_save_oss_dir = f"{save_oss_dir}/images"
self.label_save_oss_dir = f"{save_oss_dir}/labels"
self.annotations_save_oss_path = f"{save_oss_dir}/annotations.json"
self.load_screen_png_images_and_category(screen_images_path)
def load_screen_png_images_and_category(self, screen_images_dir):
screen_images_dir = Path(screen_images_dir)
category_id = self.category_id
screen_images_path = screen_images_dir / self.category_name
img_files = [p for p in screen_images_path.iterdir() if p.suffix in [
".png", ".jpg"]]
img_files.sort(key=lambda x: x.stem)
for i, img_file in enumerate(img_files):
self.screen_png_images.append(
dict(id=i, name=img_file.stem, supercategory=None, path=str(img_file)))
def add_new_images(self, bg_img_path: Path, gen_image_num=None, subset="train"):
gen_image_num = gen_image_num or self.gen_num
background_origin = cv2.imread(str(bg_img_path))
if background_origin is None:
print(f"open image {bg_img_path} failed")
return
max_box_num = 1
for gen_id in range(gen_image_num):
background = background_origin.copy()
category_id = self.category_id
overlay_img_path = self.sample_category_data()
overlay = cv2.imread(overlay_img_path, cv2.IMREAD_UNCHANGED)
if overlay.shape[2] == 3:
overlay = cv2.cvtColor(overlay, cv2.COLOR_BGR2BGRA)
if self.rotate_aug:
overlay = rotate_image(overlay, random.uniform(-180, 180))
# # 随机裁剪图片
# if random.random() < 0.5:
# origin_height = overlay.shape[0]
# min_height = origin_height // 4
# new_height = random.randint(min_height, origin_height)
# new_top = random.randint(0, origin_height - new_height)
# overlay = overlay[new_top:new_top+new_height, :, :]
box_num = random.randint(1, max_box_num)
# 获取随机缩放和位置
max_scale = self.max_scale
min_scale = self.min_scale
scale, position = random_scale_and_position(
background.shape, overlay.shape, max_scale, min_scale)
# 缩放overlay图片
overlay_resized = get_resized_overlay(overlay, scale)
# 合成后的图片
merged_img = overlay_image(background, overlay_resized, position, scale)
# 保存合成后的图片
filename = f"{bg_img_path.stem}_{category_id}_{gen_id:02d}.png"
merged_img.save(f'{output_dir}/{filename}')
# 生成COCO格式的标注数据
box = Box(*position, overlay_resized.shape[1], overlay_resized.shape[0], category_id, background.shape[1],
background.shape[0])
self.upload_image_to_oss(merged_img, filename, subset, [box])
def sample_category_data(self):
return random.choice(self.screen_png_images)["path"]
image_id = self.gen_image_id()
image_json = {
"id": image_id,
"width": image.width,
"height": image.height,
"file_name": image_name,
}
self.coco_images.append(image_json)
annotation_json = {
"id": image_id,
"image_id": image_id,
"category_id": 0,
"segmentation": None,
"area": bbox[2] * bbox[3],
"bbox": bbox,
"iscrowd": 0
}
self.coco_annotations.append(annotation_json)
def upload_image_to_oss(self, image, image_name, subset, box_list=None):
image_bytesio = io.BytesIO()
image.save(image_bytesio, format="PNG")
self.butcket.put_object(
f"{self.images_save_oss_dir}/{subset}/{image_name}", image_bytesio.getvalue())
if box_list:
label_str = "\n".join([box.to_yolo_format() for box in box_list])
label_name = image_name.split(".")[0] + ".txt"
self.butcket.put_object(
f"{self.label_save_oss_dir}/{subset}/{label_name}", label_str)
def upload_full_screen_image(self):
if not self.need_full_screen:
return
name = self.category_name
category_id = self.category_id
image_list = self.screen_png_images
subset_list = ["train" if i % 10 <= 7 else "val" if i %
10 <= 8 else "test" for i in range(len(image_list))]
for i in range(len(image_list)):
image_data = image_list[i]
subset = subset_list[i]
overlay_img_path = image_data["path"]
image = Image.open(overlay_img_path)
if random.random() < 0.5:
origin_height = image.height
min_height = origin_height // 4
new_height = random.randint(min_height, origin_height)
new_top = random.randint(0, origin_height - new_height)
image = image.crop(
(0, new_top, image.width, new_top + new_height))
filename = f"{name}_{category_id}_{i:05}.png"
box = Box(0, 0, image.width, image.height,
category_id, image.width, image.height)
self.upload_image_to_oss(image, filename, subset, [box])
class ScreenDatasetGen:
def __init__(self, cfg, output_dir):
self.output_dir = output_dir
self.screen_png_images = {}
self.coco_images = []
self.coco_annotations = []
screen_images_path = Path(
cfg.screen_images_path.format(user_root=user_root))
self.max_scale = cfg.max_scale
self.min_scale = cfg.min_scale
self.manual_scale = {}
for info in cfg.manual_scale:
self.manual_scale[info.name] = dict(
max_scale=info.max_scale, min_scale=info.min_scale)
self.category_num = 0
self.category_names = {}
self.category_id_loop = -1
self.butcket = get_oss_bucket(cfg.bucket_name)
output_dir = Path(output_dir)
save_oss_dir = f"{cfg.save_oss_dir}/{output_dir.parent.name}/{output_dir.name}"
self.save_oss_dir = save_oss_dir
self.images_save_oss_dir = f"{save_oss_dir}/images"
self.label_save_oss_dir = f"{save_oss_dir}/labels"
self.annotations_save_oss_path = f"{save_oss_dir}/annotations.json"
self.load_screen_png_images_and_category(screen_images_path)
def add_new_images(self, bg_img_path: Path, gen_image_num=1, subset="train"):
background_origin = cv2.imread(str(bg_img_path))
if background_origin is None:
print(f"open image {bg_img_path} failed")
return
max_box_num = 1
for gen_id in range(gen_image_num):
background = background_origin.copy()
category_id = self.get_category_id_loop()
overlay_img_path = self.sample_category_data(
category_id, subset=subset)
overlay = cv2.imread(overlay_img_path, cv2.IMREAD_UNCHANGED)
if overlay.shape[2] == 3:
overlay = cv2.cvtColor(overlay, cv2.COLOR_BGR2BGRA)
# # 随机裁剪图片
# if random.random() < 0.5:
# origin_height = overlay.shape[0]
# min_height = origin_height // 4
# new_height = random.randint(min_height, origin_height)
# new_top = random.randint(0, origin_height - new_height)
# overlay = overlay[new_top:new_top+new_height, :, :]
box_num = random.randint(1, max_box_num)
# 获取随机缩放和位置
category_name = self.category_names[category_id]
if category_name in self.manual_scale:
max_scale = self.manual_scale[category_name]["max_scale"]
min_scale = self.manual_scale[category_name]["min_scale"]
else:
max_scale = self.max_scale
min_scale = self.min_scale
scale, position = random_scale_and_position(
background.shape, overlay.shape, max_scale, min_scale)
# 缩放overlay图片
overlay_resized = get_resized_overlay(overlay, scale)
# 合成后的图片
merged_img = overlay_image(
background, overlay_resized, position, scale)
# 保存合成后的图片
filename = f"{bg_img_path.stem}_{category_id}_{gen_id:02d}.png"
# merged_img.save(f'{output_dir}/{filename}')
# 生成COCO格式的标注数据
box = Box(*position, overlay_resized.shape[1], overlay_resized.shape[0], category_id, background.shape[1],
background.shape[0])
self.upload_image_to_oss(merged_img, filename, subset, [box])
# self.add_image_annotion_to_coco(box, merged_img, filename)
def upload_full_screen_image(self, category_name=None):
if category_name is None:
return
if not isinstance(category_name, list):
category_name = [category_name]
for category_id in range(self.category_num):
name = self.category_names[category_id]
if name not in category_name:
continue
image_list = self.screen_png_images[category_id]
subset_list = ["train" if i % 10 <= 7 else "val" if i %
10 <= 8 else "test" for i in range(len(image_list))]
for i in range(len(image_list)):
image_data = image_list[i]
subset = subset_list[i]
overlay_img_path = image_data["path"]
image = Image.open(overlay_img_path)
if random.random() < 0.5:
origin_height = image.height
min_height = origin_height // 4
new_height = random.randint(min_height, origin_height)
new_top = random.randint(0, origin_height - new_height)
image = image.crop(
(0, new_top, image.width, new_top + new_height))
filename = f"{name}_{category_id}_{i:05}.png"
box = Box(0, 0, image.width, image.height,
category_id, image.width, image.height)
self.upload_image_to_oss(image, filename, subset, [box])
def load_screen_png_images_and_category(self, screen_images_dir):
screen_images_dir = Path(screen_images_dir)
screen_images_paths = [
f for f in screen_images_dir.iterdir() if f.is_dir()]
screen_images_paths.sort(key=lambda x: x.stem)
for category_id, screen_images_path in enumerate(screen_images_paths):
img_files = [p for p in screen_images_path.iterdir() if p.suffix in [
".png", ".jpg"]]
img_files.sort(key=lambda x: x.stem)
self.screen_png_images[category_id] = []
self.category_names[category_id] = screen_images_path.stem
print(f"{category_id}: {self.category_names[category_id]}")
for i, img_file in enumerate(img_files):
self.screen_png_images[category_id].append(
dict(id=i, name=img_file.stem, supercategory=None, path=str(img_file)))
self.category_num = len(screen_images_paths)
print(f"category_num: {self.category_num}")
def get_category_id_loop(self):
# self.category_id_loop = (self.category_id_loop + 1) % self.category_num
self.category_id_loop = random.randint(0, self.category_num - 1)
return self.category_id_loop
def sample_category_data(self, category_id, subset):
image_data = self.screen_png_images[category_id]
# valid_id = []
# if subset == "train":
# valid_id = [i for i in range(len(image_data)) if i % 10 <= 7]
# elif subset == "val":
# valid_id = [i for i in range(len(image_data)) if i % 10 == 8]
# elif subset == "test":
# valid_id = [i for i in range(len(image_data)) if i % 10 == 9]
# image_data = [image_data[i] for i in valid_id]
return random.choice(image_data)["path"]
def gen_image_id(self):
return len(self.coco_images) + 1
def add_image_annotion_to_coco(self, bbox, image: Image.Image, image_name):
image_id = self.gen_image_id()
image_json = {
"id": image_id,
"width": image.width,
"height": image.height,
"file_name": image_name,
}
self.coco_images.append(image_json)
annotation_json = {
"id": image_id,
"image_id": image_id,
"category_id": 0,
"segmentation": None,
"area": bbox[2] * bbox[3],
"bbox": bbox,
"iscrowd": 0
}
self.coco_annotations.append(annotation_json)
def upload_image_to_oss(self, image, image_name, subset, box_list=None):
image_bytesio = io.BytesIO()
image.save(image_bytesio, format="PNG")
self.butcket.put_object(
f"{self.images_save_oss_dir}/{subset}/{image_name}", image_bytesio.getvalue())
if box_list:
label_str = "\n".join([box.to_yolo_format() for box in box_list])
label_name = image_name.split(".")[0] + ".txt"
self.butcket.put_object(
f"{self.label_save_oss_dir}/{subset}/{label_name}", label_str)
def dump_coco_json(self):
categories = [{key: item[key] for key in ("id", "name", "supercategory")} for item in
self.screen_png_images.values()]
coco_json = {
"images": self.coco_images,
"annotations": self.coco_annotations,
"categories": categories
}
self.butcket.put_object(
self.annotations_save_oss_path, json.dumps(coco_json, indent=2))
# with open(f"{self.output_dir}/coco.json", "w") as fp:
# json.dump(coco_json, fp, indent=2)
@hydra.main(version_base=None, config_path=".", config_name="conf")
def main(cfg: DictConfig):
output_dir = hydra.core.hydra_config.HydraConfig.get().runtime.output_dir
# get_image_and_annotation(output_dir)
# screen_dataset_gen = ScreenDatasetGen(cfg, output_dir)
category_generators = []
for data_cfg in cfg.data_cfgs:
category_generators.append(SingleCategoryGen(cfg, data_cfg, output_dir))
bg_img_files = [f for f in Path(cfg.extract_cfg.output_dir.format(user_root=user_root)).iterdir() if
f.suffix in [".png", ".jpg"]]
if cfg.get("max_bg_img_sample"):
bg_img_files = random.sample(bg_img_files, cfg.max_bg_img_sample)
img_index = 0
for bg_img_file in tqdm(bg_img_files):
subset = "train" if img_index % 10 <= 7 else "val" if img_index % 10 == 8 else "test"
img_index += 1
for category_generator in category_generators:
category_generator.add_new_images(bg_img_path=bg_img_file, subset=subset)
for category_generator in category_generators:
category_generator.upload_full_screen_image()
if __name__ == '__main__':
main()
运行后, 可以在outputs文件夹下生成符合要求的训练数据.
image 就是背景+检测物体
labels 中的内容就是这样的文件:
1 0.6701388888888888 0.289453125 0.5736111111111111 0.57421875
# 类型 box
Step-2 训练模型
这个更简单, 在官网下载一个模型权重, 比如yolo8s.pt, 对付安全帽这种东西, 几M大的模型就够了.
训练配置文件:
names:
0: logo
1: 截屏
2: 红包
path: /outputs
test: images/test
train: images/train
val: images/val
训练代码:
没错就这么一点
from ultralytics import YOLO
model = YOLO('./yolo8s.pt')
model.train(data='dataset.yaml', epochs=100, imgsz=1280)
然后就可以自动化训练了, 结束后会自动保存模型与评估检测效果.
Step-3 检测
检测代码示意:
class Special_Obj_Detect(object):
def __init__(self, cfg) -> None:
model_path = cfg.model_path
self.model = YOLO(model_path)
self.model.requires_grad_ = False
self.cls_names = {0: 'logo', 1: '截屏', 2: '红包'}
# 单帧图像检测
def detect_image(self, img_path):
results = self.model(img_path)
objects = []
objects_cnt = dict()
objects_area_pct = dict()
for result in results:
result = result.cpu()
boxes = list(result.boxes)
for box in boxes:
if box.conf < 0.8: continue
boxcls = box.cls[0].item()
objects.append(self.cls_names[boxcls])
objects_cnt[self.cls_names[boxcls]] = objects_cnt.get(self.cls_names[boxcls], 0) + 1
area_p = sum([ (xywh[2]*xywh[3]).item() for xywh in box.xywhn])
area_p = min(1, area_p)
objects_area_pct[self.cls_names[boxcls]] = area_p
objects = list(set(objects))
return objects, objects_cnt, objects_area_pct
收工.
借用Ultralytics Yolo快速训练一个物体检测器的更多相关文章
- 教你用webgl快速创建一个小世界
收录待用,修改转载已取得腾讯云授权 作者:TAT.vorshen Webgl的魅力在于可以创造一个自己的3D世界,但相比较canvas2D来说,除了物体的移动旋转变换完全依赖矩阵增加了复杂度,就连生成 ...
- dlib下训练自己的物体检测器--手的检测
之前我们在Linux上安装了dlib(http://www.cnblogs.com/take-fetter/p/8318602.html),也成功的完成了之前的人脸检测程序, 今天我们来一起学习怎样使 ...
- 如何正确训练一个 SVM + HOG 行人检测器
这几个月一直在忙着做大论文,一个基于 SVM 的新的目标检测算法.为了做性能对比,我必须训练一个经典的 Dalal05 提出的行人检测器,我原以为这个任务很简单,但是我错了. 为了训练出一个性能达标的 ...
- 【PyTorch深度学习60分钟快速入门 】Part4：训练一个分类器
太棒啦!到目前为止,你已经了解了如何定义神经网络.计算损失,以及更新网络权重.不过,现在你可能会思考以下几个方面: 0x01 数据集 通常,当你需要处理图像.文本.音频或视频数据时,你可以使用标准 ...
- yolo回归型的物体检测
本弱又搬了另外一个博客的讲解: 缩进YOLO全称You Only Look Once: Unified, Real-Time Object Detection,是在CVPR2016提出的一种目标检测算 ...
- YOLO2 (3) 快速训练自己的目标
1快速训练自己的目标 在 YOLO2 (2) 测试自己的数据 中记录了完整的训练自己数据的过程. 训练时目标只有一类 car. 如果已经执行过第一次训练,改过一次配置文件,之后仍然训练同样的目标还是只 ...
- bert，albert的快速训练和预测
随着预训练模型越来越成熟,预训练模型也会更多的在业务中使用,本文提供了bert和albert的快速训练和部署,实际上目前的预训练模型在用起来时都大致相同. 基于不久前发布的中文数据集chineseGL ...
- jquery+flask+keras+nsfw快速搭建一个简易鉴黄工具
1. demo 地址:http://www.huchengchun.com:8127/porn_classification 接口说明: 1. http://www.huchengchun.com:8 ...
- NodeJS 最快速搭建一个HttpServer
最快速搭建一个HttpServer 在目录里放一个index.html cd D:\Web\InternalWeb start http-server -i -p 8081
- vuejsLearn---通过手脚架快速搭建一个vuejs项目
开始快速搭建一个项目 通过Webpack + vue-loader 手脚架 https://github.com/vuejs-templates/webpack 按照它的步骤一步一步来 $ npm i ...
随机推荐
- [学习笔记] 丢番图方程 & 同余 & 逆元 - 数论
首先,他们几个有着极大的相似性.另外,他们的各自的思想都能够很好的服务于另外几个,有助于加深理解. 文中有些letax公式抽风了,全屏之后应该能看得见-- 线性丢番图方程 丢番图不是个图啊!他是个ma ...
- loj6669 Nauuo and Binary Tree 题解
https://loj.ac/p/6669 赛时做法 先 \(n-1\) 次问出深度 逐层考虑.slv(vector<int> a,vector<int> b) 表示在点集 \ ...
- manim 中的三维视角
今天研究了一下 manim 中的 ThreeDAxes 和 set_camera_orientation 这里记录一下视角. manim 中的 3D 坐标系是笛卡尔的三维坐标系,属于右手坐标系,即右手 ...
- shell 删除文件内容Mac、Linux兼容方法
# 定义sedi数组 # Linux sed后面, 用 "-i" sedi=(-i) case "$(uname)" in Darwin*) # Mac sed ...
- 细说WebService
细说WebService 简介 WebService 是一个应用于客户端.服务端,基于http协议的web应用程序,他有一个非常重要的特点,那就是可以跨语言.跨平台进行远程调用,可应用于分布式系统中的 ...
- 初三年后集训测试 T2--牛吃草
初三年后集训测试 $T 2 $ 牛吃草 一言难尽 $$HZOI$$ $ Description $ 由于现代化进程的加快,农场的养殖业也趋向机械化. \(QZS\) 决定购置若干台自动喂草机来减少自己 ...
- idea关闭窗口快捷键
File->settings->keymap->main menu->window->editor tabs->close 根据自己的使用习惯将想要关闭的标签设置快 ...
- 图解Zabbix设置邮件报警
Zabbix设置邮件告警 前提条件: Zabbix Server 和 Zabbix Agent都已安装完毕,并已启动 1.添加主机 2.配置邮件告警,这里以VSFTP服务为例 yum in ...
- pyspark 数据类型及转换
spark 有哪些数据类型 https://spark.apache.org/docs/latest/sql-reference.html Spark 数据类型 Data Types Spark SQ ...
- RuleLinKClient - 再也不担心表达引擎宕机了
原来有这么多时间 六月的那么一天,天气比以往时候都更凉爽,媳妇边收拾桌子,边漫不经心的对我说:你最近好像都没怎么阅读了. 正刷着新闻我,如同被一记响亮的晴空霹雳击中一般,不知所措.是了,最近几月诸事凑 ...