前言：
项目需求，需要做一个手势识别的项目，目标对标旷视
数据集在这里,我自己已经购买了， 12W+张图片，12G+，代标注。
https://mp.weixin.qq.com/s/_VrxGzOiFJsnXgRbQ7t7hw
ref blog:https://blog.csdn.net/guyuealian/article/details/126750433
这个项目做好了，可以对标旷视的算法： https://console.faceplusplus.com.cn/documents/10065685
step1：初始化数据集
整理好的数据集，12W+ 19种手势， train：val=9:1
效果图例：
names:
0:one
1:two_up
2:two_up_inverted
3:three
4:three2
5:four
6:fist
7:palm
8:ok
9:peace
10:peace_inverted
11:like
12:dislike
13:stop
14:stop_inverted
15:call
16:mute
17:rock
18:no_gesture
数据集：
下载后的数据集比较乱，有一份coco的标注，
一份voc的标注，图片也有至少两份，一份是大图片，一份小图片：
大图片如上图，
小图片是这种：
是对大图片的裁剪，整理数据集整理了大半天才找到规律，主要是在验证。
xml标注和大图片能够契合。 就用这个组合，生成了一份yolo格式的标注。
原图片和标注我都没动，把生成好的，放到了另外的目录。
19个类别
打电话的大图，打电话的大图叫 xxxxx.jpg, 小图是 xxxx-0000.jpg， 看到这个我的心情瞬间就不好了，好几套标注，好几套图片。
验证图片和配对的关系代码。
image_path = r"D:\BaiduNetdiskDownload\Light-HaGRID\trainval\call\JPEGImages\000fa617-c7ba-47f9-9860-4971d78abc41.jpg"
xml_path = r"D:\BaiduNetdiskDownload\Light-HaGRID\trainval\call\Annotations\000fa617-c7ba-47f9-9860-4971d78abc41.xml"
in_file = open(xml_path)
basename = os.path.basename(image_path)
basename_no_ext = os.path.splitext(basename)[0]
# in_file,basename,basename_no_ext
out_file = open(basename_no_ext + '.txt', 'w')
tree = ET.parse(in_file)
root = tree.getroot()
size = root.find('size')
w = int(size.find('width').text)
h = int(size.find('height').text)
for obj in root.iter('object'):
difficult = obj.find('difficult').text
cls = obj.find('name').text
print(cls)
if cls not in classes or int(difficult)==1:
continue
cls_id = classes[cls]
xmlbox = obj.find('bndbox')
b = (float(xmlbox.find('xmin').text), float(xmlbox.find('xmax').text), float(xmlbox.find('ymin').text), float(xmlbox.find('ymax').text))
print(b)
x_top_left, y_top_left, x_bottom_right, y_bottom_right = int(b[0]), int(b[2]),int(b[1]),int(b[3])
# bb = convert((w,h), b)
# print(bb)
# out_file.write(str(cls_id) + " " + " ".join([str(a) for a in bb]) + '\n')
img = cv2.imread(image_path)
cv2.rectangle(img, (x_top_left, y_top_left), (x_bottom_right, y_bottom_right), (0, 0, 255), thickness=2)
cv2.imshow("image", img)
cv2.waitKey(0)
cv2.destroyAllWindows()
然后我有抽样，看看我的猜测对不对
def show_pic(xml_path,image_path):
in_file = open(xml_path)
basename = os.path.basename(image_path)
basename_no_ext = os.path.splitext(basename)[0]
# in_file,basename,basename_no_ext
out_file = open(basename_no_ext + '.txt', 'w')
tree = ET.parse(in_file)
root = tree.getroot()
size = root.find('size')
w = int(size.find('width').text)
h = int(size.find('height').text)
img = cv2.imread(image_path)
for obj in root.iter('object'):
difficult = obj.find('difficult').text
cls = obj.find('name').text
# print(cls)
if cls not in classes or int(difficult)==1:
continue
cls_id = classes[cls]
xmlbox = obj.find('bndbox')
b = (float(xmlbox.find('xmin').text), float(xmlbox.find('xmax').text), float(xmlbox.find('ymin').text), float(xmlbox.find('ymax').text))
# print(b)
x_top_left, y_top_left, x_bottom_right, y_bottom_right = int(b[0]), int(b[2]),int(b[1]),int(b[3])
# bb = convert((w,h), b)
# print(bb)
# out_file.write(str(cls_id) + " " + " ".join([str(a) for a in bb]) + '\n')
cv2.rectangle(img, (x_top_left, y_top_left), (x_bottom_right, y_bottom_right), (0, 0, 255), thickness=2)
cv2.imshow("image", img)
cv2.waitKey(0)
cv2.destroyAllWindows()
gen_img_path = glob.glob(r"D:\BaiduNetdiskDownload\Light-HaGRID\trainval\*\JPEGImages\*.jpg")
# 验证图片是否好使
def check_img():
count_num = 0
for img_path in gen_img_path:
if count_num%50000 == 0:
xml_path = img_path.replace("JPEGImages","Annotations").replace("jpg","xml")
print(xml_path,img_path)
show_pic(xml_path,img_path)
count_num += 1
check_img()
classes = { 'one': 0, 'two_up': 1, 'two_up_inverted': 2, 'three': 3, 'three2': 4,
'four': 5, 'fist': 6, 'palm': 7, 'ok': 8, 'peace': 9,
'peace_inverted': 10, 'like': 11, 'dislike': 12, 'stop': 13, 'stop_inverted': 14,
'call': 15, 'mute': 16, 'rock': 17, 'no_gesture': 18 }
证明配对关系没错。
这里面我是
···
if count_num%50000 == 0:
xml_path = img_path.replace(“JPEGImages”,“Annotations”).replace(“jpg”,“xml”)
print(xml_path,img_path)
show_pic(xml_path,img_path)
···
首次抽样的时候给的是count_num%10000或者1000张，看着标注都对，就放心了。
然后就是生成yolov 格式的标注文件，
# 构建图片和标注的位置
base_dist_dir = r"F:\\datasets\\20230603hand_gesture"
img_train_dir = os.path.join(base_dist_dir,"images","train")
img_val_dir = os.path.join(base_dist_dir,"images","val")
lable_train_dir = os.path.join(base_dist_dir,"labels","train")
label_val_dir = os.path.join(base_dist_dir,"labels","val")
def make_dir(img_train_dir,img_val_dir,lable_train_dir,label_val_dir):
os.makedirs(img_train_dir,exist_ok=True)
os.makedirs(img_val_dir,exist_ok=True)
os.makedirs(lable_train_dir,exist_ok=True)
os.makedirs(label_val_dir,exist_ok=True)
make_dir(img_train_dir,img_val_dir,lable_train_dir,label_val_dir)
# 1.根据xml生成yolov的txt文件，并将图片和txt放置到指定位置
# 2.写check脚本，每隔5k张图片
def create_txt_label(xml_path,img_path,dst_img_dir,dst_label_dir):
in_file = open(xml_path)
basename = os.path.basename(img_path)
basename_no_ext = os.path.splitext(basename)[0]
dst_img_path = os.path.join(dst_img_dir,basename)
# 移动图片
shutil.copy(img_path,dst_img_path)
# 转化成txt
dst_txt_path = os.path.join(dst_label_dir,basename_no_ext+".txt")
tree = ET.parse(in_file)
root = tree.getroot()
size = root.find('size')
w = int(size.find('width').text)
h = int(size.find('height').text)
with open(dst_txt_path,"w") as f:
for obj in root.iter('object'):
difficult = obj.find('difficult').text
cls = obj.find('name').text
# print(cls)
if cls not in classes or int(difficult)==1:
continue
cls_id = classes[cls]
xmlbox = obj.find('bndbox')
b = (float(xmlbox.find('xmin').text), float(xmlbox.find('xmax').text), float(xmlbox.find('ymin').text), float(xmlbox.find('ymax').text))
# print(b)
x_top_left, y_top_left, x_bottom_right, y_bottom_right = int(b[0]), int(b[2]),int(b[1]),int(b[3])
bb = convert((w,h), b)
## print(bb)
f.write(str(cls_id) + " " + " ".join([str(a) for a in bb]) + '\n')
in_file.close()
return dst_txt_path,dst_img_path
def show_img_with_label(dst_txt_path,dst_img_path):
cv2_img = cv2.imread(dst_img_path)
height, width, _ = cv2_img.shape
with open(dst_txt_path) as f:
line_list = f.readlines()
print(line_list)
for line_str in line_list:
info_list = line_str.strip().split(" ")
print(info_list)
idx = int(info_list[0])
keys = [k for k, v in classes.items() if v == idx]
print("idx",idx, keys)
assert len(keys)==1, print(dst_txt_path,dst_img_path)
label_name = keys[0]
x, y, w, h = map(float, info_list[1:]) # 目标中心点坐标和宽高比例
print("---------------",x,y,w,h)
# 计算出左上角和右下角坐标
left = int((x - w / 2) * width)
top = int((y - h / 2) * height)
right = int((x + w / 2) * width)
bottom = int((y + h / 2) * height)
print("--===---",(left, top), (right, bottom))
# 绘制矩形框和类别标签
cv2.rectangle(cv2_img, (left, top), (right, bottom), (0, 255, 0), 2)
label = label_name
cv2.putText(cv2_img, label, (left, top), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow('image', cv2_img)
cv2.waitKey(0)
cv2.destroyAllWindows()
count_num = 0
for img_path in gen_img_path:
if count_num != 0:
pass
xml_path = img_path.replace("JPEGImages","Annotations").replace("jpg","xml")
label_name = xml_path.split('\\')[4]
if count_num%10000==0:
print(xml_path)
print(img_path)
print(label_name)
idx_class = classes[label_name]
if count_num%10 == 0:
dst_img_dir = img_val_dir
dst_label_dir = label_val_dir
else:
dst_img_dir = img_train_dir
dst_label_dir = lable_train_dir
dst_txt_path,dst_img_path = create_txt_label(xml_path,img_path,dst_img_dir,dst_label_dir)
# copy_pic_and_create_txt(xml_path,img_path)
if count_num%10000 == 0:
show_img_with_label(dst_txt_path,dst_img_path)
count_num += 1
云淡清风的贴了这几个代码，其实是我在jypter写的，好多打印并没有采样，所以12W张，输出他就崩溃了，这是后来改的。
总之数据集，就处理好了
images 下面是 train和val的image
labels下面是train和val的label 他们两者是一一对应的。
数据集的配置文件我也给一下：
# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
# ImageNet-1k dataset https://www.image-net.org/index.php by Stanford University
# Simplified class names from https://github.com/anishathalye/imagenet-simple-labels
# Example usage: python classify/train.py --data imagenet
# parent
# ├── yolov5
# └── datasets
# └── imagenet ← downloads here (144 GB)
# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..]
# path: ../datasets/imagenet # dataset root dir
train: "F:\\datasets\\20230603hand_gesture\\images\\train" # train images (relative to 'path') 1281167 images
val: "F:\\datasets\\20230603hand_gesture\\images\\val" # val images (relative to 'path') 50000 images
test: # test images (optional)
# Classes
names:
0:one
1:two_up
2:two_up_inverted
3:three
4:three2
5:four
6:fist
7:palm
8:ok
9:peace
10:peace_inverted
11:like
12:dislike
13:stop
14:stop_inverted
15:call
16:mute
17:rock
18:no_gesture
# Download script/URL (optional)
download: data/scripts/get_imagenet.sh
step2：超参数
其实就是训练参数喽，因为我并没有设置这里， 在后面进阶的blog，我会研究一下这里。在这先挖个坑。
step3: 训练
C:\ProgramData\Anaconda3\python.exe train.py --data data\yolo_hand.yaml --weights yolov5s.pt --batch-size 12 --imgsz 320 --workers 4 --epochs 200 --save-period 50
翻车了
AMP: checks passed
optimizer: SGD(lr=0.01) with parameter groups 57 weight(decay=0.0), 60 weight(decay=0.00046875), 60 bias
train: Scanning 'F:\datasets\20230603hand_gesture\labels\train.cache' images and labels... 114597 found, 0 missing, 0 e
val: Scanning 'F:\datasets\20230603hand_gesture\labels\val.cache' images and labels... 12734 found, 0 missing, 0 empty,
Traceback (most recent call last):
File "<string>", line 1, in <module>
File "C:\ProgramData\Anaconda3\lib\multiprocessing\spawn.py", line 116, in spawn_main
exitcode = _main(fd, parent_sentinel)
File "C:\ProgramData\Anaconda3\lib\multiprocessing\spawn.py", line 125, in _main
prepare(preparation_data)
File "C:\ProgramData\Anaconda3\lib\multiprocessing\spawn.py", line 236, in prepare
_fixup_main_from_path(data['init_main_from_path'])
File "C:\ProgramData\Anaconda3\lib\multiprocessing\spawn.py", line 287, in _fixup_main_from_path
main_content = runpy.run_path(main_path,
File "C:\ProgramData\Anaconda3\lib\runpy.py", line 268, in run_path
return _run_module_code(code, init_globals, run_name,
File "C:\ProgramData\Anaconda3\lib\runpy.py", line 97, in _run_module_code
_run_code(code, mod_globals, init_globals,
File "C:\ProgramData\Anaconda3\lib\runpy.py", line 87, in _run_code
exec(code, run_globals)
File "D:\code\python_project\yolov5-master\train.py", line 29, in <module>
import torch
File "C:\ProgramData\Anaconda3\lib\site-packages\torch\__init__.py", line 129, in <module>
raise err
OSError: [WinError 1455] 页面文件太小，无法完成操作。 Error loading "C:\ProgramData\Anaconda3\lib\site-packages\torch\lib\shm.dll" or one of its dependencies.
方法1：调整虚拟内存 搜索关键词 win11， 虚拟内存
并无卵用
方法2：
缩减 workers = 2
结论：
翻车在生成数据集的 cache文件上，设置的workers 和 batchsizes 都有关系。
设置过大，就会造成莫名其妙的问题，比如 说找不到图片啥的
另外调整虚拟内存后要开关机，所以（20230604）今天我上一直在开关机。
小马拉大车太难受了， 我的电脑是32G内存， 8G的显卡 3070。
trick：
用它来生成数据集的cache文件
C:\ProgramData\Anaconda3\python.exe train.py --data data\yolo_hand.yaml --weights yolov5s.pt --batch-size 1 --imgsz 320 --workers 0 --epochs 200 --save-period 50
杀死进程，用它来训练
C:\ProgramData\Anaconda3\python.exe train.py --data data\yolo_hand.yaml --weights yolov5s.pt --batch-size 12 --imgsz 320 --workers 4 --epochs 200 --save-period 50
稳定后感觉设置小了，可是，在scan数据集的时候经常崩溃，感觉在scan数据集的阶段也比较吃显存和内存。当然 吃这两个，就肯定吃cpu了。。。。
资源消耗情况，我设置了4个线程，应该是1个主线程，4个worker。
其实出这个问题，就一个原因：穷！！！！ 来个牛叉的内存和显卡，什么事就都没了。贫贱夫妻百事哀的伤感。。。。
训练1 epoch，报错了
Traceback (most recent call last):
File "D:\code\python_project\yolov5-master\train.py", line 630, in <module>
main(opt)
File "D:\code\python_project\yolov5-master\train.py", line 526, in main
train(opt.hyp, opt, device, callbacks)
File "D:\code\python_project\yolov5-master\train.py", line 349, in train
results, maps, _ = validate.run(data_dict,
File "C:\ProgramData\Anaconda3\lib\site-packages\torch\autograd\grad_mode.py", line 27, in decorate_context
return func(*args, **kwargs)
File "D:\code\python_project\yolov5-master\val.py", line 268, in run
tp, fp, p, r, f1, ap, ap_class = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names)
File "D:\code\python_project\yolov5-master\utils\metrics.py", line 83, in ap_per_class
names = [v for k, v in names.items() if k in unique_classes] # list: only classes that have data
AttributeError: 'str' object has no attribute 'items'
原因在这里，少空格了。。。！！！！！， 我记得1年多前，犯过一次这个错误。
既然重新开始训练，我把bz=16 试试能不能跑动
这里是训练集和测试集的比例。
能跑起来，显卡使用率也不多 3000MB load， 40%的使用。 很奇怪为啥之前老报错页面文件太小。
页面太小用虚拟内存的方式解决的话，要确保C盘有足够大的容量，这是我导师说的，
之前咱也不懂，因为是置换的，要确保至少内存1.5倍C盘的空闲硬盘。否则很容易黑屏。
最终效果:
问题：
20230604-------------------------------
训练一晚上，第二天8点一看，卡在了第二轮，老老实实把bz 和worker 放小一点开整把。
C:\ProgramData\Anaconda3\python.exe train.py --data data\yolo_hand.yaml --weights yolov5s.pt --batch-size 2 --imgsz 320 --workers 0 --epochs 10 --save-period 1
ref:https://zhuanlan.zhihu.com/p/383522113
3090 上用的命令，显卡是24G的，是我自己电脑的3倍
python train.py --data data\yolo_hand.yaml --weights yolov5s.pt --batch-size 36 --imgsz 320 --workers 8 --epochs 10 --save-period 1
训练过程中的一些图片：
训练完成的表现，可以看到，epochs=5的时候就已经很好了。
看下效果：
python detect.py --weights weights/best.pt --source 0 --data /data/yolo_hand.yaml --device 0 --view-img --vid-stride 10
总结：
- 数据集我有，我自己的清洗好的，支持yolov格式的你也可以拿去，代码你也可以拿去。
- 进阶，好的深度学习的模型，会把，数据集中 有多少类，多少图片，每张图片有多少实例，实例的类别，目标大小，目标占整个图片的比例，以及图片的光照强度是否达标等，都拉出清单了。
其实就是跑一遍标注脚本。 光照强度是，光照分布不均匀性不超过40%
扫一扫
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
