yolov5 训练bdd100k踩坑记

两个都是好东西，不多说，说坑：

pycocotools装不上

到git上下载下来，进入setup.py，然后两行命令：

D:\pyenv\yolo5\Scripts\python.exe setup.py build_ext --inplace
D:\pyenv\yolo5\Scripts\python.exe setup.py build_ext install

数据格式转化

bdd的标注是这样的：
yolo的是这样的：

奉上转化代码：

class Bdd2yolo:
    def __init__(self):
        self.bdd_lab = "../BDD100k/labels_bdd/train"
        self.yolo_lab = "../BDD100k/labels/train"
        os.makedirs(self.bdd_lab,exist_ok=True)
        os.makedirs(self.yolo_lab,exist_ok=True)
        self.used_names=["person", "rider", "car", "bus", "truck", "bike", "motor", "train"]
    
    @property
    def all_names(self):
        return ["person", "rider", "car", "bus", "truck", "bike", "motor", "traffic light", "traffic sign",
                "train"]
    
    def bdd2yolo5(self,categorys, file):
        strs = ""
        name=file.replace(".json","")
        readpath=f"{self.bdd_lab}/{file}"
        writepath=f"{self.yolo_lab}/{name}.txt"
        print (readpath,writepath)
        f = open(readpath)
        info = json.load(f)
        # print(len(info))
        # print(info["name"])
        write = open(writepath, 'w')
        for obj in info["frames"]:
            # print(obj["objects"])
            for objects in obj["objects"]:
                # print(objects)
                if objects["category"] in categorys:
                    dw = 1.0 / 1280
                    dh = 1.0 / 720
                    strs += str(categorys.index(objects["category"]))
                    strs += " "
                    strs += str(((objects["box2d"]["x1"] + objects["box2d"]["x2"]) / 2.0) * dw)[0:8]
                    strs += " "
                    strs += str(((objects["box2d"]["y1"] + objects["box2d"]["y2"]) / 2.0) * dh)[0:8]
                    strs += " "
                    strs += str(((objects["box2d"]["x2"] - objects["box2d"]["x1"])) * dw)[0:8]
                    strs += " "
                    strs += str(((objects["box2d"]["y2"] - objects["box2d"]["y1"])) * dh)[0:8]
                    strs += "\n"
            write.writelines(strs)
            write.close()
            print("%s has been dealt!" % info["name"])


if __name__ == "__main__":
    b=Bdd2yolo()
    fileList = os.listdir(b.bdd_lab)
    for file in fileList:
        # 自己需要从BDD数据集里提取的目标类别
        b.bdd2yolo5(b.used_names, file)

yaml文件改

这个网络上有，改两个data和models下的，其中models下主要是识别类别数量变了。

计算机性能

这个数据集比较大，市场爆出内存不足错误，一般的电脑，我看还是算了，动辄以周月计算。