wkt

this happen to me too when I train coco without any pretrain weights I change the loss to: lkpt += kpt_loss_factor*( ( (1 - torch.exp(-d/(s*(4*sigmas**2)+1e-9))) + 0.05*d)*kpt_mask).mean() and it fix.

> > this happen to me too when I train coco without any pretrain weights I change the loss to: lkpt += kpt_loss_factor*( ( (1 - torch.exp(-d/(s*(4_sigmas**2)+1e-9))) + 0.05_d)*kpt_mask).mean() and...

As we have comment for this request, OK,I will work on update the FileRoller to 3.22+ @nightskydreamer @overclockedllama

运行环境贴一下，android版本，cpu类型等

> 使用您的ncnn版本正常，升级到最新版本ncnn-20240410-android-vulkan-shared.zip后，打包正常，启动也是报"libncnn.so" not found 安卓9 vivox21a CPU核数：8 CPU位数：64-bit 据说ncnn对vulkan支持不够优化，所以可以试试其它版

> 您现在项目的ncnn是基于哪个版本的呢 https://github.com/wkt/YoloMobile/blob/d4f28a524e87c9ede1ba04cd901de29136c37414/YoloMobile/src/main/cpp/ncnn/x86_64/include/ncnn/platform.h#L64 `#define NCNN_VERSION_STRING "1.0.20230816"`

> > > 您现在项目的ncnn是基于哪个版本的呢 > > > > > > https://github.com/wkt/YoloMobile/blob/d4f28a524e87c9ede1ba04cd901de29136c37414/YoloMobile/src/main/cpp/ncnn/x86_64/include/ncnn/platform.h#L64 > > > > `#define NCNN_VERSION_STRING "1.0.20230816"` > > 解决了吗，我也碰到了？没有编译进apk里 对方，没有再回应，我也不知道TA是否解决了？ 你也是因为想换ncnn的版本，导致的问题么？ 项目内置的ncnn是https://github.com/Tencent/ncnn/releases/download/20230816/ncnn-20230816-android-shared.zip

试试更新一下ncnn到最新版，替换YoloMobile/src/main/cpp/ncnn下的文件即可 如果还是有问题，贴一下相关的日志信息

使用model.export转换的模型，请使用如下参数配置： ```{ "name": "yolov8n-640", "input_size": 640, "param": "yolov8n.param", "bin": "yolov8n.bin", "box_thr": 0.25, "iou_thr": 0.35, "ver": 8, "input_name": "in0", "outputs": [ {"name": "out0","stride":0,"anchors": [0,0]} ] } ``` 然后根据需要修改 "name"，"input_size"， "param" ，"bin"等字段

目前仅支持，如下代码生成的ncnn模型： ``` from ultralytics import YOLO # Load a model model = YOLO('path/to/best.pt') # load a custom trained model # Export the model model.export(format='ncnn', imgsz=640) ``` 并input_size=imgsz。 至于处理/解析模型的直接输出，可参考：YoloMobile/src/main/cpp/Yolo.cpp