基于此代码，直接用yolov11n.bin，yolov12n.bin推理，结果是不对的！使用自己的样本集（10类），基于yolov11n.bin权重，推理结果和pyhon差异太大！

[badtiger001]

基于此代码，直接用yolov11n.bin，yolov12n.bin推理，结果是不对的！使用自己的样本集（10类），基于yolov11n.bin权重，900张图，训练100次，什么都识别不出来！同样的样本使用python训练，效果很好。YoloSharp现在支持训练自己的样本吗 ？支持从头训练吗？还有就是如果能支持直接使用pt权重作为预训练权重就好了。

[IntptrMax]

I have the same situation as you, it looks like yolo11n or yolo12n pre-trained model are not suit for every image. If you change the pretrained model to s or m, it will work much more better.
BTW, yolo11 and yolo12 have several self-atten block. Self-atten will cost much more resources than other blocks, I have changed self-atten method to SDPA to reduce resources used, it will has some bias between self-attn and SDPA them two. You can load the pretrained model and train some steps, it will have a better result.
YoloSharp can support using your own data to train a model. YoloSharp use completely random data, it will take a lot of steps to train a model from completely random, if you want to have a fast speed in training, load pretrained model and continue to train.

[badtiger001]

我用yolov11s.bin预训练权重，也用同样的样本试了一下。效果也不好，狐狸竟然识别成马了！分数还挺高和python使用官网提供的pt预训练权重的训练结果差的还是比较远！楼主有空还有什么建议吗？

[IntptrMax]

我用yolov11s.bin预训练权重，也用同样的样本试了一下。效果也不好，狐狸竟然识别成马了！分数还挺高和python使用官网提供的pt预训练权重的训练结果差的还是比较远！楼主有空还有什么建议吗？

You are right. I have reviewed my code, and find there are some bugs in attention blocks for yolov11 and yolov12. I'm trying to fix them.

[IntptrMax]

I have fixed the bugs for yolov11 attention block and yolov12n.bin, please try again.

[IntptrMax]

我用yolov11s.bin预训练权重，也用同样的样本试了一下。效果也不好，狐狸竟然识别成马了！分数还挺高和python使用官网提供的pt预训练权重的训练结果差的还是比较远！楼主有空还有什么建议吗？

I have fix the bugs in Yolov11 detect module, attention block, please try again.