Chilicyy comments

Results 107 comments of


                                            Chilicyy

--fuse_ab 参数开启后可以更快收敛么？实测每个epoch的时间加倍了

@roxroxroxrox 添加`--fuse_ab`参数后，训练过程会增加anchor-based目标框检测分支，训练时间会加长，收敛速度会加快。

how to use self-distillation, and what's the difference with finetune teacher model with new data

Hi @HaoLiuHust when you try self-distillation on your task, you can use the first trained model on your own data for teacher model, and use coco weights for finetuning which...

need yolov6 pose

Thanks for your attention! We are planning to add this feature in the near future.

from yolov6.core.engine import Trainer ModuleNotFoundError: No module named 'yolov6.core'

您好，您需要在YOLOv6项目代码主目录下运行代码，同时将项目所在目录路径添加到PYTHONPATH路径中。

关于V6的设计思路

您好，感谢关注！在小尺寸模型上backbone和neck部分采用ReLU激活函数主要是为了加快模型推理速度，而在head部分采用SiLU激活相对于ReLU激活精度更高，但对速度影响不会太大。

Why infer with onnx is slower than ckpt？

Hi，did you infer with onnx and checkpoint both on GPUs？ Besides, the process of converting the checkpoint model to ONNX format may introduce some overhead, resulting in slower inference times....

Change the number of keypoints of yolov6-face

Thanks for your attention. We'll consider it and add this feature.

flag `--patience` (epochs to wait for no observable improvement for early stopping of training)

Hi, you can check if `total_loss` stop decreasing or `self.ap` stop increasing to optimize your training epochs.

No metrics, f1 score and precision

Hi, you can add `--do_pr_metric ` to get metrics like precision and recall when doing evaluation.

Benchmark mAP value different in paper.

@qazimbhat1 Hi, it seems that you compared the results between the original paper and the latest release on github. We have a full scale reloading , and you can checkout...