mindyolo FAQ

[Ash-Lee233]

Here is a summary of common training and inference issues.

[WongGawa]

When you are training with yolo on Ascend910，you may encounter the following similar problem: the loss suddenly increases.
Solution: Remove the overflow_still_update: False parameter in yolo‘s yaml configuration and re-execute training.
e.g. https://github.com/mindspore-lab/mindyolo/blob/master/configs/yolov8/yolov8s.yaml/#L7

当你在Ascend910上使用yolo全量训练时可能会遇到以下类似问题：训练过程中loss会突然增加。
解决方法：删除yolo的yaml里的overflow_still_update: False参数配置，然后重新执行全量训练。
e.g. https://github.com/mindspore-lab/mindyolo/blob/master/configs/yolov8/yolov8s.yaml/#L7

[WongGawa]

When you are training a model on 2 NPU，you may encounter the following similar problem: compared with the benchmark, the performance results are poor.
Solution: When using mpirun command to train a model, add the --bind-to numa parameter to improve performance. For example:

  mpirun --allow-run-as-root -n 2 --bind-to numa python train.py --config ./configs/yolov7/yolov7.yaml  --is_parallel True

如果您在双卡设备上使用mpirun进行分布式模型训练时，可能会遇到以下类似问题：相比较benchmark，性能结果较差。
解决方法：使用mpirun指令启动时，请添加--bind-to numa参数以提升性能表现。例如：

  mpirun --allow-run-as-root -n 2 --bind-to numa python train.py --config ./configs/yolov7/yolov7.yaml  --is_parallel True