Laughing

@dmddmd also you can use box.cls to access the predicted cls id.

@justin-fun please add `save_json=True`.

@AdamRoden I'd like to share you my local env info: - cuda-11.1 - cudnn-8.4.1.50 - Tensorrt-8.4.1.5 - pytorch-1.10.2+cu111 - onnxruntime-gpu-1.14.0 all of them are not the latest version, but works...

@Rm1n90 Thanks for the reporting! I can reproduce it and we'll fix it soon. :)

@Rm1n90 BTW I think you can rename the `best.engine` to `best-seg.engine` for now as a temporary solution. The code is guessing the task by the suffix i.e `-seg` and `-cls`...

> @Laughing-q so this is happening because we are trying to assign a task to a model before the user has a chance to define a task, i.e. both CLI...

@alicera it also depends on your number of instances per image. Also It seems that 1080ti doesn't support amp training so it takes more memory for fp32 training. BTW please...

@haimat hi, @AyushExel has made a PR to support this, now you're able to access the metrics by the following script: ```python model.train() metrics = model.metrics metrics.box.map # map50-95 metrics.box.map50...

@Auggie3 thanks for the reporting! It'd be very helpful if you provide us your command line or script to reproduce the issue. :)

@Auggie3 I tried to reproduce your issue but actually it works fine to me. ```python model = YOLO("yolov8m.pt") model.train(data="coco128.yaml", epochs=5) model.val() ```